public_git/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/whisper.cpp.git synced 2026-10-22 03:48:19 +02:00
Code Issues Packages Projects Releases Wiki Activity

talk-llama : sync llama.cpp

Browse Source
This commit is contained in:
Georgi Gerganov
2026-06-15 10:33:53 +03:00
parent 0a3fa9ca17
commit 0ec0845110
25 changed files with 671 additions and 96 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/talk-llama/models/delta-net-base.cpp
+20 -21
View File
@@ -398,9 +398,8 @@ std::pair<ggml_tensor *, ggml_tensor *> llm_build_delta_net_base::build_delta_ne
GGML_ASSERT(b->ne[0] == 1 && b->ne[1] == H_v && b->ne[2] == n_tokens && b->ne[3] == n_seqs);
GGML_ASSERT(s->ne[0] == S_v && s->ne[1] == S_v && s->ne[2] == H_v && s->ne[3] == n_seqs);
// K=1 (final state only): reshape to 3D (S_v*S_v*H_v, 1, n_seqs) for ggml_gated_delta_net.
ggml_tensor * s_3d = ggml_reshape_3d(ctx0, s, S_v * S_v * H_v, 1, n_seqs);
ggml_tensor * result = ggml_gated_delta_net(ctx0, q, k, v, g, b, s_3d);
// K=1: output carries the final state only. state s is 4D [S_v, S_v, H_v, n_seqs].
ggml_tensor * result = ggml_gated_delta_net(ctx0, q, k, v, g, b, s, /*K=*/1);
if (n_tokens == 1) {
cb(result, LLAMA_TENSOR_NAME_FGDN_AR, il);
} else {
@@ -564,11 +563,8 @@ ggml_tensor * llm_build_delta_net_base::build_recurrent_attn(
const int64_t D = S_v * S_v * H_v;
const int64_t K = cparams.n_rs_seq + 1;
// TODO: remove pad + simplify
ggml_tensor * s_3d = ggml_reshape_3d(ctx0, s, D, 1, n_seqs);
ggml_tensor * s_3d_pad = ggml_pad (ctx0, s_3d, 0, K - 1, 0, 0);
ggml_tensor * gdn_out = ggml_gated_delta_net(ctx0, q, k, v, g, b, s_3d_pad);
// state s is 4D [S_v, S_v, H_v, n_seqs]; K snapshot slots are written into the output.
ggml_tensor * gdn_out = ggml_gated_delta_net(ctx0, q, k, v, g, b, s, K);
if (n_seq_tokens > 1) {
cb(gdn_out, LLAMA_TENSOR_NAME_FGDN_CH, il);
} else {
@@ -587,21 +583,24 @@ ggml_tensor * llm_build_delta_net_base::build_recurrent_attn(
cb(output, "attn_output", il);
const size_t row_size = hparams.n_embd_s() * ggml_element_size(ssm_states_all);
for (int64_t k_i = 0; k_i < K; ++k_i) {
const uint32_t cache_slot = (uint32_t) (K - 1 - k_i);
ggml_tensor * src = ggml_view_4d(ctx0, gdn_out,
S_v, S_v, H_v, n_seqs,
ggml_row_size(gdn_out->type, S_v),
ggml_row_size(gdn_out->type, S_v * S_v),
ggml_row_size(gdn_out->type, S_v * S_v * H_v),
ggml_row_size(gdn_out->type, attn_score_elems + k_i * state_size_per_snap));
ggml_tensor * dst = ggml_view_2d(ctx0, ssm_states_all,
hparams.n_embd_s(), n_seqs, ssm_states_all->nb[1],
((size_t) cache_slot * mem_size + kv_head) * row_size);
// op writes the last min(n_seq_tokens, K) snapshots; trailing slots are left unwritten
const int64_t n_written = std::min<int64_t>(n_seq_tokens, K);
ggml_build_forward_expand(gf, ggml_cpy(ctx0, src, dst));
}
// write the produced snapshots into the recurrent cache (snapshot slot i -> rollback group i)
ggml_tensor * src = ggml_view_3d(ctx0, gdn_out,
D, n_seqs, n_written,
ggml_row_size(gdn_out->type, D),
ggml_row_size(gdn_out->type, state_size_per_snap),
ggml_row_size(gdn_out->type, attn_score_elems));
ggml_tensor * dst = ggml_view_3d(ctx0, ssm_states_all,
D, n_seqs, n_written,
ssm_states_all->nb[1],
(size_t) mem_size * row_size,
(size_t) kv_head * row_size);
ggml_build_forward_expand(gf, ggml_cpy(ctx0, src, dst));
return output;
}
examples/talk-llama/models/eagle3.cpp
+323
View File
@@ -0,0 +1,323 @@
#include "models.h"
void llama_model_eagle3::load_arch_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
if (!ml.get_arr(LLM_KV_TARGET_LAYERS, target_layer_ids, false)) {
throw std::runtime_error("EAGLE3 model requires 'extract_layers' in GGUF metadata");
}
if (target_layer_ids.size() != 3) {
throw std::runtime_error("EAGLE3 requires exactly 3 entries in 'extract_layers'");
}
LLAMA_LOG_INFO("%s: EAGLE3 extract_layers = [%d, %d, %d]\n", __func__,
target_layer_ids[0],
target_layer_ids[1],
target_layer_ids[2]);
uint32_t n_embd_tgt = 0;
ml.get_key(LLM_KV_TARGET_HIDDEN_SIZE, n_embd_tgt);
LLAMA_LOG_INFO("%s: EAGLE3 n_embd_tgt = %u (draft n_embd = %u)\n", __func__, n_embd_tgt, hparams.n_embd);
hparams.n_embd_inp_impl = (uint32_t) target_layer_ids.size() * n_embd_tgt;
// eagle3 norm_before_residual (optional, default false)
// compatible with Readhat eagle3 speculator model
ml.get_key(LLM_KV_NORM_BEFORE_RESIDUAL, hparams.norm_before_residual, false);
if (hparams.norm_before_residual) {
LLAMA_LOG_INFO("%s: EAGLE3gnorm_before_residual = true\n", __func__);
}
type = LLM_TYPE_UNKNOWN;
}
void llama_model_eagle3::load_arch_tensors(llama_model_loader &) {
LLAMA_LOAD_LOCALS;
const int64_t n_embd_inp = hparams.n_embd_inp();
const int64_t n_embd_attn_input = 2 * n_embd;
// Get vocab size from the d2t tensor in the GGUF file (optional - only needed if eagle3 has different vocab_size than target)
// d2t: draft to target vocabulary mapping
int64_t n_draft_vocab = n_vocab; // Default: same as target vocab
const struct ggml_tensor * d2t_meta = ml->get_tensor_meta("d2t");
if (d2t_meta) {
n_draft_vocab = d2t_meta->ne[0]; // update draft vocab size
d2t = create_tensor(tn(LLM_TENSOR_D2T), {n_draft_vocab}, 0);
LLAMA_LOG_INFO("%s: EAGLE3 using d2t mapping (draft_vocab_size = %lld)\n", __func__, (long long)n_draft_vocab);
} else {
d2t = nullptr; // no d2t, use default vocab size
LLAMA_LOG_INFO("%s: EAGLE3 without d2t - sharing same vocab_size with target (vocab_size = %lld)\n", __func__, (long long)n_draft_vocab);
}
// Feature fusion layer: projects 3 target layers to draft hidden size
fc = create_tensor(tn(LLM_TENSOR_FC, "weight"), {n_embd_inp, n_embd}, 0);
// Output layer (uses draft vocab size)
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}, 0);
output = create_tensor(tn(LLM_TENSOR_OUTPUT, "weight"), {n_embd, n_draft_vocab}, TENSOR_NOT_REQUIRED);
// Token embeddings (optional - Llama 3.3 70B EAGLE3 has its own)
const struct ggml_tensor * tok_embd_meta = ml->get_tensor_meta(tn(LLM_TENSOR_TOKEN_EMBD, "weight").str().c_str());
if (tok_embd_meta) {
const int64_t n_target_vocab = tok_embd_meta->ne[1];
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_target_vocab}, 0);
LLAMA_LOG_INFO("%s: EAGLE3 using its own token_embd (vocab = %lld)\n", __func__, (long long)n_target_vocab);
}
// Single decoder layer
for (int i = 0; i < n_layer; ++i) {
auto & layer = layers[i];
// input_layernorm: applied to token embeddings
layer.attn_norm = create_tensor(tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}, 0);
// eagle3 specific: hidden_norm applied to fused target features
layer.attn_norm_2 = create_tensor(tn(LLM_TENSOR_ATTN_NORM_2, "weight", i), {n_embd}, 0);
// Attention takes input_embeds_normed + fused_target_normed as input
layer.wq = create_tensor(tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd_attn_input, n_embd_head_k * n_head}, 0);
layer.wk = create_tensor(tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd_attn_input, n_embd_k_gqa}, 0);
layer.wv = create_tensor(tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd_attn_input, 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_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);
// rope_freqs for llama3 rope scaling (optional - only if eagle3 config has rope_scaling)
layer.rope_freqs = create_tensor(tn(LLM_TENSOR_ROPE_FREQS, "weight", i), {n_rot/2}, TENSOR_NOT_REQUIRED);
}
}
std::unique_ptr<llm_graph_context> llama_model_eagle3::build_arch_graph(const llm_graph_params & params) const {
switch (params.gtype) {
case LLM_GRAPH_TYPE_ENCODER:
return std::make_unique<graph<true>>(*this, params);
case LLM_GRAPH_TYPE_DEFAULT:
case LLM_GRAPH_TYPE_DECODER:
return std::make_unique<graph<false>>(*this, params);
default:
GGML_ABORT("invalid graph type");
};
}
template <>
ggml_tensor * llama_model_eagle3::graph<true>::build_inp_embd_enc() const {
ggml_tensor * cur = nullptr;
// Input: Target model features (3 layers concatenated: low, mid, high)
// Data will be provided via ubatch->embd in encode_eagle3_features()
auto inp_target = std::make_unique<llm_graph_input_embd>(hparams.n_embd_inp());
inp_target->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32,hparams.n_embd_inp(), n_tokens);
ggml_set_input(inp_target->embd);
cur = inp_target->embd;
cb(cur, "inp_embd", -1);
res->add_input(std::move(inp_target));
return cur;
}
// eagle3 Encoder: processes target model features through feature fusion layer
// Input: target_features e.g. [12288, n_tokens] from target model layers low, middle, high
// Output: g_embeddings e.g. [4096, n_tokens] stored in context
template <>
llama_model_eagle3::graph<true>::graph(const llama_model & model, const llm_graph_params & params) : llm_graph_context(params) {
ggml_tensor * cur = nullptr;
cur = build_inp_embd_enc();
// Feature fusion layer
cur = build_lora_mm(model.fc, cur);
cb(cur, "fc_out", -1);
// Output: g_embeddings e.g. [4096, n_tokens]
// store in t_h_nextn (same as MTP) so can be read via llama_get_embeddings_nextn(ctx_dft)
ggml_set_output(cur);
res->t_h_nextn = cur;
ggml_build_forward_expand(gf, cur);
}
// eagle3 Decoder: processes draft tokens using g_embeddings from encoder
// Input: draft tokens + g_embeddings from encoder
// Output: draft logits
template <>
llama_model_eagle3::graph<false>::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_layer == 1); // eagle3 has only one decoder layer
ggml_tensor * cur;
ggml_tensor * inpL;
// eagle3 Decoder receives:
// 1. Token embeddings (e.g.from eagle3's own tok_embd for Llama 3.3 70B, or target model for Llama 3.1 8B)
// 2. g_embeddings from encoder
auto * tok_embd = model.tok_embd;
if (model.tok_embd == nullptr) {
GGML_ASSERT(cparams.ctx_other != nullptr);
const auto * model_other = llama_get_model(cparams.ctx_other);
GGML_ASSERT(model_other->tok_embd != nullptr && "EAGLE3 decoder requires token embeddings (own or from target model)");
tok_embd = model_other->tok_embd;
}
auto inp = std::make_unique<llm_graph_input_embd>(n_embd);
inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, n_tokens);
ggml_set_input(inp->tokens);
inp->embd = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, n_tokens);
ggml_set_input(inp->embd);
ggml_tensor * inp_embd = ggml_get_rows(ctx0, tok_embd, inp->tokens);
cb(inp_embd, "inp_embd", -1);
ggml_tensor * inp_g = inp->embd;
cb(inp_g, "inp_g_embeddings", -1);
res->add_input(std::move(inp));
inpL = inp_g;
// 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));
// Single decoder layer (il = 0)
const int il = 0;
{
// Apply input_layernorm to the token embeddings
ggml_tensor * embd_norm = build_norm(inp_embd,
model.layers[il].attn_norm, NULL,
LLM_NORM_RMS, il);
cb(embd_norm, "embd_norm", il);
// Apply hidden_norm to inp_g
ggml_tensor * g_norm = build_norm(inp_g,
model.layers[il].attn_norm_2, NULL,
LLM_NORM_RMS, -1);
cb(g_norm, "g_norm", il);
// norm_before_residual: determines what goes into the residual connection (compatible with Readhat eagle3 speculator model)
// - false (default): use raw inp_g for residual
// - true: use normalized g_norm for residual
// inpL is the concatenated input (normalized inp_embd + normalized inp_g)
ggml_tensor * inpSA = hparams.norm_before_residual ? g_norm : inpL;
// Concatenate normalized inp_embd and normalized inp_g
cur = ggml_concat(ctx0, embd_norm, g_norm, il);
cb(cur, "concat_embd", il);
// Self-attention with concatenated input
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, n_head, n_tokens);
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens);
Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
// rope freq factors, returns nullptr if not available
ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
// RoPE
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_rope", il);
cb(Kcur, "Kcur_rope", il);
cur = build_attn(inp_attn,
model.layers[il].wo, NULL, nullptr,
Qcur, Kcur, Vcur, nullptr, nullptr, nullptr, kq_scale, il);
// Add residual and update it
ggml_tensor * ffn_inp = ggml_add(ctx0, cur, inpSA);
cb(ffn_inp, "ffn_inp", il);
// Apply FFN norm to the sum
cur = build_norm(ffn_inp,
model.layers[il].ffn_norm, NULL,
LLM_NORM_RMS, il);
cb(cur, "post_attn_norm", il);
cur = 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, "ffn_out", il);
// Output norm with residual
cur = ggml_add(ctx0, cur, ffn_inp);
cb(cur, "eagle3_prenorm", il);
inpL = cur;
}
cur = inpL;
// Output prenorm state (for next token's g_embeddings in autoregressive generation)
ggml_set_output(cur);
res->t_h_nextn = cur;
cur = build_norm(cur,
model.output_norm, NULL,
LLM_NORM_RMS, -1);
cb(cur, "result_norm", -1);
// lm_head - projects to draft vocabulary
// if the draft has no own output projection, inherit the target model's lm_head
auto * output = model.output;
if (output == nullptr) {
GGML_ASSERT(cparams.ctx_other != nullptr);
const auto * model_other = llama_get_model(cparams.ctx_other);
GGML_ASSERT(model_other->output != nullptr && "EAGLE3 decoder requires an output projection (own or from target model)");
output = model_other->output;
}
cur = build_lora_mm(output, cur);
if (model.d2t) {
const int64_t n_draft_vocab = cur->ne[0];
const int64_t n_outputs = cur->ne[1];
const int64_t n_vocab = (int64_t) model.vocab.n_tokens();
GGML_ASSERT(model.d2t->type == GGML_TYPE_I64);
GGML_ASSERT(model.d2t->ne[0] == n_draft_vocab);
ggml_tensor * logits = ggml_fill(ctx0, ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, 1, n_vocab, n_outputs), -INFINITY);
cur = ggml_set_rows(ctx0, logits,
ggml_reshape_3d(ctx0, cur, 1, n_draft_vocab, n_outputs),
ggml_reshape_3d(ctx0, model.d2t, n_draft_vocab, 1, 1));
cur = ggml_reshape_2d(ctx0, cur, n_vocab, n_outputs);
}
cb(cur, "result_output", -1);
res->t_logits = cur;
ggml_build_forward_expand(gf, cur);
}
examples/talk-llama/models/gemma4-assistant.cpp
+3
View File
@@ -39,6 +39,9 @@ void llama_model_gemma4_assistant::load_arch_tensors(llama_model_loader &) {
output_norm = create_tensor(tn(LLM_TENSOR_OUTPUT_NORM, "weight"), { n_embd }, 0);
create_tensor(tn(LLM_TENSOR_MASKED_EMBD_CENTROIDS, "weight"), {}, TENSOR_NOT_REQUIRED);
create_tensor(tn(LLM_TENSOR_MASKED_EMBD_ORDERING), {}, TENSOR_NOT_REQUIRED);
const int64_t n_embd_backbone = hparams.n_embd_inp();
nextn_proj_post = create_tensor(tn(LLM_TENSOR_NEXTN_PROJ_POST, "weight"), { n_embd, n_embd_backbone }, 0);
examples/talk-llama/models/gemma4.cpp
+2
View File
@@ -210,6 +210,8 @@ llama_model_gemma4::graph::graph(const llama_model & model, const llm_graph_para
const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
const int n_rot_l = hparams.n_rot(il);
res->t_layer_inp[il] = inpL;
// norm
cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);
examples/talk-llama/models/llama.cpp
+2
View File
@@ -124,6 +124,8 @@ llama_model_llama::graph<embed>::graph(const llama_model & model, const llm_grap
ggml_tensor * inp_out_ids = build_inp_out_ids();
for (int il = 0; il < n_layer; ++il) {
res->t_layer_inp[il] = inpL;
ggml_tensor * inpSA = inpL;
// norm
examples/talk-llama/models/models.h
+16 -1
View File
@@ -46,7 +46,7 @@ struct llm_build_delta_net_base : public llm_graph_context {
ggml_tensor * s,
int il);
// use the ggml_gated_delta_net fused operator (K=1; state has shape (D, 1, n_seqs))
// use the ggml_gated_delta_net fused operator (K=1; state has shape [S_v, S_v, H_v, n_seqs])
std::pair<ggml_tensor *, ggml_tensor *> build_delta_net_fused(
ggml_tensor * q,
ggml_tensor * k,
@@ -1089,6 +1089,21 @@ struct llama_model_glm_dsa : public llama_model_base {
std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
};
struct llama_model_eagle3 : public llama_model_base {
llama_model_eagle3(const struct llama_model_params & params) : llama_model_base(params) {}
void load_arch_hparams(llama_model_loader & ml) override;
void load_arch_tensors(llama_model_loader & ml) override;
template <bool is_enc>
struct graph : public llm_graph_context {
graph(const llama_model & model, const llm_graph_params & params);
ggml_tensor * build_inp_embd_enc() const;
};
std::unique_ptr<llm_graph_context> build_arch_graph(const llm_graph_params & params) const override;
};
struct llama_model_mistral4 : public llama_model_deepseek2 {
llama_model_mistral4(const struct llama_model_params & params) : llama_model_deepseek2(params) {}
examples/talk-llama/models/openai-moe.cpp
+2
View File
@@ -75,6 +75,8 @@ llama_model_openai_moe::graph::graph(const llama_model & model, const llm_graph_
ggml_tensor * inp_out_ids = build_inp_out_ids();
for (int il = 0; il < n_layer; ++il) {
res->t_layer_inp[il] = inpL;
const float freq_base_l = model.get_rope_freq_base (cparams, il);
const float freq_scale_l = model.get_rope_freq_scale(cparams, il);
examples/talk-llama/models/plamo2.cpp
+5 -1
View File
@@ -11,6 +11,10 @@ void llama_model_plamo2::load_arch_hparams(llama_model_loader & ml) {
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);
// Load attention parameters
ml.get_key(LLM_KV_ATTENTION_KEY_LENGTH, hparams.n_embd_head_k_full, false);
ml.get_key(LLM_KV_ATTENTION_VALUE_LENGTH, hparams.n_embd_head_v_full, false);
for (uint32_t i = 0; i < hparams.n_layer(); ++i) {
hparams.is_recr_impl[i] = hparams.n_head_kv(i) == 0;
}
@@ -273,7 +277,7 @@ ggml_tensor * llama_model_plamo2::graph::build_plamo2_mamba_layer(llm_graph_inpu
GGML_ASSERT(n_seqs != 0);
GGML_ASSERT(ubatch.equal_seqs());
GGML_ASSERT(ubatch.n_tokens == n_seq_tokens * n_seqs);
GGML_ASSERT(d_inner % n_head == 0);
GGML_ASSERT(d_inner % n_heads == 0);
GGML_ASSERT(n_group == 0);
ggml_tensor * conv_states_all = mctx_cur->get_r_l(il);
examples/talk-llama/models/qwen3.cpp
+2
View File
@@ -69,6 +69,8 @@ llama_model_qwen3::graph::graph(const llama_model & model, const llm_graph_param
ggml_tensor * inp_out_ids = build_inp_out_ids();
for (int il = 0; il < n_layer; ++il) {
res->t_layer_inp[il] = inpL;
ggml_tensor * inpSA = inpL;
// norm
examples/talk-llama/models/qwen35.cpp
+1 -1
View File
@@ -173,7 +173,7 @@ llama_model_qwen35::graph::graph(const llama_model & model, const llm_graph_para
}
if (il == n_layer - 1 && inp_out_ids && cparams.embeddings_nextn_masked) {
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
cur = ggml_get_rows(ctx0, cur, inp_out_ids);
inpSA = ggml_get_rows(ctx0, inpSA, inp_out_ids);
}
examples/talk-llama/models/qwen3moe.cpp
+2
View File
@@ -78,6 +78,8 @@ llama_model_qwen3moe::graph::graph(const llama_model & model, const llm_graph_pa
ggml_tensor * inp_out_ids = build_inp_out_ids();
for (int il = 0; il < n_layer; ++il) {
res->t_layer_inp[il] = inpL;
ggml_tensor * inpSA = inpL;
// norm