A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 5, 23:30 UTC
Fix minimax m2.5 nvfp4 kv scales weight loading (#37214)
Signed-off-by: wzhao18 <wzhao18.sz@gmail.com>
Wei Zhao committed
Mar 25, 2026 at 20:48 UTC
74056039b776776ed29eb0649e51ac6780592a4c
1 file changed
+11
vllm/model_executor/models/minimax_m2.py
+11
@@ -439,6 +439,17 @@ class MiniMaxM2Model(nn.Module):
439
if is_pp_missing_parameter(name, self):
440
continue
441
442
+ # Remap qkv_proj.[kv]_scale to attn.[kv]_scale
443
+ if name.endswith((".k_scale", ".v_scale")):
444
+ remapped_name = maybe_remap_kv_scale_name(name, params_dict)
445
+ if remapped_name is not None and remapped_name in params_dict:
446
+ param = params_dict[remapped_name]
447
+ weight_loader = getattr(
448
+ param, "weight_loader", default_weight_loader
449
+ )
450
+ weight_loader(param, loaded_weight)
451
+ break
452
+
453
param = params_dict[name]
454
weight_loader = param.weight_loader
455
weight_loader(param, loaded_weight, shard_id)