Fix: add abstract create_weights to CuTeDSLNvfp4LinearMethod

2026-05-18 20:40:48 +00:00
parent 450793311c
commit 6326222d68
1 changed files with 5 additions and 0 deletions
--- a/vllm/cutedsl_quant_method.py
+++ b/vllm/cutedsl_quant_method.py
@@ -131,5 +131,10 @@ class CuTeDSLNvfp4Method(LinearMethodBase):
 class CuTeDSLNvfp4LinearMethod(LinearMethodBase):
    """Forward path: BF16 input → CuTeDSL NVFP4 GEMM → BF16 output."""

+    def create_weights(self, layer, input_size_per_partition,
+                       output_partition_sizes, input_size, output_size,
+                       params_dtype, **extra_weight_attrs):
+        pass
+
    def apply(self, layer, x: torch.Tensor, bias=None) -> torch.Tensor:
        return layer._cutedsl_runner(x)