diff options
| author | Kawrakow <iwankawrakow@gmail.com> | 2024-09-28 17:41:21 +0300 |
|---|---|---|
| committer | GitHub <noreply@github.com> | 2024-09-28 17:41:21 +0300 |
| commit | 7abcc6cc0b0a48b780bb0877e4720c46a7e3c255 (patch) | |
| tree | 0fbe1e7c3d02462ba3972f9be88a39ab9a3c0c30 /src/llama.cpp | |
| parent | 737514fd814d944f8ce965620293a16e5e8a285d (diff) | |
CUDA non-contiguous RoPE (#66)
In this way we can avoid the Q, K, V copies being made
after multiplication with the QKV tensor in, e.g., Phi-3.5-mini.
This results in a 6-7% speedup of PP-512(Phi-3.5-mini)
on CUDA (RTX-4080)
Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Diffstat (limited to 'src/llama.cpp')
| -rw-r--r-- | src/llama.cpp | 11 |
1 files changed, 5 insertions, 6 deletions
diff --git a/src/llama.cpp b/src/llama.cpp index d52590a6..94a939d8 100644 --- a/src/llama.cpp +++ b/src/llama.cpp @@ -10919,23 +10919,22 @@ struct llm_build_context { cur = llm_build_lora_mm(lctx, ctx0, model.layers[il].wqkv, attn_norm_output); cb(cur, "wqkv", il); - Qcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd, n_tokens, cur->nb[1], 0 * sizeof(float) * (n_embd))); - Kcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1 * sizeof(float) * (n_embd))); - Vcur = ggml_cont(ctx0, ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa))); + Qcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 0 * sizeof(float) * (n_embd)); + Kcur = ggml_view_3d(ctx0, cur, n_embd_head, n_head_kv, n_tokens, n_embd_head*sizeof(float), cur->nb[1], 1 * sizeof(float) * (n_embd)); + Vcur = ggml_view_2d(ctx0, cur, n_embd_gqa, n_tokens, cur->nb[1], 1 * sizeof(float) * (n_embd + n_embd_gqa)); } else { Qcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wq, attn_norm_output), model.layers[il].bq); Kcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wk, attn_norm_output), model.layers[il].bk); Vcur = ggml_add(ctx0, llm_build_lora_mm(lctx, ctx0, model.layers[il].wv, attn_norm_output), model.layers[il].bv); + Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); + Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); } cb(Qcur, "Qcur", il); cb(Kcur, "Kcur", il); cb(Vcur, "Vcur", il); - Qcur = ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens); - Kcur = ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens); - Qcur = ggml_rope_ext( ctx0, Qcur, inp_pos, rope_factors, n_rot, rope_type, n_ctx_orig, freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow |