Superminaren/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-27 21:46:57 +02:00
Code Issues Packages Projects Releases Wiki Activity

model: use ggml_rope_set_offset() (#27382)

Browse Source 
* model: use ggml_rope_set_offset()

* partially apply to deepseek2
This commit is contained in:
Xuan-Son Nguyen
2026-08-21 18:54:29 +02:00
committed by GitHub
parent d7fa69b7de
commit 873e5d8e39
6 changed files with 71 additions and 133 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/models/plm.cpp
+10 -18
View File
@@ -81,19 +81,9 @@ llama_model_plm::graph::graph(const llama_model & model, const llm_graph_params
q = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
cb(q, "q", il);
// split into {n_head * n_embd_head_qk_nope, n_tokens}
ggml_tensor * q_nope = ggml_view_3d(ctx0, q, n_embd_head_qk_nope, n_head, n_tokens,
ggml_row_size(q->type, hparams.n_embd_head_k()),
ggml_row_size(q->type, hparams.n_embd_head_k() * n_head),
0);
cb(q_nope, "q_nope", il);
// and {n_head * n_embd_head_qk_rope, n_tokens}
ggml_tensor * q_pe = ggml_view_3d(ctx0, q, n_embd_head_qk_rope, n_head, n_tokens,
ggml_row_size(q->type, hparams.n_embd_head_k()),
ggml_row_size(q->type, hparams.n_embd_head_k() * n_head),
ggml_row_size(q->type, n_embd_head_qk_nope));
cb(q_pe, "q_pe", il);
// {n_embd_head_k, n_head, n_tokens}, RoPE is applied to the trailing dims only
q = ggml_reshape_3d(ctx0, q, hparams.n_embd_head_k(), n_head, n_tokens);
cb(q, "q", il);
// {n_embd, kv_lora_rank + n_embd_head_qk_rope} * {n_embd, n_tokens} -> {kv_lora_rank + n_embd_head_qk_rope, n_tokens}
ggml_tensor * kv_pe_compresseed = ggml_mul_mat(ctx0, model.layers[il].wkv_a_mqa, cur);
@@ -143,12 +133,13 @@ llama_model_plm::graph::graph(const llama_model & model, const llm_graph_params
0);
cb(v_states, "v_states", il);
q_pe = ggml_rope_ext(
ctx0, q_pe, inp_pos, nullptr,
q = ggml_rope_ext(
ctx0, q, inp_pos, nullptr,
n_rot, rope_type, n_ctx_orig, freq_base, freq_scale,
ext_factor, attn_factor, beta_fast, beta_slow
);
cb(q_pe, "q_pe", il);
q = ggml_rope_set_offset(q, n_embd_head_qk_nope);
cb(q, "q_rope", il);
// shared RoPE key
k_pe = ggml_rope_ext(
@@ -158,10 +149,11 @@ llama_model_plm::graph::graph(const llama_model & model, const llm_graph_params
);
cb(k_pe, "k_pe", il);
ggml_tensor * q_states = ggml_concat(ctx0, q_nope, q_pe, 0);
ggml_tensor * q_states = q;
cb(q_states, "q_states", il);
ggml_tensor * k_states = ggml_concat(ctx0, k_nope, ggml_repeat(ctx0, k_pe, q_pe), 0);
ggml_tensor * k_states = ggml_concat(ctx0, k_nope,
ggml_repeat_4d(ctx0, k_pe, n_embd_head_qk_rope, n_head, n_tokens, 1), 0);
cb(k_states, "k_states", il);
cur = build_attn(inp_attn,