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

Kimi-K2.5: Use merged QKV for vision

This commit is contained in:
Aes Sedai 2026-02-07 21:14:17 -08:00
parent f13b383843
commit be1b0c3554
2 changed files with 5 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
convert_hf_to_gguf.py
View File

@ -11113,15 +11113,6 @@ class KimiK25Model(MmprojModel):
if not is_vision: if not is_vision:
return return
# Split fused QKV tensors in vision encoder
if "wqkv" in name:
split_dim = 0 if "weight" in name else -1
wq, wk, wv = data_torch.chunk(3, dim=split_dim)
yield from super().modify_tensors(wq, name.replace("wqkv", "wq"), bid)
yield from super().modify_tensors(wk, name.replace("wqkv", "wk"), bid)
yield from super().modify_tensors(wv, name.replace("wqkv", "wv"), bid)
return
# Temporal embeddings: (T, 1, C) → (T, C) # Temporal embeddings: (T, 1, C) → (T, C)
if "pos_emb.time_weight" in name: if "pos_emb.time_weight" in name:
T, _, C = data_torch.shape T, _, C = data_torch.shape

5
tools/mtmd/clip.cpp
View File

@ -726,6 +726,11 @@ ggml_tensor * clip_graph::build_rope_2d_interleaved(
GGML_ASSERT(n_dim % 4 == 0); // Must be divisible by 4 for interleaved x,y pairs GGML_ASSERT(n_dim % 4 == 0); // Must be divisible by 4 for interleaved x,y pairs
// Ensure input is contiguous (needed when using merged QKV with ggml_view)
if (!ggml_is_contiguous(cur)) {
cur = ggml_cont(ctx0, cur);
}
// Step 1: Reshape to expose interleaved structure // Step 1: Reshape to expose interleaved structure
// cur: [n_dim, n_head, n_pos] -> [4, n_dim/4, n_head, n_pos] // cur: [n_dim, n_head, n_pos] -> [4, n_dim/4, n_head, n_pos]
ggml_tensor * reshaped = ggml_reshape_4d(ctx0, cur, 4, n_dim/4, n_head, n_pos); ggml_tensor * reshaped = ggml_reshape_4d(ctx0, cur, 4, n_dim/4, n_head, n_pos);