Fix post-quant hook: register on inner model, fix module refs

vLLM V1 calls DeepseekV4Model.forward() directly, not
DeepseekV4ForCausalLM.forward(). Hook on the outer model never fires.
Moved hook to self.model (inner) and fixed module.model.layers →
module.layers.

vllm/patches/deepseek_v4.py

@@ -2438,7 +2438,7 @@ class DeepseekV4ForCausalLM(nn.Module):
             
             E2M1_LUT = torch.tensor([0, 0.5, 1, 1.5, 2, 3, 4, 6], dtype=torch.bfloat16)
             fixed = 0
-            for layer_idx, layer in enumerate(module.model.layers):
+            for layer_idx, layer in enumerate(module.layers):
                 attn = layer.attn
                 for proj_name in ["fused_wqa_wkv", "wq_b", "wo_b"]:
                     if not hasattr(attn, proj_name):
@@ -2447,7 +2447,7 @@ class DeepseekV4ForCausalLM(nn.Module):
                     if not hasattr(mod, "weight") or mod.weight.dtype not in (torch.uint8, torch.int8):
                         continue
                     # Dequantize to BF16
-                    module.model._dequant_nvfp4_to_bf16(mod, E2M1_LUT)
+                    module._dequant_nvfp4_to_bf16(mod, E2M1_LUT)
                     # Replace quant method (quant method already ran, won't overwrite again)
                     mod.quant_method = UnquantizedLinearMethod()
                     # Clean up NVFP4 attributes that might confuse forward
@@ -2468,6 +2468,6 @@ class DeepseekV4ForCausalLM(nn.Module):
             del module._post_quant_fix_handle
             print("  [CLAWMINE] Post-quant-init fix done ✓", flush=True)
         
-        handle = self.register_forward_pre_hook(_fix_attn_nvfp4)
+        handle = self.model.register_forward_pre_hook(_fix_attn_nvfp4)
         self._post_quant_fix_handle = handle