matt/ollama37
1
0
Fork 0
mirror of https://github.com/dogkeeper886/ollama37.git synced 2025-12-12 16:57:04 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama: update to commit e1e8e099 (#10513)

Browse Source
This commit is contained in:
Jeffrey Morgan
2025-05-01 18:24:09 -07:00
committed by GitHub
parent e6d2d04121
commit 8dd12c873d
68 changed files with 3783 additions and 1774 deletions
Download Patch File Download Diff File Expand all files Collapse all files

152
llama/patches/0006-add-mllama-support.patch
View File

@@ -5,7 +5,6 @@ Subject: [PATCH] add mllama support
adds support for the llama 3.2 vision architecture
---
examples/llava/gemma3-cli.cpp | 3 +-
examples/llava/llava.cpp | 5 +-
examples/llava/mtmd.cpp | 6 +-
ggml/src/ggml-backend-reg.cpp | 6 +-
@@ -25,34 +24,13 @@ adds support for the llama 3.2 vision architecture
src/llama-model.cpp | 309 +++++++++++++++++++++++++++++++++-
src/llama-model.h | 12 ++
src/llama-quant.cpp | 4 +-
20 files changed, 475 insertions(+), 22 deletions(-)
19 files changed, 473 insertions(+), 21 deletions(-)
diff --git a/examples/llava/gemma3-cli.cpp b/examples/llava/gemma3-cli.cpp
index 3d566475..654d1358 100644
--- a/examples/llava/gemma3-cli.cpp
+++ b/examples/llava/gemma3-cli.cpp
@@ -106,7 +106,7 @@ struct decode_embd_batch {
std::vector<llama_seq_id *> seq_ids;
std::vector<int8_t> logits;
llama_batch batch;
- decode_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
+ decode_embd_batch(float * embd, int32_t n_embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
pos .resize(n_tokens);
n_seq_id.resize(n_tokens);
seq_ids .resize(n_tokens + 1);
@@ -118,6 +118,7 @@ struct decode_embd_batch {
/*n_tokens =*/ n_tokens,
/*tokens =*/ nullptr,
/*embd =*/ embd,
+ /*n_embd =*/ n_embd,
/*pos =*/ pos.data(),
/*n_seq_id =*/ n_seq_id.data(),
/*seq_id =*/ seq_ids.data(),
diff --git a/examples/llava/llava.cpp b/examples/llava/llava.cpp
index 03a22cbb..5eb40bcd 100644
index c00d16ae..bab027b5 100644
--- a/examples/llava/llava.cpp
+++ b/examples/llava/llava.cpp
@@ -456,7 +456,7 @@ struct llava_embd_batch {
@@ -457,7 +457,7 @@ struct llava_embd_batch {
std::vector<llama_seq_id *> seq_ids;
std::vector<int8_t> logits;
llama_batch batch;
@@ -61,7 +39,7 @@ index 03a22cbb..5eb40bcd 100644
pos .resize(n_tokens);
n_seq_id.resize(n_tokens);
seq_ids .resize(n_tokens + 1);
@@ -468,6 +468,7 @@ struct llava_embd_batch {
@@ -469,6 +469,7 @@ struct llava_embd_batch {
/*n_tokens =*/ n_tokens,
/*tokens =*/ nullptr,
/*embd =*/ embd,
@@ -69,7 +47,7 @@ index 03a22cbb..5eb40bcd 100644
/*pos =*/ pos.data(),
/*n_seq_id =*/ n_seq_id.data(),
/*seq_id =*/ seq_ids.data(),
@@ -491,7 +492,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
@@ -492,7 +493,7 @@ bool llava_eval_image_embed(llama_context * ctx_llama, const struct llava_image_
n_eval = n_batch;
}
float * embd = image_embed->embed+i*n_embd;
@@ -79,19 +57,19 @@ index 03a22cbb..5eb40bcd 100644
LOG_ERR("%s : failed to eval\n", __func__);
return false;
diff --git a/examples/llava/mtmd.cpp b/examples/llava/mtmd.cpp
index 3fd5bebc..f0cec596 100644
index 7081fd73..c14ac501 100644
--- a/examples/llava/mtmd.cpp
+++ b/examples/llava/mtmd.cpp
@@ -233,7 +233,7 @@ struct decode_embd_batch {
@@ -476,7 +476,7 @@ struct decode_embd_batch {
std::vector<llama_seq_id *> seq_ids;
std::vector<int8_t> logits;
llama_batch batch;
- decode_embd_batch(float * embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
+ decode_embd_batch(float * embd, int32_t n_embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) {
pos .resize(n_tokens);
- decode_embd_batch(float * embd, int32_t n_tokens, int n_pos_per_embd, int n_mmproj_embd) : n_pos_per_embd(n_pos_per_embd), n_mmproj_embd(n_mmproj_embd) {
+ decode_embd_batch(float * embd, int32_t n_embd, int32_t n_tokens, llama_pos pos_0, llama_seq_id seq_id) : n_pos_per_embd(n_pos_per_embd), n_mmproj_embd(n_mmproj_embd) {
pos .resize(n_tokens * n_pos_per_embd);
n_seq_id.resize(n_tokens);
seq_ids .resize(n_tokens + 1);
@@ -245,6 +245,7 @@ struct decode_embd_batch {
@@ -487,6 +487,7 @@ struct decode_embd_batch {
/*n_tokens =*/ n_tokens,
/*tokens =*/ nullptr,
/*embd =*/ embd,
@@ -99,16 +77,16 @@ index 3fd5bebc..f0cec596 100644
/*pos =*/ pos.data(),
/*n_seq_id =*/ n_seq_id.data(),
/*seq_id =*/ seq_ids.data(),
@@ -311,7 +312,8 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
int32_t n_tokens = mtmd_image_tokens_get_n_tokens(chunk.tokens_image.get());
@@ -610,7 +611,8 @@ int32_t mtmd_helper_eval(mtmd_context * ctx,
int32_t i_batch = 0;
int32_t n_img_batches = GGML_PAD(n_tokens, n_batch) / n_batch;
float * embd = mtmd_get_output_embd(ctx);
- decode_embd_batch batch_img(embd, n_tokens, n_past, 0);
- decode_embd_batch batch_embd(embd, n_tokens, n_pos_per_embd, n_mmproj_embd);
+ int n_embd = llama_model_n_embd(llama_get_model(lctx));
+ decode_embd_batch batch_img(embd, n_embd, n_tokens, n_past, 0);
int64_t t1 = ggml_time_ms();
ret = llama_decode(lctx, batch_img.batch);
if (ret != 0) {
+ decode_embd_batch batch_embd(embd, n_embd, n_tokens, n_past, 0);
const int nx = mtmd_image_tokens_get_nx(chunk.tokens_image.get());
const int ny = mtmd_image_tokens_get_ny(chunk.tokens_image.get());
diff --git a/ggml/src/ggml-backend-reg.cpp b/ggml/src/ggml-backend-reg.cpp
index 405d8e31..82ae1b5b 100644
--- a/ggml/src/ggml-backend-reg.cpp
@@ -127,10 +105,10 @@ index 405d8e31..82ae1b5b 100644
register_backend(ggml_backend_rpc_reg());
#endif
diff --git a/include/llama.h b/include/llama.h
index 5657fbf0..f91896e4 100644
index 06c56395..f1628e88 100644
--- a/include/llama.h
+++ b/include/llama.h
@@ -255,6 +255,7 @@ extern "C" {
@@ -256,6 +256,7 @@ extern "C" {
llama_token * token;
float * embd;
@@ -138,7 +116,7 @@ index 5657fbf0..f91896e4 100644
llama_pos * pos;
int32_t * n_seq_id;
llama_seq_id ** seq_id;
@@ -357,6 +358,7 @@ extern "C" {
@@ -358,6 +359,7 @@ extern "C" {
bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
bool flash_attn; // whether to use flash attention [EXPERIMENTAL]
bool no_perf; // whether to measure performance timings
@@ -146,7 +124,7 @@ index 5657fbf0..f91896e4 100644
// Abort callback
// if it returns true, execution of llama_decode() will be aborted
@@ -458,6 +460,10 @@ extern "C" {
@@ -459,6 +461,10 @@ extern "C" {
struct llama_context_params params),
"use llama_init_from_model instead");
@@ -158,7 +136,7 @@ index 5657fbf0..f91896e4 100644
LLAMA_API void llama_free(struct llama_context * ctx);
diff --git a/src/llama-arch.cpp b/src/llama-arch.cpp
index f754bc8f..0568565f 100644
index 5ab3f572..eb7b5325 100644
--- a/src/llama-arch.cpp
+++ b/src/llama-arch.cpp
@@ -6,6 +6,7 @@
@@ -169,7 +147,7 @@ index f754bc8f..0568565f 100644
{ LLM_ARCH_LLAMA4, "llama4" },
{ LLM_ARCH_DECI, "deci" },
{ LLM_ARCH_FALCON, "falcon" },
@@ -142,6 +143,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
@@ -144,6 +145,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" },
{ LLM_KV_ATTENTION_SCALE, "%s.attention.scale" },
{ LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION, "%s.attention.block_skip_connection" },
@@ -177,7 +155,7 @@ index f754bc8f..0568565f 100644
{ LLM_KV_ATTENTION_KEY_LENGTH_MLA, "%s.attention.key_length_mla" },
{ LLM_KV_ATTENTION_VALUE_LENGTH_MLA, "%s.attention.value_length_mla" },
@@ -271,6 +273,40 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
@@ -273,6 +275,40 @@ static const std::map<llm_arch, std::map<llm_tensor, const char *>> LLM_TENSOR_N
{ LLM_TENSOR_FFN_UP_SHEXP, "blk.%d.ffn_up_shexp" },
},
},
@@ -218,7 +196,7 @@ index f754bc8f..0568565f 100644
{
LLM_ARCH_DECI,
{
@@ -1681,6 +1717,14 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
@@ -1701,6 +1737,14 @@ static const std::map<llm_tensor, llm_tensor_info> LLM_TENSOR_INFOS = {
// this tensor is loaded for T5, but never used
{LLM_TENSOR_DEC_CROSS_ATTN_REL_B, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_NONE}},
{LLM_TENSOR_BSKCN_TV, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
@@ -234,7 +212,7 @@ index f754bc8f..0568565f 100644
{LLM_TENSOR_POS_NET_NORM, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
{LLM_TENSOR_POS_NET_NORM1, {LLM_TENSOR_LAYER_REPEATING, GGML_OP_MUL}},
diff --git a/src/llama-arch.h b/src/llama-arch.h
index 439aaeab..6a989034 100644
index 525c1b7d..bc8a4f0b 100644
--- a/src/llama-arch.h
+++ b/src/llama-arch.h
@@ -11,6 +11,7 @@
@@ -245,7 +223,7 @@ index 439aaeab..6a989034 100644
LLM_ARCH_DECI,
LLM_ARCH_FALCON,
LLM_ARCH_BAICHUAN,
@@ -146,6 +147,7 @@ enum llm_kv {
@@ -148,6 +149,7 @@ enum llm_kv {
LLM_KV_ATTENTION_SLIDING_WINDOW,
LLM_KV_ATTENTION_SCALE,
LLM_KV_ATTENTION_BLOCK_SKIP_CONNECTION,
@@ -253,7 +231,7 @@ index 439aaeab..6a989034 100644
LLM_KV_ATTENTION_KEY_LENGTH_MLA,
LLM_KV_ATTENTION_VALUE_LENGTH_MLA,
@@ -347,6 +349,14 @@ enum llm_tensor {
@@ -349,6 +351,14 @@ enum llm_tensor {
LLM_TENSOR_CLS,
LLM_TENSOR_CLS_OUT,
LLM_TENSOR_BSKCN_TV,
@@ -297,10 +275,10 @@ index 01d5ca57..8682b0e6 100644
batch.token = (llama_token *) malloc(sizeof(llama_token) * n_tokens_alloc);
}
diff --git a/src/llama-context.cpp b/src/llama-context.cpp
index 32f59819..0343ba8a 100644
index 9c1fe93f..cd06ad91 100644
--- a/src/llama-context.cpp
+++ b/src/llama-context.cpp
@@ -862,7 +862,7 @@ float * llama_context::get_logits_ith(int32_t i) {
@@ -851,7 +851,7 @@ float * llama_context::get_logits_ith(int32_t i) {
throw std::runtime_error(format("corrupt output buffer (j=%d, n_outputs=%d)", j, n_outputs));
}
@@ -309,7 +287,7 @@ index 32f59819..0343ba8a 100644
} catch (const std::exception & err) {
LLAMA_LOG_ERROR("%s: invalid logits id %d, reason: %s\n", __func__, i, err.what());
#ifndef NDEBUG
@@ -983,6 +983,10 @@ void llama_context::set_warmup(bool value) {
@@ -972,6 +972,10 @@ void llama_context::set_warmup(bool value) {
cparams.warmup = value;
}
@@ -320,7 +298,7 @@ index 32f59819..0343ba8a 100644
void llama_context::set_adapter_lora(
llama_adapter_lora * adapter,
float scale) {
@@ -1058,7 +1062,7 @@ int llama_context::encode(llama_batch & inp_batch) {
@@ -1047,7 +1051,7 @@ int llama_context::encode(llama_batch & inp_batch) {
const int64_t n_embd = hparams.n_embd;
@@ -329,7 +307,7 @@ index 32f59819..0343ba8a 100644
const llama_ubatch ubatch = sbatch.split_simple(n_tokens);
@@ -1198,10 +1202,9 @@ int llama_context::decode(llama_batch & inp_batch) {
@@ -1187,10 +1191,9 @@ int llama_context::decode(llama_batch & inp_batch) {
const llama_batch & batch = batch_allocr.batch;
@@ -341,7 +319,7 @@ index 32f59819..0343ba8a 100644
const int64_t n_tokens_all = batch.n_tokens;
const int64_t n_embd = hparams.n_embd;
@@ -1249,7 +1252,7 @@ int llama_context::decode(llama_batch & inp_batch) {
@@ -1238,7 +1241,7 @@ int llama_context::decode(llama_batch & inp_batch) {
const bool logits_all = n_outputs_all == n_tokens_all;
@@ -350,7 +328,7 @@ index 32f59819..0343ba8a 100644
/* simple_split */ !kv_self->recurrent,
/* logits_all */ logits_all);
@@ -1483,12 +1486,11 @@ int llama_context::decode(llama_batch & inp_batch) {
@@ -1472,12 +1475,11 @@ int llama_context::decode(llama_batch & inp_batch) {
int32_t llama_context::output_reserve(int32_t n_outputs) {
const auto & hparams = model.hparams;
@@ -364,7 +342,7 @@ index 32f59819..0343ba8a 100644
const auto n_embd = hparams.n_embd;
// TODO: use a per-batch flag for logits presence instead
@@ -1558,7 +1560,7 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
@@ -1545,7 +1547,7 @@ int32_t llama_context::output_reserve(int32_t n_outputs) {
void llama_context::output_reorder() {
auto & out_ids = sbatch.out_ids;
if (!out_ids.empty()) {
@@ -373,7 +351,7 @@ index 32f59819..0343ba8a 100644
const uint32_t n_embd = model.hparams.n_embd;
GGML_ASSERT((size_t) n_outputs == out_ids.size());
@@ -2065,7 +2067,7 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
@@ -2052,7 +2054,7 @@ size_t llama_context::state_write_data(llama_io_write_i & io) {
{
LLAMA_LOG_DEBUG("%s: - writing logits\n", __func__);
@@ -382,7 +360,7 @@ index 32f59819..0343ba8a 100644
io.write(&logits_size, sizeof(logits_size));
@@ -2248,6 +2250,7 @@ llama_context_params llama_context_default_params() {
@@ -2235,6 +2237,7 @@ llama_context_params llama_context_default_params() {
/*.offload_kqv =*/ true,
/*.flash_attn =*/ false,
/*.no_perf =*/ true,
@@ -390,7 +368,7 @@ index 32f59819..0343ba8a 100644
/*.abort_callback =*/ nullptr,
/*.abort_callback_data =*/ nullptr,
};
@@ -2375,6 +2378,10 @@ void llama_set_warmup(llama_context * ctx, bool warmup) {
@@ -2362,6 +2365,10 @@ void llama_set_warmup(llama_context * ctx, bool warmup) {
ctx->set_warmup(warmup);
}
@@ -402,7 +380,7 @@ index 32f59819..0343ba8a 100644
ctx->synchronize();
}
diff --git a/src/llama-context.h b/src/llama-context.h
index 04facb54..baa03276 100644
index 5457f077..a50c4afa 100644
--- a/src/llama-context.h
+++ b/src/llama-context.h
@@ -65,6 +65,7 @@ struct llama_context {
@@ -426,10 +404,10 @@ index 30e550f0..85ad91b9 100644
enum llama_pooling_type pooling_type;
diff --git a/src/llama-graph.cpp b/src/llama-graph.cpp
index a85e9728..d740c120 100644
index fabb9ca2..b67216a4 100644
--- a/src/llama-graph.cpp
+++ b/src/llama-graph.cpp
@@ -546,6 +546,12 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
@@ -560,6 +560,12 @@ void llm_graph_input_attn_cross::set_input(const llama_ubatch * ubatch) {
}
}
@@ -442,7 +420,7 @@ index a85e9728..d740c120 100644
//
// llm_graph_context
//
@@ -1506,6 +1512,25 @@ llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const {
@@ -1532,6 +1538,25 @@ llm_graph_input_attn_cross * llm_graph_context::build_attn_inp_cross() const {
return (llm_graph_input_attn_cross *) res->add_input(std::move(inp));
}
@@ -469,7 +447,7 @@ index a85e9728..d740c120 100644
llm_graph_input_attn_cross * inp,
ggml_cgraph * gf,
diff --git a/src/llama-graph.h b/src/llama-graph.h
index d192dc14..260a2af2 100644
index d0c8d321..0fe18150 100644
--- a/src/llama-graph.h
+++ b/src/llama-graph.h
@@ -86,6 +86,7 @@ public:
@@ -480,7 +458,7 @@ index d192dc14..260a2af2 100644
};
class llm_graph_input_pos : public llm_graph_input_i {
@@ -285,6 +286,16 @@ public:
@@ -283,6 +284,16 @@ public:
const llama_cross * cross = nullptr;
};
@@ -497,7 +475,7 @@ index d192dc14..260a2af2 100644
//
// llm_graph_result
//
@@ -493,6 +504,7 @@ struct llm_graph_context {
@@ -491,6 +502,7 @@ struct llm_graph_context {
ggml_tensor * build_inp_cls() const;
ggml_tensor * build_inp_s_copy() const;
ggml_tensor * build_inp_s_mask() const;
@@ -518,7 +496,7 @@ index 8a667960..6a02de03 100644
+ return std::find(cross_attn_layers.begin(), cross_attn_layers.end(), il) != cross_attn_layers.end();
+}
diff --git a/src/llama-hparams.h b/src/llama-hparams.h
index 6e278945..c8a34d52 100644
index 48dce407..b6fc7e6d 100644
--- a/src/llama-hparams.h
+++ b/src/llama-hparams.h
@@ -2,6 +2,8 @@
@@ -546,7 +524,7 @@ index 6e278945..c8a34d52 100644
uint32_t n_layer_dense_lead = 0;
uint32_t n_lora_q = 0;
@@ -158,6 +162,9 @@ struct llama_hparams {
@@ -159,6 +163,9 @@ struct llama_hparams {
// Block skip connection
bool n_bskcn(uint32_t n, uint32_t il) const;
@@ -593,10 +571,10 @@ index a012aeae..2e11507d 100644
bool llama_model_loader::get_arr(const std::string & key, std::array<T, N_MAX> & result, bool required) {
const int kid = gguf_find_key(meta.get(), key.c_str());
diff --git a/src/llama-model.cpp b/src/llama-model.cpp
index aba42819..d051696c 100644
index 572378c9..9d099f11 100644
--- a/src/llama-model.cpp
+++ b/src/llama-model.cpp
@@ -419,6 +419,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
@@ -423,6 +423,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
// get general kv
ml.get_key(LLM_KV_GENERAL_NAME, name, false);
@@ -604,7 +582,7 @@ index aba42819..d051696c 100644
// everything past this point is not vocab-related
if (hparams.vocab_only) {
@@ -430,6 +431,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
@@ -434,6 +435,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_BLOCK_COUNT, hparams.n_layer);
ml.get_key(LLM_KV_EXPERT_COUNT, hparams.n_expert, false);
ml.get_key(LLM_KV_EXPERT_USED_COUNT, hparams.n_expert_used, false);
@@ -612,7 +590,7 @@ index aba42819..d051696c 100644
if (arch == LLM_ARCH_WAVTOKENIZER_DEC) {
ml.get_key(LLM_KV_FEATURES_LENGTH, hparams.n_embd_features);
@@ -453,9 +455,11 @@ void llama_model::load_hparams(llama_model_loader & ml) {
@@ -457,9 +459,11 @@ void llama_model::load_hparams(llama_model_loader & ml) {
std::fill(hparams.n_head_arr.begin(), hparams.n_head_arr.end(), 0);
std::fill(hparams.n_head_kv_arr.begin(), hparams.n_head_kv_arr.end(), 0);
std::fill(hparams.n_ff_arr.begin(), hparams.n_ff_arr.end(), 0);
@@ -624,7 +602,7 @@ index aba42819..d051696c 100644
// n_head_kv is optional, default to n_head
hparams.n_head_kv_arr = hparams.n_head_arr;
@@ -508,7 +512,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
@@ -512,7 +516,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
ml.get_key(LLM_KV_ROPE_DIMENSION_COUNT, hparams.n_rot, false);
@@ -633,7 +611,7 @@ index aba42819..d051696c 100644
if (hparams.n_rot != hparams.n_embd_head_k) {
throw std::runtime_error(format("invalid n_rot: %u, expected %u", hparams.n_rot, hparams.n_embd_head_k));
}
@@ -571,6 +575,16 @@ void llama_model::load_hparams(llama_model_loader & ml) {
@@ -575,6 +579,16 @@ void llama_model::load_hparams(llama_model_loader & ml) {
hparams.use_kq_norm = false;
}
} break;
@@ -650,7 +628,7 @@ index aba42819..d051696c 100644
case LLM_ARCH_DECI:
{
ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps);
@@ -1550,7 +1564,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
@@ -1562,7 +1576,7 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
const int64_t n_embd_head_v = hparams.n_embd_head_v;
const int64_t n_ff = hparams.n_ff();
const int64_t n_embd_gqa = n_embd_v_gqa;
@@ -659,7 +637,7 @@ index aba42819..d051696c 100644
const int64_t n_token_types = vocab.n_token_types();
const int64_t n_rot = hparams.n_rot;
const int64_t n_expert = hparams.n_expert;
@@ -1803,6 +1817,52 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
@@ -1815,6 +1829,52 @@ bool llama_model::load_tensors(llama_model_loader & ml) {
}
}
} break;
@@ -712,7 +690,7 @@ index aba42819..d051696c 100644
case LLM_ARCH_DECI:
{
tok_embd = create_tensor(tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, 0);
@@ -4683,6 +4743,246 @@ struct llm_build_llama : public llm_graph_context {
@@ -4707,6 +4767,246 @@ struct llm_build_llama : public llm_graph_context {
}
};
@@ -959,7 +937,7 @@ index aba42819..d051696c 100644
struct llm_build_deci : public llm_graph_context {
llm_build_deci(const llama_model & model, const llm_graph_params & params, ggml_cgraph * gf) : llm_graph_context(params) {
const int64_t n_embd_head = hparams.n_embd_head_v;
@@ -13017,6 +13317,10 @@ llm_graph_result_ptr llama_model::build_graph(
@@ -13063,6 +13363,10 @@ llm_graph_result_ptr llama_model::build_graph(
{
llm = std::make_unique<llm_build_llama>(*this, params, gf);
} break;
@@ -970,7 +948,7 @@ index aba42819..d051696c 100644
case LLM_ARCH_DECI:
{
llm = std::make_unique<llm_build_deci>(*this, params, gf);
@@ -13377,6 +13681,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
@@ -13424,6 +13728,7 @@ llama_rope_type llama_model_rope_type(const llama_model * model) {
// use what we call a normal RoPE, operating on pairs of consecutive head values
case LLM_ARCH_LLAMA:
case LLM_ARCH_LLAMA4:
@@ -979,7 +957,7 @@ index aba42819..d051696c 100644
case LLM_ARCH_BAICHUAN:
case LLM_ARCH_STARCODER:
diff --git a/src/llama-model.h b/src/llama-model.h
index 5865d5e9..72bab5be 100644
index 856e6042..6be91282 100644
--- a/src/llama-model.h
+++ b/src/llama-model.h
@@ -11,6 +11,7 @@
@@ -990,15 +968,15 @@ index 5865d5e9..72bab5be 100644
struct llama_cparams;
struct llama_ubatch;
@@ -70,6 +71,7 @@ enum llm_type {
@@ -73,6 +74,7 @@ enum llm_type {
LLM_TYPE_40B,
LLM_TYPE_65B,
LLM_TYPE_70B,
+ LLM_TYPE_90B,
LLM_TYPE_236B,
LLM_TYPE_290B,
LLM_TYPE_314B,
LLM_TYPE_671B,
@@ -310,6 +312,16 @@ struct llama_layer {
@@ -314,6 +316,16 @@ struct llama_layer {
struct ggml_tensor * bskcn_tv = nullptr;