CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-06 09:46:50 +00:00
Code Issues Packages Projects Releases Wiki Activity

mtmd: pad mask for qwen2.5vl (#16954)

Browse Source 
* mtmd: pad mask for qwen2.5vl

* improve
This commit is contained in:
Xuan-Son Nguyen
2025-11-03 10:25:55 +01:00
committed by GitHub
parent fcfce040e8
commit bf7b0c9725
1 changed files with 9 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
tools/mtmd/clip.cpp
View File

@@ -761,6 +761,15 @@ struct clip_graph {
ggml_set_name(window_mask, "window_mask"); ggml_set_name(window_mask, "window_mask");
ggml_set_input(window_mask); ggml_set_input(window_mask);
// if flash attn is used, we need to pad the mask and cast to f16
if (ctx->flash_attn_type == CLIP_FLASH_ATTN_TYPE_ENABLED) {
int n_pad = GGML_PAD(window_mask->ne[1], GGML_KQ_MASK_PAD) - window_mask->ne[1];
if (n_pad > 0) {
window_mask = ggml_pad(ctx0, window_mask, 0, n_pad, 0, 0);
}
window_mask = ggml_cast(ctx0, window_mask, GGML_TYPE_F16);
}
// inpL shape: [n_embd, n_patches_x * n_patches_y, batch_size] // inpL shape: [n_embd, n_patches_x * n_patches_y, batch_size]
GGML_ASSERT(batch_size == 1); GGML_ASSERT(batch_size == 1);
inpL = ggml_reshape_2d(ctx0, inpL, n_embd * 4, n_patches_x * n_patches_y * batch_size / 4); inpL = ggml_reshape_2d(ctx0, inpL, n_embd * 4, n_patches_x * n_patches_y * batch_size / 4);