Fix weight mapper: compressor → attn.compressor (not mla_attn), quant weights_proj

- The compressor is on attn.compressor (not attn.mla_attn.compressor)
- weights_proj in indexer is NVFP4-quantized in our checkpoint

vllm/patches/deepseek_v4.py

@@ -1707,7 +1707,7 @@ def _make_deepseek_v4_nvfp4_weights_mapper() -> WeightsMapper:
             "compressor.kv_norm.": "compressor.norm.",
             "compressor.position_bias": "compressor.ape",
             # Attention compressor (after indexer renames)
-            ".self_attn.compressor.": ".attn.mla_attn.compressor.",
+            ".self_attn.compressor.": ".attn.compressor.",
             # Attention projections (specific before .self_attn. → .attn.)
             ".self_attn.q_a_proj.": ".attn.wq_a.",
             ".self_attn.kv_proj.": ".attn.wkv.",

vllm/patches/deepseek_v4_attention.py

@@ -1143,11 +1143,13 @@ class DeepseekV4Indexer(nn.Module):
             quant_config=quant_config,
             prefix=f"{prefix}.wq_b",
         )
+        # weights_proj is NVFP4-quantized in the modelopt checkpoint.
+        # Upstream uses quant_config=None for the MXFP4 checkpoint.
         self.weights_proj = ReplicatedLinear(
             hidden_size,
             self.n_head,
             bias=False,
-            quant_config=None,
+            quant_config=quant_config,
             prefix=f"{prefix}.weights_proj",
         )
         self.softmax_scale = self.head_dim**-0.5