tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-27 20:05:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

llama : Support llama 4 text-only (#12791)

Browse Source 
* llama4 conversion

* initial support, no chat template

* clean up a bit

* fix tokenizer conversion

* correct hparams

* try this

* fix shexp

* ffn_inp_normed

* chat template

* clean up model conversion

* add_bos

* add scale_before_ffn

* fix order

* weight_before_ffn

* llm_graph_input_attn_temp

* add chunk attn mask

* build_inp_attn_scale()

* add comment about ggml_repeat

* clarify comments

* fix build
This commit is contained in:
Xuan-Son Nguyen
2025-04-07 23:06:44 +02:00
committed by GitHub
parent 82974011f3
commit 1466621e73
17 changed files with 532 additions and 22 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
src/llama-graph.h
View File

@ -100,6 +100,23 @@ public:
const int64_t n_pos_per_token = 1;
};
// temperature tuning, used by llama4
class llm_graph_input_attn_temp : public llm_graph_input_i {
public:
llm_graph_input_attn_temp(int64_t n_pos_per_token, uint32_t n_attn_temp_floor_scale, float f_attn_temp_scale)
: n_pos_per_token(n_pos_per_token), n_attn_temp_floor_scale(n_attn_temp_floor_scale), f_attn_temp_scale(f_attn_temp_scale) {}
virtual ~llm_graph_input_attn_temp() = default;
void set_input(const llama_ubatch * ubatch) override;
ggml_tensor * attn_scale = nullptr; // F32 [n_batch]
const int64_t n_pos_per_token = 1;
const uint32_t n_attn_temp_floor_scale;
const float f_attn_temp_scale;
};
class llm_graph_input_pos_bucket : public llm_graph_input_i {
public:
llm_graph_input_pos_bucket(const llama_hparams & hparams) : hparams(hparams) {}
@ -470,6 +487,7 @@ struct llm_graph_context {
ggml_tensor * build_inp_embd(ggml_tensor * tok_embd) const;
ggml_tensor * build_inp_pos() const;
ggml_tensor * build_inp_attn_scale() const;
ggml_tensor * build_inp_out_ids() const;
ggml_tensor * build_inp_mean() const;
ggml_tensor * build_inp_cls() const;