[Bugfix] [Quantization] Fix loading for CT DSV2 (#47780)
Kyle Sayers committed
Jul 6, 2026 at 22:28 UTC
445321fab445394c38aa485c27a783acc01e1b83
1 file changed
+1
-1
vllm/model_executor/models/deepseek_v2.py
+1
-1
@@ -829,7 +829,7 @@ def _try_load_fp8_indexer_wk(
829
if "indexer.wk." not in name or "wk_weights" in name:
830
return False # Weight is not an isolated WK weight for the indexer, ignore.
831
is_weight = name.endswith(".weight") and tensor.dtype == torch.float8_e4m3fn
832
- is_scale = "weight_scale_inv" in name
832
+ is_scale = "weight_scale" in name
833
if not is_weight and not is_scale:
834
return False # WK is not in FP8 format, ignore.
835
# Buffer this tensor (weight or scale) until both have arrived.