revert

Signed-off-by: Woosuk Kwon <woosuk@inferact.ai>

vllm/model_executor/models/deepseek_mtp.py

@@ -30,7 +30,6 @@ from .deepseek_v2 import (
     DeepseekV2DecoderLayer,
     DeepseekV2MixtureOfExperts,
     DeepseekV2MoE,
-    _try_load_fp8_indexer_wk,
     get_spec_layer_idx_from_weight_name,
 )
 from .utils import maybe_prefix
@@ -191,6 +190,10 @@ class DeepSeekMTP(nn.Module, DeepseekV2MixtureOfExperts):
         )
         # Set MoE hyperparameters
         self.set_moe_parameters()
+        self.is_fp4_ckpt = (
+            self.quant_config is not None
+            and self.quant_config.get_name() == "modelopt_fp4"
+        )
 
     def set_moe_parameters(self):
         self.expert_weights = []
@@ -245,12 +248,13 @@ class DeepSeekMTP(nn.Module, DeepseekV2MixtureOfExperts):
             ("fused_qkv_a_proj", "kv_a_proj_with_mqa", 1),
         ]
 
-        # Fused indexer wk + weights_proj (shard 0 = wk, shard 1 = weights_proj)
-        indexer_fused_mapping = [
-            ("wk_weights_proj", "wk", 0),
-            ("wk_weights_proj", "weights_proj", 1),
-        ]
-        stacked_params_mapping.extend(indexer_fused_mapping)
+        if self.is_fp4_ckpt:
+            # Fused indexer wk + weights_proj (shard 0 = wk, shard 1 = weights_proj)
+            indexer_fused_mapping = [
+                ("wk_weights_proj", "wk", 0),
+                ("wk_weights_proj", "weights_proj", 1),
+            ]
+            stacked_params_mapping.extend(indexer_fused_mapping)
 
         expert_params_mapping = SharedFusedMoE.make_expert_params_mapping(
             self,
@@ -267,7 +271,6 @@ class DeepSeekMTP(nn.Module, DeepseekV2MixtureOfExperts):
 
         params_dict = dict(self.named_parameters())
         loaded_params: set[str] = set()
-        _pending_wk_fp8: dict = {}  # FP8 indexer wk dequant buffer
         for name, loaded_weight in weights:
             if "rotary_emb.inv_freq" in name:
                 continue
@@ -278,12 +281,6 @@ class DeepSeekMTP(nn.Module, DeepseekV2MixtureOfExperts):
                 rocm_aiter_moe_shared_expert_enabled and ("mlp.shared_experts" in name)
             )
             name = self._rewrite_spec_layer_name(spec_layer, name)
-
-            if _try_load_fp8_indexer_wk(
-                name, loaded_weight, _pending_wk_fp8, params_dict, loaded_params
-            ):
-                continue
-
             for param_name, weight_name, shard_id in stacked_params_mapping:
                 # Skip non-stacked layers and experts (experts handled below).
                 if weight_name not in name: