matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-14 09:47:02 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama: update to commit de4c07f93 (#10655)

Browse Source
This commit is contained in:
Jeffrey Morgan
2025-05-12 12:17:26 -07:00
committed by GitHub
parent ad035ad595
commit 0cefd46f23
113 changed files with 8097 additions and 4383 deletions
Download Patch File Download Diff File Expand all files Collapse all files

58
llama/llama.cpp/src/llama-graph.cpp vendored
View File

@@ -284,24 +284,7 @@ void llm_graph_input_s_copy::set_input(const llama_ubatch * ubatch) {
// assuming copy destinations ALWAYS happen ONLY on the cells between head and head+n
for (uint32_t i = 0; i < n_kv; ++i) {
const uint32_t cell_id = i + kv_self->head;
//////////////////////////////////////////////
// TODO: this should not mutate the KV cache !
llama_kv_cell & kv_cell = const_cast<class llama_kv_cache_unified *>(kv_self)->cells[i];
// prevent out-of-bound sources
if (kv_cell.src < 0 || (uint32_t) kv_cell.src >= kv_self->size) {
kv_cell.src = cell_id;
}
data[i] = kv_cell.src;
// TODO: do not mutate the KV cache
// ensure copy only happens once
if (kv_cell.src != (int32_t) cell_id) {
kv_cell.src = cell_id;
}
data[i] = kv_self->s_copy(i);
}
}
}
@@ -317,18 +300,7 @@ void llm_graph_input_s_mask::set_input(const llama_ubatch * ubatch) {
// clear unused states
for (int i = 0; i < n_kv; ++i) {
const uint32_t cell_id = i + kv_self->head;
//////////////////////////////////////////////
// TODO: this should not mutate the KV cache !
llama_kv_cell & kv_cell = const_cast<class llama_kv_cache_unified *>(kv_self)->cells[i];
data[i] = (float) (kv_cell.src >= 0);
// only clear once
if (kv_cell.src < 0) {
kv_cell.src = cell_id;
}
data[i] = kv_self->s_mask(i);
}
}
}
@@ -816,7 +788,7 @@ ggml_tensor * llm_graph_context::build_ffn(
} break;
}
if (type_gate == LLM_FFN_PAR) {
if (gate && type_gate == LLM_FFN_PAR) {
cur = ggml_mul(ctx0, cur, tmp);
cb(cur, "ffn_gate_par", il);
}
@@ -1005,6 +977,7 @@ ggml_tensor * llm_graph_context::build_inp_embd(ggml_tensor * tok_embd) const {
inp->tokens = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, ubatch.n_tokens);
//cb(inp->tokens, "inp_tokens", -1);
ggml_set_input(inp->tokens);
res->t_tokens = inp->tokens;
cur = ggml_get_rows(ctx0, tok_embd, inp->tokens);
@@ -1111,7 +1084,7 @@ ggml_tensor * llm_graph_context::build_inp_cls() const {
}
ggml_tensor * llm_graph_context::build_inp_s_copy() const {
const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
auto inp = std::make_unique<llm_graph_input_s_copy>(kv_self);
@@ -1128,7 +1101,7 @@ ggml_tensor * llm_graph_context::build_inp_s_copy() const {
}
ggml_tensor * llm_graph_context::build_inp_s_mask() const {
const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
auto inp = std::make_unique<llm_graph_input_s_mask>(kv_self);
@@ -1261,8 +1234,19 @@ ggml_tensor * llm_graph_context::build_attn_mha(
ggml_flash_attn_ext_set_prec(cur, GGML_PREC_F32);
if (v_mla) {
#if 0
// v_mla can be applied as a matrix-vector multiplication with broadcasting across dimension 3 == n_tokens.
// However, the code is optimized for dimensions 0 and 1 being large, so this is ineffient.
cur = ggml_reshape_4d(ctx0, cur, v_mla->ne[0], 1, n_head, n_tokens);
cur = ggml_mul_mat(ctx0, v_mla, cur);
#else
// It's preferable to do the calculation as a matrix-matrix multiplication with n_tokens in dimension 1.
// The permutations are noops and only change how the tensor data is interpreted.
cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
cur = ggml_mul_mat(ctx0, v_mla, cur);
cur = ggml_permute(ctx0, cur, 0, 2, 1, 3);
cur = ggml_cont(ctx0, cur); // Needed because ggml_reshape_2d expects contiguous inputs.
#endif
}
cur = ggml_reshape_2d(ctx0, cur, cur->ne[0]*n_head, n_tokens);
@@ -1442,8 +1426,6 @@ ggml_tensor * llm_graph_context::build_attn(
// store to KV cache
{
GGML_ASSERT(!kv_self->recurrent);
const auto kv_head = kv_self->head;
GGML_ASSERT(kv_self->size == n_ctx);
@@ -1612,7 +1594,7 @@ ggml_tensor * llm_graph_context::build_copy_mask_state(
ggml_tensor * state_mask,
int32_t n_state,
int32_t n_seqs) const {
const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
const auto n_kv = kv_self->n;
const auto kv_head = kv_self->head;
@@ -1644,7 +1626,7 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_load(
ggml_tensor * state_mask,
const llama_ubatch & ubatch,
int il) const {
const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
const auto token_shift_count = hparams.token_shift_count;
@@ -1665,7 +1647,7 @@ ggml_tensor * llm_graph_context::build_rwkv_token_shift_store(
ggml_tensor * token_shift,
const llama_ubatch & ubatch,
int il) const {
const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
const llama_kv_cache_recurrent * kv_self = static_cast<const llama_kv_cache_recurrent *>(memory);
const auto token_shift_count = hparams.token_shift_count;
const auto n_embd = hparams.n_embd;