[Bugfix][Model] MiMo: keep fused fp8 qkv_proj pairing state across weight-loading calls (#58142)

Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>

vllm/model_executor/models/mimo_v2.py

@@ -742,8 +742,12 @@ class MiMoV2Model(nn.Module, EagleModelMixin):
         expert_params_mapping = self.get_expert_mapping()
         # Pro-format fused qkv_proj arrives as two tensors (weight and
         # weight_scale_inv). Store them per-layer so that they can be
-        # sharded together.
-        pending_fp8_qkv_proj: dict[str, dict[str, torch.Tensor]] = {}
+        # sharded together. The state must outlive this call: AutoWeightsLoader
+        # delegates per contiguous group of names, so a pair can straddle two
+        # calls and would otherwise be dropped silently.
+        pending_fp8_qkv_proj = getattr(self, "_pending_fp8_qkv_proj", None)
+        if pending_fp8_qkv_proj is None:
+            self._pending_fp8_qkv_proj = pending_fp8_qkv_proj = {}
         for name, loaded_weight in weights:
             if "rotary_emb.inv_freq" in name:
                 continue

vllm/model_executor/models/mimo_v2_mtp.py

@@ -266,7 +266,12 @@ class MiMoV2MTP(nn.Module):
 
         params_dict = dict(self.named_parameters())
         loaded_params: set[str] = set()
-        pending_qkv_proj: dict[str, dict[str, torch.Tensor]] = {}
+        # The pairing state must outlive this call: AutoWeightsLoader delegates
+        # per contiguous group of names, so a pair can straddle two calls and
+        # would otherwise be dropped silently.
+        pending_qkv_proj = getattr(self, "_pending_qkv_proj", None)
+        if pending_qkv_proj is None:
+            self._pending_qkv_proj = pending_qkv_proj = {}
 
         for name, loaded_weight in weights:
             if "rotary_emb.inv_freq" in name: