Fix: keep wo_a as FP8 (fp8_einsum path), dequant others to BF16

wo_a uses fp8_einsum which is weight-only FP8 (no input_scale).
Only q_a, q_b, kv, o_b need BF16 dequant to avoid broken input_scale.

vllm/patches/deepseek_v4.py

@@ -1711,8 +1711,11 @@ class DeepseekV4Model(nn.Module):
         because the stacking weight_loader corrupts NVFP4 uint8 data.
         """
         # All attention projections to dequantize to BF16
-        bf16_proj_names = {"wq_a", "wq_b", "wkv", "wo_a", "wo_b"}
+        # wo_a is excluded — it uses fp8_einsum (no input_scale, weight-only FP8)
+        bf16_proj_names = {"wq_a", "wq_b", "wkv", "wo_b"}
+        fp8_proj_names = {"wo_a"}
         bf16_converted = 0
+        fp8_converted = 0
         compressor_converted = 0
 
         _shard_index = self._build_shard_index("/model") if os.path.isdir("/model") else None
@@ -1721,7 +1724,7 @@ class DeepseekV4Model(nn.Module):
         for layer_idx, layer in tqdm(enumerate(self.layers), total=len(self.layers), desc="  (upcast)NVFP4→BF16 attn projs", unit="layer"):
             attn = layer.attn
             
-            # BF16 dequantization: all attention projections
+            # BF16 dequantization: attention projections (except wo_a)
             for proj_name in bf16_proj_names:
                 if not hasattr(attn, proj_name):
                     continue
@@ -1733,6 +1736,19 @@ class DeepseekV4Model(nn.Module):
                     self._dequant_nvfp4_to_bf16(mod, E2M1_LUT)
                     bf16_converted += 1
             
+            # FP8 conversion: wo_a (used by fp8_einsum, no input_scale)
+            FP8_MAX = torch.finfo(torch.float8_e4m3fn).max
+            for proj_name in fp8_proj_names:
+                if not hasattr(attn, proj_name):
+                    continue
+                mod = getattr(attn, proj_name)
+                if not hasattr(mod, "weight"):
+                    continue
+                if mod.weight.dtype in (torch.uint8, torch.int8):
+                    E2M1_LUT = torch.tensor([0, 0.5, 1, 1.5, 2, 3, 4, 6], dtype=torch.bfloat16)
+                    self._convert_nvfp4_to_fp8(mod, E2M1_LUT, FP8_MAX)
+                    fp8_converted += 1
+            
             # Compressor: still needs BF16 reconstruction
             mla_attn = getattr(attn, "mla_attn", None)
             if mla_attn is not None: