vrr/koboldcpp
2
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2025-09-12 09:59:41 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions 3

remove attn_window_size from gguf

Browse source
This commit is contained in:
HimariO 2025-04-26 20:19:51 +08:00
parent 77b144a8e7
commit 7e1bb0437a
2 changed files with 6 additions and 6 deletions
Download patch file Download diff file Expand all files Collapse all files

11
examples/llava/clip.cpp
View file

@ -1720,8 +1720,7 @@ struct clip_model_loader {
get_u32(KEY_IMAGE_SIZE, hparams.image_size); get_u32(KEY_IMAGE_SIZE, hparams.image_size);
get_u32(KEY_PATCH_SIZE, hparams.patch_size); get_u32(KEY_PATCH_SIZE, hparams.patch_size);
get_u32(KEY_IMAGE_CROP_RESOLUTION, hparams.image_crop_resolution, false); get_u32(KEY_IMAGE_CROP_RESOLUTION, hparams.image_crop_resolution, false);
get_u32(KEY_ATTN_WINDOW_SIZE, hparams.attn_window_size, false); get_u32(KEY_WIN_ATTN_PATTERN, hparams.n_wa_pattern, ctx_clip.proj_type == PROJECTOR_TYPE_QWEN2_5_VL);
get_u32(KEY_WIN_ATTN_PATTERN, hparams.n_wa_pattern, false);
get_arr_int(KEY_IMAGE_GRID_PINPOINTS, hparams.image_grid_pinpoints, false); get_arr_int(KEY_IMAGE_GRID_PINPOINTS, hparams.image_grid_pinpoints, false);
{ {
@ -3210,12 +3209,13 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
std::vector<int> idx(ph * pw); std::vector<int> idx(ph * pw);
std::vector<int> inv_idx(ph * pw); std::vector<int> inv_idx(ph * pw);
if (hparams.attn_window_size > 0) { if (use_window_attn) {
const int attn_window_size = 112;
struct ggml_tensor * window_idx = ggml_graph_get_tensor(gf, "window_idx"); struct ggml_tensor * window_idx = ggml_graph_get_tensor(gf, "window_idx");
struct ggml_tensor * inv_window_idx = ggml_graph_get_tensor(gf, "inv_window_idx"); struct ggml_tensor * inv_window_idx = ggml_graph_get_tensor(gf, "inv_window_idx");
struct ggml_tensor * window_mask = ggml_graph_get_tensor(gf, "window_mask"); struct ggml_tensor * window_mask = ggml_graph_get_tensor(gf, "window_mask");
const int grid_window = hparams.attn_window_size / patch_size / merge_ratio; const int grid_window = attn_window_size / patch_size / merge_ratio;
int dst = 0; int dst = 0;
// [num_vision_tokens, num_vision_tokens] attention mask tensor // [num_vision_tokens, num_vision_tokens] attention mask tensor
std::vector<float> mask(pow(ipw * iph, 2), std::numeric_limits<float>::lowest()); std::vector<float> mask(pow(ipw * iph, 2), std::numeric_limits<float>::lowest());
@ -3342,9 +3342,10 @@ bool clip_image_batch_encode(clip_ctx * ctx, const int n_threads, const clip_ima
struct ggml_tensor * window_mask = ggml_graph_get_tensor(gf, "window_mask"); struct ggml_tensor * window_mask = ggml_graph_get_tensor(gf, "window_mask");
const int merge_ratio = 2; const int merge_ratio = 2;
const int attn_window_size = 112;
const int pw = image_size_width / patch_size / merge_ratio; const int pw = image_size_width / patch_size / merge_ratio;
const int ph = image_size_height / patch_size / merge_ratio; const int ph = image_size_height / patch_size / merge_ratio;
const int grid_window = hparams.attn_window_size / patch_size / merge_ratio; const int grid_window = attn_window_size / patch_size / merge_ratio;
const int ipw = image_size_width / patch_size; const int ipw = image_size_width / patch_size;
const int iph = image_size_height / patch_size; const int iph = image_size_height / patch_size;
/* /*

1
examples/llava/qwen2_vl_surgery.py
View file

@ -167,7 +167,6 @@ def main(args):
if args.model_type == "qwen2.5vl": if args.model_type == "qwen2.5vl":
fout.add_uint32("clip.vision.n_wa_pattern", get_n_wa_pattern(vcfg.fullatt_block_indexes)) fout.add_uint32("clip.vision.n_wa_pattern", get_n_wa_pattern(vcfg.fullatt_block_indexes))
fout.add_uint32("clip.vision.window_size", vcfg.window_size)
fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.hidden_size) fout.add_uint32(k(KEY_EMBEDDING_LENGTH, VISION), vcfg.hidden_size)
fout.add_uint32("clip.vision.projection_dim", vcfg.out_hidden_size) fout.add_uint32("clip.vision.projection_dim", vcfg.out_hidden_size)
fout.add_string("clip.projector_type", "qwen2.5vl_merger") fout.add_string("clip.projector_type", "qwen2.5vl_merger")