解决gptq的不能开启graph的问题

vllm/_custom_ops.py

@@ -614,33 +614,72 @@ def gptq_gemm(
     #     use_v2_format,
     #     bit,
     # )
-    return quant_ops.gptq_gemm(
+    
+    # return quant_ops.gptq_gemm(
+    #     a, 
+    #     b_q_weight, 
+    #     b_gptq_qzeros, 
+    #     b_gptq_scales,
+    #     b_g_idx, 
+    #     use_exllama, 
+    #     bit,
+    # )
+    return torch.ops.vllm.gptq_gemm_(
         a, 
         b_q_weight, 
         b_gptq_qzeros, 
         b_gptq_scales,
         b_g_idx, 
         use_exllama, 
-        bit)
-
-
-if hasattr(torch.ops._C, "gptq_gemm"):
+        bit,
+    )
 
-    @register_fake("_C::gptq_gemm")
-    def _gptq_gemm_fake(
+def gptq_gemm_(
     a: torch.Tensor, 
     b_q_weight: torch.Tensor,
     b_gptq_qzeros: torch.Tensor, 
     b_gptq_scales: torch.Tensor,
     b_g_idx: torch.Tensor, 
     use_exllama: bool,
-        use_v2_format: bool,
-        bit: int,
-    ) -> torch.Tensor:
-        return torch.empty(
-            (a.size(0), b_q_weight.size(1)), dtype=a.dtype, device=a.device
+    bit: int) -> torch.Tensor:
+    return quant_ops.gptq_gemm(
+        a, 
+        b_q_weight, 
+        b_gptq_qzeros, 
+        b_gptq_scales,
+        b_g_idx, 
+        use_exllama, 
+        bit,
     )
 
+def gptq_gemm_fake_(
+    a: torch.Tensor, 
+    b_q_weight: torch.Tensor,
+    b_gptq_qzeros: torch.Tensor, 
+    b_gptq_scales: torch.Tensor,
+    b_g_idx: torch.Tensor, 
+    use_exllama: bool,
+     bit: int) -> torch.Tensor:
+    return torch.empty((a.shape[0], b_gptq_scales.shape[1]), dtype=a.dtype, device=a.device)
+
+
+# if hasattr(torch.ops._C, "gptq_gemm"):
+
+#     @register_fake("_C::gptq_gemm")
+#     def _gptq_gemm_fake(
+#         a: torch.Tensor,
+#         b_q_weight: torch.Tensor,
+#         b_gptq_qzeros: torch.Tensor,
+#         b_gptq_scales: torch.Tensor,
+#         b_g_idx: torch.Tensor,
+#         use_exllama: bool,
+#         use_v2_format: bool,
+#         bit: int,
+#     ) -> torch.Tensor:
+#         return torch.empty(
+#             (a.size(0), b_q_weight.size(1)), dtype=a.dtype, device=a.device
+#         )
+
 
 def gptq_shuffle(q_weight: torch.Tensor, q_perm: torch.Tensor, bit: int) -> None:
     # torch.ops._C.gptq_shuffle(q_weight, q_perm, bit)
@@ -3352,3 +3391,10 @@ direct_register_custom_op(
     mutates_args=[],
     fake_impl=awq_gemm_fake,
 )
+
+direct_register_custom_op(
+    op_name="gptq_gemm_",
+    op_func=gptq_gemm_,
+    mutates_args=[],
+    fake_impl=gptq_gemm_fake_,
+)