happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

up.

This commit is contained in:
Vaibhavs10 2025-07-07 23:42:40 +02:00
parent 97c64a0974
commit 996195299e
3 changed files with 12 additions and 17 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
convert_hf_to_gguf.py
View File

@ -6305,11 +6305,9 @@ class SmolLM3Model(LlamaModel):
def set_gguf_parameters(self): def set_gguf_parameters(self):
super().set_gguf_parameters() super().set_gguf_parameters()
# if self.model.config.no_rope_layers is not None: no_rope_layer_interval = self.hparams.get("no_rope_layer_interval")
# self.gguf_writer.add_array("smollm3.no_rope_layers", self.model.config.no_rope_layers, gguf.GGUFValueType.INT32) if no_rope_layer_interval is not None:
no_rope_layers = self.hparams.get("no_rope_layers") self.gguf_writer.add_uint32("no_rope_layer_interval", no_rope_layer_interval)
if no_rope_layers is not None:
self.gguf_writer.add_array("smollm3.no_rope_layers", no_rope_layers)
###### CONVERSION LOGIC ###### ###### CONVERSION LOGIC ######

3
src/llama-hparams.h
View File

@ -186,6 +186,9 @@ struct llama_hparams {
// dimension of the recurrent state embeddings // dimension of the recurrent state embeddings
uint32_t n_embd_v_s() const; uint32_t n_embd_v_s() const;
// for NoPE interval
uint32_t no_rope_layer_interval = 0;
bool is_swa(uint32_t il) const; bool is_swa(uint32_t il) const;
}; };

18
src/llama-model.cpp
View File

@ -443,6 +443,10 @@ void llama_model::load_hparams(llama_model_loader & ml) {
return; return;
} }
if (arch == LLM_ARCH_SMOLLM3) {
ml.get_key("no_rope_layer_interval", hparams.no_rope_layer_interval);
}
ml.get_key(LLM_KV_CONTEXT_LENGTH, hparams.n_ctx_train); ml.get_key(LLM_KV_CONTEXT_LENGTH, hparams.n_ctx_train);
ml.get_key(LLM_KV_EMBEDDING_LENGTH, hparams.n_embd); ml.get_key(LLM_KV_EMBEDDING_LENGTH, hparams.n_embd);
ml.get_key(LLM_KV_BLOCK_COUNT, hparams.n_layer); ml.get_key(LLM_KV_BLOCK_COUNT, hparams.n_layer);
@ -13740,17 +13744,7 @@ struct llm_build_smollm3 : public llm_graph_context {
GGML_ASSERT(n_embd_head == hparams.n_embd_head_k); GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
GGML_ASSERT(n_embd_head == hparams.n_rot); GGML_ASSERT(n_embd_head == hparams.n_rot);
// collect layers for which RoPE is disabled (metadata key: "smollm3.no_rope_layers") const uint32_t interval = hparams.no_rope_layer_interval;
std::vector<int32_t> no_rope_layers;
if (arch == LLM_ARCH_SMOLLM3) {
const int kid = gguf_find_key(model.meta, "smollm3.no_rope_layers");
if (kid != -1) {
const uint32_t n = gguf_get_arr_n(model.meta, kid);
no_rope_layers.resize(n);
const int nb = gguf_get_arr_data(model.meta, kid, no_rope_layers.data(), n * sizeof(int32_t));
GGML_ASSERT(nb == int(n * sizeof(int32_t)));
}
}
// token embeddings // token embeddings
ggml_tensor * inpL = build_inp_embd(model.tok_embd); ggml_tensor * inpL = build_inp_embd(model.tok_embd);
@ -13793,7 +13787,7 @@ struct llm_build_smollm3 : public llm_graph_context {
Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, 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); Vcur = ggml_reshape_3d(ctx0, Vcur, n_embd_head, n_head_kv, n_tokens);
if (std::find(no_rope_layers.begin(), no_rope_layers.end(), il) == no_rope_layers.end()) { if (interval == 0 || il % interval != 0) {
ggml_tensor * rope_factors = model.get_rope_factors(cparams, il); ggml_tensor * rope_factors = model.get_rope_factors(cparams, il);
Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors, Qcur = ggml_rope_ext(ctx0, Qcur, inp_pos, rope_factors,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,