talk-llama : sync llama.cpp

examples/talk-llama/llama-arch.cpp

@@ -232,6 +232,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
     { LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN,       "%s.attention.sliding_window_pattern"       },
     { LLM_KV_ATTENTION_SCALE,                        "%s.attention.scale"                        },
     { LLM_KV_ATTENTION_OUTPUT_SCALE,                 "%s.attention.output_scale"                 },
+    { LLM_KV_ATTENTION_VALUE_SCALE,                  "%s.attention.value_scale"                  },
     { LLM_KV_ATTENTION_TEMPERATURE_LENGTH,           "%s.attention.temperature_length"           },
     { LLM_KV_ATTENTION_TEMPERATURE_SCALE,            "%s.attention.temperature_scale"            },
     { LLM_KV_ATTENTION_KEY_LENGTH_MLA,               "%s.attention.key_length_mla"               },

examples/talk-llama/llama-arch.h

@@ -236,6 +236,7 @@ enum llm_kv {
     LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN,
     LLM_KV_ATTENTION_SCALE,
     LLM_KV_ATTENTION_OUTPUT_SCALE,
+    LLM_KV_ATTENTION_VALUE_SCALE,
     LLM_KV_ATTENTION_TEMPERATURE_LENGTH,
     LLM_KV_ATTENTION_TEMPERATURE_SCALE,
     LLM_KV_ATTENTION_KEY_LENGTH_MLA,

examples/talk-llama/llama-context.cpp

@@ -2230,13 +2230,17 @@ llm_graph_cb llama_context::graph_get_cb() const {
 
 class llama_io_write_dummy : public llama_io_write_i {
 public:
-    llama_io_write_dummy() = default;
+    llama_io_write_dummy(bool skip_tensors) : skip_tensors(skip_tensors) {}
 
     void write(const void * /* src */, size_t size) override {
         size_written += size;
     }
 
-    void write_tensor(const ggml_tensor * /* tensor */, size_t /* offset */, size_t size) override {
+    void write_tensor(ggml_tensor * /* tensor */, size_t /* offset */, size_t size) override {
+        if (skip_tensors) {
+            return;
+        }
+
         size_written += size;
     }
 
@@ -2245,14 +2249,23 @@ public:
     }
 
 private:
+    const bool skip_tensors;
+
     size_t size_written = 0;
 };
 
-class llama_io_write_buffer : public llama_io_write_i {
+class llama_io_write_host : public llama_io_write_i {
 public:
-    llama_io_write_buffer(
+    llama_io_write_host(
             uint8_t * p, size_t len) : ptr(p), buf_size(len) {}
 
+    ~llama_io_write_host() {
+        // TODO: add backend support to batch tensor_get? or some other way to speed this up
+        for (const auto & winfo : winfos) {
+            ggml_backend_tensor_get(winfo.tensor, winfo.ptr, winfo.offset, winfo.size);
+        }
+    }
+
     void write(const void * src, size_t size) override {
         if (size > buf_size) {
             throw std::runtime_error("unexpectedly reached end of buffer");
@@ -2263,11 +2276,14 @@ public:
         buf_size -= size;
     }
 
-    void write_tensor(const ggml_tensor * tensor, size_t offset, size_t size) override {
+    void write_tensor(ggml_tensor * tensor, size_t offset, size_t size) override {
         if (size > buf_size) {
             throw std::runtime_error("unexpectedly reached end of buffer");
         }
-        ggml_backend_tensor_get(tensor, ptr, offset, size);
+
+        // save the write for later during destruction
+        winfos.push_back({tensor, ptr, size, offset});
+
         ptr += size;
         size_written += size;
         buf_size -= size;
@@ -2281,25 +2297,48 @@ private:
     uint8_t * ptr;
     size_t buf_size = 0;
     size_t size_written = 0;
+
+    struct write_info {
+        ggml_tensor * tensor;
+        uint8_t * ptr;
+        size_t size;
+        size_t offset;
+    };
+    std::vector<write_info> winfos;
 };
 
-class llama_io_read_buffer : public llama_io_read_i {
+class llama_io_read_host : public llama_io_read_i {
 public:
-    llama_io_read_buffer(const uint8_t * p, size_t len) : ptr(p), buf_size(len) {}
+    llama_io_read_host(const uint8_t * p, size_t len) : ptr(p), buf_size(len) {}
 
-    const uint8_t * read(size_t size) override {
-        const uint8_t * base_ptr = ptr;
+    ~llama_io_read_host() {
+        // flush the reads
+        for (const auto & rinfo : rinfos) {
+            ggml_backend_tensor_set(rinfo.tensor, rinfo.ptr, rinfo.offset, rinfo.size);
+        }
+    }
+
+    void read(void * dst, size_t size) override {
         if (size > buf_size) {
             throw std::runtime_error("unexpectedly reached end of buffer");
         }
+        memcpy(dst, ptr, size);
         ptr += size;
         size_read += size;
         buf_size -= size;
-        return base_ptr;
     }
 
-    void read_to(void * dst, size_t size) override {
-        memcpy(dst, read(size), size);
+    void read_tensor(ggml_tensor * tensor, size_t offset, size_t size) override {
+        if (size > buf_size) {
+            throw std::runtime_error("unexpectedly reached end of buffer");
+        }
+
+        // save for later during destruction
+        rinfos.push_back({tensor, ptr, size, offset});
+
+        ptr += size;
+        size_read += size;
+        buf_size -= size;
     }
 
     size_t n_bytes() override {
@@ -2310,6 +2349,14 @@ private:
     const uint8_t * ptr;
     size_t buf_size = 0;
     size_t size_read = 0;
+
+    struct read_info {
+        ggml_tensor * tensor;
+        const uint8_t * ptr;
+        size_t size;
+        size_t offset;
+    };
+    std::vector<read_info> rinfos;
 };
 
 class llama_io_write_file : public llama_io_write_i {
@@ -2321,7 +2368,7 @@ public:
         size_written += size;
     }
 
-    void write_tensor(const ggml_tensor * tensor, size_t offset, size_t size) override {
+    void write_tensor(ggml_tensor * tensor, size_t offset, size_t size) override {
         temp_buffer.resize(size);
         ggml_backend_tensor_get(tensor, temp_buffer.data(), offset, size);
         write(temp_buffer.data(), temp_buffer.size());
@@ -2341,15 +2388,15 @@ class llama_io_read_file : public llama_io_read_i {
 public:
     llama_io_read_file(llama_file * f) : file(f) {}
 
-    void read_to(void * dst, size_t size) override {
+    void read(void * dst, size_t size) override {
         file->read_raw(dst, size);
         size_read += size;
     }
 
-    const uint8_t * read(size_t size) override {
+    void read_tensor(ggml_tensor * tensor, size_t offset, size_t size) override {
         temp_buffer.resize(size);
-        read_to(temp_buffer.data(), size);
-        return temp_buffer.data();
+        read(temp_buffer.data(), size);
+        ggml_backend_tensor_set(tensor, temp_buffer.data(), offset, size);
     }
 
     size_t n_bytes() override {
@@ -2362,8 +2409,212 @@ private:
     std::vector<uint8_t> temp_buffer;
 };
 
+class llama_io_write_device : public llama_io_write_i {
+public:
+    llama_io_write_device(uint8_t * p, size_t len, llama_memory_buffers & mbufs) : ptr(p), buf_size(len), mbufs(mbufs)  {
+    }
+
+    ~llama_io_write_device() {
+        llama_memory_buffers mbufs_new;
+
+        for (const auto & winfo : winfos) {
+            auto * buft = ggml_backend_buffer_get_type(winfo.tensor->buffer);
+
+            mbufs_new[buft].n_tensors++;
+            mbufs_new[buft].total_size += winfo.size;
+        }
+
+        for (auto & [buft, mbuf] : mbufs_new) {
+            ggml_init_params params = {
+                /*.mem_size   =*/ 2*mbuf.n_tensors*ggml_tensor_overhead(),
+                /*.mem_buffer =*/ NULL,
+                /*.no_alloc   =*/ true,
+            };
+
+            mbuf.ctx.reset(ggml_init(params));
+
+            mbuf.org.reserve(mbuf.n_tensors);
+            mbuf.cpy.reserve(mbuf.n_tensors);
+        }
+
+        for (const auto & winfo : winfos) {
+            auto * buft = ggml_backend_buffer_get_type(winfo.tensor->buffer);
+
+            const int64_t n = winfo.size/ggml_element_size(winfo.tensor);
+
+            auto & mbuf = mbufs_new[buft];
+
+            mbuf.org.push_back(ggml_view_1d      (mbuf.ctx.get(), winfo.tensor, n, winfo.offset));
+            mbuf.cpy.push_back(ggml_new_tensor_1d(mbuf.ctx.get(), winfo.tensor->type, n));
+        }
+
+        for (auto & [buft, mbuf] : mbufs_new) {
+            auto & mbuf_cur = mbufs[buft];
+
+            bool need_alloc = false;
+
+            need_alloc = need_alloc || (!mbuf_cur.buf);
+            need_alloc = need_alloc || (mbuf_cur.org.size() != mbuf.org.size());
+            need_alloc = need_alloc || (mbuf_cur.total_size != mbuf.total_size);
+
+            if (!need_alloc) {
+                for (size_t i = 0; i < mbuf_cur.org.size(); ++i) {
+                    auto * org0 = mbuf_cur.org[i];
+                    auto * org1 = mbuf.org[i];
+
+                    if (!ggml_are_same_shape(org0, org1)) {
+                        need_alloc = true;
+                        break;
+                    }
+
+                    if (org0->view_src != org1->view_src || org0->view_offs != org1->view_offs) {
+                        need_alloc = true;
+                        break;
+                    }
+                }
+            }
+
+            if (need_alloc) {
+                mbuf_cur = std::move(mbuf);
+
+                mbuf_cur.buf.reset(ggml_backend_alloc_ctx_tensors_from_buft(mbuf_cur.ctx.get(), buft));
+
+                LLAMA_LOG_INFO("%s: allocated '%s' buffer %.3f MiB\n", __func__, ggml_backend_buft_name(buft), mbuf.total_size/1024.0/1024.0);
+            }
+
+            for (size_t i = 0; i < mbuf_cur.org.size(); ++i) {
+                ggml_backend_tensor_copy(mbuf_cur.org[i], mbuf_cur.cpy[i]);
+            }
+        }
+    }
+
+    void write(const void * src, size_t size) override {
+        if (size > buf_size) {
+            throw std::runtime_error("unexpectedly reached end of buffer");
+        }
+        memcpy(ptr, src, size);
+        ptr += size;
+        size_written += size;
+        buf_size -= size;
+    }
+
+    void write_tensor(ggml_tensor * tensor, size_t offset, size_t size) override {
+        // save the write for later during destruction
+        winfos.push_back({tensor, ptr, size, offset});
+    }
+
+    size_t n_bytes() override {
+        return size_written;
+    }
+
+private:
+    uint8_t * ptr;
+    size_t buf_size = 0;
+    size_t size_written = 0;
+
+    struct write_info {
+        ggml_tensor * tensor;
+        uint8_t * ptr;
+        size_t size;
+        size_t offset;
+    };
+    std::vector<write_info> winfos;
+
+    llama_memory_buffers & mbufs;
+};
+
+class llama_io_read_device : public llama_io_read_i {
+public:
+    llama_io_read_device(const uint8_t * p, size_t len, const llama_memory_buffers & mbufs) : ptr(p), buf_size(len), mbufs(mbufs) {
+    }
+
+    ~llama_io_read_device() {
+        llama_memory_buffers mbufs_new;
+
+        for (const auto & rinfo : rinfos) {
+            auto * buft = ggml_backend_buffer_get_type(rinfo.tensor->buffer);
+
+            mbufs_new[buft].n_tensors++;
+            mbufs_new[buft].total_size += rinfo.size;
+        }
+
+        for (auto & [buft, mbuf] : mbufs_new) {
+            ggml_init_params params = {
+                /*.mem_size   =*/ mbuf.n_tensors*ggml_tensor_overhead(),
+                /*.mem_buffer =*/ NULL,
+                /*.no_alloc   =*/ true,
+            };
+
+            mbuf.ctx.reset(ggml_init(params));
+
+            mbuf.org.reserve(mbuf.n_tensors);
+        }
+
+        for (const auto & rinfo : rinfos) {
+            auto * buft = ggml_backend_buffer_get_type(rinfo.tensor->buffer);
+
+            const int64_t n = rinfo.size/ggml_element_size(rinfo.tensor);
+
+            auto & mbuf = mbufs_new[buft];
+
+            mbuf.org.push_back(ggml_view_1d(mbuf.ctx.get(), rinfo.tensor, n, rinfo.offset));
+
+            auto & view = mbuf.org.back();
+            view->buffer = rinfo.tensor->buffer;
+        }
+
+        for (auto & [buft, mbuf] : mbufs_new) {
+            const auto & mbuf_cur = mbufs.at(buft);
+
+            if (!mbuf_cur.buf || mbuf_cur.n_tensors != mbuf.n_tensors || mbuf_cur.total_size != mbuf.total_size) {
+                GGML_ABORT("%s: memory buffer mismatch\n", __func__);
+            }
+
+            for (size_t i = 0; i < mbuf_cur.org.size(); ++i) {
+                ggml_backend_tensor_copy(mbuf_cur.cpy[i], mbuf.org[i]);
+            }
+        }
+
+        GGML_ASSERT(buf_size == 0);
+    }
+
+    void read(void * dst, size_t size) override {
+        if (size > buf_size) {
+            throw std::runtime_error("unexpectedly reached end of buffer");
+        }
+        memcpy(dst, ptr, size);
+        ptr += size;
+        size_read += size;
+        buf_size -= size;
+    }
+
+    void read_tensor(ggml_tensor * tensor, size_t offset, size_t size) override {
+        // save for later during destruction
+        rinfos.push_back({tensor, ptr, size, offset});
+    }
+
+    size_t n_bytes() override {
+        return size_read;
+    }
+
+private:
+    const uint8_t * ptr;
+    size_t buf_size = 0;
+    size_t size_read = 0;
+
+    struct read_info {
+        ggml_tensor * tensor;
+        const uint8_t * ptr;
+        size_t size;
+        size_t offset;
+    };
+    std::vector<read_info> rinfos;
+
+    const llama_memory_buffers & mbufs;
+};
+
 size_t llama_context::state_get_size() {
-    llama_io_write_dummy io;
+    llama_io_write_dummy io(false);
     try {
         return state_write_data(io);
     } catch (const std::exception & err) {
@@ -2373,7 +2624,7 @@ size_t llama_context::state_get_size() {
 }
 
 size_t llama_context::state_get_data(uint8_t * dst, size_t size) {
-    llama_io_write_buffer io(dst, size);
+    llama_io_write_host io(dst, size);
     try {
         return state_write_data(io);
     } catch (const std::exception & err) {
@@ -2383,7 +2634,7 @@ size_t llama_context::state_get_data(uint8_t * dst, size_t size) {
 }
 
 size_t llama_context::state_set_data(const uint8_t * src, size_t size) {
-    llama_io_read_buffer io(src, size);
+    llama_io_read_host io(src, size);
     try {
         return state_read_data(io);
     } catch (const std::exception & err) {
@@ -2392,9 +2643,14 @@ size_t llama_context::state_set_data(const uint8_t * src, size_t size) {
     }
 }
 
+static constexpr uint32_t io_magic = 0xaf143cd8;
+
 size_t llama_context::state_seq_get_size(llama_seq_id seq_id, llama_state_seq_flags flags) {
-    llama_io_write_dummy io;
+    llama_io_write_dummy io(flags & LLAMA_STATE_SEQ_FLAGS_ON_DEVICE);
     try {
+        io.write(&io_magic, sizeof(io_magic));
+        io.write(&seq_id, sizeof(seq_id));
+
         return state_seq_write_data(io, seq_id, flags);
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: error getting state size: %s\n", __func__, err.what());
@@ -2403,9 +2659,18 @@ size_t llama_context::state_seq_get_size(llama_seq_id seq_id, llama_state_seq_fl
 }
 
 size_t llama_context::state_seq_get_data(llama_seq_id seq_id, uint8_t * dst, size_t size, llama_state_seq_flags flags) {
-    llama_io_write_buffer io(dst, size);
+    std::unique_ptr<llama_io_write_i> io;
+    if (flags & LLAMA_STATE_SEQ_FLAGS_ON_DEVICE) {
+        io = std::make_unique<llama_io_write_device>(dst, size, mem_storage[seq_id]);
+    } else {
+        io = std::make_unique<llama_io_write_host>(dst, size);
+    }
+
     try {
-        return state_seq_write_data(io, seq_id, flags);
+        io->write(&io_magic, sizeof(io_magic));
+        io->write(&seq_id, sizeof(seq_id));
+
+        return state_seq_write_data(*io, seq_id, flags);
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: error saving state: %s\n", __func__, err.what());
         return 0;
@@ -2413,9 +2678,38 @@ size_t llama_context::state_seq_get_data(llama_seq_id seq_id, uint8_t * dst, siz
 }
 
 size_t llama_context::state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size, llama_state_seq_flags flags) {
-    llama_io_read_buffer io(src, size);
+    std::unique_ptr<llama_io_read_i> io;
+    if (flags & LLAMA_STATE_SEQ_FLAGS_ON_DEVICE) {
+        // create a temporary io to read the magic and the src seq_id
+        io = std::make_unique<llama_io_read_host>(src, size);
+
+        uint32_t magic_read;
+        io->read(&magic_read, sizeof(magic_read));
+        if (io_magic != magic_read) {
+            throw std::runtime_error("wrong sequence state magic");
+        }
+
+        llama_seq_id seq_id_read;
+        io->read(&seq_id_read, sizeof(seq_id_read));
+
+        GGML_ASSERT(mem_storage.find(seq_id_read) != mem_storage.end());
+
+        io = std::make_unique<llama_io_read_device>(src, size, mem_storage[seq_id_read]);
+    } else {
+        io = std::make_unique<llama_io_read_host>(src, size);
+    }
+
     try {
-        return state_seq_read_data(io, seq_id, flags);
+        uint32_t magic_read;
+        io->read(&magic_read, sizeof(magic_read));
+        if (io_magic != magic_read) {
+            throw std::runtime_error("wrong sequence state magic");
+        }
+
+        llama_seq_id seq_id_read;
+        io->read(&seq_id_read, sizeof(seq_id_read));
+
+        return state_seq_read_data(*io, seq_id, flags);
     } catch (const std::exception & err) {
         LLAMA_LOG_ERROR("%s: error loading state: %s\n", __func__, err.what());
         return 0;
@@ -3406,7 +3700,6 @@ size_t llama_state_seq_get_data_ext(llama_context * ctx, uint8_t * dst, size_t s
 
     return ctx->state_seq_get_data(seq_id, dst, size, flags);
 }
-
 size_t llama_state_seq_set_data_ext(llama_context * ctx, const uint8_t * src, size_t size, llama_seq_id seq_id, llama_state_seq_flags flags) {
     ctx->synchronize();
 

examples/talk-llama/llama-context.h

@@ -23,6 +23,21 @@ class llama_io_write_i;
 struct llama_memory_i;
 struct llama_memory_context_i;
 
+// stores copy of the memory in device buffer. used for fast state save/load
+struct llama_memory_buffer {
+    int n_tensors = 0;
+    size_t total_size = 0;
+
+    ggml_backend_buffer_ptr buf;
+
+    ggml_context_ptr ctx;
+
+    std::vector<ggml_tensor *> org;
+    std::vector<ggml_tensor *> cpy;
+};
+
+using llama_memory_buffers = std::map<ggml_backend_buffer_type_t, llama_memory_buffer>;
+
 struct llama_context {
     // init scheduler and compute buffers, reserve worst-case graphs
     llama_context(
@@ -128,6 +143,7 @@ struct llama_context {
     size_t state_set_data(const uint8_t * src, size_t size);
 
     size_t state_seq_get_size(llama_seq_id seq_id, llama_state_seq_flags flags);
+
     size_t state_seq_get_data(llama_seq_id seq_id,       uint8_t * dst, size_t size, llama_state_seq_flags flags);
     size_t state_seq_set_data(llama_seq_id seq_id, const uint8_t * src, size_t size, llama_state_seq_flags flags);
 
@@ -328,6 +344,9 @@ private:
     // host buffer for the model output (logits and embeddings)
     ggml_backend_buffer_ptr buf_output;
 
+    // keep copies of the per-sequence memory on the device
+    std::map<llama_seq_id, llama_memory_buffers> mem_storage;
+
     bool has_evaluated_once = false;
 
     // env: LLAMA_GRAPH_REUSE_DISABLE

examples/talk-llama/llama-graph.cpp

@@ -65,8 +65,13 @@ static ggml_tensor * ggml_mul_mat_aux(
 
     ggml_tensor * res;
 
-    res = ggml_reshape_2d(ctx, cur, n, ggml_nelements(cur)/n);
+    if (!ggml_is_contiguous(cur)) {
+        res = ggml_cont_2d   (ctx, cur, n, ggml_nelements(cur)/n);
+    } else {
+        res = ggml_reshape_2d(ctx, cur, n, ggml_nelements(cur)/n);
+    }
     res = ggml_mul_mat   (ctx, rot, res);
+    ggml_mul_mat_set_hint(res, GGML_HINT_SRC0_IS_HADAMARD);
     res = ggml_reshape_4d(ctx, res, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3]);
 
     return res;

examples/talk-llama/llama-hparams.h

@@ -166,6 +166,8 @@ struct llama_hparams {
     float    f_attn_out_scale = 0.0f;
     uint32_t attn_temp_length = 0;
 
+    float    f_attn_value_scale = 0.0f;
+
     bool causal_attn   = true;
     bool use_alibi     = false;
     bool attn_soft_cap = false;

examples/talk-llama/llama-io.cpp

@@ -1,5 +1,7 @@
 #include "llama-io.h"
 
+#include <vector>
+
 void llama_io_write_i::write_string(const std::string & str) {
     uint32_t str_size = str.size();
 
@@ -9,7 +11,10 @@ void llama_io_write_i::write_string(const std::string & str) {
 
 void llama_io_read_i::read_string(std::string & str) {
     uint32_t str_size;
-    read_to(&str_size, sizeof(str_size));
+    read(&str_size, sizeof(str_size));
 
-    str.assign((const char *) read(str_size), str_size);
+    std::vector<char> buf(str_size);
+    read(buf.data(), str_size);
+
+    str.assign(buf.data(), str_size);
 }

examples/talk-llama/llama-io.h

@@ -12,7 +12,7 @@ public:
     virtual ~llama_io_write_i() = default;
 
     virtual void write(const void * src, size_t size) = 0;
-    virtual void write_tensor(const ggml_tensor * tensor, size_t offset, size_t size) = 0;
+    virtual void write_tensor(ggml_tensor * tensor, size_t offset, size_t size) = 0;
 
     // bytes written so far
     virtual size_t n_bytes() = 0;
@@ -25,8 +25,8 @@ public:
     llama_io_read_i() = default;
     virtual ~llama_io_read_i() = default;
 
-    virtual const uint8_t * read(size_t size) = 0;
-    virtual void read_to(void * dst, size_t size) = 0;
+    virtual void read(void * dst, size_t size) = 0;
+    virtual void read_tensor(ggml_tensor * tensor, size_t offset, size_t size) = 0;
 
     // bytes read so far
     virtual size_t n_bytes() = 0;

examples/talk-llama/llama-kv-cache.cpp

@@ -67,6 +67,7 @@ static ggml_tensor * ggml_mul_mat_aux(
 
     res = ggml_reshape_2d(ctx, cur, n, ggml_nelements(cur)/n);
     res = ggml_mul_mat   (ctx, rot, res);
+    ggml_mul_mat_set_hint(res, GGML_HINT_SRC0_IS_HADAMARD);
     res = ggml_reshape_4d(ctx, res, cur->ne[0], cur->ne[1], cur->ne[2], cur->ne[3]);
 
     return res;
@@ -1900,14 +1901,14 @@ void llama_kv_cache::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama
     GGML_ASSERT(seq_id == -1 || (seq_id >= 0 && (size_t) seq_id < seq_to_stream.size()));
 
     uint32_t n_stream_cur;
-    io.read_to(&n_stream_cur, sizeof(n_stream_cur));
+    io.read(&n_stream_cur, sizeof(n_stream_cur));
     if (n_stream_cur != n_stream) {
         throw std::runtime_error("n_stream mismatch");
     }
 
     for (uint32_t s = 0; s < n_stream; ++s) {
         uint32_t cell_count;
-        io.read_to(&cell_count, sizeof(cell_count));
+        io.read(&cell_count, sizeof(cell_count));
 
         if (cell_count == 0) {
             continue;
@@ -2082,8 +2083,8 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
             llama_pos pos;
             uint32_t n_seq_id;
 
-            io.read_to(&pos,      sizeof(pos));
-            io.read_to(&n_seq_id, sizeof(n_seq_id));
+            io.read(&pos,      sizeof(pos));
+            io.read(&n_seq_id, sizeof(n_seq_id));
 
             if (n_seq_id != 1) {
                 LLAMA_LOG_ERROR("%s: invalid seq_id-agnostic kv cell\n", __func__);
@@ -2092,7 +2093,7 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
 
             if (hparams.n_pos_per_embd() > 1) {
                 llama_kv_cell_ext ext;
-                io.read_to(&ext, sizeof(ext));
+                io.read(&ext, sizeof(ext));
 
                 ubatch.pos[i + ubatch.n_tokens]   = ext.y;
                 ubatch.pos[i + ubatch.n_tokens*2] = ext.x;
@@ -2101,7 +2102,7 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
             // read the sequence id, but directly discard it - we will use dest_seq_id instead
             {
                 llama_seq_id seq_id;
-                io.read_to(&seq_id, sizeof(seq_id));
+                io.read(&seq_id, sizeof(seq_id));
             }
 
             ubatch.pos[i]      = pos;
@@ -2143,20 +2144,20 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
             llama_pos pos;
             uint32_t  n_seq_id;
 
-            io.read_to(&pos,      sizeof(pos));
-            io.read_to(&n_seq_id, sizeof(n_seq_id));
+            io.read(&pos,      sizeof(pos));
+            io.read(&n_seq_id, sizeof(n_seq_id));
 
             cells.pos_set(i, pos);
 
             if (hparams.n_pos_per_embd() > 1) {
                 llama_kv_cell_ext ext;
-                io.read_to(&ext, sizeof(ext));
+                io.read(&ext, sizeof(ext));
                 cells.ext_set(i, ext);
             }
 
             for (uint32_t j = 0; j < n_seq_id; ++j) {
                 llama_seq_id seq_id;
-                io.read_to(&seq_id, sizeof(seq_id));
+                io.read(&seq_id, sizeof(seq_id));
 
                 if (seq_id < 0 || (uint32_t) seq_id >= n_seq_max) {
                     LLAMA_LOG_ERROR("%s: invalid seq_id, %d is out of range [0, %u)\n", __func__, seq_id, n_seq_max);
@@ -2189,8 +2190,8 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
     uint32_t v_trans;
     uint32_t n_layer;
 
-    io.read_to(&v_trans, sizeof(v_trans));
-    io.read_to(&n_layer, sizeof(n_layer));
+    io.read(&v_trans, sizeof(v_trans));
+    io.read(&n_layer, sizeof(n_layer));
 
     if (n_layer != layers.size()) {
         LLAMA_LOG_ERROR("%s: mismatched layer count (%u instead of %u)\n", __func__, n_layer, (uint32_t) layers.size());
@@ -2217,7 +2218,7 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
 
         // Read type of key
         int32_t k_type_i_ref;
-        io.read_to(&k_type_i_ref, sizeof(k_type_i_ref));
+        io.read(&k_type_i_ref, sizeof(k_type_i_ref));
         const int32_t k_type_i = (int32_t) k->type;
         if (k_type_i != k_type_i_ref) {
             LLAMA_LOG_ERROR("%s: mismatched key type (%d != %d, layer %d)\n", __func__, k_type_i, k_type_i_ref, il);
@@ -2226,7 +2227,7 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
 
         // Read row size of key
         uint64_t k_size_row_ref;
-        io.read_to(&k_size_row_ref, sizeof(k_size_row_ref));
+        io.read(&k_size_row_ref, sizeof(k_size_row_ref));
         const size_t k_size_row = ggml_row_size(k->type, n_embd_k_gqa);
         if (k_size_row != k_size_row_ref) {
             LLAMA_LOG_ERROR("%s: mismatched key row size (%zu != %zu, layer %d)\n", __func__, k_size_row, (size_t) k_size_row_ref, il);
@@ -2236,13 +2237,12 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
         if (cell_count) {
             if (sinfo.is_contiguous()) {
                 // Fast path: contiguous cells, single memcpy
-                ggml_backend_tensor_set(k, io.read(cell_count * k_size_row), sinfo.head() * k_size_row, cell_count * k_size_row);
+                io.read_tensor(k, sinfo.head() * k_size_row, cell_count * k_size_row);
             } else {
                 // Slow path: scatter to non-contiguous positions
-                const void * src = io.read(cell_count * k_size_row);
                 for (uint32_t i = 0; i < cell_count; ++i) {
                     const size_t dst_offset = sinfo.idxs[0][i] * k_size_row;
-                    ggml_backend_tensor_set(k, (const char*)src + i * k_size_row, dst_offset, k_size_row);
+                    io.read_tensor(k, dst_offset, k_size_row);
                 }
             }
         }
@@ -2261,7 +2261,7 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
 
             // Read type of value
             int32_t v_type_i_ref;
-            io.read_to(&v_type_i_ref, sizeof(v_type_i_ref));
+            io.read(&v_type_i_ref, sizeof(v_type_i_ref));
             const int32_t v_type_i = (int32_t) v->type;
             if (v_type_i != v_type_i_ref) {
                 LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il);
@@ -2270,7 +2270,7 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
 
             // Read row size of value
             uint64_t v_size_row_ref;
-            io.read_to(&v_size_row_ref, sizeof(v_size_row_ref));
+            io.read(&v_size_row_ref, sizeof(v_size_row_ref));
             const size_t v_size_row = ggml_row_size(v->type, n_embd_v_gqa);
             if (v_size_row != v_size_row_ref) {
                 LLAMA_LOG_ERROR("%s: mismatched value row size (%zu != %zu, layer %d)\n", __func__, v_size_row, (size_t) v_size_row_ref, il);
@@ -2280,13 +2280,12 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
             if (cell_count) {
                 if (sinfo.is_contiguous()) {
                     // Fast path: contiguous cells, single memcpy
-                    ggml_backend_tensor_set(v, io.read(cell_count * v_size_row), sinfo.head() * v_size_row, cell_count * v_size_row);
+                    io.read_tensor(v, sinfo.head() * v_size_row, cell_count * v_size_row);
                 } else {
                     // Slow path: scatter to non-contiguous positions
-                    const void * src = io.read(cell_count * v_size_row);
                     for (uint32_t i = 0; i < cell_count; ++i) {
                         const size_t dst_offset = sinfo.idxs[0][i] * v_size_row;
-                        ggml_backend_tensor_set(v, (const char*)src + i * v_size_row, dst_offset, v_size_row);
+                        io.read_tensor(v, dst_offset, v_size_row);
                     }
                 }
             }
@@ -2305,7 +2304,7 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
 
             // Read type of value
             int32_t v_type_i_ref;
-            io.read_to(&v_type_i_ref, sizeof(v_type_i_ref));
+            io.read(&v_type_i_ref, sizeof(v_type_i_ref));
             const int32_t v_type_i = (int32_t) v->type;
             if (v_type_i != v_type_i_ref) {
                 LLAMA_LOG_ERROR("%s: mismatched value type (%d != %d, layer %d)\n", __func__, v_type_i, v_type_i_ref, il);
@@ -2314,7 +2313,7 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
 
             // Read element size of value
             uint32_t v_size_el_ref;
-            io.read_to(&v_size_el_ref, sizeof(v_size_el_ref));
+            io.read(&v_size_el_ref, sizeof(v_size_el_ref));
             const size_t v_size_el = ggml_type_size(v->type);
             if (v_size_el != v_size_el_ref) {
                 LLAMA_LOG_ERROR("%s: mismatched value element size (%zu != %zu, layer %d)\n", __func__, v_size_el, (size_t) v_size_el_ref, il);
@@ -2323,7 +2322,7 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
 
             // Read GQA embedding size
             uint32_t n_embd_v_gqa_ref;
-            io.read_to(&n_embd_v_gqa_ref, sizeof(n_embd_v_gqa_ref));
+            io.read(&n_embd_v_gqa_ref, sizeof(n_embd_v_gqa_ref));
             if (n_embd_v_gqa != n_embd_v_gqa_ref) {
                 LLAMA_LOG_ERROR("%s: mismatched GQA embedding size (%u != %u, layer %d)\n", __func__, n_embd_v_gqa, n_embd_v_gqa_ref, il);
                 return false;
@@ -2335,15 +2334,14 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
                     const uint32_t h = sinfo.head();
                     for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
                         const size_t dst_offset = (h + j * cells.size()) * v_size_el;
-                        ggml_backend_tensor_set(v, io.read(cell_count * v_size_el), dst_offset, cell_count * v_size_el);
+                        io.read_tensor(v, dst_offset, cell_count * v_size_el);
                     }
                 } else {
                     // Slow path: scatter to non-contiguous positions
                     for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
-                        const void * src = io.read(cell_count * v_size_el);
                         for (uint32_t i = 0; i < cell_count; ++i) {
                             const size_t dst_offset = (sinfo.idxs[0][i] + j * cells.size()) * v_size_el;
-                            ggml_backend_tensor_set(v, (const char*)src + i * v_size_el, dst_offset, v_size_el);
+                            io.read_tensor(v, dst_offset, v_size_el);
                         }
                     }
                 }

examples/talk-llama/llama-memory-recurrent.cpp

@@ -726,6 +726,10 @@ void llama_memory_recurrent::state_write(llama_io_write_i & io, llama_seq_id seq
         cell_ranges.emplace_back(cell_range_begin, size);
     }
 
+    if (flags % LLAMA_STATE_SEQ_FLAGS_ON_DEVICE && cell_ranges.size() > 1) {
+        GGML_ABORT("cannot save/load multiple ranges of cells to/from device memory\n");
+    }
+
     // DEBUG CHECK: Sum of cell counts in ranges should equal the total cell count
     uint32_t cell_count_check = 0;
     for (const auto & range : cell_ranges) {
@@ -743,7 +747,7 @@ void llama_memory_recurrent::state_read(llama_io_read_i & io, llama_seq_id seq_i
     GGML_UNUSED(flags);
 
     uint32_t cell_count;
-    io.read_to(&cell_count, sizeof(cell_count));
+    io.read(&cell_count, sizeof(cell_count));
 
     bool res = true;
 
@@ -784,7 +788,7 @@ void llama_memory_recurrent::state_write_data(llama_io_write_i & io, const std::
     const uint32_t n_layer = hparams.n_layer;
 
     io.write(&s_trans, sizeof(s_trans));
-    io.write(&n_layer,   sizeof(n_layer));
+    io.write(&n_layer, sizeof(n_layer));
 
     // Iterate and write all the R tensors first, each row is a cell
     // Get whole range at a time
@@ -879,8 +883,8 @@ bool llama_memory_recurrent::state_read_meta(llama_io_read_i & io, uint32_t cell
             llama_pos pos;
             uint32_t n_seq_id;
 
-            io.read_to(&pos,      sizeof(pos));
-            io.read_to(&n_seq_id, sizeof(n_seq_id));
+            io.read(&pos,      sizeof(pos));
+            io.read(&n_seq_id, sizeof(n_seq_id));
 
             if (n_seq_id != 0) {
                 LLAMA_LOG_ERROR("%s: invalid seq_id-agnostic kv cell\n", __func__);
@@ -920,14 +924,14 @@ bool llama_memory_recurrent::state_read_meta(llama_io_read_i & io, uint32_t cell
             llama_pos pos;
             uint32_t  n_seq_id;
 
-            io.read_to(&pos,      sizeof(pos));
-            io.read_to(&n_seq_id, sizeof(n_seq_id));
+            io.read(&pos,      sizeof(pos));
+            io.read(&n_seq_id, sizeof(n_seq_id));
 
             cell.pos = pos;
 
             for (uint32_t j = 0; j < n_seq_id; ++j) {
                 llama_seq_id seq_id;
-                io.read_to(&seq_id, sizeof(seq_id));
+                io.read(&seq_id, sizeof(seq_id));
 
                 if (seq_id < 0 || (uint32_t) seq_id >= this->n_seq_max) {
                     LLAMA_LOG_ERROR("%s: invalid seq_id, %d is out of range [0, %u)\n", __func__, seq_id, this->n_seq_max);
@@ -961,8 +965,8 @@ bool llama_memory_recurrent::state_read_meta(llama_io_read_i & io, uint32_t cell
 bool llama_memory_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell_count) {
     uint32_t s_trans;
     uint32_t n_layer;
-    io.read_to(&s_trans, sizeof(s_trans));
-    io.read_to(&n_layer, sizeof(n_layer));
+    io.read(&s_trans, sizeof(s_trans));
+    io.read(&n_layer, sizeof(n_layer));
 
     if (n_layer != hparams.n_layer) {
         LLAMA_LOG_ERROR("%s: mismatched layer count (%u instead of %u)\n", __func__, n_layer, hparams.n_layer);
@@ -984,7 +988,7 @@ bool llama_memory_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell
 
         // Read type of key
         int32_t r_type_i_ref;
-        io.read_to(&r_type_i_ref, sizeof(r_type_i_ref));
+        io.read(&r_type_i_ref, sizeof(r_type_i_ref));
         const int32_t r_type_i = (int32_t) r_l[il]->type;
         if (r_type_i != r_type_i_ref) {
             LLAMA_LOG_ERROR("%s: mismatched r type (%d != %d, layer %d)\n", __func__, r_type_i, r_type_i_ref, il);
@@ -993,7 +997,7 @@ bool llama_memory_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell
 
         // Read row size of key
         uint64_t r_size_row_ref;
-        io.read_to(&r_size_row_ref, sizeof(r_size_row_ref));
+        io.read(&r_size_row_ref, sizeof(r_size_row_ref));
         const size_t r_size_row = ggml_row_size(r_l[il]->type, hparams.n_embd_r());
         if (r_size_row != r_size_row_ref) {
             LLAMA_LOG_ERROR("%s: mismatched r row size (%zu != %zu, layer %d)\n", __func__, r_size_row, (size_t) r_size_row_ref, il);
@@ -1002,7 +1006,7 @@ bool llama_memory_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell
 
         if (cell_count) {
             // Read and set the keys for the whole cell range
-            ggml_backend_tensor_set(r_l[il], io.read(cell_count * r_size_row), head * r_size_row, cell_count * r_size_row);
+            io.read_tensor(r_l[il], head * r_size_row, cell_count * r_size_row);
         }
     }
 
@@ -1013,7 +1017,7 @@ bool llama_memory_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell
 
             // Read type of value
             int32_t s_type_i_ref;
-            io.read_to(&s_type_i_ref, sizeof(s_type_i_ref));
+            io.read(&s_type_i_ref, sizeof(s_type_i_ref));
             const int32_t s_type_i = (int32_t)s_l[il]->type;
 
             if (s_type_i != s_type_i_ref) {
@@ -1023,7 +1027,7 @@ bool llama_memory_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell
 
             // Read row size of value
             uint64_t s_size_row_ref;
-            io.read_to(&s_size_row_ref, sizeof(s_size_row_ref));
+            io.read(&s_size_row_ref, sizeof(s_size_row_ref));
             const size_t s_size_row = ggml_row_size(s_l[il]->type, hparams.n_embd_s());
             if (s_size_row != s_size_row_ref) {
                 LLAMA_LOG_ERROR("%s: mismatched s row size (%zu != %zu, layer %d)\n", __func__, s_size_row, (size_t) s_size_row_ref, il);
@@ -1032,7 +1036,7 @@ bool llama_memory_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell
 
             if (cell_count) {
                 // Read and set the values for the whole cell range
-                ggml_backend_tensor_set(s_l[il], io.read(cell_count * s_size_row), head * s_size_row, cell_count * s_size_row);
+                io.read_tensor(s_l[il], head * s_size_row, cell_count * s_size_row);
             }
         }
     } else {
@@ -1045,7 +1049,7 @@ bool llama_memory_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell
 
             // Read type of value
             int32_t s_type_i_ref;
-            io.read_to(&s_type_i_ref, sizeof(s_type_i_ref));
+            io.read(&s_type_i_ref, sizeof(s_type_i_ref));
             const int32_t s_type_i = (int32_t)s_l[il]->type;
             if (s_type_i != s_type_i_ref) {
                 LLAMA_LOG_ERROR("%s: mismatched s type (%d != %d, layer %d)\n", __func__, s_type_i, s_type_i_ref, il);
@@ -1054,7 +1058,7 @@ bool llama_memory_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell
 
             // Read element size of value
             uint32_t s_size_el_ref;
-            io.read_to(&s_size_el_ref, sizeof(s_size_el_ref));
+            io.read(&s_size_el_ref, sizeof(s_size_el_ref));
             const size_t s_size_el = ggml_type_size(s_l[il]->type);
             if (s_size_el != s_size_el_ref) {
                 LLAMA_LOG_ERROR("%s: mismatched s element size (%zu != %zu, layer %d)\n", __func__, s_size_el, (size_t) s_size_el_ref, il);
@@ -1063,7 +1067,7 @@ bool llama_memory_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell
 
             // Read state embedding size
             uint32_t n_embd_s_ref;
-            io.read_to(&n_embd_s_ref, sizeof(n_embd_s_ref));
+            io.read(&n_embd_s_ref, sizeof(n_embd_s_ref));
             if (n_embd_s != n_embd_s_ref) {
                 LLAMA_LOG_ERROR("%s: mismatched s embedding size (%u != %u, layer %d)\n", __func__, n_embd_s, n_embd_s_ref, il);
                 return false;
@@ -1073,7 +1077,7 @@ bool llama_memory_recurrent::state_read_data(llama_io_read_i & io, uint32_t cell
                 // For each row in the transposed matrix, read the values for the whole cell range
                 for (uint32_t j = 0; j < n_embd_s; ++j) {
                     const size_t dst_offset = (head + j * size) * s_size_el;
-                    ggml_backend_tensor_set(s_l[il], io.read(cell_count * s_size_el), dst_offset, cell_count * s_size_el);
+                    io.read_tensor(s_l[il], dst_offset, cell_count * s_size_el);
                 }
             }
         }

examples/talk-llama/llama-model-saver.cpp

@@ -268,6 +268,7 @@ void llama_model_saver::add_kv_from_model() {
     // add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN,  ???);
     add_kv(LLM_KV_ATTENTION_SCALE,                   hparams.f_attention_scale);
     add_kv(LLM_KV_ATTENTION_OUTPUT_SCALE,            hparams.f_attn_out_scale);
+    add_kv(LLM_KV_ATTENTION_VALUE_SCALE,             hparams.f_attn_value_scale);
     add_kv(LLM_KV_ATTENTION_TEMPERATURE_LENGTH,      hparams.attn_temp_length);
     add_kv(LLM_KV_ATTENTION_TEMPERATURE_SCALE,       hparams.f_attn_temp_scale);
     add_kv(LLM_KV_ATTENTION_KEY_LENGTH_MLA,          hparams.n_embd_head_k_mla_impl);

examples/talk-llama/llama-model.h

@@ -577,14 +577,8 @@ struct llama_model {
     int64_t t_load_us  = 0;
     int64_t t_start_us = 0;
 
-    explicit llama_model(const struct llama_model_params & params);
-    ~llama_model();
-
-    void load_stats  (llama_model_loader & ml);
-    void load_arch   (llama_model_loader & ml);
-    void load_hparams(llama_model_loader & ml);
-    void load_vocab  (llama_model_loader & ml);
-    bool load_tensors(llama_model_loader & ml); // returns false if cancelled by progress_callback
+    explicit llama_model(const llama_model_params & params);
+    virtual ~llama_model();
 
     std::string arch_name() const;
     std::string type_name() const;
@@ -620,21 +614,94 @@ struct llama_model {
 
     ggml_tensor * get_rope_factors(const llama_cparams & cparams, int il) const;
 
-    // TODO: move this to new llm_arch_model_i interface
     llama_memory_i * create_memory(const llama_memory_params & params, const llama_cparams & cparams) const;
 
-    // TODO: move this to new llm_arch_model_i interface
     ggml_cgraph * build_graph(const llm_graph_params & params) const;
 
-private:
+    virtual void load_stats  (llama_model_loader & ml) = 0;
+    virtual void load_hparams(llama_model_loader & ml) = 0;
+    virtual void load_vocab  (llama_model_loader & ml) = 0;
+    virtual bool load_tensors(llama_model_loader & ml) = 0; // returns false if cancelled by progress_callback
+
+    // model must define these
+    virtual void load_arch_hparams(llama_model_loader & ml) = 0;
+    virtual void load_arch_tensors(llama_model_loader & ml) = 0;
+    virtual std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const = 0;
+
+protected:
     llama_model_params params;
 
     struct impl;
     std::unique_ptr<impl> pimpl;
 };
 
+llama_model * llama_model_create(llm_arch arch, const llama_model_params & params);
+llama_model * llama_model_create(llama_model_loader & ml, const llama_model_params & params);
+
+// model must inherit from this
+struct llama_model_base : public llama_model {
+    friend struct llama_model;
+
+    llama_model * model;
+    llama_model_loader * ml = nullptr;
+    const LLM_TN tn;
+
+    // llama_model_loader is not yet defined at this point, so we will set it after construction
+    const int TENSOR_DUPLICATED;
+    const int TENSOR_NOT_REQUIRED;
+    const int TENSOR_SKIP;
+    const int TENSOR_SKIP_IF_VIRTUAL;
+
+    explicit llama_model_base(const llama_model_params & params);
+    virtual ~llama_model_base() = default;
+
+    ggml_tensor * create_tensor(llama_model_loader & ml, const LLM_TN_IMPL & tn, const std::initializer_list<int64_t> & ne, int flags);
+
+    // convenience overload of create_tensor that doesn't require llama_model_loader
+    ggml_tensor * create_tensor(const LLM_TN_IMPL & tn, const std::initializer_list<int64_t> & ne, int flags);
+
+    // helper: try merged gate_up_exps first, fall back to separate gate and up
+    void create_tensor_gate_up_exps(llama_layer & layer, int bid, int64_t n_embd_,
+                int64_t n_ff_, int64_t n_expert_, int flags);
+
+    // helper: try to load merged qkv first, fall back to separate q, k, v
+    void create_tensor_qkv(llama_layer & layer, int bid,
+                int64_t n_embd_, int64_t n_embd_q_, int64_t n_embd_k_, int64_t n_embd_v_,
+                int flags);
+
+    void load_stats  (llama_model_loader & ml) override;
+    void load_hparams(llama_model_loader & ml) override;
+    void load_vocab  (llama_model_loader & ml) override;
+    bool load_tensors(llama_model_loader & ml) override;
+
+    // model must define these
+    void load_arch_hparams(llama_model_loader & ml) override = 0;
+    void load_arch_tensors(llama_model_loader & ml) override = 0;
+    std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override = 0;
+};
+
 const char * llm_type_name(llm_type type);
 
+// convenience macro for loading local variables for load_tensors() in llama_model_base
+// note: cast to int64_t since we will use these for the tensor dimensions
+#define LLAMA_LOAD_LOCALS \
+    const int     n_layer        = hparams.n_layer;          GGML_UNUSED(n_layer); \
+    const int64_t n_head         = hparams.n_head();         GGML_UNUSED(n_head); \
+    const int64_t n_head_kv      = hparams.n_head_kv();      GGML_UNUSED(n_head_kv); \
+    const int64_t n_embd         = hparams.n_embd;           GGML_UNUSED(n_embd); \
+    const int64_t n_embd_k_gqa   = hparams.n_embd_k_gqa();   GGML_UNUSED(n_embd_k_gqa); \
+    const int64_t n_embd_v_gqa   = hparams.n_embd_v_gqa();   GGML_UNUSED(n_embd_v_gqa); \
+    const int64_t n_embd_head_k  = hparams.n_embd_head_k();  GGML_UNUSED(n_embd_head_k); \
+    const int64_t n_embd_head_v  = hparams.n_embd_head_v();  GGML_UNUSED(n_embd_head_v); \
+    const int64_t n_ff           = hparams.n_ff();           GGML_UNUSED(n_ff); \
+    const int64_t n_embd_gqa     = n_embd_v_gqa;             GGML_UNUSED(n_embd_gqa); \
+    const int64_t n_vocab        = vocab.n_tokens();         GGML_UNUSED(n_vocab); \
+    const int64_t n_token_types  = vocab.n_token_types();    GGML_UNUSED(n_token_types); \
+    const int64_t n_rot          = hparams.n_rot();          GGML_UNUSED(n_rot); \
+    const int64_t n_expert       = hparams.n_expert;         GGML_UNUSED(n_expert); \
+    const int64_t n_expert_used  = hparams.n_expert_used;    GGML_UNUSED(n_expert_used); \
+    const int64_t n_ctx_train    = hparams.n_ctx_train;      GGML_UNUSED(n_ctx_train);
+
 // For internal test use
 // TODO: remove
 const std::vector<std::pair<std::string, ggml_tensor *>> & llama_internal_get_tensor_map(const llama_model * model);

examples/talk-llama/llama-quant.cpp

@@ -882,13 +882,18 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         fname_inp, splits, /*file*/ nullptr, use_mmap, /*use_direct_io*/ false, /*check_tensors*/ true, /*no_alloc*/ false, kv_overrides, nullptr);
     ml.init_mappings(false); // no prefetching
 
-    llama_model model(llama_model_default_params());
+    auto mparams = llama_model_default_params();
+    std::unique_ptr<llama_model> model_ptr(llama_model_create(ml, mparams));
 
-    model.load_arch   (ml);
-    model.load_hparams(ml);
-    model.load_stats  (ml);
+    auto * model = dynamic_cast<llama_model_base *>(model_ptr.get());
+    if (model == nullptr) {
+        GGML_ABORT("fatal error: model does not implement llama_model_base");
+    }
 
-    quantize_state_impl qs(model, params);
+    model->load_hparams(ml);
+    model->load_stats  (ml);
+
+    quantize_state_impl qs(*model, params);
 
     if (params->only_copy) {
         ftype = ml.ftype;
@@ -1023,7 +1028,7 @@ static void llama_model_quantize_impl(const std::string & fname_inp, const std::
         }
         gguf_add_tensor(ctx_outs[i_split].get(), tensor);
 
-        metadata[i].allows_quantization = tensor_allows_quantization(params, model.arch, tensor);
+        metadata[i].allows_quantization = tensor_allows_quantization(params, model->arch, tensor);
 
         if (metadata[i].allows_quantization) {
             metadata[i].target_type = llama_tensor_get_type(qs, params, tensor, default_type, metadata[i]);
@@ -1331,9 +1336,9 @@ void llama_quant_free(quantize_state_impl * qs) {
 
 llama_model * llama_quant_model_from_metadata(const llama_quant_model_desc * desc) {
     struct llama_model_params mparams = llama_model_default_params();
-    auto * model = new llama_model(mparams);
-
-    model->arch = llm_arch_from_string(desc->architecture);
+    auto arch = llm_arch_from_string(desc->architecture);
+    auto * model = llama_model_create(arch, mparams);
+    model->arch = arch;
 
     // infer llm_type: only LLM_TYPE_70B matters for quantization logic
     if (model->arch == LLM_ARCH_LLAMA && desc->n_layer == 80 && desc->n_head != desc->n_head_kv) {

examples/talk-llama/llama-vocab.cpp

@@ -503,6 +503,14 @@ struct llm_tokenizer_bpe : llm_tokenizer {
                 };
                 byte_encode = false; // uses raw UTF-8, not GPT-2 byte encoding
                 break;
+            case LLAMA_VOCAB_PRE_TYPE_SARVAM_MOE:
+                // Sarvam uses SPM-style BPE (same shape as Gemma4): spaces replaced with U+2581
+                // by the normalizer, BPE merges over the whole text on raw UTF-8.
+                regex_exprs = {
+                    "[^\\n]+|[\\n]+",
+                };
+                byte_encode = false;
+                break;
             default:
                 // default regex for BPE tokenization pre-processing
                 regex_exprs = {
@@ -2005,6 +2013,11 @@ void llama_vocab::impl::load(llama_model_loader & ml, const LLM_KV & kv) {
                     tokenizer_pre == "gemma4") {
                 pre_type = LLAMA_VOCAB_PRE_TYPE_GEMMA4;
                 escape_whitespaces = true;
+            } else if (
+                    tokenizer_pre == "sarvam-moe") {
+                pre_type = LLAMA_VOCAB_PRE_TYPE_SARVAM_MOE;
+                escape_whitespaces = true;
+                clean_spaces = false;
             } else if (
                     tokenizer_pre == "jina-v1-en" ||
                     tokenizer_pre == "jina-v2-code" ||

examples/talk-llama/llama-vocab.h

@@ -59,6 +59,7 @@ enum llama_vocab_pre_type {
     LLAMA_VOCAB_PRE_TYPE_JOYAI_LLM       = 48,
     LLAMA_VOCAB_PRE_TYPE_JAIS2           = 49,
     LLAMA_VOCAB_PRE_TYPE_GEMMA4          = 50,
+    LLAMA_VOCAB_PRE_TYPE_SARVAM_MOE      = 51,
 };
 
 struct LLM_KV;

examples/talk-llama/llama.cpp

@@ -71,12 +71,18 @@ bool llama_supports_mlock(void) {
 }
 
 bool llama_supports_gpu_offload(void) {
+    if (!ggml_backend_reg_count()) {
+        ggml_backend_load_all();
+    }
     return ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_GPU) != nullptr ||
            ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_IGPU) != nullptr ||
            llama_supports_rpc();
 }
 
 bool llama_supports_rpc(void) {
+    if (!ggml_backend_reg_count()) {
+        ggml_backend_load_all();
+    }
     return ggml_backend_reg_by_name("RPC") != nullptr;
 }
 
@@ -89,6 +95,10 @@ void llama_backend_init(void) {
         struct ggml_context * ctx = ggml_init(params);
         ggml_free(ctx);
     }
+
+    if (!ggml_backend_reg_count()) {
+        ggml_backend_load_all();
+    }
 }
 
 void llama_numa_init(enum ggml_numa_strategy numa) {
@@ -111,113 +121,8 @@ int64_t llama_time_us(void) {
     return ggml_time_us();
 }
 
-// Returns 0 on success, -1 on error, and -2 on cancellation via llama_progress_callback
-static int llama_model_load(struct gguf_context * metadata, llama_model_set_tensor_data_t set_tensor_data, void * set_tensor_data_ud,
-        const std::string & fname, std::vector<std::string> & splits, FILE * file, llama_model & model, llama_model_params & params) {
-    // loading time will be recalculated after the first eval, so
-    // we take page faults deferred by mmap() into consideration
-    model.t_load_us = 0;
-    time_meas tm(model.t_load_us);
-
-    model.t_start_us = tm.t_start_us;
-
-    try {
-        llama_model_loader ml(metadata, set_tensor_data, set_tensor_data_ud, fname, splits, file, params.use_mmap, params.use_direct_io,
-            params.check_tensors, params.no_alloc, params.kv_overrides, params.tensor_buft_overrides);
-
-        ml.print_info();
-
-        model.hparams.vocab_only = params.vocab_only;
-        model.hparams.no_alloc   = params.no_alloc;
-
-        try {
-            model.load_arch(ml);
-        } catch(const std::exception & e) {
-            throw std::runtime_error("error loading model architecture: " + std::string(e.what()));
-        }
-        try {
-            model.load_hparams(ml);
-        } catch(const std::exception & e) {
-            throw std::runtime_error("error loading model hyperparameters: " + std::string(e.what()));
-        }
-        if (model.arch == LLM_ARCH_CLIP) {
-            throw std::runtime_error("CLIP cannot be used as main model, use it with --mmproj instead");
-        }
-        try {
-            model.load_vocab(ml);
-        } catch(const std::exception & e) {
-            throw std::runtime_error("error loading model vocabulary: " + std::string(e.what()));
-        }
-
-        model.load_stats(ml);
-        model.print_info();
-
-        if (params.vocab_only) {
-            LLAMA_LOG_INFO("%s: vocab only - skipping tensors\n", __func__);
-            return 0;
-        }
-
-        if (!model.load_tensors(ml)) {
-            return -2;
-        }
-    } catch (const std::exception & err) {
-        LLAMA_LOG_ERROR("%s: error loading model: %s\n", __func__, err.what());
-        return -1;
-    }
-
-    return 0;
-}
-
-static struct llama_model * llama_model_load_from_file_impl(
-        struct gguf_context * metadata,
-        llama_model_set_tensor_data_t set_tensor_data,
-        void * set_tensor_data_ud,
-        const std::string & path_model,
-        std::vector<std::string> & splits,
-        FILE * file,
-        struct llama_model_params params) {
-    {
-        int n_sources_defined = 0;
-        if (metadata != nullptr) {
-            n_sources_defined++;
-        }
-        if (!path_model.empty()) {
-            n_sources_defined++;
-        }
-        if (file != nullptr) {
-            n_sources_defined++;
-        }
-        if (n_sources_defined != 1) {
-            LLAMA_LOG_ERROR("%s: exactly one out metadata, path_model, and file must be defined\n", __func__);
-            return nullptr;
-        }
-    }
-    ggml_time_init();
-
-    if (!params.vocab_only && ggml_backend_reg_count() == 0) {
-        LLAMA_LOG_ERROR("%s: no backends are loaded. hint: use ggml_backend_load() or ggml_backend_load_all() to load a backend before calling this function\n", __func__);
-        return nullptr;
-    }
-
-    unsigned cur_percentage = 0;
-    if (params.progress_callback == NULL) {
-        params.progress_callback_user_data = &cur_percentage;
-        params.progress_callback = [](float progress, void * ctx) {
-            unsigned * cur_percentage_p = (unsigned *) ctx;
-            unsigned percentage = (unsigned) (100 * progress);
-            while (percentage > *cur_percentage_p) {
-                *cur_percentage_p = percentage;
-                LLAMA_LOG_CONT(".");
-                if (percentage >= 100) {
-                    LLAMA_LOG_CONT("\n");
-                }
-            }
-            return true;
-        };
-    }
-
-    llama_model * model = new llama_model(params);
-
+// returns true on success
+static bool llama_prepare_model_devices(const llama_model_params & params, llama_model * model) {
     // create list of devices to use with this model
     if (params.devices) {
         if (params.split_mode == LLAMA_SPLIT_MODE_TENSOR) {
@@ -227,7 +132,7 @@ static struct llama_model * llama_model_load_from_file_impl(
             }
             if (n_devs == 0) {
                 LLAMA_LOG_ERROR("%s: LLAMA_SPLIT_MODE_TENSOR needs >= 1 devices\n", __func__);
-                return nullptr;
+                return false;
             }
             LLAMA_LOG_INFO("%s: creating a Meta device with %zu devices\n", __func__, n_devs);
             for (size_t i = 0; i < n_devs; ++i) {
@@ -265,7 +170,7 @@ static struct llama_model * llama_model_load_from_file_impl(
             }
             if (devs.empty()) {
                 LLAMA_LOG_ERROR("%s: LLAMA_SPLIT_MODE_TENSOR needs >= 1 devices\n", __func__);
-                return nullptr;
+                return false;
             }
 
             LLAMA_LOG_INFO("%s: creating a Meta device for tensor parallelism from %zu devices:\n", __func__, devs.size());
@@ -347,8 +252,7 @@ static struct llama_model * llama_model_load_from_file_impl(
         } else {
             if (params.main_gpu >= (int)model->devices.size()) {
                 LLAMA_LOG_ERROR("%s: invalid value for main_gpu: %d (available devices: %zu)\n", __func__, params.main_gpu, model->devices.size());
-                llama_model_free(model);
-                return nullptr;
+                return false;
             }
             llama_device main_gpu = model->devices[params.main_gpu];
             model->devices.clear();
@@ -365,7 +269,121 @@ static struct llama_model * llama_model_load_from_file_impl(
                 props.memory_free/1024/1024);
     }
 
-    const int status = llama_model_load(metadata, set_tensor_data, set_tensor_data_ud, path_model, splits, file, *model, params);
+    return true;
+}
+
+// Returns 0 on success, -1 on error, and -2 on cancellation via llama_progress_callback
+static std::pair<int, llama_model *> llama_model_load(struct gguf_context * metadata, llama_model_set_tensor_data_t set_tensor_data, void * set_tensor_data_ud,
+        const std::string & fname, std::vector<std::string> & splits, FILE * file, llama_model_params & params) {
+    try {
+        llama_model_loader ml(metadata, set_tensor_data, set_tensor_data_ud, fname, splits, file, params.use_mmap, params.use_direct_io,
+            params.check_tensors, params.no_alloc, params.kv_overrides, params.tensor_buft_overrides);
+
+        ml.print_info();
+        std::unique_ptr<llama_model> model_ptr(llama_model_create(ml, params));
+
+        bool ok = llama_prepare_model_devices(params, model_ptr.get());
+        if (!ok) {
+            return {-1, nullptr};
+        }
+
+        auto * model = dynamic_cast<llama_model_base *>(model_ptr.get());
+        if (model == nullptr) {
+            GGML_ABORT("fatal error: model does not implement llama_model_base");
+        }
+
+        // loading time will be recalculated after the first eval, so
+        // we take page faults deferred by mmap() into consideration
+        model->t_load_us = 0;
+        time_meas tm(model->t_load_us);
+
+        model->t_start_us = tm.t_start_us;
+
+        model->hparams.vocab_only = params.vocab_only;
+        model->hparams.no_alloc   = params.no_alloc;
+
+        try {
+            model->load_hparams(ml);
+        } catch(const std::exception & e) {
+            throw std::runtime_error("error loading model hyperparameters: " + std::string(e.what()));
+        }
+        if (model->arch == LLM_ARCH_CLIP) {
+            throw std::runtime_error("CLIP cannot be used as main model, use it with --mmproj instead");
+        }
+        try {
+            model->load_vocab(ml);
+        } catch(const std::exception & e) {
+            throw std::runtime_error("error loading model vocabulary: " + std::string(e.what()));
+        }
+
+        model->load_stats(ml);
+        model->print_info();
+
+        if (params.vocab_only) {
+            LLAMA_LOG_INFO("%s: vocab only - skipping tensors\n", __func__);
+            return {0, model_ptr.release()};
+        }
+
+        if (!model->load_tensors(ml)) {
+            return {-2, nullptr};
+        }
+
+        return {0, model_ptr.release()};
+    } catch (const std::exception & err) {
+        LLAMA_LOG_ERROR("%s: error loading model: %s\n", __func__, err.what());
+        return {-1, nullptr};
+    }
+}
+
+static struct llama_model * llama_model_load_from_file_impl(
+        struct gguf_context * metadata,
+        llama_model_set_tensor_data_t set_tensor_data,
+        void * set_tensor_data_ud,
+        const std::string & path_model,
+        std::vector<std::string> & splits,
+        FILE * file,
+        struct llama_model_params params) {
+    {
+        int n_sources_defined = 0;
+        if (metadata != nullptr) {
+            n_sources_defined++;
+        }
+        if (!path_model.empty()) {
+            n_sources_defined++;
+        }
+        if (file != nullptr) {
+            n_sources_defined++;
+        }
+        if (n_sources_defined != 1) {
+            LLAMA_LOG_ERROR("%s: exactly one out metadata, path_model, and file must be defined\n", __func__);
+            return nullptr;
+        }
+    }
+    ggml_time_init();
+
+    if (!params.vocab_only && ggml_backend_reg_count() == 0) {
+        LLAMA_LOG_ERROR("%s: no backends are loaded. hint: use ggml_backend_load() or ggml_backend_load_all() to load a backend before calling this function\n", __func__);
+        return nullptr;
+    }
+
+    unsigned cur_percentage = 0;
+    if (params.progress_callback == NULL) {
+        params.progress_callback_user_data = &cur_percentage;
+        params.progress_callback = [](float progress, void * ctx) {
+            unsigned * cur_percentage_p = (unsigned *) ctx;
+            unsigned percentage = (unsigned) (100 * progress);
+            while (percentage > *cur_percentage_p) {
+                *cur_percentage_p = percentage;
+                LLAMA_LOG_CONT(".");
+                if (percentage >= 100) {
+                    LLAMA_LOG_CONT("\n");
+                }
+            }
+            return true;
+        };
+    }
+
+    const auto [status, model] = llama_model_load(metadata, set_tensor_data, set_tensor_data_ud, path_model, splits, file, params);
     GGML_ASSERT(status <= 0);
     if (status < 0) {
         if (status == -1) {
@@ -374,7 +392,9 @@ static struct llama_model * llama_model_load_from_file_impl(
             LLAMA_LOG_INFO("%s: cancelled model load\n", __func__);
         }
 
-        llama_model_free(model);
+        if (model) {
+            llama_model_free(model);
+        }
         return nullptr;
     }
 

examples/talk-llama/llama.h

@@ -864,6 +864,9 @@ extern "C" {
 // work only with partial states, such as SWA KV cache or recurrent cache (e.g. Mamba)
 #define LLAMA_STATE_SEQ_FLAGS_PARTIAL_ONLY 1
 
+// keeps the tensor data on device buffers (i.e. not accessible in host memory, but faster save/load)
+#define LLAMA_STATE_SEQ_FLAGS_ON_DEVICE 2
+
     typedef uint32_t llama_state_seq_flags;
 
     LLAMA_API size_t llama_state_seq_get_size_ext(

examples/talk-llama/models/afmoe.cpp

@@ -1,6 +1,112 @@
 #include "models.h"
 
-llm_build_afmoe::llm_build_afmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_afmoe::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
+    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);
+    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa, false);
+
+    // Set up interleaved sliding window attention (ISWA)
+    // Pattern: 3 sliding - 1 full (global_attn_every_n_layers = 4)
+    if (hparams.n_swa > 0) {
+        hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+        uint32_t swa_period = 4;
+        ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
+        hparams.set_swa_pattern(swa_period);
+
+        hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+        hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
+        ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
+    } else {
+        hparams.swa_type = LLAMA_SWA_TYPE_NONE;
+    }
+
+    // Default to sigmoid if not set
+    if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {
+        hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;
+    }
+
+    switch (hparams.n_layer) {
+        case 56: type = LLM_TYPE_6B; break;
+        case 32: type = LLM_TYPE_26B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_afmoe::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+    const int64_t n_expert_shared = hparams.n_expert_shared;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    const int64_t n_ff_exp = hparams.n_ff_exp;
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        // dual attention normalization
+        layer.attn_norm      = create_tensor(tn(LLM_TENSOR_ATTN_NORM,      "weight", i), {n_embd}, 0);
+        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+
+        // attention projections
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        // Q/K normalization
+        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
+        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
+
+        // attention gating
+        layer.wqkv_gate = create_tensor(tn(LLM_TENSOR_ATTN_GATE, "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+
+        // dual ffn normalization
+        layer.ffn_norm      = create_tensor(tn(LLM_TENSOR_FFN_NORM,      "weight", i), {n_embd}, 0);
+        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
+
+        if (static_cast<uint32_t>(i) >= hparams.n_layer_dense_lead) {
+            // MoE layers
+            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, 0);
+
+            // grouped expert weights
+            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, 0);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, 0);
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff_exp, n_expert}, 0);
+
+            // shared expert
+            if (n_expert_shared > 0) {
+                const int64_t n_ff_shexp = n_ff_exp * n_expert_shared;
+                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp}, 0);
+                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, 0);
+                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, 0);
+            }
+        } else {
+            // Dense layers
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_afmoe::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_afmoe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
 

examples/talk-llama/models/apertus.cpp

@@ -1,6 +1,62 @@
 #include "models.h"
 
-llm_build_apertus::llm_build_apertus(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_apertus::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key_or_arr(LLM_KV_XIELU_ALPHA_N,        hparams.xielu_alpha_n, hparams.n_layer);
+    ml.get_key_or_arr(LLM_KV_XIELU_ALPHA_P,        hparams.xielu_alpha_p, hparams.n_layer);
+    ml.get_key_or_arr(LLM_KV_XIELU_BETA,           hparams.xielu_beta,    hparams.n_layer);
+    ml.get_key_or_arr(LLM_KV_XIELU_EPS,            hparams.xielu_eps,     hparams.n_layer);
+
+    switch (hparams.n_layer) {
+        case 32: type = LLM_TYPE_8B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_apertus::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), { n_embd, n_vocab }, 0);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);
+
+        if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
+            layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), { n_rot/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+            layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), { n_rot/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+        } else {
+            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), { n_rot/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+        }
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0);
+
+        // optional bias tensors
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), { n_embd }, TENSOR_NOT_REQUIRED);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd }, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }, 0);
+        layer.ffn_up = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, n_ff }, 0);
+
+        // Q and K layernorms for Apertus
+        layer.attn_q_norm   = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), { n_embd_head_k }, 0);
+        layer.attn_q_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "bias",   i), { n_embd_head_k }, TENSOR_NOT_REQUIRED);
+        layer.attn_k_norm   = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), { n_embd_head_k }, 0);
+        layer.attn_k_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "bias",   i), { n_embd_head_k }, TENSOR_NOT_REQUIRED);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_apertus::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_apertus::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/arcee.cpp

@@ -1,6 +1,51 @@
 #include "models.h"
 
-llm_build_arcee::llm_build_arcee(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_arcee::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    // Arcee uses the same structure as Llama
+    switch (hparams.n_layer) {
+        case 36: type = LLM_TYPE_4B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_arcee::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_arcee::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_arcee::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/arctic.cpp

@@ -1,6 +1,59 @@
 #include "models.h"
 
-llm_build_arctic::llm_build_arctic(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_arctic::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    if (hparams.n_expert == 128) {
+        switch (hparams.n_layer) {
+            case 35: type = LLM_TYPE_10B_128x3_66B; break;
+            default: type = LLM_TYPE_UNKNOWN;
+        }
+    } else {
+        type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_arctic::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_embd}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_embd, n_embd}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_embd}, 0);
+
+        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+        layer.ffn_norm_exps = create_tensor(tn(LLM_TENSOR_FFN_NORM_EXPS, "weight", i), {n_embd}, 0);
+        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, false);
+        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);
+        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_arctic::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_arctic::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/arwkv7.cpp

@@ -1,7 +1,123 @@
 #include "models.h"
 
+void llama_model_arwkv7::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,                hparams.f_norm_eps, false);
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,            hparams.f_norm_rms_eps, false);
+    ml.get_key(LLM_KV_WKV_HEAD_SIZE,                          hparams.wkv_head_size);
+    ml.get_key(LLM_KV_ATTENTION_DECAY_LORA_RANK,              hparams.n_lora_decay);
+    ml.get_key(LLM_KV_ATTENTION_ICLR_LORA_RANK,               hparams.n_lora_iclr);
+    ml.get_key(LLM_KV_ATTENTION_VALUE_RESIDUAL_MIX_LORA_RANK, hparams.n_lora_value_res_mix);
+    ml.get_key(LLM_KV_ATTENTION_GATE_LORA_RANK,               hparams.n_lora_gate, false);
+    ml.get_key(LLM_KV_TOKEN_SHIFT_COUNT,                      hparams.token_shift_count, false);
 
-llm_build_arwkv7::llm_build_arwkv7(const llama_model & model, const llm_graph_params & params) : llm_build_rwkv7_base(model, params) {
+    switch (hparams.n_layer) {
+        case 12:
+            switch (hparams.n_embd) {
+                case 768: type = LLM_TYPE_190M; break;
+                default: type = LLM_TYPE_UNKNOWN;
+            } break;
+        case 24:
+            switch (hparams.n_embd) {
+                case 1024: type = LLM_TYPE_450M; break;
+                case 2048: type = LLM_TYPE_1_5B; break;
+                default: type = LLM_TYPE_UNKNOWN;
+            } break;
+        case 28:
+            switch (hparams.n_embd) {
+                case 1536: type = LLM_TYPE_1_5B; break;
+                case 3584: type = LLM_TYPE_7B; break;
+                default: type = LLM_TYPE_UNKNOWN;
+            } break;
+        case 32:
+            switch (hparams.n_embd) {
+                case 2560: type = LLM_TYPE_2_9B; break;
+                case 4096: type = LLM_TYPE_7B; break;
+                default: type = LLM_TYPE_UNKNOWN;
+            } break;
+        case 61:
+            switch (hparams.n_embd) {
+                case 4096: type = LLM_TYPE_14B; break;
+                default: type = LLM_TYPE_UNKNOWN;
+            } break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_arwkv7::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0);
+
+    const int n_lora_decay = hparams.n_lora_decay;
+    const int n_lora_iclr = hparams.n_lora_iclr;
+    const int n_lora_value_res_mix = hparams.n_lora_value_res_mix;
+    const int n_lora_gate = hparams.n_lora_gate;
+    const int attn_hidden_size = n_embd;
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.time_mix_w0 = create_tensor(tn(LLM_TENSOR_TIME_MIX_W0, "weight", i), {n_embd}, 0);
+        layer.time_mix_w1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_W1, "weight", i), {n_embd, n_lora_decay}, 0);
+        layer.time_mix_w2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_W2, "weight", i), {n_lora_decay, n_embd}, 0);
+
+        layer.time_mix_a0 = create_tensor(tn(LLM_TENSOR_TIME_MIX_A0, "weight", i), {n_embd}, 0);
+        layer.time_mix_a1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_A1, "weight", i), {n_embd, n_lora_iclr}, 0);
+        layer.time_mix_a2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_A2, "weight", i), {n_lora_iclr, n_embd}, 0);
+
+        if (i == 0) {
+            // actually not used
+            layer.time_mix_v0 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V0, "weight", i), {n_embd}, 0);
+            layer.time_mix_v1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V1, "weight", i), {n_embd, n_lora_iclr}, 0);
+            layer.time_mix_v2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V2, "weight", i), {n_lora_iclr, n_embd}, 0);
+        } else {
+            layer.time_mix_v0 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V0, "weight", i), {n_embd}, 0);
+            layer.time_mix_v1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V1, "weight", i), {n_embd, n_lora_value_res_mix}, 0);
+            layer.time_mix_v2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_V2, "weight", i), {n_lora_value_res_mix, n_embd}, 0);
+        }
+
+        layer.time_mix_g1 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G1, "weight", i), {n_embd, n_lora_gate}, TENSOR_NOT_REQUIRED);
+        layer.time_mix_g2 = create_tensor(tn(LLM_TENSOR_TIME_MIX_G2, "weight", i), {n_lora_gate, n_embd}, TENSOR_NOT_REQUIRED);
+
+        try {
+            layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 6}, 0);
+        } catch(std::runtime_error & e) {
+            // ARWKV models may not have gate tensors
+            layer.time_mix_lerp_fused = create_tensor(tn(LLM_TENSOR_TIME_MIX_LERP_FUSED, "weight", i), {n_embd, 1, 1, 5}, 0);
+        }
+
+        layer.time_mix_k_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_K_K, "weight", i), {attn_hidden_size}, 0);
+        layer.time_mix_k_a = create_tensor(tn(LLM_TENSOR_TIME_MIX_K_A, "weight", i), {attn_hidden_size}, 0);
+        layer.time_mix_r_k = create_tensor(tn(LLM_TENSOR_TIME_MIX_R_K, "weight", i), {attn_hidden_size}, 0);
+
+        layer.time_mix_key = create_tensor(tn(LLM_TENSOR_TIME_MIX_KEY, "weight", i), {attn_hidden_size, n_embd}, 0);
+        layer.time_mix_value = create_tensor(tn(LLM_TENSOR_TIME_MIX_VALUE, "weight", i), {attn_hidden_size, n_embd}, 0);
+        layer.time_mix_receptance = create_tensor(tn(LLM_TENSOR_TIME_MIX_RECEPTANCE, "weight", i), {attn_hidden_size, n_embd}, 0);
+
+        layer.time_mix_ln = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
+        layer.time_mix_ln_b = create_tensor(tn(LLM_TENSOR_TIME_MIX_LN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+        layer.time_mix_output = create_tensor(tn(LLM_TENSOR_TIME_MIX_OUTPUT, "weight", i), {n_embd, attn_hidden_size}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+    }
+
+}
+
+std::unique_ptr<llm_graph_context> llama_model_arwkv7::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_arwkv7::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_build_rwkv7_base(model, params) {
     GGML_ASSERT(n_embd == hparams.n_embd_r());
 
     ggml_tensor * cur;

examples/talk-llama/models/baichuan.cpp

@@ -1,6 +1,49 @@
 #include "models.h"
 
-llm_build_baichuan::llm_build_baichuan(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_baichuan::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    switch (hparams.n_layer) {
+        case 32: type = LLM_TYPE_7B; break;
+        case 40: type = LLM_TYPE_13B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+
+    if (type == LLM_TYPE_13B) {
+        // TODO: become GGUF KV parameter
+        hparams.f_max_alibi_bias = 8.0f;
+    }
+}
+
+void llama_model_baichuan::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+    {
+        output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+        output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_baichuan::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_baichuan::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/bailingmoe.cpp

@@ -1,6 +1,65 @@
 #include "models.h"
 
-llm_build_bailingmoe::llm_build_bailingmoe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_bailingmoe::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);
+
+    switch (hparams.n_layer) {
+        case 28: type = LLM_TYPE_16B; break;
+        case 88: type = LLM_TYPE_290B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_bailingmoe::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+    const int64_t n_expert_shared = hparams.n_expert_shared;
+
+    const int64_t n_ff_exp            = hparams.n_ff_exp;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_head * n_rot, n_head_kv * n_rot, n_head_kv * n_rot, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_rot, n_embd}, 0);
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+
+        if (n_expert == 0) {
+            throw std::runtime_error("n_expert must be > 0");
+        }
+        if (n_expert_used == 0) {
+            throw std::runtime_error("n_expert_used must be > 0");
+        }
+
+        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
+        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+
+        layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+        layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd}, 0);
+        layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_bailingmoe::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_bailingmoe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     ggml_tensor * cur;
     ggml_tensor * inpL;
 

examples/talk-llama/models/bailingmoe2.cpp

@@ -1,6 +1,100 @@
 #include "models.h"
 
-llm_build_bailingmoe2::llm_build_bailingmoe2(const llama_model & model, const llm_graph_params & params) :
+void llama_model_bailingmoe2::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead, false);
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
+    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
+    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,               hparams.expert_weights_norm, false);
+    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,                hparams.expert_gating_func);
+    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS,              hparams.nextn_predict_layers, false);
+    GGML_ASSERT(hparams.nextn_predict_layers < hparams.n_layer && "nextn_predict_layers must be < n_layer");
+
+    // TODO: when MTP is implemented, this should probably be updated if needed
+    hparams.n_layer_kv_from_start = hparams.n_layer - hparams.nextn_predict_layers;
+
+    switch (hparams.n_layer) {
+        case 20: type = LLM_TYPE_16B_A1B; break;
+        case 21: type = LLM_TYPE_16B_A1B; break;
+        case 32: type = LLM_TYPE_100B_A6B; break;
+        case 33: type = LLM_TYPE_100B_A6B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_bailingmoe2::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+    const int64_t n_expert_shared = hparams.n_expert_shared;
+
+    const int64_t n_ff_exp        = hparams.n_ff_exp;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+    GGML_ASSERT(n_expert > 0 && "n_expert must be > 0 for bailingmoe2");
+    GGML_ASSERT(n_expert_used > 0 && "n_expert_used must be > 0 for bailingmoe2");
+
+    for (int i = 0; i < n_layer; ++i) {
+        int flags = 0;
+        if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+            // skip all tensors in the NextN layers
+            flags |= TENSOR_SKIP;
+        }
+
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, flags);
+
+        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, flags);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, flags);
+
+        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, flags);
+        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, flags);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);
+
+        if (static_cast<uint32_t>(i) >= hparams.n_layer_dense_lead) { // MoE layers
+            const int64_t n_ff_shexp = (hparams.n_ff_shexp ? hparams.n_ff_shexp : n_ff_exp) * n_expert_shared;
+
+            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, flags);
+            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED | flags);
+
+            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, flags);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, flags);
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, flags);
+
+            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp}, flags);
+            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, flags);
+            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, flags);
+        } else { // Dense layers
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, flags);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, flags);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, flags);
+        }
+
+        // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
+        if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags);
+            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED | flags);
+            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags);
+            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags);
+            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED | flags);
+            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, TENSOR_NOT_REQUIRED | flags);
+            layer.layer_out_norm         = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, flags);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_bailingmoe2::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_bailingmoe2::graph::graph(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 

examples/talk-llama/models/bert.cpp

@@ -1,6 +1,83 @@
 #include "models.h"
 
-llm_build_bert::llm_build_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_bert::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
+
+    switch (hparams.n_layer) {
+        case 3:
+            type = LLM_TYPE_17M; break; // bge-micro
+        case 6:
+            type = LLM_TYPE_22M; break; // MiniLM-L6
+        case 12:
+            switch (hparams.n_embd) {
+                case 384: type = LLM_TYPE_33M; break; // MiniLM-L12, bge-small
+                case 768: type = LLM_TYPE_109M; break; // bge-base
+                default: type = LLM_TYPE_UNKNOWN;
+            } break;
+        case 24:
+            type = LLM_TYPE_335M; break; // bge-large
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_bert::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    if (n_token_types == 0) {
+        throw std::runtime_error(arch_name() + " model needs to define token type count");
+    }
+    tok_embd     = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);
+    type_embd    = create_tensor(tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_token_types}, TENSOR_NOT_REQUIRED);
+
+    if (arch == LLM_ARCH_BERT) {
+        pos_embd = create_tensor(tn(LLM_TENSOR_POS_EMBD,    "weight"), {n_embd, n_ctx_train}, 0);
+
+        cls   = create_tensor(tn(LLM_TENSOR_CLS, "weight"), {n_embd, n_embd}, TENSOR_NOT_REQUIRED);
+        cls_b = create_tensor(tn(LLM_TENSOR_CLS, "bias"),   {n_embd},         TENSOR_NOT_REQUIRED);
+
+        cls_out   = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
+        cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT, "bias"),   {hparams.n_cls_out},         TENSOR_NOT_REQUIRED);
+    }
+
+    tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
+    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias",   0), {n_embd}, 0);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
+
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+
+        layer.attn_out_norm   = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}, 0);
+        layer.attn_out_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i),   {n_embd}, 0);
+
+        if (hparams.moe_every_n_layers > 0 && i % hparams.moe_every_n_layers == 1) {
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff,   n_expert}, 0);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff,   n_embd, n_expert}, 0);
+            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,   "weight", i), {n_embd, n_expert}, 0);
+        } else {
+            layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, TENSOR_NOT_REQUIRED);
+            layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);
+
+            if (arch == LLM_ARCH_NOMIC_BERT) {
+                layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+            }
+        }
+
+        layer.layer_out_norm   = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, 0);
+        layer.layer_out_norm_b = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "bias", i),   {n_embd}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_bert::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_bert::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/bitnet.cpp

@@ -1,7 +1,54 @@
 #include "models.h"
 
+void llama_model_bitnet::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
 
-llm_build_bitnet::llm_build_bitnet(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    switch (hparams.n_layer) {
+        case 26: type = LLM_TYPE_3B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_bitnet::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm     = create_tensor(tn(LLM_TENSOR_ATTN_NORM,     "weight", i), {n_embd}, 0);
+        layer.attn_sub_norm = create_tensor(tn(LLM_TENSOR_ATTN_SUB_NORM, "weight", i), {n_embd}, 0);
+
+        layer.wq       = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd}, 0);
+        layer.wq_s     = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "scale",  i), {1}, TENSOR_NOT_REQUIRED);
+        layer.wk       = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_gqa}, 0);
+        layer.wk_s     = create_tensor(tn(LLM_TENSOR_ATTN_K,   "scale",  i), {1}, TENSOR_NOT_REQUIRED);
+        layer.wv       = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_gqa}, 0);
+        layer.wv_s     = create_tensor(tn(LLM_TENSOR_ATTN_V,   "scale",  i), {1}, TENSOR_NOT_REQUIRED);
+        layer.wo       = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+        layer.wo_s     = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "scale",  i), {1}, TENSOR_NOT_REQUIRED);
+
+        layer.ffn_norm     = create_tensor(tn(LLM_TENSOR_FFN_NORM,     "weight", i), {n_embd}, 0);
+        layer.ffn_sub_norm = create_tensor(tn(LLM_TENSOR_FFN_SUB_NORM, "weight", i), {n_ff}, 0);
+
+        layer.ffn_gate       = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+        layer.ffn_gate_s = create_tensor(tn(LLM_TENSOR_FFN_GATE, "scale",  i), {1}, TENSOR_NOT_REQUIRED);
+        layer.ffn_down       = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+        layer.ffn_down_s = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "scale",  i), {1}, TENSOR_NOT_REQUIRED);
+        layer.ffn_up         = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+        layer.ffn_up_s   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "scale",  i), {1}, TENSOR_NOT_REQUIRED);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_bitnet::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_bitnet::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/bloom.cpp

@@ -1,6 +1,68 @@
 #include "models.h"
 
-llm_build_bloom::llm_build_bloom(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_bloom::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+
+    switch (hparams.n_layer) {
+        case 24: type = LLM_TYPE_1B; break;
+        case 30:
+            switch (hparams.n_embd) {
+                case 2560: type = LLM_TYPE_3B; break;
+                case 4096: type = LLM_TYPE_7B; break;
+                default: type = LLM_TYPE_UNKNOWN;
+            } break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+
+    // TODO: become GGUF KV parameter
+    hparams.f_max_alibi_bias = 8.0f;
+}
+
+void llama_model_bloom::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,      "weight"), {n_embd, n_vocab}, 0);
+    tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
+    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias",   0), {n_embd}, 0);
+
+    // output
+    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);
+    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias",   i), {n_embd}, 0);
+
+        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
+        layer.wqkv_b = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, 0);
+
+        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias",   i), {n_embd}, 0);
+
+        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias",   i), {n_embd}, 0);
+
+        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias",   i), {n_embd}, 0);
+
+        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}, 0);
+        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias",   i), {n_ff}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_bloom::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_bloom::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/chameleon.cpp

@@ -1,8 +1,56 @@
 #include "models.h"
-
 #include <float.h>
 
-llm_build_chameleon::llm_build_chameleon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_chameleon::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    hparams.f_norm_eps = 1e-5;  // eps for qk-norm, torch default
+    ml.get_key(LLM_KV_SWIN_NORM, hparams.swin_norm, false);
+
+    switch (hparams.n_layer) {
+        case 32: type = LLM_TYPE_7B; break;
+        case 48: type = LLM_TYPE_34B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+   }
+}
+
+void llama_model_chameleon::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k, n_head}, 0);
+        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k, n_head_kv}, 0);
+        layer.attn_q_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "bias", i),  {n_embd_head_k, n_head}, TENSOR_NOT_REQUIRED);
+        layer.attn_k_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "bias", i),  {n_embd_head_k, n_head_kv}, TENSOR_NOT_REQUIRED);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_chameleon::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_chameleon::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/chatglm.cpp

@@ -1,7 +1,60 @@
 #include "models.h"
 
+void llama_model_chatglm::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    switch (hparams.n_layer) {
+        case 28: {
+            if (hparams.n_head(0) == 16) {
+                type = LLM_TYPE_1_5B;
+            } else {
+                type = LLM_TYPE_6B;
+            }
+        } break;
+        case 40: {
+            if (hparams.n_head(0) == 24) {
+                type = LLM_TYPE_4B;
+            } else {
+                type = LLM_TYPE_9B;
+            }
+        } break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
 
-llm_build_chatglm::llm_build_chatglm(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_chatglm::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,      "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+
+        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff * 2}, 0);
+
+        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_chatglm::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_chatglm::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/codeshell.cpp

@@ -1,6 +1,55 @@
 #include "models.h"
 
-llm_build_codeshell::llm_build_codeshell(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_codeshell::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+    switch (hparams.n_layer) {
+        case 42: type = LLM_TYPE_7B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_codeshell::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    // if tok embd is NULL, init from output
+    if (tok_embd == NULL) {
+        tok_embd = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    // output
+    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);
+    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
+
+        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd}, 0);
+
+        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, 0);
+
+        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, 0);
+
+        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i),   {n_embd, n_ff}, 0);
+        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias", i),     {n_ff}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_codeshell::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_codeshell::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/cogvlm.cpp

@@ -1,6 +1,55 @@
 #include "models.h"
 
-llm_build_cogvlm::llm_build_cogvlm(const llama_model & model, const llm_graph_params & params) :
+void llama_model_cogvlm::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    switch (hparams.n_layer) {
+        case 32: type = LLM_TYPE_13B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_cogvlm::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd_head_k * n_head * 3}, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        layer.visexp_attn_wqkv = create_tensor(tn(LLM_TENSOR_VISEXP_ATTN_QKV, "weight", i), {n_embd, n_embd_head_k * n_head * 3}, 0);
+        layer.visexp_attn_wo = create_tensor(tn(LLM_TENSOR_VISEXP_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+
+        layer.visexp_ffn_gate = create_tensor(tn(LLM_TENSOR_VISEXP_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+        layer.visexp_ffn_down = create_tensor(tn(LLM_TENSOR_VISEXP_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.visexp_ffn_up   = create_tensor(tn(LLM_TENSOR_VISEXP_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_cogvlm::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_cogvlm::graph::graph(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
     const float   kq_scale    = 1.0f / sqrtf(float(n_embd_head));

examples/talk-llama/models/cohere2.cpp

@@ -1,6 +1,53 @@
 #include "models.h"
 
-llm_build_cohere2_iswa::llm_build_cohere2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_cohere2::load_arch_hparams(llama_model_loader & ml) {
+    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+    uint32_t swa_period = 4;
+    ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
+    hparams.set_swa_pattern(swa_period);
+    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
+
+    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,       hparams.rope_freq_base_train_swa, false);
+    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
+    ml.get_key(LLM_KV_LOGIT_SCALE,              hparams.f_logit_scale);
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,  hparams.f_norm_eps);
+    switch (hparams.n_layer) {
+        case 32: type = LLM_TYPE_8B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_cohere2::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);
+    // init output from the input tok embed
+    output      = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab },
+                                      TENSOR_DUPLICATED);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd, n_embd }, 0);
+
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), { n_embd, n_ff }, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, n_ff }, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_cohere2::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_cohere2::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/command-r.cpp

@@ -1,8 +1,48 @@
 #include "models.h"
 
+void llama_model_command_r::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_LOGIT_SCALE,             hparams.f_logit_scale, false);
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+    switch (hparams.n_layer) {
+        case 40: type = LLM_TYPE_35B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
 
+void llama_model_command_r::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
 
-llm_build_command_r::llm_build_command_r(const llama_model & model, const llm_graph_params & params) :
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    // init output from the input tok embed
+    output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        if (n_layer >= 64){
+            layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k, n_head}, 0);
+            layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k, n_head_kv}, 0);
+        }
+
+        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_command_r::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_command_r::graph::graph(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 

examples/talk-llama/models/dbrx.cpp

@@ -1,6 +1,50 @@
 #include "models.h"
 
-llm_build_dbrx::llm_build_dbrx(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_dbrx::load_arch_hparams(llama_model_loader & ml) {
+ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+ml.get_key(LLM_KV_ATTENTION_CLAMP_KQV,     hparams.f_clamp_kqv);
+
+switch (hparams.n_layer) {
+    case 40: type = LLM_TYPE_16x12B; break;
+    default: type = LLM_TYPE_UNKNOWN;
+}
+        }
+
+void llama_model_dbrx::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    if (n_expert == 0) {
+        throw std::runtime_error("DBRX model cannot have zero experts");
+    }
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
+        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+        layer.attn_out_norm = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
+        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff,   n_expert}, 0);
+        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff,   n_embd, n_expert}, 0);
+        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff,   n_expert}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_dbrx::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_dbrx::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/deci.cpp

@@ -1,6 +1,82 @@
 #include "models.h"
 
-llm_build_deci::llm_build_deci(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_deci::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    switch (hparams.n_layer) {
+        case 32: type = LLM_TYPE_7B; break;
+        case 80: type = LLM_TYPE_70B; break;
+        case 162: type = LLM_TYPE_405B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_deci::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+        const int64_t n_embd_k_gqa  = hparams.n_embd_k_gqa(i);
+        const int64_t n_embd_v_gqa  = hparams.n_embd_v_gqa(i);
+        const int64_t n_ff          = hparams.n_ff(i);
+        const int64_t n_head        = hparams.n_head(i);
+        const int64_t n_head_kv     = hparams.n_head_kv(i);
+
+        if (n_head_kv == 0 && n_head > 0) {
+            // linear attention for DeciLMCausalModel
+            layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+        }
+        else if (n_head_kv > 0) {
+            layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+            create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+        }
+
+        // optional bias tensors
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+
+        if (n_ff > 0) {
+            layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        }
+
+        if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
+            layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+            layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+        }
+        else {
+            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+        }
+
+        if (n_ff > 0) {
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+        }
+
+        // optional MLP bias
+        layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
+        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_deci::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_deci::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/deepseek.cpp

@@ -1,6 +1,77 @@
 #include "models.h"
 
-llm_build_deepseek::llm_build_deepseek(const llama_model & model, const llm_graph_params & params) :
+void llama_model_deepseek::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);
+
+    switch (hparams.n_ff_exp) {
+        case 1408: type = LLM_TYPE_16B; break;
+        case 1792: type = LLM_TYPE_20B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_deepseek::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+    const int64_t n_expert_shared = hparams.n_expert_shared;
+
+
+    const int64_t n_ff_exp        = hparams.n_ff_exp;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    // try to load output.weight, if not found, use token_embd (tied embeddings)
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    if (!output) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        if (i < (int) hparams.n_layer_dense_lead) {
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+        } else {
+            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+
+            if (n_expert == 0) {
+                throw std::runtime_error("n_expert must be > 0");
+            }
+            if (n_expert_used == 0) {
+                throw std::runtime_error("n_expert_used must be > 0");
+            }
+
+            // MoE branch
+            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+
+            // Shared expert branch
+            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd}, 0);
+            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_deepseek::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_deepseek::graph::graph(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 

examples/talk-llama/models/deepseek2.cpp

@@ -1,6 +1,149 @@
 #include "models.h"
 
-llm_build_deepseek2::llm_build_deepseek2(const llama_model & model, const llm_graph_params & params) :
+void llama_model_deepseek2::load_arch_hparams(llama_model_loader & ml) {
+    uint32_t n_vocab = 0;
+    ml.get_key(LLM_KV_VOCAB_SIZE, n_vocab, false) || ml.get_arr_n(LLM_KV_TOKENIZER_LIST, n_vocab, false);
+
+    // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B, Kanana-2-30B-A3B
+    const bool is_lite = (hparams.n_layer == 27 || hparams.n_layer == 26 || (hparams.n_layer == 48 && n_vocab == 128256));
+
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
+    if (!is_lite) {
+        ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK, hparams.n_lora_q);
+    }
+    ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,     hparams.n_lora_kv);
+    ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,   hparams.n_embd_head_k_mla_impl, false);
+    ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla_impl, false);
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,        hparams.n_expert_shared);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,       hparams.expert_weights_scale, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,        hparams.expert_weights_norm, false);
+    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,         hparams.expert_gating_func, false);
+    if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {
+        // for compatibility with existing DeepSeek V2 and V2.5 GGUFs
+        // that have no expert_gating_func model parameter set
+        if ((hparams.n_layer == 47 || hparams.n_layer == 48) && n_vocab == 154880) {
+            // GLM 4.7 Lite
+            hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;
+        } else {
+            hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX;
+        }
+    }
+
+    if (ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul, false)) {
+        // [TAG_DEEPSEEK2_YARN_LOG_MUL_FIX]
+        // cancel the factor from the convert script
+        hparams.rope_yarn_log_mul /= 0.1f;
+    }
+
+    // (optional) temperature tuning - used by mistral-large
+    ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_SCALE,  hparams.f_attn_temp_scale,       false);
+    ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_LENGTH, hparams.n_attn_temp_floor_scale, false); // FIXME why not use temperature_length?
+
+    hparams.f_attn_temp_offset = 0.0f;
+
+    switch (hparams.n_layer) {
+        case 27: type = LLM_TYPE_16B; break;
+        case 47: type = LLM_TYPE_30B_A3B; break;
+        case 60: type = LLM_TYPE_236B; break;
+        case 61: type = LLM_TYPE_671B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_deepseek2::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+    const int64_t n_expert_shared = hparams.n_expert_shared;
+
+    const bool is_mla = hparams.is_mla();
+
+    // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
+    const int64_t n_embd_head_k_mla = hparams.n_embd_head_k_mla();
+    const int64_t n_embd_head_v_mla = hparams.n_embd_head_v_mla();
+
+    const int64_t n_embd_head_qk_rope = hparams.n_rot();
+    const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
+    GGML_ASSERT(n_embd_head_qk_nope >= 1);
+
+    const int64_t q_lora_rank  = hparams.n_lora_q;
+    const int64_t kv_lora_rank = hparams.n_lora_kv;
+
+    const int64_t n_ff_exp        = hparams.n_ff_exp;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    // try to load output.weight, if not found, use token_embd (tied embeddings)
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    if (!output) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+        if (q_lora_rank > 0) {
+            layer.attn_q_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}, 0);
+        }
+
+        layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, 0);
+
+        if (q_lora_rank > 0) {
+            layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, 0);
+            layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, 0);
+        } else {
+            layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_head * n_embd_head_k_mla}, 0);
+        }
+
+        layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}, 0);
+
+        // note: only old legacy GGUF files will have the unsplit wkv_b tensor in
+        if (is_mla) {
+            layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head}, 0);
+            layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head}, 0);
+        } else {
+            layer.wkv_b = create_tensor(tn(LLM_TENSOR_ATTN_KV_B, "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v_mla)}, 0);
+        }
+
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_embd_head_v_mla, n_embd}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        if (i < (int) hparams.n_layer_dense_lead) {
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+        } else {
+            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);
+
+            if (n_expert == 0) {
+                throw std::runtime_error("n_expert must be > 0");
+            }
+            if (n_expert_used == 0) {
+                throw std::runtime_error("n_expert_used must be > 0");
+            }
+
+            // MoE branch
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
+            create_tensor_gate_up_exps(layer, i, n_embd, n_ff_exp, n_expert, 0);
+
+            // Shared expert branch
+            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd}, 0);
+            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_deepseek2::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_deepseek2::graph::graph(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     // lite variants include DeepSeek-V2-Lite, GigaChat3-10B-A1.8B
     bool is_ocr = model.arch == LLM_ARCH_DEEPSEEK2OCR;

examples/talk-llama/models/deepseek2ocr.cpp

@@ -0,0 +1,82 @@
+#include "models.h"
+
+void llama_model_deepseek2ocr::load_arch_hparams(llama_model_loader & ml) {
+    // similar to deepseek2, but without MLA
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);
+    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);
+
+    if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {
+        hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX;
+    }
+
+    switch (hparams.n_layer) {
+        case 12: type = LLM_TYPE_3B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_deepseek2ocr::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+    const int64_t n_expert_shared = hparams.n_expert_shared;
+
+    // similar to deepseek2, but without MLA
+    const int64_t n_ff_exp        = hparams.n_ff_exp;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    // try to load output.weight, if not found, use token_embd (tied embeddings)
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    if (!output) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd}, 0);
+        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd}, 0);
+        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd}, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+        // norm
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        if (i < (int) hparams.n_layer_dense_lead) {
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+        } else {
+            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);
+
+            if (n_expert == 0) {
+                throw std::runtime_error("n_expert must be > 0");
+            }
+            if (n_expert_used == 0) {
+                throw std::runtime_error("n_expert_used must be > 0");
+            }
+
+            // MoE branch
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
+            create_tensor_gate_up_exps(layer, i, n_embd, n_ff_exp, n_expert, 0);
+
+            // Shared expert branch
+            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd}, 0);
+            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_deepseek2ocr::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+

examples/talk-llama/models/dots1.cpp

@@ -1,6 +1,76 @@
 #include "models.h"
 
-llm_build_dots1::llm_build_dots1(const llama_model & model, const llm_graph_params & params) :
+void llama_model_dots1::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);
+    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);
+    switch (hparams.n_layer) {
+        case 62: type = LLM_TYPE_142B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_dots1::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+    const int64_t n_expert_shared = hparams.n_expert_shared;
+
+    const int64_t n_ff_exp        = hparams.n_ff_exp;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_head_k * n_head, n_embd_head_k * n_head, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
+        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        if (i < (int) hparams.n_layer_dense_lead) {
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+        } else {
+            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);
+
+            if (n_expert == 0) {
+                throw std::runtime_error("n_expert must be > 0");
+            }
+            if (n_expert_used == 0) {
+                throw std::runtime_error("n_expert_used must be > 0");
+            }
+
+            // MoE branch
+            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+
+            // Shared expert branch
+            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd}, 0);
+            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, 0);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_dots1::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_dots1::graph::graph(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 

examples/talk-llama/models/dream.cpp

@@ -1,6 +1,54 @@
 #include "models.h"
 
-llm_build_dream::llm_build_dream(const llama_model & model, const llm_graph_params & params) :
+void llama_model_dream::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    // Dream models are primarily 7B with 28 layers
+    switch (hparams.n_layer) {
+        case 28:
+            type = LLM_TYPE_7B;
+            break;
+        default:
+            type = LLM_TYPE_UNKNOWN;
+    }
+    // Set non-causal attention for diffusion models
+    hparams.causal_attn = false;
+}
+
+void llama_model_dream::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    output_b    = create_tensor(tn(LLM_TENSOR_OUTPUT,      "bias"),   {n_vocab}, TENSOR_NOT_REQUIRED);
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_dream::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_dream::graph::graph(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     //copied from qwen2
     const int64_t n_embd_head = hparams.n_embd_head_v();

examples/talk-llama/models/ernie4-5-moe.cpp

@@ -1,6 +1,10 @@
 #include "models.h"
 
-llm_build_ernie4_5_moe::llm_build_ernie4_5_moe(const llama_model & model, const llm_graph_params & params) :
+std::unique_ptr<llm_graph_context> llama_model_ernie4_5_moe::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_ernie4_5_moe::graph::graph(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 

examples/talk-llama/models/ernie4-5.cpp

@@ -1,6 +1,79 @@
 #include "models.h"
 
-llm_build_ernie4_5::llm_build_ernie4_5(const llama_model & model, const llm_graph_params & params) :
+void llama_model_ernie4_5::load_arch_hparams(llama_model_loader & ml) {
+    // paddleocr need mrope_section
+    ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
+
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    if (arch == LLM_ARCH_ERNIE4_5_MOE) {
+        ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
+        ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
+        ml.get_key(LLM_KV_INTERLEAVE_MOE_LAYER_STEP,         hparams.n_moe_layer_step);
+        ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead, false);
+    }
+
+    switch (hparams.n_layer) {
+        case 18: type = LLM_TYPE_0_3B; break;
+        case 28: type = LLM_TYPE_21B_A3B; break;
+        case 54: type = LLM_TYPE_300B_A47B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_ernie4_5::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        // optional bias tensors
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        if (arch == LLM_ARCH_ERNIE4_5_MOE && static_cast<uint32_t>(i) >= hparams.n_layer_dense_lead) { // MoE layers
+            int n_ff_exp = hparams.n_ff_exp;
+
+            layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
+            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);
+            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff_exp, n_embd, n_expert}, 0);
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff_exp, n_expert}, 0);
+
+            // Shared expert (if present)
+            if (hparams.n_ff_shexp > 0) {
+                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {    n_embd, hparams.n_ff_shexp}, 0);
+                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd    }, 0);
+                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {    n_embd, hparams.n_ff_shexp}, 0);
+            }
+        } else { // Dense layers
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_ernie4_5::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_ernie4_5::graph::graph(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 

examples/talk-llama/models/eurobert.cpp

@@ -1,6 +1,41 @@
 #include "models.h"
 
-llm_build_eurobert::llm_build_eurobert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_eurobert::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    if (hparams.n_layer == 12) {
+        type = LLM_TYPE_SMALL;  // 0.2B
+    }
+}
+
+void llama_model_eurobert::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_eurobert::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_eurobert::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/exaone-moe.cpp

@@ -1,6 +1,117 @@
 #include "models.h"
 
-llm_build_exaone_moe::llm_build_exaone_moe(const llama_model & model, const llm_graph_params & params) :
+void llama_model_exaone_moe::load_arch_hparams(llama_model_loader & ml) {
+    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+    hparams.n_swa = 128;
+    uint32_t swa_period = 4;
+    ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
+    hparams.set_swa_pattern(swa_period);
+    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
+
+    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,                hparams.rope_freq_base_train_swa, false);
+    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,          hparams.n_swa);
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared, false);
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
+    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
+    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,                hparams.expert_gating_func);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,               hparams.expert_weights_norm, false);
+    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead, false);
+
+    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS,              hparams.nextn_predict_layers, false);
+    GGML_ASSERT(hparams.nextn_predict_layers < hparams.n_layer && "nextn_predict_layers must be < n_layer");
+
+    switch (hparams.n_layer) {
+        case 32: type = LLM_TYPE_30B_A3B; break;
+        case 48:
+        case 49: type = LLM_TYPE_235B_A22B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_exaone_moe::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    const int64_t n_ff_exp       = hparams.n_ff_exp;
+    const int64_t n_ff_shexp     = hparams.n_ff_shexp > 0 ? hparams.n_ff_shexp : n_ff_exp;
+    const int64_t head_dim       = hparams.n_embd_head_k();
+    const int64_t n_qo_dim       = n_head * head_dim;
+    const int64_t n_kv_dim       = n_head_kv * head_dim;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        int flags = 0;
+        if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+            // skip all tensors in the NextN layers
+            flags |= TENSOR_SKIP;
+        }
+
+        auto & layer = layers[i];
+        create_tensor_qkv(layer, i, n_embd, n_qo_dim, n_kv_dim, n_kv_dim, flags);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_qo_dim, n_embd}, flags);
+
+        layer.rope_freqs   = create_tensor(tn(LLM_TENSOR_ROPE_FREQS,  "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0) | flags);
+
+        layer.attn_norm    = create_tensor(tn(LLM_TENSOR_ATTN_NORM,   "weight", i), {n_embd}, flags);
+        layer.attn_q_norm  = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, flags);
+        layer.attn_k_norm  = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, flags);
+
+        layer.ffn_norm     = create_tensor(tn(LLM_TENSOR_FFN_NORM,    "weight", i), {n_embd}, flags);
+
+        // dense layers for first n_layer_dense_lead layers or nextn_predict_layers layers at the end
+        if (i < (int) hparams.n_layer_dense_lead || (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers)) {
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, flags);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, flags);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, flags);
+        } else {
+            layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, flags);
+            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED | flags);
+
+            if (n_expert == 0) {
+                throw std::runtime_error("n_expert must be > 0");
+            }
+            if (n_expert_used == 0) {
+                throw std::runtime_error("n_expert_used must be > 0");
+            }
+
+            layer.ffn_gate_exps  = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS,  "weight", i), {n_embd, n_ff_exp, n_expert}, flags);
+            layer.ffn_down_exps  = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS,  "weight", i), {n_ff_exp, n_embd, n_expert}, flags);
+            layer.ffn_up_exps    = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,    "weight", i), {n_embd, n_ff_exp, n_expert}, flags);
+
+            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp}, flags);
+            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, flags);
+            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, flags);
+        }
+
+        // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
+        if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), {2 * n_embd, n_embd}, flags);
+            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM,   "weight", i), {n_embd}, flags);
+            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM,   "weight", i), {n_embd}, flags);
+
+            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), {n_embd}, flags | TENSOR_NOT_REQUIRED);
+            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS,     "weight", i), {n_embd, n_vocab}, flags | TENSOR_NOT_REQUIRED);
+            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), {n_embd, n_vocab}, flags | TENSOR_NOT_REQUIRED);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_exaone_moe::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_exaone_moe::graph::graph(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_k();
 

examples/talk-llama/models/exaone.cpp

@@ -1,6 +1,49 @@
 #include "models.h"
 
-llm_build_exaone::llm_build_exaone(const llama_model & model, const llm_graph_params & params) :
+void llama_model_exaone::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    switch (hparams.n_layer) {
+        case 32: type = LLM_TYPE_8B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_exaone::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM,   "weight", i), {n_embd}, 0);
+        layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+        layer.ffn_gate   = create_tensor(tn(LLM_TENSOR_FFN_GATE,   "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN,   "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,     "weight", i), {n_embd,   n_ff}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_exaone::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_exaone::graph::graph(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 

examples/talk-llama/models/exaone4.cpp

@@ -1,7 +1,71 @@
 #include "models.h"
 
+void llama_model_exaone4::load_arch_hparams(llama_model_loader & ml) {
+    if (hparams.n_layer == 64) {    // 32B
+        hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+        hparams.n_swa = 4096;
+        uint32_t swa_period = 4;
+        ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
+        hparams.set_swa_pattern(swa_period);
+
+        hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+        hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
+        ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
+    }
+
+    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa, false);
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    switch (hparams.n_layer) {
+        case 30: type = LLM_TYPE_1_2B; break;
+        case 64: type = LLM_TYPE_32B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_exaone4::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+        layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+
+        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
+        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
+
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_post_norm  = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_exaone4::build_arch_graph(const llm_graph_params & params) const {
+    if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) {
+        return std::make_unique<graph<true>>(*this, params);
+    } else {
+        return std::make_unique<graph<false>>(*this, params);
+    }
+}
+
 template <bool iswa>
-llm_build_exaone4<iswa>::llm_build_exaone4(const llama_model & model, const llm_graph_params & params) :
+llama_model_exaone4::graph<iswa>::graph(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_k();
 
@@ -108,5 +172,5 @@ llm_build_exaone4<iswa>::llm_build_exaone4(const llama_model & model, const llm_
 }
 
 // Explicit template instantiations
-template struct llm_build_exaone4<false>;
-template struct llm_build_exaone4<true>;
+template struct llama_model_exaone4::graph<false>;
+template struct llama_model_exaone4::graph<true>;

examples/talk-llama/models/falcon-h1.cpp

@@ -1,6 +1,115 @@
 #include "models.h"
 
-llm_build_falcon_h1::llm_build_falcon_h1(const llama_model & model, const llm_graph_params & params) :
+void llama_model_falcon_h1::load_arch_hparams(llama_model_loader & ml) {
+    // Common parameters
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    // SSM parameters
+    ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);
+    ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);
+    ml.get_key(LLM_KV_SSM_STATE_SIZE,     hparams.ssm_d_state);
+    ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
+    ml.get_key(LLM_KV_SSM_GROUP_COUNT,    hparams.ssm_n_group);
+
+    std::fill(hparams.recurrent_layer_arr.begin(), hparams.recurrent_layer_arr.end(), true);
+
+    switch (hparams.n_layer) {
+        case 36:
+            type = LLM_TYPE_0_5B; break;
+        case 24:
+            type = LLM_TYPE_1_5B; break;
+        case 66:
+            type = LLM_TYPE_1B; break;
+        case 32:
+            type = LLM_TYPE_3B; break;
+        case 44:
+            type = LLM_TYPE_7B; break;
+        case 72:
+            type = LLM_TYPE_34B; break;
+        default:
+            type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_falcon_h1::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    // Common
+    const int64_t hidden_size = hparams.n_embd; // hidden_size
+
+    // mamba2 Mixer SSM params
+    const int64_t ssm_conv_kernel_size  = hparams.ssm_d_conv; // ssm_conv_kernel_size
+    const int64_t ssm_n_groups          = hparams.ssm_n_group; // ssm_n_groups
+    const int64_t ssm_state_size        = hparams.ssm_d_state; // ssm_state_size
+    const int64_t ssm_intermediate_size = hparams.ssm_d_inner; // TODO expand
+    const int64_t ssm_num_heads         = hparams.ssm_dt_rank; // ssm_num_heads
+    const int64_t ssm_conv_dim          = ssm_intermediate_size + 2 * ssm_n_groups * ssm_state_size;
+    const int64_t ssm_projection_size   = ssm_intermediate_size + ssm_conv_dim + ssm_num_heads;
+
+    // attn params
+    const int64_t attn_num_attention_head = hparams.n_head(0); // rename to: attn_num_attention_head
+    const int64_t attn_num_key_value_head = hparams.n_head_kv(0);
+
+    // ffn params
+    const int64_t ffn_intermediate_size = hparams.n_ff(0);
+
+    // embeddings
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {hidden_size, n_vocab}, 0);
+
+    // output
+    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {hidden_size, n_vocab}, TENSOR_NOT_REQUIRED);
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {hidden_size}, 0);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {hidden_size, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        /*SSM LAYERS*/
+        // ssm in
+        layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {hidden_size, ssm_projection_size}, 0);
+        // ssm 1d conv
+        layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {ssm_conv_kernel_size, ssm_conv_dim}, 0);
+        layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {ssm_conv_dim}, TENSOR_NOT_REQUIRED);
+        // ssm_dt
+        layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {ssm_num_heads}, 0);
+        // no "weight" suffix for these
+        layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, ssm_num_heads}, 0);
+        layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, ssm_num_heads}, 0);
+        // ssm_norm
+        layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {ssm_intermediate_size / ssm_n_groups, ssm_n_groups}, TENSOR_NOT_REQUIRED);
+        // out_proj
+        layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {ssm_intermediate_size, hidden_size}, 0);
+
+        /*ATTENTION LAYERS*/
+        // attention layers (with optional bias)
+        create_tensor_qkv(layer, i, hidden_size, n_embd_head_k * attn_num_attention_head, attn_num_key_value_head * n_embd_head_k, attn_num_key_value_head * n_embd_head_v, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * attn_num_attention_head, hidden_size}, 0);
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {hidden_size}, TENSOR_NOT_REQUIRED);
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {hidden_size}, 0);
+
+
+        // feed forward (w/ optional biases)
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, i), {hidden_size}, 0);
+        layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {hidden_size,   ffn_intermediate_size}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  ffn_intermediate_size, hidden_size}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {hidden_size,   ffn_intermediate_size}, 0);
+
+        layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {ffn_intermediate_size}, TENSOR_NOT_REQUIRED);
+        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {hidden_size}, TENSOR_NOT_REQUIRED);
+        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {ffn_intermediate_size}, TENSOR_NOT_REQUIRED);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_falcon_h1::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_falcon_h1::graph::graph(const llama_model & model, const llm_graph_params & params) :
     llm_build_mamba_base(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 

examples/talk-llama/models/falcon.cpp

@@ -1,6 +1,53 @@
 #include "models.h"
 
-llm_build_falcon::llm_build_falcon(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_falcon::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+
+    switch (hparams.n_layer) {
+        case 32: type = LLM_TYPE_7B; break;
+        case 60: type = LLM_TYPE_40B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_falcon::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    {
+        output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+        output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);
+
+        output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+        if (!output) {
+            output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); // needs to be on GPU
+        }
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, 0);
+
+        layer.attn_norm_2   = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
+        layer.attn_norm_2_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);
+
+        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
+        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_falcon::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_falcon::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/gemma-embedding.cpp

@@ -1,6 +1,78 @@
 #include "models.h"
 
-llm_build_gemma_embedding::llm_build_gemma_embedding(const llama_model & model, const llm_graph_params & params) :
+void llama_model_gemma_embedding::load_arch_hparams(llama_model_loader & ml) {
+    hparams.swa_type = LLAMA_SWA_TYPE_SYMMETRIC;
+    uint32_t swa_period = 6;
+    ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
+    hparams.set_swa_pattern(swa_period);
+
+    hparams.causal_attn = false; // embeddings do not use causal attention
+
+    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
+    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    //applied only if model converted with --sentence-transformers-dense-modules
+    ml.get_key(LLM_KV_DENSE_2_FEAT_IN, hparams.dense_2_feat_in, false);
+    ml.get_key(LLM_KV_DENSE_2_FEAT_OUT, hparams.dense_2_feat_out, false);
+    ml.get_key(LLM_KV_DENSE_3_FEAT_IN, hparams.dense_3_feat_in, false);
+    ml.get_key(LLM_KV_DENSE_3_FEAT_OUT, hparams.dense_3_feat_out, false);
+
+    GGML_ASSERT((hparams.dense_2_feat_in == 0 || hparams.dense_2_feat_in == hparams.n_embd) && "dense_2_feat_in must be equal to n_embd");
+    GGML_ASSERT((hparams.dense_3_feat_out == 0 || hparams.dense_3_feat_out == hparams.n_embd) && "dense_3_feat_out must be equal to n_embd");
+
+    switch (hparams.n_layer) {
+        case 24: type = LLM_TYPE_0_3B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+    hparams.f_attention_scale = 1.0f / std::sqrt(float(hparams.n_embd_head_k()));
+
+}
+
+void llama_model_gemma_embedding::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,   "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    // Dense linear weights
+    dense_2_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.dense_2_feat_out}, TENSOR_NOT_REQUIRED);
+    dense_3_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_3_OUT, "weight"), {hparams.dense_3_feat_in, n_embd}, TENSOR_NOT_REQUIRED);
+
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+        layer.attn_k_norm    = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM,    "weight", i), {n_embd_head_k}, 0);
+        layer.attn_q_norm    = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM,    "weight", i), {n_embd_head_k}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_gemma_embedding::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_gemma_embedding::graph::graph(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_k();
 

examples/talk-llama/models/gemma.cpp

@@ -1,6 +1,44 @@
 #include "models.h"
 
-llm_build_gemma::llm_build_gemma(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_gemma::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    switch (hparams.n_layer) {
+        case 18: type = LLM_TYPE_2B; break;
+        case 28: type = LLM_TYPE_7B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+   }
+}
+
+void llama_model_gemma::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_gemma::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_gemma::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     ggml_tensor * cur;

examples/talk-llama/models/gemma2.cpp

@@ -1,6 +1,65 @@
 #include "models.h"
 
-llm_build_gemma2_iswa::llm_build_gemma2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_gemma2::load_arch_hparams(llama_model_loader & ml) {
+    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+    hparams.n_swa = 4096; // default value of gemma 2
+    uint32_t swa_period = 2;
+    ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
+    hparams.set_swa_pattern(swa_period);
+    hparams.attn_soft_cap = true;
+    hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+    hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
+
+    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,          hparams.rope_freq_base_train_swa, false);
+    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa, false);
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_ATTN_LOGIT_SOFTCAPPING,      hparams.f_attn_logit_softcapping, false);
+    ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING,     hparams.f_final_logit_softcapping, false);
+
+    switch (hparams.n_layer) {
+        case 26: type = LLM_TYPE_2B; break;
+        case 42: type = LLM_TYPE_9B; break;
+        case 46: type = LLM_TYPE_27B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+   }
+
+    // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/config.py#L173
+    hparams.f_attention_scale = type == LLM_TYPE_27B
+        ? 1.0f / std::sqrt(float(hparams.n_embd / hparams.n_head(0)))
+        : 1.0f / std::sqrt(float(hparams.n_embd_head_k()));
+}
+
+void llama_model_gemma2::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_gemma2::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_gemma2::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_k();
 
     ggml_tensor * cur;

examples/talk-llama/models/gemma3.cpp

@@ -1,7 +1,87 @@
 #include "models.h"
 
+void llama_model_gemma3::load_arch_hparams(llama_model_loader & ml) {
+    const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
+    if (found_swa && hparams.n_swa > 0) {
+        hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+        uint32_t swa_period = 6;
+        ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
+        hparams.set_swa_pattern(swa_period);
+
+        ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
+    } else {
+        hparams.swa_type = LLAMA_SWA_TYPE_NONE;
+    }
+
+    hparams.f_final_logit_softcapping = 0.0f;
+    ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING, hparams.f_final_logit_softcapping, false);
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    switch (hparams.n_layer) {
+        case 18: type = LLM_TYPE_270M; break;
+        case 26: type = LLM_TYPE_1B; break;
+        case 32: type = LLM_TYPE_8B; break; // Rnj-1
+        case 34: type = LLM_TYPE_4B; break;
+        case 48: type = LLM_TYPE_12B; break;
+        case 62: type = LLM_TYPE_27B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+
+    // ref: https://github.com/google/gemma_pytorch/blob/014acb7ac4563a5f77c76d7ff98f31b568c16508/gemma/config.py#L289
+    hparams.f_attention_scale = type == LLM_TYPE_27B
+        ? 1.0f / std::sqrt(float(hparams.n_embd / hparams.n_head(0)))
+        : 1.0f / std::sqrt(float(hparams.n_embd_head_k()));
+}
+
+void llama_model_gemma3::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,   "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    // Dense linear weights
+    dense_2_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.dense_2_feat_out}, TENSOR_NOT_REQUIRED);
+    dense_3_out_layers = create_tensor(tn(LLM_TENSOR_DENSE_3_OUT, "weight"), {hparams.dense_3_feat_in, n_embd}, TENSOR_NOT_REQUIRED);
+
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+        layer.attn_k_norm    = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM,    "weight", i), {n_embd_head_k}, 0);
+        layer.attn_q_norm    = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM,    "weight", i), {n_embd_head_k}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_gemma3::build_arch_graph(const llm_graph_params & params) const {
+    if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) {
+        return std::make_unique<graph<true>>(*this, params);
+    } else {
+        return std::make_unique<graph<false>>(*this, params);
+    }
+}
+
 template <bool iswa>
-llm_build_gemma3<iswa>::llm_build_gemma3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+llama_model_gemma3::graph<iswa>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_k();
 
     ggml_tensor * cur;
@@ -141,5 +221,5 @@ llm_build_gemma3<iswa>::llm_build_gemma3(const llama_model & model, const llm_gr
     ggml_build_forward_expand(gf, cur);
 }
 
-template struct llm_build_gemma3<false>;
-template struct llm_build_gemma3<true>;
+template struct llama_model_gemma3::graph<false>;
+template struct llama_model_gemma3::graph<true>;

examples/talk-llama/models/gemma3n.cpp

@@ -1,5 +1,86 @@
 #include "models.h"
 
+void llama_model_gemma3n::load_arch_hparams(llama_model_loader & ml) {
+    uint32_t swa_period = 5;
+    ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
+    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+    hparams.set_swa_pattern(swa_period);
+
+    hparams.n_layer_kv_from_start     = 20;
+    hparams.f_attention_scale         = 1.0f;
+
+    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,          hparams.rope_freq_base_train_swa, false);
+    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa);
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    switch (hparams.n_layer) {
+        case 30: type = LLM_TYPE_E2B; break;
+        case 35: type = LLM_TYPE_E4B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_gemma3n::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    const int64_t n_altup      = hparams.n_altup;
+    const int64_t laurel_rank  = hparams.laurel_rank;
+    const int64_t n_embd_altup = hparams.n_embd_altup;
+
+    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    altup_proj        = create_tensor(tn(LLM_TENSOR_ALTUP_PROJ,        "weight"), {n_embd, n_embd, n_altup - 1}, 0);
+    altup_unembd_proj = create_tensor(tn(LLM_TENSOR_ALTUP_UNEMBD_PROJ, "weight"), {n_embd, n_embd, n_altup - 1}, 0);
+
+    per_layer_tok_embd   = create_tensor(tn(LLM_TENSOR_PER_LAYER_TOKEN_EMBD, "weight"), {n_embd_altup * n_layer, n_vocab}, 0);
+    per_layer_model_proj = create_tensor(tn(LLM_TENSOR_PER_LAYER_MODEL_PROJ, "weight", 0), {n_embd, n_embd_altup * n_layer}, 0);
+    per_layer_proj_norm  = create_tensor(tn(LLM_TENSOR_PER_LAYER_PROJ_NORM,  "weight", 0), {n_embd_altup}, 0);
+
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        layer.attn_q_norm    = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM,    "weight", i), {n_embd_head_k}, 0);
+        layer.attn_k_norm    = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM,    "weight", i), {n_embd_head_k}, 0);
+        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
+
+        // altup & laurel
+        layer.per_layer_inp_gate   = create_tensor(tn(LLM_TENSOR_PER_LAYER_INP_GATE,  "weight", i), {n_embd, n_embd_altup}, 0);
+        layer.per_layer_proj       = create_tensor(tn(LLM_TENSOR_PER_LAYER_PROJ,      "weight", i), {n_embd_altup, n_embd}, 0);
+        layer.per_layer_post_norm  = create_tensor(tn(LLM_TENSOR_PER_LAYER_POST_NORM, "weight", i), {n_embd}, 0);
+        layer.altup_correct_coef   = create_tensor(tn(LLM_TENSOR_ALTUP_CORRECT_COEF,  "weight", i), {n_altup, n_altup}, 0);
+        layer.altup_correct_scale  = create_tensor(tn(LLM_TENSOR_ALTUP_CORRECT_SCALE, "weight", i), {n_embd}, 0);
+        layer.altup_predict_coef   = create_tensor(tn(LLM_TENSOR_ALTUP_PREDICT_COEF,  "weight", i), {n_altup, n_altup * n_altup}, 0);
+        layer.altup_router         = create_tensor(tn(LLM_TENSOR_ALTUP_ROUTER,        "weight", i), {n_embd, n_altup}, 0);
+        layer.altup_router_norm    = create_tensor(tn(LLM_TENSOR_ALTUP_ROUTER_NORM,   "weight", i), {n_embd}, 0);
+        layer.laurel_l             = create_tensor(tn(LLM_TENSOR_LAUREL_L,            "weight", i), {n_embd, laurel_rank}, 0);
+        layer.laurel_r             = create_tensor(tn(LLM_TENSOR_LAUREL_R,            "weight", i), {laurel_rank, n_embd}, 0);
+        layer.laurel_post_norm     = create_tensor(tn(LLM_TENSOR_LAUREL_POST_NORM,    "weight", i), {n_embd}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_gemma3n::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
 // get 2D slice view from a 3D tensor, the idx corresponds to the 3rd dim
 static ggml_tensor * ggml_view_2d_slice(ggml_context * ctx0, ggml_tensor * x, int idx) {
     GGML_ASSERT(idx < (int) x->ne[2]);
@@ -7,7 +88,7 @@ static ggml_tensor * ggml_view_2d_slice(ggml_context * ctx0, ggml_tensor * x, in
                         idx * x->ne[0] * x->ne[1] * ggml_element_size(x));
 }
 
-llm_build_gemma3n_iswa::llm_build_gemma3n_iswa(const llama_model & model, const llm_graph_params & params) :
+llama_model_gemma3n::graph::graph(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params),
     model(model),
     n_embd_head(model.hparams.n_embd_head_k()),
@@ -229,13 +310,13 @@ llm_build_gemma3n_iswa::llm_build_gemma3n_iswa(const llama_model & model, const
     ggml_build_forward_expand(gf, cur);
 }
 
-ggml_tensor * llm_build_gemma3n_iswa::calc_magnitude(ggml_tensor * x) {
+ggml_tensor * llama_model_gemma3n::graph::calc_magnitude(ggml_tensor * x) {
     return ggml_sqrt(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, x)));
 }
 
 // equivalent to get_per_layer_inputs() in python code
 // output shape: [n_embd_altup, n_layer, n_tokens]
-ggml_tensor * llm_build_gemma3n_iswa::build_inp_per_layer() {
+ggml_tensor * llama_model_gemma3n::graph::build_inp_per_layer() {
     auto inp = std::make_unique<llm_graph_input_embd>(n_embd);
     ggml_tensor * inp_per_layer;
     float tok_embd_scale = sqrtf((float) n_embd_altup);
@@ -268,7 +349,7 @@ ggml_tensor * llm_build_gemma3n_iswa::build_inp_per_layer() {
 // equivalent to project_per_layer_inputs() in python code
 // this calculates the per-layer inputs, so the final tensor shape will have n_layer as the last dim
 // output shape: [n_embd_altup, n_tokens, n_layer]
-ggml_tensor * llm_build_gemma3n_iswa::project_per_layer_inputs(ggml_tensor * inp_batch, ggml_tensor * inp_per_layer) {
+ggml_tensor * llama_model_gemma3n::graph::project_per_layer_inputs(ggml_tensor * inp_batch, ggml_tensor * inp_per_layer) {
     const float per_layer_projection_scale = 1.0f / sqrtf((float) n_embd);
     const float per_layer_input_scale      = 1.0f / sqrtf(2.0f);
 
@@ -291,7 +372,7 @@ ggml_tensor * llm_build_gemma3n_iswa::project_per_layer_inputs(ggml_tensor * inp
 
 // input cur shape: [n_altup, n_tokens]
 // output    shape: [n_altup, n_tokens]
-ggml_tensor * llm_build_gemma3n_iswa::laurel(ggml_tensor * cur, int il) {
+ggml_tensor * llama_model_gemma3n::graph::laurel(ggml_tensor * cur, int il) {
     ggml_tensor * tmp = cur;
     tmp               = build_lora_mm(model.layers[il].laurel_l, tmp);
     tmp               = build_lora_mm(model.layers[il].laurel_r, tmp);
@@ -303,7 +384,7 @@ ggml_tensor * llm_build_gemma3n_iswa::laurel(ggml_tensor * cur, int il) {
 
 // input x shape: [n_embd, n_tokens]
 // output  shape: [n_embd, n_tokens]
-ggml_tensor * llm_build_gemma3n_iswa::gaussian_topk(ggml_tensor * x) {
+ggml_tensor * llama_model_gemma3n::graph::gaussian_topk(ggml_tensor * x) {
     ggml_tensor * mean = ggml_mean(ctx0, x);
     ggml_tensor * std  = ggml_sqrt(ctx0, ggml_scale(ctx0, ggml_sum_rows(ctx0, ggml_sqr(ctx0, ggml_sub(ctx0, x, mean))),
                                                     1.0f / (float) (x->ne[0] - 1)));
@@ -318,7 +399,7 @@ ggml_tensor * llm_build_gemma3n_iswa::gaussian_topk(ggml_tensor * x) {
 // equivalent to compute_router_modalities() in python code
 // input x shape: [n_embd,  n_tokens]
 // output  shape: [n_altup, n_tokens]
-ggml_tensor * llm_build_gemma3n_iswa::altup_compute_router_modalities(ggml_tensor * x, int il) {
+ggml_tensor * llama_model_gemma3n::graph::altup_compute_router_modalities(ggml_tensor * x, int il) {
     ggml_tensor * router_inputs = build_norm(x, model.layers[il].altup_router_norm, NULL, LLM_NORM_RMS, il);
 
     // router_input_scale
@@ -330,7 +411,7 @@ ggml_tensor * llm_build_gemma3n_iswa::altup_compute_router_modalities(ggml_tenso
 
 // input cur shape: [n_embd, n_tokens, n_altup]
 // output    shape: [n_embd, n_tokens, n_altup]
-ggml_tensor * llm_build_gemma3n_iswa::altup_predict(ggml_tensor * cur, int il) {
+ggml_tensor * llama_model_gemma3n::graph::altup_predict(ggml_tensor * cur, int il) {
     ggml_tensor * activated  = ggml_view_2d_slice(ctx0, cur, i_altup_act);      // [n_embd, n_tokens]
     ggml_tensor * modalities = altup_compute_router_modalities(activated, il);  // [n_altup, n_tokens]
     cb(modalities, "modalities", il);
@@ -355,7 +436,7 @@ ggml_tensor * llm_build_gemma3n_iswa::altup_predict(ggml_tensor * cur, int il) {
 // input predictions       shape: [n_embd, n_tokens, n_altup]
 // input activated         shape: [n_embd, n_tokens]
 // output                  shape: [n_embd, n_tokens, n_altup]
-ggml_tensor * llm_build_gemma3n_iswa::altup_correct(ggml_tensor * predictions, ggml_tensor * activated, int il) {
+ggml_tensor * llama_model_gemma3n::graph::altup_correct(ggml_tensor * predictions, ggml_tensor * activated, int il) {
     ggml_tensor * modalities = altup_compute_router_modalities(activated, il);  // [n_altup, n_tokens]
     cb(modalities, "modalities", il);
 

examples/talk-llama/models/gemma4.cpp

@@ -1,5 +1,140 @@
 #include "models.h"
 
+void llama_model_gemma4::load_arch_hparams(llama_model_loader & ml) {
+    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+    ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.swa_layers, hparams.n_layer);
+
+    uint32_t n_kv_shared_layers = 0;
+    ml.get_key(LLM_KV_ATTENTION_SHARED_KV_LAYERS, n_kv_shared_layers, false);
+
+    hparams.n_layer_kv_from_start = hparams.n_layer - (int32_t)n_kv_shared_layers;
+    hparams.f_attention_scale     = 1.0f; // Gemma4 uses self.scaling = 1.0 (no pre-attn scaling)
+
+    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,          hparams.rope_freq_base_train_swa, false);
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp, false);
+    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,    hparams.n_swa);
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_EMBEDDING_LENGTH_PER_LAYER,  hparams.n_embd_per_layer);
+    ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_SWA,    hparams.n_embd_head_k_swa);
+    ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_SWA,  hparams.n_embd_head_v_swa);
+    ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING,     hparams.f_final_logit_softcapping, false);
+
+    switch (hparams.n_layer) {
+        case 30: type = LLM_TYPE_26B_A4B; break;
+        case 35: type = LLM_TYPE_E2B; break;
+        case 42: type = LLM_TYPE_E4B; break;
+        case 60: type = LLM_TYPE_31B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_gemma4::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    const uint32_t n_embd_per_layer = hparams.n_embd_per_layer;
+    const int64_t  n_ff_exp         = hparams.n_ff_exp;
+
+    if (n_embd_head_k != n_embd_head_v) {
+        throw std::runtime_error("Gemma 4 requires n_embd_head_k == n_embd_head_v");
+    }
+    if (hparams.n_embd_head_k_swa != hparams.n_embd_head_v_swa) {
+        throw std::runtime_error("Gemma 4 requires n_embd_head_k_swa == n_embd_head_v_swa");
+    }
+
+    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    if (n_embd_per_layer > 0) {
+        per_layer_tok_embd   = create_tensor(tn(LLM_TENSOR_PER_LAYER_TOKEN_EMBD, "weight"),    {n_embd_per_layer * n_layer, n_vocab}, 0);
+        per_layer_model_proj = create_tensor(tn(LLM_TENSOR_PER_LAYER_MODEL_PROJ, "weight", 0), {n_embd, n_embd_per_layer * n_layer}, 0);
+        per_layer_proj_norm  = create_tensor(tn(LLM_TENSOR_PER_LAYER_PROJ_NORM,  "weight", 0), {n_embd_per_layer}, 0);
+    }
+
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+
+    int rope_freqs_flag = 0;
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+        const int64_t n_head      = hparams.n_head(i);
+        const int64_t n_embd_head = hparams.n_embd_head_k(i);
+        const int64_t n_embd_k    = hparams.n_embd_k_gqa(i);
+        const int64_t n_embd_v    = hparams.n_embd_v_gqa(i);
+        const int     kv_flags    = hparams.has_kv(i) ? 0 : TENSOR_NOT_REQUIRED;
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        // note: use_alternative_attention (v_proj is optional, if it's not present, use k_proj)
+        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q,   "weight", i), {n_embd, n_embd_head * n_head}, 0);
+        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K,   "weight", i), {n_embd, n_embd_k}, kv_flags);
+        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V,   "weight", i), {n_embd, n_embd_v}, TENSOR_NOT_REQUIRED);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head * n_head, n_embd}, 0);
+
+        layer.attn_q_norm    = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM,    "weight", i), {n_embd_head}, 0);
+        layer.attn_k_norm    = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM,    "weight", i), {n_embd_head}, kv_flags);
+        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, 0);
+
+        layer.out_scale = create_tensor(tn(LLM_TENSOR_LAYER_OUT_SCALE, "weight", i), {1u}, TENSOR_NOT_REQUIRED);
+
+        if (!hparams.is_swa(i)) {
+            // full_attention layers use rope_freqs for proportional rope
+            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_embd_head/2}, rope_freqs_flag);
+            rope_freqs_flag = TENSOR_DUPLICATED;
+        }
+
+        // handle use_double_wide_mlp
+        int64_t n_ff_cur = hparams.n_ff(i);
+
+        // for expert layers, we use normal FFN as shared expert (same as python code)
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff_cur}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff_cur}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff_cur, n_embd}, 0);
+        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
+
+        // MoE router
+        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED);
+        bool has_expert = layer.ffn_gate_inp != nullptr;
+
+        // norm
+        if (has_expert) {
+            layer.ffn_gate_inp_s = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "scale", i), {n_embd}, 0);
+
+            layer.ffn_pre_norm_2  = create_tensor(tn(LLM_TENSOR_FFN_PRE_NORM_2,  "weight", i), {n_embd}, 0);
+            layer.ffn_post_norm_1 = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM_1, "weight", i), {n_embd}, 0);
+            layer.ffn_post_norm_2 = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM_2, "weight", i), {n_embd}, 0);
+
+            // MoE FFN
+            layer.ffn_gate_up_exps  = create_tensor(tn(LLM_TENSOR_FFN_GATE_UP_EXPS,  "weight", i), {n_embd, n_ff_exp * 2, n_expert}, TENSOR_NOT_REQUIRED);
+
+            if (layer.ffn_gate_up_exps == nullptr) {
+                layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, 0);
+                layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff_exp, n_expert}, 0);
+            }
+
+            layer.ffn_down_exps     = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS,     "weight", i), {n_ff_exp, n_embd, n_expert}, 0);
+
+            // per-expert scale will be loaded as down_exps_s at the end of the current switch case
+        }
+
+        // per-layer embeddings
+        if (n_embd_per_layer > 0) {
+            layer.per_layer_inp_gate   = create_tensor(tn(LLM_TENSOR_PER_LAYER_INP_GATE,  "weight", i), {n_embd, n_embd_per_layer}, 0);
+            layer.per_layer_proj       = create_tensor(tn(LLM_TENSOR_PER_LAYER_PROJ,      "weight", i), {n_embd_per_layer, n_embd}, 0);
+            layer.per_layer_post_norm  = create_tensor(tn(LLM_TENSOR_PER_LAYER_POST_NORM, "weight", i), {n_embd}, 0);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_gemma4::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
 // get 2D slice view from a 3D tensor, the idx corresponds to the 3rd dim
 static ggml_tensor * ggml_view_2d_slice(ggml_context * ctx0, ggml_tensor * x, int idx) {
     GGML_ASSERT(idx < (int) x->ne[2]);
@@ -7,7 +142,7 @@ static ggml_tensor * ggml_view_2d_slice(ggml_context * ctx0, ggml_tensor * x, in
                         idx * x->ne[0] * x->ne[1] * ggml_element_size(x));
 }
 
-llm_build_gemma4_iswa::llm_build_gemma4_iswa(const llama_model & model, const llm_graph_params & params) :
+llama_model_gemma4::graph::graph(const llama_model & model, const llm_graph_params & params) :
         llm_graph_context(params),
         model(model),
         n_embd_per_layer(model.hparams.n_embd_per_layer) {
@@ -157,8 +292,8 @@ llm_build_gemma4_iswa::llm_build_gemma4_iswa(const llama_model & model, const ll
 
             cur_moe = build_moe_ffn(cur_moe,
                     nullptr, // gate_inp
-                    nullptr, // up_exps
-                    nullptr, // gate_exps
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
                     model.layers[il].ffn_down_exps,
                     nullptr, // exp_probs_b (not used for gemma4)
                     n_expert, n_expert_used,
@@ -167,8 +302,8 @@ llm_build_gemma4_iswa::llm_build_gemma4_iswa(const llama_model & model, const ll
                     LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX,
                     il, logits,
                     model.layers[il].ffn_gate_up_exps,
-                    nullptr, // up_exps_s
-                    nullptr, // gate_exps_s
+                    model.layers[il].ffn_up_exps_s,
+                    model.layers[il].ffn_gate_exps_s,
                     model.layers[il].ffn_down_exps_s);
             cur_moe = build_norm(cur_moe,
                     model.layers[il].ffn_post_norm_2, nullptr,
@@ -261,7 +396,7 @@ llm_build_gemma4_iswa::llm_build_gemma4_iswa(const llama_model & model, const ll
 
 // equivalent to get_per_layer_inputs() in python code
 // output shape: [n_embd_per_layer, n_layer, n_tokens]
-ggml_tensor * llm_build_gemma4_iswa::build_inp_per_layer() {
+ggml_tensor * llama_model_gemma4::graph::build_inp_per_layer() {
     auto inp = std::make_unique<llm_graph_input_embd>(n_embd);
 
     ggml_tensor * inp_per_layer;
@@ -299,7 +434,7 @@ ggml_tensor * llm_build_gemma4_iswa::build_inp_per_layer() {
 // inp_batch     shape: [n_embd, n_tokens]
 // inp_per_layer shape: [n_embd_per_layer, n_layer, n_tokens] (from build_inp_per_layer)
 // output shape: [n_embd_per_layer, n_tokens, n_layer]
-ggml_tensor * llm_build_gemma4_iswa::project_per_layer_inputs(ggml_tensor * inp_batch, ggml_tensor * inp_per_layer) {
+ggml_tensor * llama_model_gemma4::graph::project_per_layer_inputs(ggml_tensor * inp_batch, ggml_tensor * inp_per_layer) {
     const float per_layer_projection_scale = 1.0f / sqrtf((float) n_embd);
     const float per_layer_input_scale      = 1.0f / sqrtf(2.0f);
 

examples/talk-llama/models/glm-dsa.cpp

@@ -0,0 +1,155 @@
+#include "models.h"
+
+void llama_model_glm_dsa::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,     hparams.n_ff_exp);
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,    hparams.f_norm_rms_eps);
+    ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
+
+    // MoE parameters
+    ml.get_key(LLM_KV_EXPERT_COUNT,                hparams.n_expert);
+    ml.get_key(LLM_KV_EXPERT_USED_COUNT,           hparams.n_expert_used);
+    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
+    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);
+
+    // deepseek MLA parameters
+    ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK,      hparams.n_lora_q);
+    ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,     hparams.n_lora_kv);
+    ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,   hparams.n_embd_head_k_mla_impl, false);
+    ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA, hparams.n_embd_head_v_mla_impl, false);
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,        hparams.n_expert_shared);
+
+    // DSA parameters
+    ml.get_key(LLM_KV_ATTENTION_INDEXER_HEAD_COUNT, hparams.indexer_n_head);
+    ml.get_key(LLM_KV_ATTENTION_INDEXER_KEY_LENGTH, hparams.indexer_head_size);
+    ml.get_key(LLM_KV_ATTENTION_INDEXER_TOP_K,      hparams.indexer_top_k);
+
+    // Expert gating function (GLM-4.5 uses sigmoid)
+    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);
+    if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {
+        hparams.expert_gating_func =  LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;
+    }
+
+    // NextN/MTP parameters
+    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS,        hparams.nextn_predict_layers, false);
+    GGML_ASSERT(hparams.nextn_predict_layers < hparams.n_layer && "nextn_predict_layers must be < n_layer");
+
+    // TODO: when MTP is implemented, this should probably be updated if needed
+    hparams.n_layer_kv_from_start = hparams.n_layer - hparams.nextn_predict_layers;
+
+    switch (hparams.n_layer) {
+        case 79: type = LLM_TYPE_744B_A40B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_glm_dsa::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+    const int64_t n_expert_shared = hparams.n_expert_shared;
+
+    const bool is_mla = hparams.is_mla();
+    if (!is_mla) {
+        throw std::runtime_error("GLM_DSA architecture requires MLA");
+    }
+
+    // note: these are the actual head sizes you get when treating as MHA or after "decompression" using wv_b for MLA
+    const int64_t n_embd_head_k_mla = hparams.n_embd_head_k_mla();
+    const int64_t n_embd_head_v_mla = hparams.n_embd_head_v_mla();
+
+    const int64_t n_embd_head_qk_rope = hparams.n_rot();
+    const int64_t n_embd_head_qk_nope = n_embd_head_k_mla - n_embd_head_qk_rope;
+
+    const int64_t q_lora_rank  = hparams.n_lora_q;
+    const int64_t kv_lora_rank = hparams.n_lora_kv;
+
+    const int64_t n_ff_exp        = hparams.n_ff_exp;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    // try to load output.weight, if not found, use token_embd (tied embeddings)
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    if (!output) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        int flags = 0;
+        if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+            // skip all tensors in the NextN layers
+            // TODO @ngxson : TENSOR_NOT_REQUIRED was a hack, need to remove it later
+            flags |= TENSOR_SKIP | TENSOR_NOT_REQUIRED;
+        }
+
+        auto & layer = layers[i];
+
+        layer.attn_norm      = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, flags);
+        layer.attn_q_a_norm  = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}, flags);
+        layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, flags);
+
+        layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, flags);
+        layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, flags);
+
+        layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + n_embd_head_qk_rope}, flags);
+
+        // note: only old legacy GGUF files will have the unsplit wkv_b tensor in
+        layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_qk_nope, kv_lora_rank, n_head}, flags);
+        layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head}, flags);
+
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_embd_head_v_mla, n_embd}, flags);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);
+
+        // DSA indexer
+        layer.indexer_k_norm   = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM,   "weight", i), {hparams.indexer_head_size}, flags);
+        layer.indexer_k_norm_b = create_tensor(tn(LLM_TENSOR_INDEXER_K_NORM,   "bias",   i), {hparams.indexer_head_size}, flags);
+        layer.indexer_proj     = create_tensor(tn(LLM_TENSOR_INDEXER_PROJ,     "weight", i), {n_embd, hparams.indexer_n_head}, flags);
+        layer.indexer_attn_k   = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_K,   "weight", i), {n_embd, hparams.indexer_head_size}, flags);
+        layer.indexer_attn_q_b = create_tensor(tn(LLM_TENSOR_INDEXER_ATTN_Q_B, "weight", i), {q_lora_rank, hparams.indexer_n_head * hparams.indexer_head_size}, flags);
+        if (i < (int) hparams.n_layer_dense_lead) {
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, flags);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, flags);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, flags);
+        } else {
+            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, flags);
+            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED);
+
+            if (n_expert == 0) {
+                throw std::runtime_error("n_expert must be > 0");
+            }
+            if (n_expert_used == 0) {
+                throw std::runtime_error("n_expert_used must be > 0");
+            }
+
+            // MoE branch
+            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, flags);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, flags);
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, flags);
+
+            // Shared expert branch
+            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags);
+            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {        n_ff_exp * n_expert_shared, n_embd}, flags);
+            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_exp * n_expert_shared}, flags);
+        }
+
+        // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
+        if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags);
+            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags);
+            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags);
+
+            // Optional tensors
+            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
+            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
+            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags | TENSOR_NOT_REQUIRED);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_glm_dsa::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+

examples/talk-llama/models/glm4-moe.cpp

@@ -1,6 +1,139 @@
 #include "models.h"
 
-llm_build_glm4_moe::llm_build_glm4_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_glm4_moe::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,     hparams.n_ff_exp);
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,    hparams.f_norm_rms_eps);
+    ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
+
+    // MoE parameters
+    ml.get_key(LLM_KV_EXPERT_COUNT,                hparams.n_expert);
+    ml.get_key(LLM_KV_EXPERT_USED_COUNT,           hparams.n_expert_used);
+    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,         hparams.n_expert_shared);
+    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,        hparams.expert_weights_scale, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,         hparams.expert_weights_norm, false);
+
+    // Expert gating function (GLM-4.5 uses sigmoid)
+    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func, false);
+    if (hparams.expert_gating_func == LLAMA_EXPERT_GATING_FUNC_TYPE_NONE) {
+        hparams.expert_gating_func =  LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID;
+    }
+
+    // NextN/MTP parameters
+    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS,        hparams.nextn_predict_layers, false);
+    GGML_ASSERT(hparams.nextn_predict_layers < hparams.n_layer && "nextn_predict_layers must be < n_layer");
+
+    // TODO: when MTP is implemented, this should probably be updated if needed
+    hparams.n_layer_kv_from_start = hparams.n_layer - hparams.nextn_predict_layers;
+
+    switch (hparams.n_layer) {
+        case 47: type = LLM_TYPE_106B_A12B; break; // GLM-4.5-Air (46 layers + 1 NextN layer)
+        case 48: type = LLM_TYPE_102B_A12B; break; // Solar Open
+        case 93: type = LLM_TYPE_355B_A32B; break; // GLM-4.5 (92 layers + 1 NextN layer)
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_glm4_moe::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+    const int64_t n_expert_shared = hparams.n_expert_shared;
+
+
+    GGML_ASSERT(hparams.n_expert > 0 && "n_expert must be > 0 for GLM4_MOE MoE layers");
+    GGML_ASSERT(hparams.n_expert_used > 0 && "n_expert_used must be > 0 for GLM4_MOE MoE layers");
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);
+    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED);
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED);
+    }
+
+    // Load ALL tensors including NextN layer to satisfy total tensor count
+    // but only PROCESS up to last layer (skipping final NextN layer) in forward pass
+    for (int i = 0; i < n_layer; ++i) {
+        int flags = 0;
+        if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+            // skip all tensors in the NextN layers
+            flags |= TENSOR_SKIP;
+        }
+
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, flags);
+
+        // GLM-style attention with bias terms
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, flags);
+
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, flags);
+
+        // K/Q norm tensors (optional for GLM-4.5 355B variant)
+        layer.attn_q_norm = create_tensor(
+            tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), { n_embd_head_k }, TENSOR_NOT_REQUIRED | flags);
+        layer.attn_k_norm = create_tensor(
+            tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), { n_embd_head_k }, TENSOR_NOT_REQUIRED | flags);
+
+        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), { n_embd }, flags);
+
+        // Check if this layer uses MoE or dense FFN based on n_layer_dense_lead
+        // GLM 4.5 uses hybrid architecture: layer 0 is dense, layers 1+ are MoE
+        const bool use_moe = (static_cast<uint32_t>(i) >= hparams.n_layer_dense_lead);
+
+        if (use_moe) {
+            // MoE layers
+            layer.ffn_gate_inp =
+                create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), { n_embd, n_expert }, flags);
+            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), { n_expert }, flags);
+
+            // MoE branch
+            const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+
+            layer.ffn_gate_exps = create_tensor(
+                tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, flags);
+            layer.ffn_down_exps = create_tensor(
+                tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), { n_ff_exp, n_embd, n_expert }, flags);
+            layer.ffn_up_exps = create_tensor(
+                tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i), { n_embd, n_ff_exp, n_expert }, flags);
+
+            // Shared expert
+            if (n_expert_shared > 0) {
+                const int64_t n_ff_shexp = n_ff_exp * n_expert_shared;
+                layer.ffn_gate_shexp = create_tensor(
+                    tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), { n_embd, n_ff_shexp }, flags);
+                layer.ffn_down_shexp = create_tensor(
+                    tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), { n_ff_shexp, n_embd }, flags);
+                layer.ffn_up_shexp = create_tensor(
+                    tn(LLM_TENSOR_FFN_UP_SHEXP, "weight", i), { n_embd, n_ff_shexp }, flags);
+            }
+        } else {
+            // Dense layers (first k layers) - GLM uses separate gate/up projections
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), { n_embd, n_ff }, flags);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }, flags);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), { n_embd, n_ff }, flags);
+        }
+
+        // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
+        if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags);
+            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags);
+            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags);
+
+            // Optional tensors
+            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
+            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
+            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags | TENSOR_NOT_REQUIRED);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_glm4_moe::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_glm4_moe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/glm4.cpp

@@ -1,6 +1,78 @@
 #include "models.h"
 
-llm_build_glm4::llm_build_glm4(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_glm4::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,    hparams.f_norm_rms_eps);
+    ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
+
+    // NextN/MTP parameters (GLM-OCR)
+    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.nextn_predict_layers, false);
+    GGML_ASSERT(hparams.nextn_predict_layers < hparams.n_layer && "nextn_predict_layers must be < n_layer");
+
+    // TODO: when MTP is implemented, this should probably be updated if needed
+    hparams.n_layer_kv_from_start = hparams.n_layer - hparams.nextn_predict_layers;
+
+    switch (hparams.n_layer) {
+        case 17: type = LLM_TYPE_1B; break; // GLM-OCR
+        case 40: type = LLM_TYPE_9B; break;
+        case 61: type = LLM_TYPE_32B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_glm4::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        int flags = 0;
+        if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+            // skip all tensors in the NextN layers
+            flags |= TENSOR_SKIP;
+        }
+
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, flags);
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, flags);
+
+        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, flags);
+
+        layer.attn_post_norm = create_tensor(tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}, flags);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, flags);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, flags);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff * 2}, flags);
+
+        layer.ffn_post_norm  = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, flags);
+
+        // NextN/MTP tensors (preserved but unused) - conditionally load for last nextn_predict_layers
+        if (hparams.nextn_predict_layers > 0 && static_cast<uint32_t>(i) >= n_layer - hparams.nextn_predict_layers) {
+            layer.nextn.eh_proj          = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), { 2 * n_embd, n_embd }, flags);
+            layer.nextn.enorm            = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM, "weight", i), { n_embd }, flags);
+            layer.nextn.hnorm            = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM, "weight", i), { n_embd }, flags);
+
+            // Optional tensors
+            layer.nextn.embed_tokens     = create_tensor(tn(LLM_TENSOR_NEXTN_EMBED_TOKENS, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
+            layer.nextn.shared_head_head = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_HEAD, "weight", i), { n_embd, n_vocab }, flags | TENSOR_NOT_REQUIRED);
+            layer.nextn.shared_head_norm = create_tensor(tn(LLM_TENSOR_NEXTN_SHARED_HEAD_NORM, "weight", i), { n_embd }, flags | TENSOR_NOT_REQUIRED);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_glm4::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_glm4::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/gpt2.cpp

@@ -1,6 +1,60 @@
 #include "models.h"
 
-llm_build_gpt2::llm_build_gpt2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_gpt2::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+    switch (hparams.n_layer) {
+        case 12: type = LLM_TYPE_SMALL; break;
+        case 24: type = LLM_TYPE_MEDIUM; break;
+        case 36: type = LLM_TYPE_LARGE; break;
+        case 48: type = LLM_TYPE_XL; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_gpt2::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+    pos_embd = create_tensor(tn(LLM_TENSOR_POS_EMBD,   "weight"), {n_embd, n_ctx_train}, 0);
+
+    // output
+    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);
+    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM,   "weight", i), {n_embd}, 0);
+        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM,   "bias", i),   {n_embd}, 0);
+
+        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
+        layer.wqkv_b = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, 0);
+
+        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd}, 0);
+
+        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, 0);
+
+        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, 0);
+
+        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_gpt2::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_gpt2::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/gptneox.cpp

@@ -1,6 +1,89 @@
 #include "models.h"
 
-llm_build_gptneox::llm_build_gptneox(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_gptneox::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+    ml.get_key(LLM_KV_USE_PARALLEL_RESIDUAL,   hparams.use_par_res);
+    switch (hparams.n_layer) {
+        case 6:
+            switch (hparams.n_ff()) {
+                case 512:  type = LLM_TYPE_14M; break;
+                case 2048: type = LLM_TYPE_70M; break;
+                default:   type = LLM_TYPE_UNKNOWN;
+            } break;
+        case 12:
+            switch (hparams.n_ff()) {
+                case 3072: type = LLM_TYPE_160M; break;
+                default: type = LLM_TYPE_UNKNOWN;
+            } break;
+        case 16:
+            switch (hparams.n_ff()) {
+                case 8192: type = LLM_TYPE_1B; break;
+                default: type = LLM_TYPE_UNKNOWN;
+            } break;
+        case 24:
+            switch (hparams.n_ff()) {
+                case 4096: type = LLM_TYPE_410M; break;
+                case 8192: type = LLM_TYPE_1_4B; break;
+                default: type = LLM_TYPE_UNKNOWN;
+            } break;
+        case 32:
+            switch (hparams.n_ff()) {
+                case 10240: type = LLM_TYPE_2_8B; break;
+                case 16384: type = LLM_TYPE_6_9B; break;
+                default: type = LLM_TYPE_UNKNOWN;
+            } break;
+        case 36:
+            switch (hparams.n_ff()) {
+                case 20480: type = LLM_TYPE_12B; break;
+                default: type = LLM_TYPE_UNKNOWN;
+            } break;
+        case 44:
+            switch (hparams.n_ff()) {
+                case 24576: type = LLM_TYPE_20B; break;
+                default: type = LLM_TYPE_UNKNOWN;
+            } break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_gptneox::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);
+    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, 0);
+
+        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
+        layer.wqkv_b = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, 0);
+
+        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd}, 0);
+
+        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, 0);
+
+        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, 0);
+
+        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_gptneox::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_gptneox::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/granite-hybrid.cpp

@@ -1,6 +1,137 @@
 #include "models.h"
 
-llm_build_granite_hybrid::llm_build_granite_hybrid(const llama_model & model, const llm_graph_params & params) :
+void llama_model_granite_hybrid::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_LOGIT_SCALE,                 hparams.f_logit_scale, /* required */ false);
+    ml.get_key(LLM_KV_RESIDUAL_SCALE,              hparams.f_residual_scale, /* required */ false);
+    ml.get_key(LLM_KV_EMBEDDING_SCALE,             hparams.f_embedding_scale, /* required */ false);
+    ml.get_key(LLM_KV_ATTENTION_SCALE,             hparams.f_attention_scale, /* required */ false);
+
+    ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);
+    ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);
+    ml.get_key(LLM_KV_SSM_STATE_SIZE,     hparams.ssm_d_state);
+    ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
+    ml.get_key(LLM_KV_SSM_GROUP_COUNT,    hparams.ssm_n_group);
+
+    // Granite uses rope_finetuned as a switch for rope, so default to true
+    bool rope_finetuned = true;
+    ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false);
+    hparams.rope_finetuned = rope_finetuned;
+
+    // A layer is recurrent IFF the n_head_kv value is set to 0
+    for (uint32_t i = 0; i < hparams.n_layer; ++i) {
+        hparams.recurrent_layer_arr[i] = hparams.n_head_kv(i) == 0;
+    }
+
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    switch (hparams.n_embd) {
+        case 768: type = LLM_TYPE_350M; break;
+        case 1536: type = (hparams.n_ff() == 512 ? LLM_TYPE_7B_A1B : LLM_TYPE_1B); break;
+        case 2048: case 2560: type = LLM_TYPE_3B; break;
+        case 4096: type = LLM_TYPE_32B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+
+    // For Granite MoE Shared
+    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false);
+}
+
+void llama_model_granite_hybrid::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    // mamba2 Mixer SSM params
+    // NOTE: int64_t for tensor dimensions
+    const int64_t d_conv     = hparams.ssm_d_conv;
+    const int64_t d_inner    = hparams.ssm_d_inner;
+    const int64_t d_state    = hparams.ssm_d_state;
+    const int64_t n_ssm_head = hparams.ssm_dt_rank;
+    const int64_t n_group    = hparams.ssm_n_group;
+    const int64_t d_in_proj  = 2*d_inner + 2*n_group*d_state + n_ssm_head;
+
+    // only an expansion factor of 2 is supported for now
+    GGML_ASSERT(2 * n_embd == d_inner);
+
+    // embeddings
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    {
+        output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+        output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+        // if output is NULL, init from the input tok embed, duplicated to allow offloading
+        if (output == NULL) {
+            output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+        }
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        // norm
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        if (hparams.is_recurrent(i)) {
+            // ssm layers
+            layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, d_in_proj}, 0);
+
+            layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner + 2*n_group*d_state}, 0);
+            layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner + 2*n_group*d_state}, TENSOR_NOT_REQUIRED);
+
+            layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {n_ssm_head}, 0);
+
+            // no "weight" suffix for these
+            layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_ssm_head}, 0);
+            layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, n_ssm_head}, 0);
+
+            layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {d_inner / n_group, n_group}, 0);
+
+            // out_proj
+            layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0);
+        } else {
+            // attention layers (with optional bias)
+            const int64_t n_head_i = hparams.n_head(i);
+            const int64_t n_embd_k_gqa_i = hparams.n_embd_k_gqa(i);
+            const int64_t n_embd_v_gqa_i = hparams.n_embd_v_gqa(i);
+            create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head_i, n_embd_k_gqa_i, n_embd_v_gqa_i, 0);
+            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head_i, n_embd}, 0);
+            layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+        }
+
+        // feed forward (w/ optional biases)
+        if (n_expert > 0) {
+            // MoE FFN
+            layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+            layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
+            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, TENSOR_NOT_REQUIRED);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);
+
+            // For Granite MoE Shared
+            if (hparams.n_ff_shexp > 0) {
+                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
+                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
+                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
+            }
+        } else {
+            layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+            layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
+            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_granite_hybrid::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_granite_hybrid::graph::graph(const llama_model & model, const llm_graph_params & params) :
     llm_build_mamba_base(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
@@ -67,7 +198,7 @@ llm_build_granite_hybrid::llm_build_granite_hybrid(const llama_model & model, co
     ggml_build_forward_expand(gf, cur);
 }
 
-ggml_tensor * llm_build_granite_hybrid::build_attention_layer(ggml_tensor *             cur,
+ggml_tensor * llama_model_granite_hybrid::graph::build_attention_layer(ggml_tensor *             cur,
                                                               ggml_tensor *             inp_pos,
                                                               llm_graph_input_attn_kv * inp_attn,
                                                               const llama_model &       model,
@@ -98,7 +229,7 @@ ggml_tensor * llm_build_granite_hybrid::build_attention_layer(ggml_tensor *
     return cur;
 }
 
-ggml_tensor * llm_build_granite_hybrid::build_layer_ffn(ggml_tensor *       cur,
+ggml_tensor * llama_model_granite_hybrid::graph::build_layer_ffn(ggml_tensor *       cur,
                                                         ggml_tensor *       inpSA,
                                                         const llama_model & model,
                                                         const int           il) {

examples/talk-llama/models/granite-moe.cpp

@@ -0,0 +1,89 @@
+#include "models.h"
+
+void llama_model_granite_moe::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_LOGIT_SCALE,                 hparams.f_logit_scale);
+    ml.get_key(LLM_KV_RESIDUAL_SCALE,              hparams.f_residual_scale, false);
+    ml.get_key(LLM_KV_EMBEDDING_SCALE,             hparams.f_embedding_scale, false);
+    ml.get_key(LLM_KV_ATTENTION_SCALE,             hparams.f_attention_scale, false);
+
+    // Granite uses rope_finetuned as a switch for rope, so default to true
+    bool rope_finetuned = true;
+    ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false);
+    hparams.rope_finetuned = rope_finetuned;
+
+    switch (hparams.n_layer) {
+        case 32: type = LLM_TYPE_3B; break;
+        case 40: type = LLM_TYPE_3B; break;
+        // Add additional layer/vocab/etc checks here for other model sizes
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+
+    // For Granite MoE Shared
+    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false);
+}
+
+void llama_model_granite_moe::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        // optional bias tensors
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
+            layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+            layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+        }
+        else {
+            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+        }
+
+        if (n_expert == 0) {
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+
+            // optional MLP bias
+            layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
+            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
+        } else {
+            layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
+            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, TENSOR_NOT_REQUIRED);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);
+
+            // For Granite MoE Shared
+            if (hparams.n_ff_shexp > 0) {
+                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
+                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
+                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
+            }
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_granite_moe::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+

examples/talk-llama/models/granite.cpp

@@ -1,6 +1,93 @@
 #include "models.h"
 
-llm_build_granite::llm_build_granite(
+void llama_model_granite::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_LOGIT_SCALE,                 hparams.f_logit_scale);
+    ml.get_key(LLM_KV_RESIDUAL_SCALE,              hparams.f_residual_scale, false);
+    ml.get_key(LLM_KV_EMBEDDING_SCALE,             hparams.f_embedding_scale, false);
+    ml.get_key(LLM_KV_ATTENTION_SCALE,             hparams.f_attention_scale, false);
+
+    // Granite uses rope_finetuned as a switch for rope, so default to true
+    bool rope_finetuned = true;
+    ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false);
+    hparams.rope_finetuned = rope_finetuned;
+
+    switch (hparams.n_layer) {
+        case 32: type = LLM_TYPE_3B; break;
+        case 40: type = LLM_TYPE_3B; break;
+        // Add additional layer/vocab/etc checks here for other model sizes
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+
+    // For Granite MoE Shared
+    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, /* required */ false);
+}
+
+void llama_model_granite::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        // optional bias tensors
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
+            layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+            layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+        }
+        else {
+            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+        }
+
+        if (n_expert == 0) {
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+
+            // optional MLP bias
+            layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
+            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
+        } else {
+            layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
+            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, TENSOR_NOT_REQUIRED);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);
+
+            // For Granite MoE Shared
+            if (hparams.n_ff_shexp > 0) {
+                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
+                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
+                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
+            }
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_granite::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_granite::graph::graph(
     const llama_model & model,
     const llm_graph_params & params)
     : llm_graph_context(params) {
@@ -68,7 +155,7 @@ llm_build_granite::llm_build_granite(
     ggml_build_forward_expand(gf, cur);
 }
 
-ggml_tensor * llm_build_granite::build_attention_layer(
+ggml_tensor * llama_model_granite::graph::build_attention_layer(
           ggml_tensor             * cur,
           ggml_tensor             * inp_pos,
           llm_graph_input_attn_kv * inp_attn,
@@ -107,7 +194,7 @@ ggml_tensor * llm_build_granite::build_attention_layer(
     return cur;
 }
 
-ggml_tensor * llm_build_granite::build_layer_ffn(
+ggml_tensor * llama_model_granite::graph::build_layer_ffn(
           ggml_tensor       * cur,
           ggml_tensor       * inpSA,
     const llama_model       & model,

examples/talk-llama/models/grok.cpp

@@ -1,6 +1,89 @@
 #include "models.h"
 
-llm_build_grok::llm_build_grok(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_grok::load_arch_hparams(llama_model_loader & ml) {
+    // defaults for old GGUFs
+    hparams.yarn_beta_fast = 8.0f;
+    hparams.f_logit_scale = 0.5773502691896257f;
+    hparams.f_embedding_scale = 78.38367176906169f;
+    hparams.f_attn_out_scale = 0.08838834764831845f;
+    hparams.f_attn_logit_softcapping = 30.0f;
+    hparams.f_router_logit_softcapping = 30.0f;
+    // no final_logit_softcapping in grok-1
+    hparams.f_final_logit_softcapping = 0.0f;
+
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,  hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,   hparams.n_ff_exp, false);
+    ml.get_key(LLM_KV_LOGIT_SCALE,                  hparams.f_logit_scale, false);
+    ml.get_key(LLM_KV_EMBEDDING_SCALE,              hparams.f_embedding_scale, false);
+    ml.get_key(LLM_KV_ATTENTION_OUTPUT_SCALE,       hparams.f_attn_out_scale, false);
+    ml.get_key(LLM_KV_ATTN_LOGIT_SOFTCAPPING,       hparams.f_attn_logit_softcapping, false);
+    ml.get_key(LLM_KV_ROUTER_LOGIT_SOFTCAPPING,     hparams.f_router_logit_softcapping, false);
+    ml.get_key(LLM_KV_FINAL_LOGIT_SOFTCAPPING,      hparams.f_final_logit_softcapping, false);
+
+    ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_LENGTH,  hparams.attn_temp_length, false);
+    ml.get_key(LLM_KV_ROPE_SCALING_YARN_EXT_FACTOR,  hparams.yarn_ext_factor, false);
+    ml.get_key(LLM_KV_ROPE_SCALING_YARN_ATTN_FACTOR, hparams.yarn_attn_factor, false);
+    ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_FAST,   hparams.yarn_beta_fast, false);
+    ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_SLOW,   hparams.yarn_beta_slow, false);
+
+    switch (hparams.n_layer) {
+        case 64: type = LLM_TYPE_314B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_grok::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    if (n_expert == 0) {
+        throw std::runtime_error(arch_name() + " model cannot have zero experts");
+    }
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff/* / n_expert_used*/; // grok-1 n_ff_exp == n_ff
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+        layer.attn_out_norm   = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff,   n_embd}, TENSOR_NOT_REQUIRED);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);
+
+        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
+        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff_exp, n_expert}, TENSOR_NOT_REQUIRED);
+        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd,   n_expert}, 0);
+        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff_exp, n_expert}, 0);
+
+        layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
+        if (!layer.ffn_post_norm) {
+            layer.ffn_post_norm = create_tensor(tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}, 0);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_grok::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_grok::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/grovemoe.cpp

@@ -1,6 +1,70 @@
 #include "models.h"
 
-llm_build_grovemoe::llm_build_grovemoe(const llama_model & model, const llm_graph_params & params) :
+void llama_model_grovemoe::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
+    ml.get_key(LLM_KV_EXPERT_CHUNK_FEED_FORWARD_LENGTH,  hparams.n_ff_chexp, false);
+    ml.get_key(LLM_KV_EXPERT_GROUP_SCALE,                hparams.expert_group_scale);
+    ml.get_key(LLM_KV_EXPERTS_PER_GROUP,                 hparams.n_group_experts);
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);
+
+    switch (hparams.n_layer) {
+        case 48: type = LLM_TYPE_30B_A3B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_grovemoe::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    GGML_ASSERT(n_expert > 0 && "n_expert must be > 0 for GROVEMOE");
+    GGML_ASSERT(n_expert_used > 0 && "n_expert_used must be > 0 for GROVEMOE");
+    GGML_ASSERT(hparams.n_group_experts > 0 && "n_group_experts must be > 0 for GROVEMOE");
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
+        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+
+        // MoE branch
+        const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+        const int64_t n_ff_chexp = hparams.n_ff_chexp ? hparams.n_ff_chexp : n_embd_head_k;
+        const int64_t n_chunk_expert = n_expert / hparams.n_group_experts;
+
+        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
+        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+
+        layer.ffn_gate_chexps = create_tensor(tn(LLM_TENSOR_FFN_GATE_CHEXPS, "weight", i), {  n_embd, n_ff_chexp, n_chunk_expert}, 0);
+        layer.ffn_down_chexps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_CHEXPS, "weight", i), {n_ff_chexp,   n_embd, n_chunk_expert}, 0);
+        layer.ffn_up_chexps   = create_tensor(tn(LLM_TENSOR_FFN_UP_CHEXPS,   "weight", i), {  n_embd, n_ff_chexp, n_chunk_expert}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_grovemoe::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_grovemoe::graph::graph(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     const int64_t n_embd_head    = hparams.n_embd_head_v();
     const int64_t n_chunk_expert = n_expert / hparams.n_group_experts;

examples/talk-llama/models/hunyuan-dense.cpp

@@ -1,132 +1,6 @@
 #include "models.h"
 
-llm_build_hunyuan_dense::llm_build_hunyuan_dense(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    const int64_t n_embd_head = hparams.n_embd_head_v();
-
-    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
-    GGML_ASSERT(n_embd_head == n_rot);
-
-    const bool use_mrope = hparams.use_mrope();
-
-    int sections[4];
-    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
-
-    ggml_tensor * cur;
-    ggml_tensor * inpL;
-
-    inpL = build_inp_embd(model.tok_embd);
-
-    // inp_pos - contains the positions
-    ggml_tensor * inp_pos = build_inp_pos();
-
-    auto * inp_attn = build_attn_inp_kv();
-
-    const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
-
-    ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-    for (int il = 0; il < n_layer; ++il) {
-        ggml_tensor * inpSA = inpL;
-
-        // norm
-        cur = build_norm(inpL,
-                model.layers[il].attn_norm, NULL,
-                LLM_NORM_RMS, il);
-        cb(cur, "attn_norm", il);
-        // self-attention
-        {
-            // rope freq factors for llama3; may return nullptr for llama2 and other models
-            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
-
-            // compute Q and K and RoPE them
-            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
-                    n_embd_head, n_head, n_head_kv, il);
-
-            if (use_mrope) {
-                Qcur = ggml_rope_multi(
-                            ctx0, Qcur, inp_pos, rope_factors,
-                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
-                            ext_factor, attn_factor, beta_fast, beta_slow
-                            );
-
-                Kcur = ggml_rope_multi(
-                            ctx0, Kcur, inp_pos, rope_factors,
-                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
-                            ext_factor, attn_factor, beta_fast, beta_slow
-                            );
-            } else {
-                Qcur = ggml_rope_ext(
-                            ctx0, Qcur, inp_pos, rope_factors,
-                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                            ext_factor, attn_factor, beta_fast, beta_slow
-                            );
-
-                Kcur = ggml_rope_ext(
-                            ctx0, Kcur, inp_pos, rope_factors,
-                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
-                            ext_factor, attn_factor, beta_fast, beta_slow
-                            );
-            }
-
-            cb(Qcur, "Qcur", il);
-            cb(Kcur, "Kcur", il);
-            cb(Vcur, "Vcur", il);
-
-            Kcur = build_norm(Kcur,
-                        model.layers[il].attn_k_norm, nullptr,
-                        LLM_NORM_RMS, il);
-            cb(Kcur, "Kcur_norm", il);
-
-            Qcur = build_norm(Qcur,
-                        model.layers[il].attn_q_norm, nullptr,
-                        LLM_NORM_RMS, il);
-            cb(Qcur, "Qcur_norm", il);
-
-            cur = build_attn(inp_attn,
-                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
-                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
-            cb(cur, "attn_out", il);
-        }
-        if (il == n_layer - 1 && inp_out_ids) {
-            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-        }
-        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-        cb(ffn_inp, "ffn_inp", il);
-
-        cur = build_norm(ffn_inp,
-                model.layers[il].ffn_norm, NULL,
-                LLM_NORM_RMS, il);
-        cb(cur, "ffn_norm", il);
-        // feed-forward network (non-MoE)
-        ggml_tensor * cur_mlp = build_ffn(cur,
-                    model.layers[il].ffn_up,   NULL, NULL,
-                    model.layers[il].ffn_gate, NULL, NULL,
-                    model.layers[il].ffn_down, NULL, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-        cb(cur_mlp, "ffn_out", il);
-
-        cur = ggml_add(ctx0, cur_mlp, ffn_inp);
-
-        cur = build_cvec(cur, il);
-        cb(cur, "l_out", il);
-
-        // input for next layer
-        inpL = cur;
-    }
-    cur = inpL;
-
-    cur = build_norm(cur,
-            model.output_norm, NULL,
-            LLM_NORM_RMS, -1);
-
-    cb(cur, "result_norm", -1);
-    res->t_embd = cur;
-    // lm_head
-    cur = build_lora_mm(model.output, cur);
-    cb(cur, "result_output", -1);
-    res->t_logits = cur;
-
-    ggml_build_forward_expand(gf, cur);
+std::unique_ptr<llm_graph_context> llama_model_hunyuan_dense::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
 }
+

examples/talk-llama/models/hunyuan-moe.cpp

@@ -1,6 +1,59 @@
 #include "models.h"
 
-llm_build_hunyuan_moe::llm_build_hunyuan_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_hunyuan_moe::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,       hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
+    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp, false);
+
+    switch (hparams.n_layer) {
+        case 32: type = LLM_TYPE_A13B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_hunyuan_moe::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+        const uint32_t n_ff_shexp = hparams.n_ff_shexp > 0 ? hparams.n_ff_shexp : hparams.n_ff(i);
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
+        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
+        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, 0);
+        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);
+        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);
+
+        layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp}, 0);
+        layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, 0);
+        layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_hunyuan_moe::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_hunyuan_moe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/hunyuan-vl.cpp

@@ -0,0 +1,189 @@
+#include "models.h"
+
+void llama_model_hunyuan_vl::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key_or_arr(LLM_KV_ROPE_DIMENSION_SECTIONS, hparams.rope_sections, 4, false);
+
+    // XDRoPE / NTK-aware scaling: base = rope_theta * alpha^(dim / (dim - 2))
+    if (hparams.rope_scaling_alpha > 0.0f) {
+        const int dim = hparams.n_embd_head_k();
+        hparams.rope_freq_base_train = hparams.rope_freq_base_train
+            * powf(hparams.rope_scaling_alpha, (float)dim / (float)(dim - 2));
+    }
+
+    switch (hparams.n_embd) {
+        case 1024: type = LLM_TYPE_0_5B; break;
+        case 2048: type = LLM_TYPE_1_8B; break;
+        case 3072: type = LLM_TYPE_4B; break;
+        case 4096: type = LLM_TYPE_7B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_hunyuan_vl::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
+        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_hunyuan_vl::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_hunyuan_vl::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    const int64_t n_embd_head = hparams.n_embd_head_v();
+
+    GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
+    GGML_ASSERT(n_embd_head == n_rot);
+
+    const bool use_mrope = hparams.use_mrope();
+
+    int sections[4];
+    std::copy(std::begin(hparams.rope_sections), std::begin(hparams.rope_sections) + 4, sections);
+
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    // inp_pos - contains the positions
+    ggml_tensor * inp_pos = build_inp_pos();
+
+    auto * inp_attn = build_attn_inp_kv();
+
+    const float kq_scale = 1.0f / sqrtf(float(n_embd_head));
+
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    for (int il = 0; il < n_layer; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        // norm
+        cur = build_norm(inpL,
+                model.layers[il].attn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "attn_norm", il);
+        // self-attention
+        {
+            // rope freq factors for llama3; may return nullptr for llama2 and other models
+            ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
+
+            // compute Q and K and RoPE them
+            auto [Qcur, Kcur, Vcur] = build_qkv(model.layers[il], cur,
+                    n_embd_head, n_head, n_head_kv, il);
+
+            if (use_mrope) {
+                Qcur = ggml_rope_multi(
+                            ctx0, Qcur, inp_pos, rope_factors,
+                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow
+                            );
+
+                Kcur = ggml_rope_multi(
+                            ctx0, Kcur, inp_pos, rope_factors,
+                            n_rot, sections, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow
+                            );
+            } else {
+                Qcur = ggml_rope_ext(
+                            ctx0, Qcur, inp_pos, rope_factors,
+                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow
+                            );
+
+                Kcur = ggml_rope_ext(
+                            ctx0, Kcur, inp_pos, rope_factors,
+                            n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
+                            ext_factor, attn_factor, beta_fast, beta_slow
+                            );
+            }
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            Kcur = build_norm(Kcur,
+                        model.layers[il].attn_k_norm, nullptr,
+                        LLM_NORM_RMS, il);
+            cb(Kcur, "Kcur_norm", il);
+
+            Qcur = build_norm(Qcur,
+                        model.layers[il].attn_q_norm, nullptr,
+                        LLM_NORM_RMS, il);
+            cb(Qcur, "Qcur_norm", il);
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, model.layers[il].wo_b, model.layers[il].wo_s,
+                    Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
+            cb(cur, "attn_out", il);
+        }
+        if (il == n_layer - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+        // feed-forward network (non-MoE)
+        ggml_tensor * cur_mlp = build_ffn(cur,
+                    model.layers[il].ffn_up,   NULL, NULL,
+                    model.layers[il].ffn_gate, NULL, NULL,
+                    model.layers[il].ffn_down, NULL, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+        cb(cur_mlp, "ffn_out", il);
+
+        cur = ggml_add(ctx0, cur_mlp, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

examples/talk-llama/models/internlm2.cpp

@@ -1,6 +1,43 @@
 #include "models.h"
 
-llm_build_internlm2::llm_build_internlm2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_internlm2::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    switch (hparams.n_layer) {
+        case 32: type = LLM_TYPE_7B; break;
+        case 48: type = LLM_TYPE_20B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_internlm2::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+        // layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
+        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
+
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_internlm2::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_internlm2::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/jais.cpp

@@ -1,6 +1,58 @@
 #include "models.h"
 
-llm_build_jais::llm_build_jais(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_jais::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+    ml.get_key(LLM_KV_ATTENTION_MAX_ALIBI_BIAS, hparams.f_max_alibi_bias, false);
+
+    switch (hparams.n_layer) {
+        case 24: type = LLM_TYPE_1_3B; break;
+        case 40: type = LLM_TYPE_13B; break;
+        /* TODO: add variants */
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_jais::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);
+    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM,   "weight", i), {n_embd}, 0);
+        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM,   "bias", i),   {n_embd}, 0);
+
+        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
+        layer.wqkv_b = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, 0);
+
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
+
+        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, 0);
+
+        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, 0);
+
+        layer.ffn_gate   = create_tensor(tn(LLM_TENSOR_FFN_GATE,   "weight", i), {n_embd, n_ff}, 0);
+        layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE,   "bias", i),   {n_ff}, 0);
+
+        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_jais::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_jais::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/jais2.cpp

@@ -1,8 +1,63 @@
 #include "models.h"
 
+void llama_model_jais2::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+
+    switch (hparams.n_layer) {
+        case 32: type = LLM_TYPE_8B; break;
+        case 68: type = LLM_TYPE_70B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_jais2::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, 0);
+    output        = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    if (!output) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, 0);
+
+        layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * n_head}, 0);
+        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}, 0);
+        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        // attention biases - all have shape n_embd (output dimension of projections)
+        layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q, "bias", i), {n_embd}, 0);
+        layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K, "bias", i), {n_embd}, 0);
+        layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V, "bias", i), {n_embd}, 0);
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0);
+
+        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, 0);
+
+        // Jais-2 uses simple MLP (no gate) with biases
+        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, 0);
+        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_jais2::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
 // JAIS-2 model graph builder
 // Uses: LayerNorm (not RMSNorm), relu2 activation, separate Q/K/V, RoPE embeddings
-llm_build_jais2::llm_build_jais2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+llama_model_jais2::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/jamba.cpp

@@ -1,6 +1,111 @@
 #include "models.h"
 
-llm_build_jamba::llm_build_jamba(const llama_model & model, const llm_graph_params & params) : llm_build_mamba_base(params) {
+void llama_model_jamba::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);
+    ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);
+    ml.get_key(LLM_KV_SSM_STATE_SIZE,     hparams.ssm_d_state);
+    ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
+
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    for (uint32_t i = 0; i < hparams.n_layer; ++i) {
+        hparams.recurrent_layer_arr[i] = hparams.n_head_kv(i) == 0;
+    }
+
+    switch (hparams.n_layer) {
+        // TODO: Jamba layers are a bit heterogeneous, so naming this is hard.
+        case 12: // 900M  8x???M
+        case 32: // 51B  16x?B
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_jamba::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    const int64_t d_conv  = hparams.ssm_d_conv;
+    const int64_t d_inner = hparams.ssm_d_inner;
+    const int64_t d_state = hparams.ssm_d_state;
+    const int64_t dt_rank = hparams.ssm_dt_rank;
+
+    // only an expansion factor of 2 is supported for now
+    GGML_ASSERT(2 * n_embd == d_inner);
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    {
+        output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+
+        output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+        // if output is NULL, init from the input tok embed, duplicated to allow offloading
+        if (output == NULL) {
+            output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+        }
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        const int64_t n_head_kv = hparams.n_head_kv(i);
+        const int64_t n_embd_gqa = hparams.n_embd_v_gqa(i);
+
+        auto & layer = layers[i];
+
+        // norm
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        if (n_head_kv == 0) {
+            // Mamba layer
+            layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, 2*d_inner}, 0);
+
+            layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner}, 0);
+            layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner}, 0);
+
+            layer.ssm_x = create_tensor(tn(LLM_TENSOR_SSM_X, "weight", i), {d_inner, dt_rank + 2*d_state}, 0);
+
+            layer.ssm_dt_norm = create_tensor(tn(LLM_TENSOR_SSM_DT_NORM, "weight", i), {dt_rank}, 0);
+
+            layer.ssm_dt = create_tensor(tn(LLM_TENSOR_SSM_DT, "weight", i), {dt_rank, d_inner}, 0);
+            layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {d_inner}, 0);
+
+            layer.ssm_b_norm = create_tensor(tn(LLM_TENSOR_SSM_B_NORM, "weight", i), {d_state}, 0);
+            layer.ssm_c_norm = create_tensor(tn(LLM_TENSOR_SSM_C_NORM, "weight", i), {d_state}, 0);
+
+            // no "weight" suffix for these
+            layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {d_state, d_inner}, 0);
+            layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {d_inner}, 0);
+
+            // out_proj
+            layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0);
+        } else {
+            // Attention layers
+
+            create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
+            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+        }
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED);
+
+        if (layer.ffn_gate_inp) {
+            // MoE
+            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff, n_expert}, 0);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff, n_embd, n_expert}, 0);
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff, n_expert}, 0);
+        } else {
+            // FFN (no MoE)
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_jamba::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_jamba::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_build_mamba_base(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     ggml_tensor * cur;

examples/talk-llama/models/jina-bert-v2.cpp

@@ -0,0 +1,66 @@
+#include "models.h"
+
+void llama_model_jina_bert_v2::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
+    hparams.f_max_alibi_bias = 8.0f;
+
+    switch (hparams.n_layer) {
+        case 4:  type = LLM_TYPE_33M;  break; // jina-embeddings-small
+        case 12: type = LLM_TYPE_137M; break; // jina-embeddings-base
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_jina_bert_v2::load_arch_tensors(llama_model_loader & ml) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd  = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0); // word_embeddings
+    type_embd = create_tensor(tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_token_types}, 0); // token_type_embeddings
+
+    tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0); // LayerNorm
+    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias",   0), {n_embd}, 0); // LayerNorm bias
+
+    cls   = create_tensor(tn(LLM_TENSOR_CLS, "weight"), {n_embd, 1}, TENSOR_NOT_REQUIRED);
+    cls_b = create_tensor(tn(LLM_TENSOR_CLS, "bias"),   {1},         TENSOR_NOT_REQUIRED);
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i]; // JinaBertLayer
+
+        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
+
+        layer.attn_q_norm   = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
+        layer.attn_q_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "bias",   i), {n_embd}, TENSOR_NOT_REQUIRED);
+
+        layer.attn_k_norm   = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
+        layer.attn_k_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "bias",   i), {n_embd}, TENSOR_NOT_REQUIRED);
+
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0); //output_dens
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, 0); //output_dens
+
+        layer.attn_out_norm   = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}, 0); //output_norm
+        layer.attn_out_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "bias",   i), {n_embd}, 0);
+
+        layer.attn_norm_2   = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
+        layer.attn_norm_2_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "bias",   i), {n_embd}, TENSOR_NOT_REQUIRED);
+
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, TENSOR_NOT_REQUIRED);
+
+        const auto tn_ffn_up_weight = tn(LLM_TENSOR_FFN_UP, "weight", i);
+        ggml_tensor * t_ffn_up = ml.get_tensor_meta(tn_ffn_up_weight.str().c_str());
+        const int64_t n_ffn_up = t_ffn_up ? t_ffn_up->ne[1] : n_ff;
+
+        GGML_ASSERT(n_ffn_up == n_ff || n_ffn_up == n_ff * 2);
+        layer.ffn_up   = create_tensor(tn_ffn_up_weight, {n_embd, n_ffn_up}, 0);
+        layer.ffn_up_b = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias", i), {n_ffn_up}, TENSOR_NOT_REQUIRED);
+
+        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias",   i), {n_embd}, 0);
+
+        layer.layer_out_norm   = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, 0);
+        layer.layer_out_norm_b = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "bias",   i), {n_embd}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_jina_bert_v2::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+

examples/talk-llama/models/jina-bert-v3.cpp

@@ -0,0 +1,69 @@
+#include "models.h"
+
+void llama_model_jina_bert_v3::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,    hparams.f_norm_eps);
+
+    switch (hparams.n_layer) {
+        case 24:
+            type = LLM_TYPE_558M; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_jina_bert_v3::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    if (n_token_types == 0) {
+        throw std::runtime_error(arch_name() + " model needs to define token type count");
+    }
+    tok_embd     = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);
+    type_embd    = create_tensor(tn(LLM_TENSOR_TOKEN_TYPES, "weight"), {n_embd, n_token_types}, TENSOR_NOT_REQUIRED);
+
+    if (arch == LLM_ARCH_BERT) {
+        pos_embd = create_tensor(tn(LLM_TENSOR_POS_EMBD,    "weight"), {n_embd, n_ctx_train}, 0);
+
+        cls   = create_tensor(tn(LLM_TENSOR_CLS, "weight"), {n_embd, n_embd}, TENSOR_NOT_REQUIRED);
+        cls_b = create_tensor(tn(LLM_TENSOR_CLS, "bias"),   {n_embd},         TENSOR_NOT_REQUIRED);
+
+        cls_out   = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
+        cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT, "bias"),   {hparams.n_cls_out},         TENSOR_NOT_REQUIRED);
+    }
+
+    tok_norm   = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
+    tok_norm_b = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "bias",   0), {n_embd}, 0);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
+
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+
+        layer.attn_out_norm   = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "weight", i), {n_embd}, 0);
+        layer.attn_out_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT_NORM, "bias", i),   {n_embd}, 0);
+
+        if (hparams.moe_every_n_layers > 0 && i % hparams.moe_every_n_layers == 1) {
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff,   n_expert}, 0);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff,   n_embd, n_expert}, 0);
+            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,   "weight", i), {n_embd, n_expert}, 0);
+        } else {
+            layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, TENSOR_NOT_REQUIRED);
+            layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);
+
+            if (arch == LLM_ARCH_NOMIC_BERT) {
+                layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+            }
+        }
+
+        layer.layer_out_norm   = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, 0);
+        layer.layer_out_norm_b = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "bias", i),   {n_embd}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_jina_bert_v3::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+

examples/talk-llama/models/kimi-linear.cpp

@@ -1,7 +1,175 @@
 #include "models.h"
-
 #include "llama-memory-recurrent.h"
 
+void llama_model_kimi_linear::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH_MLA,    hparams.n_embd_head_k_mla_impl);
+    ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH_MLA,  hparams.n_embd_head_v_mla_impl);
+    ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,      hparams.n_lora_kv);
+    ml.get_key(LLM_KV_SSM_CONV_KERNEL,             hparams.ssm_d_conv);
+    ml.get_key(LLM_KV_KDA_HEAD_DIM,                hparams.n_embd_head_kda);
+
+    // MLA qk_rope_head_dim (for reference)
+    // qk_rope_head_dim = 64, qk_nope_head_dim = 128, qk_head_dim = 192
+
+    // Mark KDA layers as recurrent using n_head_kv pattern (like Jamba)
+    // Set n_head_kv = 0 for KDA layers (recurrent), n_head_kv = n_head for MLA layers (attention)
+    for (uint32_t i = 0; i < hparams.n_layer; ++i) {
+        hparams.recurrent_layer_arr[i] = hparams.n_head_kv(i) == 0;  // KDA layers are recurrent
+    }
+
+    // MoE parameters - Kimi uses moe_intermediate_size = 1024
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp);
+    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared);
+    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,         hparams.n_layer_dense_lead, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);
+    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,                hparams.expert_gating_func);
+
+    switch (hparams.n_layer) {
+        case 27: type = LLM_TYPE_48B_A3B; break; // Kimi-Linear-48B-A3B
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_kimi_linear::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        // Check for KDA specific tensors to determine layer type or if it's a mixed model
+        // Assuming KDA layer if KDA tensors are present
+
+        // KDA uses head_dim = 128 (from linear_attn_config.head_dim)
+        const int64_t n_embd_head_k_kda = hparams.n_embd_head_kda;
+        const int64_t n_embd_head_v_kda = hparams.n_embd_head_kda;
+        const int64_t ssm_d_conv = hparams.ssm_d_conv;
+
+        if (hparams.is_recurrent(i)) {
+            // Conv1d weights: try 4D first, then 3D (quantization may remove trailing 1)
+            // 4D: [d_conv, 1, d_inner, 1], 3D: [d_conv, 1, d_inner]
+            layer.ssm_q_conv = create_tensor(tn(LLM_TENSOR_SSM_CONV1D_Q, "weight", i), {ssm_d_conv, 1, n_embd_head_k_kda * n_head, 1}, TENSOR_NOT_REQUIRED);
+            if (!layer.ssm_q_conv) {
+                layer.ssm_q_conv = create_tensor(tn(LLM_TENSOR_SSM_CONV1D_Q, "weight", i), {ssm_d_conv, 1, n_embd_head_k_kda * n_head}, 0);
+            }
+
+             // KDA Layer - Conv1d weights may be 3D or 4D
+             layer.ssm_k_conv = create_tensor(tn(LLM_TENSOR_SSM_CONV1D_K, "weight", i), {ssm_d_conv, 1, n_embd_head_k_kda * n_head, 1}, TENSOR_NOT_REQUIRED);
+             if (!layer.ssm_k_conv) {
+                 layer.ssm_k_conv = create_tensor(tn(LLM_TENSOR_SSM_CONV1D_K, "weight", i), {ssm_d_conv, 1, n_embd_head_k_kda * n_head}, 0);
+             }
+             layer.ssm_v_conv = create_tensor(tn(LLM_TENSOR_SSM_CONV1D_V, "weight", i), {ssm_d_conv, 1, n_embd_head_v_kda * n_head, 1}, TENSOR_NOT_REQUIRED);
+             if (!layer.ssm_v_conv) {
+                 layer.ssm_v_conv = create_tensor(tn(LLM_TENSOR_SSM_CONV1D_V, "weight", i), {ssm_d_conv, 1, n_embd_head_v_kda * n_head}, 0);
+             }
+
+             // q, k, v projections
+             // Python: q_proj, k_proj, v_proj
+             create_tensor_qkv(layer, i, n_embd, n_embd_head_k_kda * n_head, n_embd_head_k_kda * n_head, n_embd_head_v_kda * n_head, 0);
+
+             // KDA specific projections
+             // f_a_proj, f_b_proj
+             layer.ssm_f_a = create_tensor(tn(LLM_TENSOR_SSM_F_A, "weight", i), {n_embd, n_embd_head_k_kda}, 0); // head_dim
+             layer.ssm_f_b = create_tensor(tn(LLM_TENSOR_SSM_F_B, "weight", i), {n_embd_head_k_kda, n_embd_head_k_kda * n_head}, 0); // projection_size
+
+             // b_proj (beta mixing coefficient)
+             layer.ssm_beta = create_tensor(tn(LLM_TENSOR_SSM_BETA, "weight", i), {n_embd, n_head}, 0);
+
+             // A_log - Shape in GGUF: [1, num_heads, 1, 1] (4D) or [1, num_heads] (2D after quantization) Note: -exp(A_log) is applied in convert_hf_to_gguf.py
+             layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_head, 1, 1}, TENSOR_NOT_REQUIRED);
+             if (!layer.ssm_a) {
+                 layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_head}, 0);
+             }
+
+             // dt_bias - shape [n_embd_head_k_kda * n_head] = [4096]
+             layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {n_embd_head_k_kda * n_head}, 0);
+
+             // g_a_proj, g_b_proj (output gate)
+             layer.ssm_g_a = create_tensor(tn(LLM_TENSOR_SSM_G_A, "weight", i), {n_embd, n_embd_head_k_kda}, 0);
+             layer.ssm_g_b = create_tensor(tn(LLM_TENSOR_SSM_G_B, "weight", i), {n_embd_head_k_kda, n_embd_head_k_kda * n_head}, 0);
+
+             // o_norm (reusing SSM_NORM)
+             layer.ssm_o_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {n_embd_head_k_kda}, 0); // FusedRMSNormGated
+
+             // o_proj
+             layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_v_kda * n_head, n_embd}, 0);
+
+        } else {
+             // MLA Layer - use MLA-specific head dimensions
+             const int64_t q_lora_rank  = hparams.n_lora_q;
+             const int64_t kv_lora_rank = hparams.n_lora_kv;
+             const int64_t n_embd_head_k_mla = hparams.n_embd_head_k_mla();
+             const int64_t n_embd_head_v_mla = hparams.n_embd_head_v_mla();
+
+             layer.attn_q_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}, TENSOR_NOT_REQUIRED);
+             layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, 0);
+
+             if (layer.attn_q_a_norm) {
+                 layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, 0);
+                 layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k_mla}, 0);
+             } else {
+                 // Kimi MLA without Q compression: wq = [n_embd, n_head * n_embd_head_k_mla]
+                 layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_head * n_embd_head_k_mla}, 0);
+             }
+
+             // Kimi: qk_rope_head_dim = 64 (actual RoPE dimension for MLA)
+             // Note: hparams.n_rot may be 72 (from conversion) but actual is 64
+             const int64_t qk_rope_head_dim = hparams.n_rot();  // From config: qk_rope_head_dim
+             layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + qk_rope_head_dim}, 0);
+             // Support Legacy GGUFs that don't split wkv_b (MLA KV cache disabled)
+             layer.wkv_b = create_tensor(tn(LLM_TENSOR_ATTN_KV_B, "weight", i),
+                {kv_lora_rank, n_head * (n_embd_head_k_mla - qk_rope_head_dim + n_embd_head_v_mla)}, TENSOR_NOT_REQUIRED | TENSOR_SKIP_IF_VIRTUAL);
+             if (!layer.wkv_b) { // MLA KV cache enabled
+                 layer.wk_b = create_tensor(tn(LLM_TENSOR_ATTN_K_B, "weight", i), {n_embd_head_k_mla - qk_rope_head_dim, kv_lora_rank, n_head}, 0);
+                 layer.wv_b = create_tensor(tn(LLM_TENSOR_ATTN_V_B, "weight", i), {kv_lora_rank, n_embd_head_v_mla, n_head}, 0);
+             }
+             layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_head * n_embd_head_v_mla, n_embd}, 0);
+        }
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        // MoE intermediate size (different from dense FFN)
+        const int64_t n_ff_exp = hparams.n_ff_exp;
+
+        // Kimi uses n_layer_dense_lead to determine which layers use dense FFN vs MoE
+        // first_k_dense_replace = 1 means layer 0 uses dense FFN, layers 1+ use MoE
+        if (i < (int) hparams.n_layer_dense_lead) {
+            // Dense FFN layer - use normal n_ff
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}, 0);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff}, 0);
+        } else {
+            // MoE layer - use n_ff_exp (1024) instead of n_ff (9216)
+            layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp, n_expert}, 0);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp, n_embd, n_expert}, 0);
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff_exp, n_expert}, 0);
+
+            // Shared experts use moe_intermediate_size * num_shared_experts
+            // Kimi: shared_expert_intermediate_size = 1024 * 1 = 1024
+            // Tensors are 2D: [n_embd, n_ff_shexp] or [n_ff_shexp, n_embd]
+            const int64_t n_ff_shexp_actual = n_ff_exp * (hparams.n_expert_shared > 0 ? hparams.n_expert_shared : 1);
+            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, n_ff_shexp_actual}, TENSOR_NOT_REQUIRED);
+            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp_actual, n_embd}, TENSOR_NOT_REQUIRED);
+            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp_actual}, TENSOR_NOT_REQUIRED);
+
+            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, 0);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_kimi_linear::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
 // Causal Conv1d function for Q,K,V
 // When qkv is 0, it is Q, 1 is K, 2 is V
 static ggml_tensor * causal_conv1d(ggml_cgraph * gf, ggml_context * ctx0, ggml_tensor * conv_states_all, ggml_tensor * conv_state_all, int64_t qkv, ggml_tensor * x, ggml_tensor * proj_w, ggml_tensor * conv_w, int64_t d_conv, int64_t head_dim, int64_t n_head, int64_t n_seq_tokens, int64_t n_seqs, int64_t n_tokens, int64_t kv_head) {
@@ -63,7 +231,7 @@ static ggml_tensor * causal_conv1d(ggml_cgraph * gf, ggml_context * ctx0, ggml_t
     return ggml_reshape_4d(ctx0, Xcur, head_dim, n_head, n_seq_tokens, n_seqs);
 }
 
-llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const llm_graph_params & params) :
+llama_model_kimi_linear::graph::graph(const llama_model & model, const llm_graph_params & params) :
     llm_build_delta_net_base(params), model(model) {
     ggml_tensor * cur;
     ggml_tensor * inpL;

examples/talk-llama/models/lfm2.cpp

@@ -1,10 +1,94 @@
 #include "models.h"
-
 #include "../llama-memory-hybrid-iswa.h"
 #include "../llama-memory-hybrid.h"
 
+void llama_model_lfm2::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_SHORTCONV_L_CACHE,           hparams.n_shortconv_l_cache);
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    for (uint32_t il = 0; il < hparams.n_layer; ++il) {
+        hparams.recurrent_layer_arr[il] = hparams.n_head_kv(il) == 0;
+    }
+    hparams.n_layer_dense_lead = hparams.n_layer;
+    switch (hparams.n_ff()) {
+        case  4608: type = LLM_TYPE_350M; break;
+        case  6912: type = LLM_TYPE_700M; break;
+        case  8192: type = LLM_TYPE_1_2B; break;
+        case 10752: type = LLM_TYPE_2_6B; break;
+        default:    type = LLM_TYPE_UNKNOWN;
+    }
+    if (const auto is_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false); is_swa && hparams.n_swa > 0) {
+        hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+        for (uint32_t il = 0; il < hparams.n_layer; ++il) {
+            hparams.swa_layers[il] = !hparams.recurrent_layer_arr[il];
+        }
+    }
+}
+
+void llama_model_lfm2::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM_LFM2, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,           "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        const bool is_moe_layer = i >= static_cast<int>(hparams.n_layer_dense_lead);
+
+        // ffn/moe is same for transformer and conv layers
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        if (is_moe_layer) {
+            GGML_ASSERT(n_expert && n_expert_used);
+            layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i),  {n_embd, n_expert}, 0);
+            layer.ffn_gate_exps   = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, hparams.n_ff_exp, n_expert}, 0);
+            layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {hparams.n_ff_exp,   n_embd, n_expert}, 0);
+            layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i),   {n_embd, hparams.n_ff_exp, n_expert}, 0);
+            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, 0);
+        } else {  // dense
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+        }
+
+        // for operator_norm
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        if (!hparams.is_recurrent(i)) {
+            layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
+            layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
+            GGML_ASSERT(n_embd_v_gqa == n_embd_k_gqa);
+
+            create_tensor_qkv(layer, i, n_embd, n_embd, hparams.n_embd_k_gqa(i), hparams.n_embd_v_gqa(i), 0);
+
+            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+        } else {
+            layer.shortconv.conv     = create_tensor(tn(LLM_TENSOR_SHORTCONV_CONV,    "weight", i), {hparams.n_shortconv_l_cache, n_embd}, 0);
+            layer.shortconv.in_proj  = create_tensor(tn(LLM_TENSOR_SHORTCONV_INPROJ,  "weight", i), {n_embd, 3 * n_embd}, 0);
+            layer.shortconv.out_proj = create_tensor(tn(LLM_TENSOR_SHORTCONV_OUTPROJ, "weight", i), {n_embd, n_embd}, 0);
+        }
+    }
+
+    // for LFM2-ColBert-350M
+    dense_2_out_layers   = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.n_embd_out()}, TENSOR_NOT_REQUIRED);
+    dense_2_out_layers_b = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "bias"),   {hparams.n_embd_out()        }, TENSOR_NOT_REQUIRED);
+}
+
+std::unique_ptr<llm_graph_context> llama_model_lfm2::build_arch_graph(const llm_graph_params & params) const {
+    if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) {
+        return std::make_unique<graph<true>>(*this, params);
+    } else {
+        return std::make_unique<graph<false>>(*this, params);
+    }
+}
+
 template <bool iswa>
-llm_build_lfm2<iswa>::llm_build_lfm2(const llama_model & model, const llm_graph_params & params) :
+llama_model_lfm2::graph<iswa>::graph(const llama_model & model, const llm_graph_params & params) :
     llm_graph_context(params) {
     using inp_hybrid_type = std::conditional_t<iswa, llm_graph_input_mem_hybrid_iswa,  llm_graph_input_mem_hybrid>;
     using inp_attn_type   = std::conditional_t<iswa, llm_graph_input_attn_kv_iswa,     llm_graph_input_attn_kv>;
@@ -187,5 +271,5 @@ llm_build_lfm2<iswa>::llm_build_lfm2(const llama_model & model, const llm_graph_
 }
 
 // Explicit template instantiations
-template struct llm_build_lfm2<true>;
-template struct llm_build_lfm2<false>;
+template struct llama_model_lfm2::graph<true>;
+template struct llama_model_lfm2::graph<false>;

examples/talk-llama/models/lfm2moe.cpp

@@ -0,0 +1,85 @@
+#include "models.h"
+#include "../llama-memory-hybrid-iswa.h"
+#include "../llama-memory-hybrid.h"
+
+void llama_model_lfm2moe::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_SHORTCONV_L_CACHE,           hparams.n_shortconv_l_cache);
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_LEADING_DENSE_BLOCK_COUNT,   hparams.n_layer_dense_lead, false);
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,          hparams.expert_gating_func);
+
+    for (uint32_t il = 0; il < hparams.n_layer; ++il) {
+        hparams.recurrent_layer_arr[il] = hparams.n_head_kv(il) == 0;
+    }
+
+    switch (hparams.n_layer) {
+        case 24: type = LLM_TYPE_8B_A1B;  break;
+        case 40: type = LLM_TYPE_24B_A2B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_lfm2moe::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM_LFM2, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,           "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        const bool is_moe_layer = i >= static_cast<int>(hparams.n_layer_dense_lead);
+
+        // ffn/moe is same for transformer and conv layers
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        if (is_moe_layer) {
+            GGML_ASSERT(n_expert && n_expert_used);
+            layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i),  {n_embd, n_expert}, 0);
+            layer.ffn_gate_exps   = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, hparams.n_ff_exp, n_expert}, 0);
+            layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {hparams.n_ff_exp,   n_embd, n_expert}, 0);
+            layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS, "weight", i),   {n_embd, hparams.n_ff_exp, n_expert}, 0);
+            layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, 0);
+        } else {  // dense
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+        }
+
+        // for operator_norm
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        if (!hparams.is_recurrent(i)) {
+            layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
+            layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
+            GGML_ASSERT(n_embd_v_gqa == n_embd_k_gqa);
+
+            create_tensor_qkv(layer, i, n_embd, n_embd, hparams.n_embd_k_gqa(i), hparams.n_embd_v_gqa(i), 0);
+
+            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+        } else {
+            layer.shortconv.conv     = create_tensor(tn(LLM_TENSOR_SHORTCONV_CONV,    "weight", i), {hparams.n_shortconv_l_cache, n_embd}, 0);
+            layer.shortconv.in_proj  = create_tensor(tn(LLM_TENSOR_SHORTCONV_INPROJ,  "weight", i), {n_embd, 3 * n_embd}, 0);
+            layer.shortconv.out_proj = create_tensor(tn(LLM_TENSOR_SHORTCONV_OUTPROJ, "weight", i), {n_embd, n_embd}, 0);
+        }
+    }
+
+    // for LFM2-ColBert-350M
+    dense_2_out_layers   = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "weight"), {n_embd, hparams.n_embd_out()}, TENSOR_NOT_REQUIRED);
+    dense_2_out_layers_b = create_tensor(tn(LLM_TENSOR_DENSE_2_OUT, "bias"),   {hparams.n_embd_out()        }, TENSOR_NOT_REQUIRED);
+}
+
+std::unique_ptr<llm_graph_context> llama_model_lfm2moe::build_arch_graph(const llm_graph_params & params) const {
+    if (hparams.swa_type == LLAMA_SWA_TYPE_STANDARD) {
+        return std::make_unique<graph<true>>(*this, params);
+    } else {
+        return std::make_unique<graph<false>>(*this, params);
+    }
+}
+

examples/talk-llama/models/llada-moe.cpp

@@ -1,6 +1,56 @@
 #include "models.h"
 
-llm_build_llada_moe::llm_build_llada_moe(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_llada_moe::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp, false);
+
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    // diffusion language model uses non-causal attention
+    hparams.causal_attn = false;
+    switch (hparams.n_layer) {
+        case 16: type = LLM_TYPE_A1_7B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_llada_moe::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+    GGML_ASSERT(n_expert > 0 && "n_expert must be > 0 for llada-moe");
+    GGML_ASSERT(n_expert_used > 0 && "n_expert_used must be > 0 for llada-moe");
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
+        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+
+        const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+
+        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, 0);
+        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {  n_embd, n_ff_exp, n_expert}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_llada_moe::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_llada_moe::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/llada.cpp

@@ -1,6 +1,72 @@
 #include "models.h"
 
-llm_build_llada::llm_build_llada(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_llada::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    // LLaDA-8B has 32 layers, similar to LLaMA but for diffusion
+    switch (hparams.n_layer) {
+        case 32:
+            type = LLM_TYPE_8B;
+            break;
+        default:
+            type = LLM_TYPE_UNKNOWN;
+    }
+    // Set non-causal attention for diffusion models
+    hparams.causal_attn = false;
+}
+
+void llama_model_llada::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);
+    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), { n_embd, n_vocab }, TENSOR_NOT_REQUIRED);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output =
+            create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), { n_embd, n_vocab }, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), { n_embd }, 0);
+
+        // Use separate Q, K, V projections without bias, matching LLaDALlamaBlock
+        layer.wq =
+            create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), { n_embd, n_embd_head_k * n_head }, 0);
+        layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), { n_embd, n_embd_k_gqa }, 0);
+        layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), { n_embd, n_embd_v_gqa }, 0);
+        // No bias for QKV projections as per config: include_bias=false, include_qkv_bias=false
+        layer.wo =
+            create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0);
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), { n_embd }, TENSOR_NOT_REQUIRED);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), { n_embd }, 0);
+
+        layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), { n_rot / 2 },
+                                         TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), { n_embd, n_ff }, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd }, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP, "weight", i), { n_embd, n_ff }, 0);
+
+        // optional MLP bias
+        layer.ffn_gate_b =
+            create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), { n_ff }, TENSOR_NOT_REQUIRED);
+        layer.ffn_down_b =
+            create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), { n_embd }, TENSOR_NOT_REQUIRED);
+        layer.ffn_up_b = create_tensor(tn(LLM_TENSOR_FFN_UP, "bias", i), { n_ff }, TENSOR_NOT_REQUIRED);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_llada::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_llada::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     // LLaDA is similar to LLaMA but uses non-causal attention for diffusion
     const int64_t n_embd_head = hparams.n_embd_head_v();
 

examples/talk-llama/models/llama-embed.cpp

@@ -0,0 +1,6 @@
+#include "models.h"
+
+std::unique_ptr<llm_graph_context> llama_model_llama_embed::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph<true>>(*this, params);
+}
+

examples/talk-llama/models/llama.cpp

@@ -1,7 +1,102 @@
 #include "models.h"
 
+void llama_model_llama::load_arch_hparams(llama_model_loader & ml) {
+    uint32_t n_vocab = 0;
+    ml.get_key(LLM_KV_VOCAB_SIZE, n_vocab, false) || ml.get_arr_n(LLM_KV_TOKENIZER_LIST, n_vocab, false);
+
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    if (hparams.n_expert == 8) {
+        switch (hparams.n_layer) {
+            case 32: type = LLM_TYPE_8x7B; break;
+            case 56: type = LLM_TYPE_8x22B; break;
+            default: type = LLM_TYPE_UNKNOWN;
+        }
+    } else {
+        switch (hparams.n_layer) {
+            case 16: type = LLM_TYPE_1B; break; // Llama 3.2 1B
+            case 22: type = LLM_TYPE_1B; break;
+            case 26: type = LLM_TYPE_3B; break;
+            case 28: type = LLM_TYPE_3B; break; // Llama 3.2 3B
+            case 30: type = LLM_TYPE_256M; break; // smoldocling 256M
+            // granite uses a vocab with len 49152
+            case 32: type = n_vocab == 49152 ? LLM_TYPE_3B : (n_vocab < 40000 ? LLM_TYPE_7B : LLM_TYPE_8B); break;
+            case 36: type = LLM_TYPE_8B; break; // granite
+            case 40: type = LLM_TYPE_13B; break;
+            case 48: type = LLM_TYPE_34B; break;
+            case 60: type = LLM_TYPE_30B; break;
+            case 80: type = hparams.n_head() == hparams.n_head_kv() ? LLM_TYPE_65B : LLM_TYPE_70B; break;
+            default: type = LLM_TYPE_UNKNOWN;
+        }
+    }
+}
+
+void llama_model_llama::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        // optional bias tensors
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
+            layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+            layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+        }
+        else {
+            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+        }
+
+        if (n_expert == 0) {
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+
+            // optional MLP bias
+            layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
+            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
+        } else {
+            layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
+            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, TENSOR_NOT_REQUIRED);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);
+
+            // For Granite MoE Shared
+            if (hparams.n_ff_shexp > 0) {
+                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
+                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
+                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
+            }
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_llama::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph<false>>(*this, params);
+}
+
 template <bool embed>
-llm_build_llama<embed>::llm_build_llama(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+llama_model_llama::graph<embed>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
@@ -149,5 +244,5 @@ llm_build_llama<embed>::llm_build_llama(const llama_model & model, const llm_gra
     ggml_build_forward_expand(gf, cur);
 }
 
-template struct llm_build_llama<false>;
-template struct llm_build_llama<true>;
+template struct llama_model_llama::graph<false>;
+template struct llama_model_llama::graph<true>;

examples/talk-llama/models/llama4.cpp

@@ -1,7 +1,109 @@
 #include "models.h"
 
+void llama_model_llama4::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,  hparams.n_ff_exp);
+    ml.get_key(LLM_KV_INTERLEAVE_MOE_LAYER_STEP,   hparams.n_moe_layer_step);
+
+    const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
+    if (found_swa && hparams.n_swa == 0) {
+        hparams.swa_type             = LLAMA_SWA_TYPE_NONE;
+        hparams.n_no_rope_layer_step = hparams.n_layer; // always use rope
+    } else {
+        hparams.swa_type                = LLAMA_SWA_TYPE_CHUNKED;
+        hparams.n_swa                   = 8192;
+        hparams.n_attn_temp_floor_scale = 8192;
+        hparams.f_attn_temp_scale       = 0.1f;
+        hparams.f_attn_temp_offset      = 1.0f;
+        uint32_t swa_period             = 4; // pattern: 3 chunked - 1 full
+        ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
+        hparams.set_swa_pattern(swa_period);
+
+        hparams.rope_freq_base_train_swa  = hparams.rope_freq_base_train;
+        hparams.rope_freq_scale_train_swa = hparams.rope_freq_scale_train;
+        ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
+    }
+
+    switch (hparams.n_expert) {
+        case 0: {
+            // MobileLLM (no MoE)
+            switch (hparams.n_embd) {
+                case 2048: type = LLM_TYPE_140M; break;
+                case 4096: type = LLM_TYPE_360M; break;
+                case 6144: type = LLM_TYPE_950M; break;
+                default:   type = LLM_TYPE_UNKNOWN;
+            }
+        } break;
+        case 16:  type = LLM_TYPE_17B_16E; break;
+        case 128: type = LLM_TYPE_17B_128E; break;
+        default:  type = LLM_TYPE_UNKNOWN;
+    }
+
+    hparams.use_kq_norm = type != LLM_TYPE_17B_128E;
+}
+
+void llama_model_llama4::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    if (n_expert == 0) {
+        throw std::runtime_error(arch_name() + " model cannot have zero experts");
+    }
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        const bool is_moe_layer = hparams.n_moe_layer_step > 0 && (i + 1) % hparams.n_moe_layer_step == 0;
+
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+
+        if (is_moe_layer) {
+            const int64_t n_ff_exp = hparams.n_ff_exp;
+
+            layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
+            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff_exp, n_expert}, 0);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff_exp, n_embd, n_expert}, 0);
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff_exp, n_expert}, 0);
+
+            // Shared expert
+            const int64_t n_ff_shexp = n_ff_exp;
+            layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {    n_embd, n_ff_shexp}, 0);
+            layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd    }, 0);
+            layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {    n_embd, n_ff_shexp}, 0);
+        } else {
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_llama4::build_arch_graph(const llm_graph_params & params) const {
+    if (hparams.swa_type == LLAMA_SWA_TYPE_NONE) {
+        return std::make_unique<graph<false>>(*this, params);
+    } else {
+        return std::make_unique<graph<true>>(*this, params);
+    }
+}
+
 template <bool iswa>
-llm_build_llama4<iswa>::llm_build_llama4(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+llama_model_llama4::graph<iswa>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
@@ -167,5 +269,5 @@ llm_build_llama4<iswa>::llm_build_llama4(const llama_model & model, const llm_gr
 }
 
 // Explicit template instantiations
-template struct llm_build_llama4<false>;
-template struct llm_build_llama4<true>;
+template struct llama_model_llama4::graph<false>;
+template struct llama_model_llama4::graph<true>;

examples/talk-llama/models/maincoder.cpp

@@ -1,6 +1,49 @@
 #include "models.h"
 
-llm_build_maincoder::llm_build_maincoder(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_maincoder::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    switch (hparams.n_layer) {
+        case 32: type = LLM_TYPE_1B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_maincoder::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_head_k}, 0);
+        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_maincoder::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_maincoder::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/mamba.cpp

@@ -1,6 +1,90 @@
 #include "models.h"
 
-llm_build_mamba::llm_build_mamba(const llama_model & model, const llm_graph_params & params) : llm_build_mamba_base(params) {
+void llama_model_mamba::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);
+    ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);
+    ml.get_key(LLM_KV_SSM_STATE_SIZE,     hparams.ssm_d_state);
+    ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
+    ml.get_key(LLM_KV_SSM_DT_B_C_RMS,     hparams.ssm_dt_b_c_rms, false);
+
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    switch (hparams.n_layer) {
+        case 24:
+            switch (hparams.n_embd) {
+                case 768: type = LLM_TYPE_SMALL; break;
+                default: type = LLM_TYPE_UNKNOWN;
+            } break;
+        case 48:
+            switch (hparams.n_embd) {
+                case 1024: type = LLM_TYPE_MEDIUM; break;
+                case 1536: type = LLM_TYPE_LARGE; break;
+                case 2048: type = LLM_TYPE_XL; break;
+                default:   type = LLM_TYPE_UNKNOWN;
+            } break;
+        case 64:
+            switch (hparams.n_embd) {
+                case 2560: type = LLM_TYPE_3B; break;
+                default: type = LLM_TYPE_UNKNOWN;
+            } break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_mamba::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    const int64_t d_conv  = hparams.ssm_d_conv;
+    const int64_t d_inner = hparams.ssm_d_inner;
+    const int64_t d_state = hparams.ssm_d_state;
+    const int64_t dt_rank = hparams.ssm_dt_rank;
+
+    // only an expansion factor of 2 is supported for now
+    if (2 * n_embd != d_inner) {
+        throw std::runtime_error("only an expansion factor of 2 is supported for now");
+    }
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+
+    output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    // if output is NULL, init from the input tok embed, duplicated to allow offloading
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        // norm
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, 2*d_inner}, 0);
+
+        layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner}, 0);
+        layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner}, 0);
+
+        layer.ssm_x = create_tensor(tn(LLM_TENSOR_SSM_X, "weight", i), {d_inner, dt_rank + 2*d_state}, 0);
+
+        layer.ssm_dt = create_tensor(tn(LLM_TENSOR_SSM_DT, "weight", i), {dt_rank, d_inner}, 0);
+        layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {d_inner}, 0);
+
+        // no "weight" suffix for these
+        layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {d_state, d_inner}, 0);
+        layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {d_inner}, 0);
+
+        // out_proj
+        layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_mamba::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_mamba::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_build_mamba_base(params) {
     ggml_tensor * cur;
     ggml_tensor * inpL;
 
@@ -51,4 +135,3 @@ llm_build_mamba::llm_build_mamba(const llama_model & model, const llm_graph_para
 
     ggml_build_forward_expand(gf, cur);
 }
-

examples/talk-llama/models/mamba2.cpp

@@ -0,0 +1,87 @@
+#include "models.h"
+
+void llama_model_mamba2::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);
+    ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);
+    ml.get_key(LLM_KV_SSM_STATE_SIZE,     hparams.ssm_d_state);
+    ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
+    ml.get_key(LLM_KV_SSM_GROUP_COUNT,    hparams.ssm_n_group);
+
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    switch (hparams.n_layer) {
+        case 24:
+            switch (hparams.n_embd) {
+                case 768: type = LLM_TYPE_SMALL; break;
+                default: type = LLM_TYPE_UNKNOWN;
+            } break;
+        case 48:
+            switch (hparams.n_embd) {
+                case 1024: type = LLM_TYPE_MEDIUM; break;
+                case 1536: type = LLM_TYPE_LARGE; break;
+                case 2048: type = LLM_TYPE_XL; break;
+                default: type = LLM_TYPE_UNKNOWN;
+            } break;
+        case 64:
+            switch (hparams.n_embd) {
+                case 2560: type = LLM_TYPE_3B; break;
+                case 4096: type = LLM_TYPE_7B; break;
+                default: type = LLM_TYPE_UNKNOWN;
+            } break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_mamba2::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    const int64_t d_conv  = hparams.ssm_d_conv;
+    const int64_t d_inner = hparams.ssm_d_inner;
+    const int64_t d_state = hparams.ssm_d_state;
+    const int64_t n_group = hparams.ssm_n_group;
+    const int64_t d_in_proj = 2*d_inner + 2*n_group*d_state + n_head;
+
+    // only an expansion factor of 2 is supported for now
+    GGML_ASSERT(2 * n_embd == d_inner);
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    {
+        output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+
+        output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+        // if output is NULL, init from the input tok embed, duplicated to allow offloading
+        if (output == NULL) {
+            output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+        }
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        // norm
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, d_in_proj}, 0);
+
+        layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner + 2*n_group*d_state}, 0);
+        layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner + 2*n_group*d_state}, 0);
+
+        layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {n_head}, 0);
+
+        // no "weight" suffix for these
+        layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_head}, 0);
+        layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, n_head}, 0);
+
+        layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {d_inner / n_group, n_group}, 0);
+
+        // out_proj
+        layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_mamba2::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+

examples/talk-llama/models/mimo2-iswa.cpp

@@ -1,129 +0,0 @@
-#include "models.h"
-
-llm_build_mimo2_iswa::llm_build_mimo2_iswa(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
-    ggml_tensor * cur;
-    ggml_tensor * inpL;
-
-    inpL = build_inp_embd(model.tok_embd);
-
-    ggml_tensor * inp_pos = build_inp_pos();
-    auto * inp_attn = build_attn_inp_kv_iswa();
-    ggml_tensor * inp_out_ids = build_inp_out_ids();
-
-    for (int il = 0; il < n_layer; ++il) {
-        ggml_tensor * inpSA = inpL;
-
-        uint32_t n_head_l    = hparams.n_head(il);
-        uint32_t n_head_kv_l = hparams.n_head_kv(il);
-        const float freq_base_l  = model.get_rope_freq_base(cparams, il);
-        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
-
-        cur = inpL;
-
-        // self_attention
-        {
-            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
-            cb(cur, "attn_norm", il);
-
-            // compute Q and K and RoPE them
-            ggml_tensor * Qcur = build_lora_mm(model.layers[il].wq, cur);
-            cb(Qcur, "Qcur", il);
-
-            ggml_tensor * Kcur = build_lora_mm(model.layers[il].wk, cur);
-            cb(Kcur, "Kcur", il);
-
-            ggml_tensor * Vcur = build_lora_mm(model.layers[il].wv, cur);
-            cb(Vcur, "Vcur", il);
-
-            Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head_l,    n_tokens);
-            Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv_l, n_tokens);
-            Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head_v, n_head_kv_l, n_tokens);
-
-            Qcur = ggml_rope_ext(
-                ctx0, Qcur, inp_pos, nullptr,
-                n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
-                ext_factor, attn_factor, beta_fast, beta_slow
-                );
-
-            Kcur = ggml_rope_ext(
-                ctx0, Kcur, inp_pos, nullptr,
-                n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
-                ext_factor, attn_factor, beta_fast, beta_slow
-                );
-
-            cb(Qcur, "Qcur", il);
-            cb(Kcur, "Kcur", il);
-            cb(Vcur, "Vcur", il);
-
-            ggml_tensor * sinks = model.layers[il].attn_sinks;
-
-            cur = build_attn(inp_attn,
-                    model.layers[il].wo, NULL, model.layers[il].wo_s,
-                    Qcur, Kcur, Vcur, nullptr, sinks, nullptr, 1.0f/sqrtf(float(n_embd_head_k)), il);
-        }
-
-        if (il == n_layer - 1 && inp_out_ids) {
-            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
-            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
-        }
-
-        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
-        cb(ffn_inp, "ffn_inp", il);
-
-        cur = build_norm(ffn_inp,
-                model.layers[il].ffn_norm, NULL,
-                LLM_NORM_RMS, il);
-        cb(cur, "ffn_norm", il);
-
-        // feed-forward network
-        if (model.layers[il].ffn_gate_inp == nullptr) {
-            // dense branch
-            cur = build_ffn(cur,
-                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
-                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
-                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
-                    NULL,
-                    LLM_FFN_SILU, LLM_FFN_PAR, il);
-            cb(cur, "ffn_out", il);
-        } else {
-            // MoE branch
-            cur = build_moe_ffn(cur,
-                    model.layers[il].ffn_gate_inp,
-                    model.layers[il].ffn_up_exps,
-                    model.layers[il].ffn_gate_exps,
-                    model.layers[il].ffn_down_exps,
-                    model.layers[il].ffn_exp_probs_b,
-                    n_expert, n_expert_used,
-                    LLM_FFN_SILU, true,
-                    hparams.expert_weights_scale,
-                    LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,
-                    il);
-            cb(cur, "ffn_moe_out", il);
-        }
-
-        cur = ggml_add(ctx0, cur, ffn_inp);
-
-        cur = build_cvec(cur, il);
-        cb(cur, "l_out", il);
-
-        // input for next layer
-        inpL = cur;
-    }
-
-    cur = inpL;
-
-    cur = build_norm(cur,
-            model.output_norm, NULL,
-            LLM_NORM_RMS, -1);
-
-    cb(cur, "result_norm", -1);
-    res->t_embd = cur;
-
-    // lm_head
-    cur = build_lora_mm(model.output, cur);
-
-    cb(cur, "result_output", -1);
-    res->t_logits = cur;
-
-    ggml_build_forward_expand(gf, cur);
-}

examples/talk-llama/models/mimo2.cpp

@@ -0,0 +1,240 @@
+#include "models.h"
+
+void llama_model_mimo2::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
+
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH, hparams.n_ff_exp);
+    ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW,   hparams.n_swa);
+    ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA,         hparams.rope_freq_base_train_swa, false);
+    ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.swa_layers, hparams.n_layer);
+
+    float value_scale = 0.0f;
+    if (ml.get_key(LLM_KV_ATTENTION_VALUE_SCALE, value_scale, false) && value_scale != 1.0f) {
+        hparams.f_attn_value_scale = value_scale;
+    }
+
+    ml.get_key(LLM_KV_NEXTN_PREDICT_LAYERS, hparams.nextn_predict_layers, false);
+    GGML_ASSERT(hparams.nextn_predict_layers < hparams.n_layer && "nextn_predict_layers must be < n_layer");
+    hparams.n_layer_kv_from_start = hparams.n_layer - hparams.nextn_predict_layers;
+
+    switch (hparams.n_layer - hparams.nextn_predict_layers) {
+        case 48: type = LLM_TYPE_310B_A15B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_mimo2::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+    const uint32_t n_nextn = hparams.nextn_predict_layers;
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+        uint32_t n_embd_k_gqa = hparams.n_embd_k_gqa(i);
+        uint32_t n_embd_v_gqa = hparams.n_embd_v_gqa(i);
+        uint32_t n_head = hparams.n_head(i);
+
+        // NextN/MTP layers (the last n_nextn blocks) are preserved but disabled pending support
+        const bool is_nextn = (n_nextn > 0) && (static_cast<uint32_t>(i) >= n_layer - n_nextn);
+        const int  skip     = is_nextn ? TENSOR_SKIP : 0;
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, skip);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_v * n_head, n_embd }, skip);
+
+        layer.attn_norm  = create_tensor(tn(LLM_TENSOR_ATTN_NORM,  "weight", i), {n_embd}, skip);
+        layer.attn_sinks = create_tensor(tn(LLM_TENSOR_ATTN_SINKS, "weight", i), {n_head}, TENSOR_NOT_REQUIRED | skip);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, skip);
+
+        // non-MoE branch
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, TENSOR_NOT_REQUIRED | skip);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, TENSOR_NOT_REQUIRED | skip);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, TENSOR_NOT_REQUIRED | skip);
+
+        // MoE branch
+        int64_t n_ff_exp = hparams.n_ff_exp;
+        layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, TENSOR_NOT_REQUIRED | skip);
+        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff_exp,   n_expert}, TENSOR_NOT_REQUIRED | skip);
+        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   n_embd, n_expert}, TENSOR_NOT_REQUIRED | skip);
+        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff_exp,   n_expert}, TENSOR_NOT_REQUIRED | skip);
+        layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, TENSOR_NOT_REQUIRED | skip);
+
+        if (is_nextn) {
+            layer.nextn.eh_proj  = create_tensor(tn(LLM_TENSOR_NEXTN_EH_PROJ, "weight", i), {2 * n_embd, n_embd}, skip);
+            layer.nextn.enorm    = create_tensor(tn(LLM_TENSOR_NEXTN_ENORM,   "weight", i), {n_embd}, skip);
+            layer.nextn.hnorm    = create_tensor(tn(LLM_TENSOR_NEXTN_HNORM,   "weight", i), {n_embd}, skip);
+            layer.layer_out_norm = create_tensor(tn(LLM_TENSOR_LAYER_OUT_NORM, "weight", i), {n_embd}, skip);
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_mimo2::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_mimo2::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+    ggml_tensor * cur;
+    ggml_tensor * inpL;
+
+    inpL = build_inp_embd(model.tok_embd);
+
+    ggml_tensor * inp_pos = build_inp_pos();
+    auto * inp_attn = build_attn_inp_kv_iswa();
+    ggml_tensor * inp_out_ids = build_inp_out_ids();
+
+    const float v_scale = hparams.f_attn_value_scale;
+
+    // The last hparams.nextn_predict_layers blocks are MTP heads, currently inactive
+    const int n_transformer_layers = n_layer - hparams.nextn_predict_layers;
+
+    for (int il = 0; il < n_transformer_layers; ++il) {
+        ggml_tensor * inpSA = inpL;
+
+        uint32_t n_head_l    = hparams.n_head(il);
+        uint32_t n_head_kv_l = hparams.n_head_kv(il);
+        const float freq_base_l  = model.get_rope_freq_base(cparams, il);
+        const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
+
+        cur = inpL;
+
+        // self_attention
+        {
+            cur = build_norm(inpL, model.layers[il].attn_norm, NULL, LLM_NORM_RMS, il);
+            cb(cur, "attn_norm", il);
+
+            ggml_tensor * Qcur;
+            ggml_tensor * Kcur;
+            ggml_tensor * Vcur;
+
+            if (model.layers[il].wqkv) {
+                // Fused qkv_proj - Q/K share head_dim_k, V uses head_dim_v
+                ggml_tensor * qkv = build_lora_mm(model.layers[il].wqkv, cur);
+                cb(qkv, "wqkv", il);
+
+                const size_t row_k    = ggml_row_size(qkv->type, n_embd_head_k);
+                const size_t row_v    = ggml_row_size(qkv->type, n_embd_head_v);
+                const size_t row_full = qkv->nb[1];
+                const size_t k_off    = row_k * n_head_l;
+                const size_t v_off    = k_off + row_k * n_head_kv_l;
+
+                Qcur = ggml_view_3d(ctx0, qkv, n_embd_head_k, n_head_l,    n_tokens, row_k, row_full, 0);
+                Kcur = ggml_view_3d(ctx0, qkv, n_embd_head_k, n_head_kv_l, n_tokens, row_k, row_full, k_off);
+                Vcur = ggml_view_3d(ctx0, qkv, n_embd_head_v, n_head_kv_l, n_tokens, row_v, row_full, v_off);
+            } else {
+                // Split path
+                Qcur = build_lora_mm(model.layers[il].wq, cur);
+                cb(Qcur, "Qcur", il);
+
+                Kcur = build_lora_mm(model.layers[il].wk, cur);
+                cb(Kcur, "Kcur", il);
+
+                Vcur = build_lora_mm(model.layers[il].wv, cur);
+                cb(Vcur, "Vcur", il);
+
+                Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head_l,    n_tokens);
+                Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv_l, n_tokens);
+                Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head_v, n_head_kv_l, n_tokens);
+            }
+
+            Qcur = ggml_rope_ext(
+                ctx0, Qcur, inp_pos, nullptr,
+                n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                ext_factor, attn_factor, beta_fast, beta_slow
+                );
+
+            Kcur = ggml_rope_ext(
+                ctx0, Kcur, inp_pos, nullptr,
+                n_rot, rope_type, n_ctx_orig, freq_base_l, freq_scale_l,
+                ext_factor, attn_factor, beta_fast, beta_slow
+                );
+
+            cb(Qcur, "Qcur", il);
+            cb(Kcur, "Kcur", il);
+            cb(Vcur, "Vcur", il);
+
+            ggml_tensor * sinks = model.layers[il].attn_sinks;
+
+            cur = build_attn(inp_attn,
+                    model.layers[il].wo, NULL, model.layers[il].wo_s,
+                    Qcur, Kcur, Vcur, nullptr, sinks, nullptr, 1.0f/sqrtf(float(n_embd_head_k)), il);
+            cb(cur, "attn_out", il);
+
+            if (v_scale) {
+                cur = ggml_scale(ctx0, cur, v_scale);
+                cb(cur, "attn_out_scaled", il);
+            }
+        }
+
+        if (il == n_transformer_layers - 1 && inp_out_ids) {
+            cur   = ggml_get_rows(ctx0,   cur, inp_out_ids);
+            inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
+        }
+
+        ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
+        cb(ffn_inp, "ffn_inp", il);
+
+        cur = build_norm(ffn_inp,
+                model.layers[il].ffn_norm, NULL,
+                LLM_NORM_RMS, il);
+        cb(cur, "ffn_norm", il);
+
+        // feed-forward network
+        if (model.layers[il].ffn_gate_inp == nullptr) {
+            // dense branch
+            cur = build_ffn(cur,
+                    model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   NULL,
+                    model.layers[il].ffn_gate, model.layers[il].ffn_gate_b, NULL,
+                    model.layers[il].ffn_down, model.layers[il].ffn_down_b, NULL,
+                    NULL,
+                    LLM_FFN_SILU, LLM_FFN_PAR, il);
+            cb(cur, "ffn_out", il);
+        } else {
+            // MoE branch
+            cur = build_moe_ffn(cur,
+                    model.layers[il].ffn_gate_inp,
+                    model.layers[il].ffn_up_exps,
+                    model.layers[il].ffn_gate_exps,
+                    model.layers[il].ffn_down_exps,
+                    model.layers[il].ffn_exp_probs_b,
+                    n_expert, n_expert_used,
+                    LLM_FFN_SILU, true,
+                    hparams.expert_weights_scale,
+                    LLAMA_EXPERT_GATING_FUNC_TYPE_SIGMOID,
+                    il);
+            cb(cur, "ffn_moe_out", il);
+        }
+
+        cur = ggml_add(ctx0, cur, ffn_inp);
+
+        cur = build_cvec(cur, il);
+        cb(cur, "l_out", il);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    cur = inpL;
+
+    cur = build_norm(cur,
+            model.output_norm, NULL,
+            LLM_NORM_RMS, -1);
+
+    cb(cur, "result_norm", -1);
+    res->t_embd = cur;
+
+    // lm_head
+    cur = build_lora_mm(model.output, cur);
+
+    cb(cur, "result_output", -1);
+    res->t_logits = cur;
+
+    ggml_build_forward_expand(gf, cur);
+}

examples/talk-llama/models/minicpm.cpp

@@ -0,0 +1,89 @@
+#include "models.h"
+
+void llama_model_minicpm::load_arch_hparams(llama_model_loader & ml) {
+    // Backward-compatible defaults for older MiniCPM GGUFs
+    hparams.f_embedding_scale = 12.0f;
+    hparams.f_residual_scale  = 1.4f / sqrtf(float(hparams.n_layer));
+    hparams.f_logit_scale     = hparams.n_embd ? (256.0f / float(hparams.n_embd)) : 1.0f;
+
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    // Optional KV reads, override defaults if present in newer GGUF exports
+    ml.get_key(LLM_KV_EMBEDDING_SCALE, hparams.f_embedding_scale, /*required=*/false);
+    ml.get_key(LLM_KV_RESIDUAL_SCALE, hparams.f_residual_scale, /*required=*/false);
+    ml.get_key(LLM_KV_LOGIT_SCALE, hparams.f_logit_scale, /*required=*/false);
+
+    // MiniCPM uses rope by default, unlike Granite which uses it as a switch
+    hparams.rope_finetuned = true;
+
+    switch (hparams.n_layer) {
+        case 52: type = LLM_TYPE_1B; break;
+        case 40: type = LLM_TYPE_2B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_minicpm::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        // optional bias tensors
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
+            layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+            layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+        }
+        else {
+            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+        }
+
+        if (n_expert == 0) {
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+
+            // optional MLP bias
+            layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
+            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
+        } else {
+            layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
+            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, TENSOR_NOT_REQUIRED);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);
+
+            // For Granite MoE Shared
+            if (hparams.n_ff_shexp > 0) {
+                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
+                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
+                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
+            }
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_minicpm::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+

examples/talk-llama/models/minicpm3.cpp

@@ -1,6 +1,66 @@
 #include "models.h"
 
-llm_build_minicpm3::llm_build_minicpm3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_minicpm3::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_ATTENTION_Q_LORA_RANK,       hparams.n_lora_q);
+    ml.get_key(LLM_KV_ATTENTION_KV_LORA_RANK,      hparams.n_lora_kv);
+
+    switch (hparams.n_layer) {
+        case 62: type = LLM_TYPE_4B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_minicpm3::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    const int64_t n_embd_head_qk_rope = hparams.n_rot();
+    const int64_t n_embd_head_qk_nope = hparams.n_embd_head_k() - hparams.n_rot();
+
+    const int64_t q_lora_rank  = hparams.n_lora_q;
+    const int64_t kv_lora_rank = hparams.n_lora_kv;
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+        layer.attn_q_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_A_NORM, "weight", i), {q_lora_rank}, 0);
+
+        layer.attn_kv_a_norm = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_NORM, "weight", i), {kv_lora_rank}, 0);
+
+        layer.wq_a = create_tensor(tn(LLM_TENSOR_ATTN_Q_A, "weight", i), {n_embd, q_lora_rank}, 0);
+        layer.wq_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_B, "weight", i), {q_lora_rank, n_head * n_embd_head_k}, 0);
+
+        layer.wkv_a_mqa = create_tensor(tn(LLM_TENSOR_ATTN_KV_A_MQA, "weight", i), {n_embd, kv_lora_rank + (n_embd_head_qk_rope)}, 0);
+        layer.wkv_b     = create_tensor(tn(LLM_TENSOR_ATTN_KV_B,     "weight", i), {kv_lora_rank, n_head * (n_embd_head_qk_nope + n_embd_head_v)}, 0);
+        layer.wo        = create_tensor(tn(LLM_TENSOR_ATTN_OUT,      "weight", i), {              n_head * (                      n_embd_head_v), n_embd}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+
+        layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), { n_embd_head_qk_rope/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+        layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), { n_embd_head_qk_rope/2 }, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_minicpm3::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_minicpm3::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     //TODO: if the model varies, these parameters need to be read from the model
     const int64_t n_embd_base = 256;
     const float scale_embd  = 12.0f;

examples/talk-llama/models/minimax-m2.cpp

@@ -1,6 +1,50 @@
 #include "models.h"
 
-llm_build_minimax_m2::llm_build_minimax_m2(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_minimax_m2::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS,  hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,   hparams.n_ff_exp);
+    ml.get_key(LLM_KV_EXPERT_GATING_FUNC,           hparams.expert_gating_func, false);
+
+    switch (hparams.n_layer) {
+        case 62: type = LLM_TYPE_230B_A10B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_minimax_m2::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, 0);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_gqa, n_embd_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), { n_embd_head_k * n_head, n_embd }, 0);
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+        layer.attn_q_norm = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd_head_k * n_head}, 0);
+        layer.attn_k_norm = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd_k_gqa}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_gate_inp = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP, "weight", i), {n_embd, n_expert}, 0);
+        layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd, n_ff,   n_expert}, 0);
+        layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff,   n_embd, n_expert}, 0);
+        layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd, n_ff,   n_expert}, 0);
+        layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_minimax_m2::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_minimax_m2::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/mistral3.cpp

@@ -1,6 +1,96 @@
 #include "models.h"
 
-llm_build_mistral3::llm_build_mistral3(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_mistral3::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+    ml.get_key(LLM_KV_ATTENTION_TEMPERATURE_SCALE, hparams.f_attn_temp_scale, false);
+
+    ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_FAST, hparams.yarn_beta_fast,    false);
+    ml.get_key(LLM_KV_ROPE_SCALING_YARN_BETA_SLOW, hparams.yarn_beta_slow,    false);
+    ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL,   hparams.rope_yarn_log_mul, false);
+
+    hparams.f_attn_temp_offset = 0.0f;
+
+    // TODO: maybe add n_attn_temp_floor_scale as a separate KV?
+    if (hparams.f_attn_temp_scale != 0.0f) {
+        hparams.n_attn_temp_floor_scale = hparams.n_ctx_orig_yarn;
+        if (hparams.n_attn_temp_floor_scale == 0) {
+            throw std::runtime_error("invalid n_ctx_orig_yarn for attention temperature scaling");
+        }
+    }
+
+    switch (hparams.n_layer) {
+        case 26: type = LLM_TYPE_3B; break;
+        case 34: type = LLM_TYPE_8B; break;
+        case 40: type = LLM_TYPE_14B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_mistral3::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output      = create_tensor(tn(LLM_TENSOR_OUTPUT,      "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+
+    // if output is NULL, init from the input tok embed
+    if (output == NULL) {
+        output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head, n_embd_k_gqa, n_embd_v_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head, n_embd}, 0);
+
+        // optional bias tensors
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        if (hparams.rope_scaling_type_train == LLAMA_ROPE_SCALING_TYPE_LONGROPE) {
+            layer.rope_long  = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_LONG,  "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+            layer.rope_short = create_tensor(tn(LLM_TENSOR_ROPE_FACTORS_SHORT, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+        }
+        else {
+            layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED | (i != 0 ? TENSOR_DUPLICATED : 0));
+        }
+
+        if (n_expert == 0) {
+            layer.ffn_gate = create_tensor(tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd,   n_ff}, 0);
+            layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+            layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+
+            // optional MLP bias
+            layer.ffn_gate_b = create_tensor(tn(LLM_TENSOR_FFN_GATE, "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
+            layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+            layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
+        } else {
+            layer.ffn_gate_inp  = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), {n_embd, n_expert}, 0);
+            layer.ffn_gate_exps = create_tensor(tn(LLM_TENSOR_FFN_GATE_EXPS, "weight", i), {n_embd,   n_ff, n_expert}, TENSOR_NOT_REQUIRED);
+            layer.ffn_down_exps = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {  n_ff, n_embd, n_expert}, 0);
+            layer.ffn_up_exps   = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {n_embd,   n_ff, n_expert}, 0);
+
+            // For Granite MoE Shared
+            if (hparams.n_ff_shexp > 0) {
+                layer.ffn_gate_shexp = create_tensor(tn(LLM_TENSOR_FFN_GATE_SHEXP, "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
+                layer.ffn_up_shexp   = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, hparams.n_ff_shexp}, 0);
+                layer.ffn_down_shexp = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {hparams.n_ff_shexp, n_embd}, 0);
+            }
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_mistral3::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_mistral3::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/mistral4.cpp

@@ -0,0 +1,6 @@
+#include "models.h"
+
+std::unique_ptr<llm_graph_context> llama_model_mistral4::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+

examples/talk-llama/models/modern-bert.cpp

@@ -1,6 +1,69 @@
 #include "models.h"
 
-llm_build_modern_bert::llm_build_modern_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_modern_bert::load_arch_hparams(llama_model_loader & ml) {
+    const bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
+    if (found_swa && hparams.n_swa > 0) {
+        hparams.swa_type = LLAMA_SWA_TYPE_SYMMETRIC;
+        ml.get_key(LLM_KV_ROPE_FREQ_BASE_SWA, hparams.rope_freq_base_train_swa, false);
+        uint32_t swa_period = 3;
+        ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, swa_period, false);
+        hparams.set_swa_pattern(swa_period, true);
+    } else {
+        hparams.swa_type = LLAMA_SWA_TYPE_NONE;
+    }
+
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+
+    switch (hparams.n_layer) {
+        case 12:
+            type = LLM_TYPE_47M; break; // granite-embedding-small
+        case 22:
+            type = LLM_TYPE_149M; break; // modern-bert-base
+        case 28:
+            type = LLM_TYPE_395M; break; // modern-bert-large
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_modern_bert::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+    tok_norm = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD_NORM, "weight", 0), {n_embd}, 0);
+
+    output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+
+    for(int i = 0; i < n_layer; ++i) {
+        auto& layer = layers[i];
+
+        if ( i != 0 ) {
+            layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+        } else{
+            // layer 0 uses identity
+            layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
+        }
+
+
+        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, 3 * n_embd }, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT,   "weight", i), {n_embd, n_embd}, 0);
+
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, 2 * n_ff}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+    }
+
+    cls_out   = create_tensor(tn(LLM_TENSOR_CLS_OUT,  "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
+    cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT,  "bias"),   {hparams.n_cls_out},         TENSOR_NOT_REQUIRED);
+    cls       = create_tensor(tn(LLM_TENSOR_CLS,      "weight"), {n_embd, n_embd},            TENSOR_NOT_REQUIRED);
+    cls_norm  = create_tensor(tn(LLM_TENSOR_CLS_NORM, "weight"), {n_embd},                    TENSOR_NOT_REQUIRED);
+
+}
+
+std::unique_ptr<llm_graph_context> llama_model_modern_bert::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_modern_bert::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/mpt.cpp

@@ -1,6 +1,70 @@
 #include "models.h"
 
-llm_build_mpt::llm_build_mpt(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_mpt::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS,  hparams.f_norm_eps);
+    ml.get_key(LLM_KV_ATTENTION_CLAMP_KQV,      hparams.f_clamp_kqv, false);
+    ml.get_key(LLM_KV_ATTENTION_MAX_ALIBI_BIAS, hparams.f_max_alibi_bias, false);
+
+    switch (hparams.n_layer) {
+        case 32: type = LLM_TYPE_7B; break;
+        case 48: type = LLM_TYPE_30B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_mpt::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+    pos_embd = create_tensor(tn(LLM_TENSOR_POS_EMBD,   "weight"), {n_embd, n_ctx_train}, TENSOR_NOT_REQUIRED);
+
+    // output
+    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"),   {n_embd}, TENSOR_NOT_REQUIRED);
+
+    output        = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+    if (!output) {
+        output    = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED); // needs to be on GPU
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);
+
+        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
+        layer.wqkv_b = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "bias", i), {n_embd + 2*n_embd_gqa}, TENSOR_NOT_REQUIRED);
+
+        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);
+
+        layer.ffn_norm   = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);
+
+        layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i),   {n_embd}, TENSOR_NOT_REQUIRED);
+
+        layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i),   {n_ff}, TENSOR_NOT_REQUIRED);
+
+        // FIXME test-llama-archs crashes if q_norm is created
+        layer.attn_q_norm   = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED | TENSOR_SKIP_IF_VIRTUAL);
+        layer.attn_q_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_Q_NORM, "bias",   i), {n_embd}, TENSOR_NOT_REQUIRED | TENSOR_SKIP_IF_VIRTUAL);
+
+        layer.attn_k_norm   = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "weight", i), {n_embd}, TENSOR_NOT_REQUIRED);
+        layer.attn_k_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_K_NORM, "bias",   i), {n_embd}, TENSOR_NOT_REQUIRED);
+
+        // AWQ ScaleActivation layer
+        layer.ffn_act = create_tensor(tn(LLM_TENSOR_FFN_ACT, "scales", i), {n_ff}, TENSOR_NOT_REQUIRED);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_mpt::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_mpt::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/nemotron-h-moe.cpp

@@ -0,0 +1,6 @@
+#include "models.h"
+
+std::unique_ptr<llm_graph_context> llama_model_nemotron_h_moe::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+

examples/talk-llama/models/nemotron-h.cpp

@@ -1,6 +1,127 @@
 #include "models.h"
 
-llm_build_nemotron_h::llm_build_nemotron_h(const llama_model & model, const llm_graph_params & params) :
+void llama_model_nemotron_h::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_SSM_CONV_KERNEL,    hparams.ssm_d_conv);
+    ml.get_key(LLM_KV_SSM_INNER_SIZE,     hparams.ssm_d_inner);
+    ml.get_key(LLM_KV_SSM_STATE_SIZE,     hparams.ssm_d_state);
+    ml.get_key(LLM_KV_SSM_TIME_STEP_RANK, hparams.ssm_dt_rank);
+    ml.get_key(LLM_KV_SSM_GROUP_COUNT,    hparams.ssm_n_group);
+
+    // A layer is recurrent IFF the n_head_kv value is set to 0 and
+    // the n_ff value is set to 0
+    for (uint32_t i = 0; i < hparams.n_layer; ++i) {
+        hparams.recurrent_layer_arr[i] = (hparams.n_head_kv(i) == 0 && hparams.n_ff(i) == 0);
+    }
+
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    ml.get_key(LLM_KV_EXPERT_FEED_FORWARD_LENGTH,        hparams.n_ff_exp,        false);
+    ml.get_key(LLM_KV_EXPERT_SHARED_FEED_FORWARD_LENGTH, hparams.n_ff_shexp,      false);
+    ml.get_key(LLM_KV_EXPERT_SHARED_COUNT,               hparams.n_expert_shared, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_NORM,               hparams.expert_weights_norm, false);
+    ml.get_key(LLM_KV_EXPERT_WEIGHTS_SCALE,              hparams.expert_weights_scale, false);
+    ml.get_key(LLM_KV_MOE_LATENT_SIZE,                   hparams.moe_latent_size, false);
+
+    switch (hparams.n_layer) {
+        case 52: type = LLM_TYPE_31B_A3_5B; break; // Nemotron-H_MOE 31B
+        case 56: type = LLM_TYPE_9B; break;
+        case 88: type = LLM_TYPE_120B_A12B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_nemotron_h::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    // mamba2 Mixer SSM params
+    // NOTE: int64_t for tensor dimensions
+    const int64_t d_conv     = hparams.ssm_d_conv;
+    const int64_t d_inner    = hparams.ssm_d_inner;
+    const int64_t d_state    = hparams.ssm_d_state;
+    const int64_t n_ssm_head = hparams.ssm_dt_rank;
+    const int64_t n_group    = hparams.ssm_n_group;
+    const int64_t d_in_proj  = 2*d_inner + 2*n_group*d_state + n_ssm_head;
+    const int64_t moe_n_embd = hparams.moe_latent_size > 0 ? hparams.moe_latent_size : n_embd;
+
+    // embeddings
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    {
+        output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+        output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, TENSOR_NOT_REQUIRED);
+        // if output is NULL, init from the input tok embed, duplicated to allow offloading
+        if (output == NULL) {
+            output = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, TENSOR_DUPLICATED);
+        }
+    }
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        // all blocks use the attn norm
+        layer.attn_norm  = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        if (hparams.is_recurrent(i)) {
+            // ssm layers
+            layer.ssm_in = create_tensor(tn(LLM_TENSOR_SSM_IN, "weight", i), {n_embd, d_in_proj}, 0);
+
+            layer.ssm_conv1d = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "weight", i), {d_conv, d_inner + 2*n_group*d_state}, 0);
+            layer.ssm_conv1d_b = create_tensor(tn(LLM_TENSOR_SSM_CONV1D, "bias", i), {d_inner + 2*n_group*d_state}, TENSOR_NOT_REQUIRED);
+
+            layer.ssm_dt_b = create_tensor(tn(LLM_TENSOR_SSM_DT, "bias", i), {n_ssm_head}, 0);
+
+            // no "weight" suffix for these
+            layer.ssm_a = create_tensor(tn(LLM_TENSOR_SSM_A, i), {1, n_ssm_head}, 0);
+            layer.ssm_d = create_tensor(tn(LLM_TENSOR_SSM_D, i), {1, n_ssm_head}, 0);
+
+            layer.ssm_norm = create_tensor(tn(LLM_TENSOR_SSM_NORM, "weight", i), {d_inner / n_group, n_group}, 0);
+
+            // out_proj
+            layer.ssm_out = create_tensor(tn(LLM_TENSOR_SSM_OUT, "weight", i), {d_inner, n_embd}, 0);
+        } else if (hparams.n_ff(i) == 0) {
+            // attention layers (with optional bias)
+            const int64_t n_head_i = hparams.n_head(i);
+            const int64_t n_embd_k_gqa_i = hparams.n_embd_k_gqa(i);
+            const int64_t n_embd_v_gqa_i = hparams.n_embd_v_gqa(i);
+            create_tensor_qkv(layer, i, n_embd, n_embd_head_k * n_head_i, n_embd_k_gqa_i, n_embd_v_gqa_i, 0);
+            layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * n_head_i, n_embd}, 0);
+            layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+        }  else {
+            if (n_expert != 0) {
+                const int64_t n_ff_exp = hparams.n_ff_exp ? hparams.n_ff_exp : n_ff / n_expert_used;
+                const int64_t n_ff_shexp = hparams.n_ff_shexp;
+
+                layer.ffn_gate_inp    = create_tensor(tn(LLM_TENSOR_FFN_GATE_INP,  "weight", i), { n_embd, n_expert}, 0);
+                layer.ffn_exp_probs_b = create_tensor(tn(LLM_TENSOR_FFN_EXP_PROBS_B, "bias", i), {n_expert         }, 0);
+
+                // MoE branch
+                layer.ffn_latent_down = create_tensor(tn(LLM_TENSOR_FFN_LATENT_DOWN, "weight", i), {n_embd, moe_n_embd}, TENSOR_NOT_REQUIRED);
+                layer.ffn_latent_up   = create_tensor(tn(LLM_TENSOR_FFN_LATENT_UP,   "weight", i), {moe_n_embd, n_embd}, TENSOR_NOT_REQUIRED);
+
+                layer.ffn_down_exps   = create_tensor(tn(LLM_TENSOR_FFN_DOWN_EXPS, "weight", i), {n_ff_exp,   moe_n_embd, n_expert}, 0);
+                layer.ffn_up_exps     = create_tensor(tn(LLM_TENSOR_FFN_UP_EXPS,   "weight", i), {moe_n_embd, n_ff_exp, n_expert}, 0);
+
+                // Shared expert branch
+                layer.ffn_down_shexp  = create_tensor(tn(LLM_TENSOR_FFN_DOWN_SHEXP, "weight", i), {n_ff_shexp, n_embd}, 0);
+                layer.ffn_up_shexp    = create_tensor(tn(LLM_TENSOR_FFN_UP_SHEXP,   "weight", i), {n_embd, n_ff_shexp}, 0);
+
+            } else {
+                // mlp layers
+                layer.ffn_down   = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  hparams.n_ff(i), n_embd}, 0);
+                layer.ffn_up     = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   hparams.n_ff(i)}, 0);
+                layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias",   i), {n_embd}, TENSOR_NOT_REQUIRED);
+                layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias",   i), {hparams.n_ff(i)}, TENSOR_NOT_REQUIRED);
+            }
+        }
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_nemotron_h::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_nemotron_h::graph::graph(const llama_model & model, const llm_graph_params & params) :
     llm_build_mamba_base(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());
@@ -60,7 +181,7 @@ llm_build_nemotron_h::llm_build_nemotron_h(const llama_model & model, const llm_
     ggml_build_forward_expand(gf, cur);
 }
 
-ggml_tensor * llm_build_nemotron_h::build_attention_layer(ggml_tensor *             cur,
+ggml_tensor * llama_model_nemotron_h::graph::build_attention_layer(ggml_tensor *             cur,
                                                           llm_graph_input_attn_kv * inp_attn,
                                                           const llama_model &       model,
                                                                 int64_t             n_embd_head,
@@ -76,7 +197,7 @@ ggml_tensor * llm_build_nemotron_h::build_attention_layer(ggml_tensor *
     return cur;
 }
 
-ggml_tensor * llm_build_nemotron_h::build_ffn_layer(ggml_tensor * cur, const llama_model & model, int il) {
+ggml_tensor * llama_model_nemotron_h::graph::build_ffn_layer(ggml_tensor * cur, const llama_model & model, int il) {
     if (model.layers[il].ffn_gate_inp == nullptr) {
         cur = build_ffn(cur,
                 model.layers[il].ffn_up,   model.layers[il].ffn_up_b,   model.layers[il].ffn_up_s,

examples/talk-llama/models/nemotron.cpp

@@ -1,6 +1,52 @@
 #include "models.h"
 
-llm_build_nemotron::llm_build_nemotron(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_nemotron::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
+    switch (hparams.n_layer) {
+        case 32: type = LLM_TYPE_4B; break;
+        default: type = LLM_TYPE_UNKNOWN;
+    }
+}
+
+void llama_model_nemotron::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
+
+    // output
+    output_norm   = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
+    output_norm_b = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "bias"), {n_embd}, 0);
+    output        = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_vocab}, 0);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm   = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+        layer.attn_norm_b = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "bias", i), {n_embd}, 0);
+
+        create_tensor_qkv(layer, i, n_embd, n_embd, n_embd_gqa, n_embd_gqa, 0);
+        layer.wo = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+        // optional bias tensors
+        layer.wo_b = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+        layer.ffn_norm_b = create_tensor(tn(LLM_TENSOR_FFN_NORM, "bias", i), {n_embd}, 0);
+
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {  n_ff, n_embd}, 0);
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd,   n_ff}, 0);
+
+        // optional MLP bias
+        layer.ffn_down_b = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "bias", i), {n_embd}, TENSOR_NOT_REQUIRED);
+        layer.ffn_up_b   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "bias", i), {n_ff}, TENSOR_NOT_REQUIRED);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_nemotron::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_nemotron::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());

examples/talk-llama/models/neo-bert.cpp

@@ -1,6 +1,46 @@
 #include "models.h"
 
-llm_build_neo_bert::llm_build_neo_bert(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
+void llama_model_neo_bert::load_arch_hparams(llama_model_loader & ml) {
+    ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
+
+    if (hparams.n_layer == 28) {
+        type = LLM_TYPE_250M;
+    }
+}
+
+void llama_model_neo_bert::load_arch_tensors(llama_model_loader &) {
+    LLAMA_LOAD_LOCALS;
+
+    tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD,  "weight"), {n_embd, n_vocab}, 0);
+
+    cls   = create_tensor(tn(LLM_TENSOR_CLS, "weight"), {n_embd, n_embd}, TENSOR_NOT_REQUIRED);
+    cls_b = create_tensor(tn(LLM_TENSOR_CLS, "bias"),   {n_embd},         TENSOR_NOT_REQUIRED);
+
+    cls_out   = create_tensor(tn(LLM_TENSOR_CLS_OUT, "weight"), {n_embd, hparams.n_cls_out}, TENSOR_NOT_REQUIRED);
+    cls_out_b = create_tensor(tn(LLM_TENSOR_CLS_OUT, "bias"),   {hparams.n_cls_out},         TENSOR_NOT_REQUIRED);
+
+    output_norm_enc = create_tensor(tn(LLM_TENSOR_ENC_OUTPUT_NORM, "weight"), {n_embd}, 0);
+
+    for (int i = 0; i < n_layer; ++i) {
+        auto & layer = layers[i];
+
+        layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.wqkv = create_tensor(tn(LLM_TENSOR_ATTN_QKV, "weight", i), {n_embd, n_embd + 2*n_embd_gqa}, 0);
+        layer.wo   = create_tensor(tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd, n_embd}, 0);
+
+        layer.ffn_norm = create_tensor(tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}, 0);
+
+        layer.ffn_up   = create_tensor(tn(LLM_TENSOR_FFN_UP,   "weight", i), {n_embd, n_ff*2}, 0);
+        layer.ffn_down = create_tensor(tn(LLM_TENSOR_FFN_DOWN, "weight", i), {n_ff, n_embd}, 0);
+    }
+}
+
+std::unique_ptr<llm_graph_context> llama_model_neo_bert::build_arch_graph(const llm_graph_params & params) const {
+    return std::make_unique<graph>(*this, params);
+}
+
+llama_model_neo_bert::graph::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
     const int64_t n_embd_head = hparams.n_embd_head_v();
 
     GGML_ASSERT(n_embd_head == hparams.n_embd_head_k());