[FP8] SM120: route FP8 linear to per-tensor (cudnn/nvjet) instead of channelwise cutlass (#38006)

Co-authored-by: BBuf <bbuf@example.com>

python/sglang/srt/layers/quantization/fp8.py

@@ -87,6 +87,7 @@ from sglang.srt.runtime_context import (
 from sglang.srt.utils import (
     cpu_has_amx_support,
     get_bool_env_var,
+    get_device_capability,
     is_cpu,
     is_cuda,
     is_flashinfer_available,
@@ -911,9 +912,20 @@ class Fp8LinearMethod(LinearMethodBase):
                         layer.input_scale.data, requires_grad=False
                     )
 
+                # On SM120 (Blackwell RTX 50) the per-tensor FP8 path dispatches
+                # to the fast cudnn/nvjet SM120 kernel, which beats the
+                # channelwise cutlass GemmUniversal by ~1.2-2.7x across the
+                # decode/prefill M range (matches vLLM's per-tensor SM120
+                # choice). Requantize per-channel -> per-tensor (max scale)
+                # there instead.
+                use_sm120_fp8_pertensor = (
+                    self.cutlass_fp8_supported
+                    and not self.use_marlin
+                    and get_device_capability()[0] == 12
+                )
                 # cutlass sgl-kernel and marlin only support per-channel scale; aiter supports per-channel scale
                 if (
-                    self.cutlass_fp8_supported
+                    (self.cutlass_fp8_supported and not use_sm120_fp8_pertensor)
                     or self.use_marlin
                     or (_use_aiter and self.use_aiter_fp8_per_token)
                 ):