[5/n]decouple quantization implementation from vLLM dependency (#9454)

2025-08-22 03:52:07 +08:00
parent 78ae175866
commit 9c8e4f69c3
4 changed files with 562 additions and 1 deletions
--- a/python/sglang/srt/layers/quantization/fp8_utils.py
+++ b/python/sglang/srt/layers/quantization/fp8_utils.py
@@ -557,7 +557,10 @@ def apply_fp8_linear(
    # We also don't pad when using torch.compile,
    # as it breaks with dynamic shapes.
    if pad_output is None:
-        pad_output = not get_bool_env_var("SGLANG_ENABLE_TORCH_COMPILE")
+        pad_output = (
+            not get_bool_env_var("SGLANG_ENABLE_TORCH_COMPILE")
+            and not cutlass_fp8_supported
+        )
    output_padding = 17 if pad_output else None

    # View input as 2D matrix for fp8 methods