allow FA for clip but with wmma disabled for turing on bad sizes

This commit is contained in:
Concedo
2025-11-27 01:03:29 +08:00
parent e6ad29341b
commit 6770767d8a
+1 -1
View File
@@ -2470,7 +2470,7 @@ ModelLoadResult gpttype_load_model(const load_model_inputs inputs, FileFormat in
}
clip_context_params ctx_clip_params {
/* use_gpu */ true,
/* flash_attn_type */ CLIP_FLASH_ATTN_TYPE_DISABLED,
/* flash_attn_type */ (kcpp_data->flash_attn?CLIP_FLASH_ATTN_TYPE_ENABLED:CLIP_FLASH_ATTN_TYPE_DISABLED),
/* image_min_tokens */ -1,
/* image_max_tokens */ -1,
};