public_git/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/whisper.cpp.git synced 2026-10-17 11:45:38 +02:00
Code Issues Packages Projects Releases Wiki Activity

talk-llama : sync llama.cpp

Browse Source
This commit is contained in:
Georgi Gerganov
2025-12-18 08:20:56 +02:00
parent 698348aadc
commit 6c22e792cb
27 changed files with 3211 additions and 2665 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/talk-llama/llama-kv-cache.cpp
+88 -39
View File
@@ -175,7 +175,15 @@ llama_kv_cache::llama_kv_cache(
// allocate tensors and initialize the buffers to avoid NaNs in the padding
for (auto & [buft, ctx] : ctx_map) {
ggml_backend_buffer_t buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft);
ggml_backend_buffer_t buf;
if (model.hparams.no_alloc) {
buf = ggml_backend_buft_alloc_buffer(buft, /*size =*/ 0); // dummy buffer
for (ggml_tensor * t = ggml_get_first_tensor(ctx.get()); t != nullptr; t = ggml_get_next_tensor(ctx.get(), t)) {
t->buffer = buf; // set dummy buffer for KV cache so that the backend scheduler won't try to allocate it
}
} else {
buf = ggml_backend_alloc_ctx_tensors_from_buft(ctx.get(), buft); // real buffer
}
if (!buf) {
throw std::runtime_error("failed to allocate buffer for kv cache");
}
@@ -482,9 +490,18 @@ llama_pos llama_kv_cache::seq_pos_max(llama_seq_id seq_id) const {
std::map<ggml_backend_buffer_type_t, size_t> llama_kv_cache::memory_breakdown() const {
std::map<ggml_backend_buffer_type_t, size_t> ret;
for (const auto & [_, buf] : ctxs_bufs) {
ret[ggml_backend_buffer_get_type(buf.get())] += ggml_backend_buffer_get_size(buf.get());
for (const auto & [ctx, buf] : ctxs_bufs) {
ggml_backend_buffer_type_t buft = ggml_backend_buffer_get_type(buf.get());
if (hparams.no_alloc) {
GGML_ASSERT(ggml_backend_buffer_get_base(buf.get()) == nullptr);
ret[buft] += ggml_backend_alloc_ctx_tensors_from_buft_size(ctx.get(), buft);
} else {
// GGML_ASSERT(ggml_backend_buffer_get_base(buf.get()) != nullptr); // multi_buffer does not have a defined base
ret[buft] += ggml_backend_buffer_get_size(buf.get());
}
}
return ret;
}
@@ -1369,9 +1386,10 @@ ggml_tensor * llama_kv_cache::build_rope_shift(
float freq_scale) const {
const auto & n_ctx_orig = cparams.n_ctx_orig_yarn;
const auto & yarn_ext_factor = cparams.yarn_ext_factor;
const auto & yarn_beta_fast = cparams.yarn_beta_fast;
const auto & yarn_beta_slow = cparams.yarn_beta_slow;
const auto & yarn_ext_factor = cparams.yarn_ext_factor;
const auto & yarn_beta_fast = cparams.yarn_beta_fast;
const auto & yarn_beta_slow = cparams.yarn_beta_slow;
const auto & yarn_attn_factor = cparams.yarn_attn_factor;
const auto & n_rot = hparams.n_rot;
const auto & rope_type = hparams.rope_type == LLAMA_ROPE_TYPE_MROPE || hparams.rope_type == LLAMA_ROPE_TYPE_IMROPE
@@ -1382,12 +1400,6 @@ ggml_tensor * llama_kv_cache::build_rope_shift(
? LLAMA_ROPE_TYPE_NEOX
: hparams.rope_type;
// See llm_build_deepseek2() for why attn_factor has to be scaled for YaRN RoPE to work correctly.
// See https://github.com/ggerganov/llama.cpp/discussions/7416 for detailed explanation.
const float yarn_attn_factor = model.arch == LLM_ARCH_DEEPSEEK2
? 1.0f / (1.0f + 0.1f * logf(1.0f / freq_scale))
: cparams.yarn_attn_factor;
ggml_tensor * tmp;
if (ggml_is_quantized(cur->type)) {
@@ -1549,9 +1561,11 @@ void llama_kv_cache::state_read(llama_io_read_i & io, llama_seq_id seq_id, llama
const uint32_t strm = seq_id == -1 ? s : seq_to_stream[seq_id];
slot_info sinfo;
bool res = true;
res = res && state_read_meta(io, strm, cell_count, seq_id);
res = res && state_read_data(io, strm, cell_count);
res = res && state_read_meta(io, strm, cell_count, sinfo, seq_id);
res = res && state_read_data(io, strm, cell_count, sinfo);
if (!res) {
if (seq_id == -1) {
@@ -1690,7 +1704,7 @@ void llama_kv_cache::state_write_data(llama_io_write_i & io, const cell_ranges_t
}
}
bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, llama_seq_id dest_seq_id) {
bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, slot_info & sinfo, llama_seq_id dest_seq_id) {
auto & cells = v_cells[strm];
auto & head = v_heads[strm];
@@ -1727,7 +1741,7 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
ubatch.seq_id[i] = &dest_seq_id;
}
const auto sinfo = find_slot(ubatch, true);
sinfo = find_slot(ubatch, false);
if (sinfo.empty()) {
LLAMA_LOG_ERROR("%s: failed to find available cells in kv cache\n", __func__);
return false;
@@ -1737,20 +1751,16 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
// see: https://github.com/ggml-org/llama.cpp/pull/16825#issuecomment-3460868350
apply_ubatch(sinfo, ubatch);
const auto head_cur = sinfo.head();
LLAMA_LOG_DEBUG("%s: cell_count = %d, dest_seq_id = %d\n", __func__, cell_count, dest_seq_id);
// keep the head at the old position because we will read the KV data into it in state_read_data()
head = head_cur;
LLAMA_LOG_DEBUG("%s: head_cur = %d, head = %d, cell_count = %d, dest_seq_id = %d\n", __func__, head_cur, head, cell_count, dest_seq_id);
// DEBUG CHECK: head_cur should be our first cell, head_cur + cell_count - 1 should be our last cell (verify seq_id and pos values)
// Assume that this is one contiguous block of cells
GGML_ASSERT(head_cur + cell_count <= cells.size());
GGML_ASSERT(cells.pos_get(head_cur) == ubatch.pos[0]);
GGML_ASSERT(cells.pos_get(head_cur + cell_count - 1) == ubatch.pos[cell_count - 1]);
GGML_ASSERT(cells.seq_has(head_cur, dest_seq_id));
GGML_ASSERT(cells.seq_has(head_cur + cell_count - 1, dest_seq_id));
// DEBUG CHECK: verify that all cells were allocated and have correct seq_id and pos values
GGML_ASSERT(sinfo.n_stream() == 1);
GGML_ASSERT(sinfo.idxs[0].size() == cell_count);
for (uint32_t i = 0; i < cell_count; ++i) {
const uint32_t idx = sinfo.idxs[0][i];
GGML_ASSERT(cells.pos_get(idx) == ubatch.pos[i]);
GGML_ASSERT(cells.seq_has(idx, dest_seq_id));
}
} else {
// whole KV cache restore
@@ -1783,15 +1793,24 @@ bool llama_kv_cache::state_read_meta(llama_io_read_i & io, uint32_t strm, uint32
}
}
// Create contiguous slot_info for whole cache restore
sinfo.s0 = strm;
sinfo.s1 = strm;
sinfo.resize(1);
sinfo.strm[0] = strm;
sinfo.idxs[0].resize(cell_count);
for (uint32_t i = 0; i < cell_count; ++i) {
sinfo.idxs[0][i] = i;
}
head = 0;
}
return true;
}
bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count) {
bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32_t cell_count, const slot_info & sinfo) {
auto & cells = v_cells[strm];
auto & head = v_heads[strm];
uint32_t v_trans;
uint32_t n_layer;
@@ -1841,8 +1860,17 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
}
if (cell_count) {
// Read and set the keys for the whole cell range
ggml_backend_tensor_set(k, io.read(cell_count * k_size_row), head * k_size_row, cell_count * k_size_row);
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);
} 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);
}
}
}
}
@@ -1873,8 +1901,17 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
}
if (cell_count) {
// Read and set the values for the whole cell range
ggml_backend_tensor_set(v, io.read(cell_count * v_size_row), head * v_size_row, cell_count * v_size_row);
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);
} 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);
}
}
}
}
} else {
@@ -1913,10 +1950,22 @@ bool llama_kv_cache::state_read_data(llama_io_read_i & io, uint32_t strm, uint32
}
if (cell_count) {
// For each row in the transposed matrix, read the values for the whole cell range
for (uint32_t j = 0; j < n_embd_v_gqa; ++j) {
const size_t dst_offset = (head + 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);
if (sinfo.is_contiguous()) {
// Fast path: contiguous cells
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);
}
} 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);
}
}
}
}
}