[Piecewise CUDA Graph] Support W4A8 (#13179)

2025-11-15 19:53:50 -08:00
parent 12c789ebd1
commit f35f7f1245
1 changed files with 4 additions and 0 deletions
--- a/python/sglang/srt/layers/quantization/fp8_kernel.py
+++ b/python/sglang/srt/layers/quantization/fp8_kernel.py
@@ -45,6 +45,10 @@ _use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
 if _is_cuda:
    from sgl_kernel import sgl_per_tensor_quant_fp8, sgl_per_token_quant_fp8

+    @torch.library.register_fake("sgl_kernel::sgl_per_tensor_quant_fp8")
+    def _sgl_per_tensor_quant_fp8(input, output_q, output_s, is_static):
+        return
+
    # Temporary
    try:
        from sgl_kernel import sgl_per_token_group_quant_8bit