support Kimi-K2.5-w4a8 on ascend

This commit is contained in:
khalilzhk
2026-03-10 14:43:27 +08:00
committed by GitHub
parent dadd4dde83
commit 5f717913a0
3 changed files with 91 additions and 15 deletions

View File

@@ -152,6 +152,9 @@ class ModelSlimConfig(QuantizationConfig):
key = "vision_model"
elif "visual" in prefix:
key = "visual"
if "vision_tower" in prefix or "mm_projector" in prefix:
prefix = prefix.replace(r"attn.qkv_proj", r"wqkv")
prefix = prefix.replace(r"attn.proj", r"wo")
packed_modules_mapping_subset = self.packed_modules_mapping.get(key, {})
prefix_in_quant_config = prefix
proj_name = prefix.split(".")[-1]