Enable SGLANG_OPT_FP8_WO_A_GEMM by default (#25181)

python/sglang/srt/environ.py

@@ -619,7 +619,7 @@ class Envs:
     SGLANG_OPT_USE_TOPK_V2 = EnvBool(False)
 
     # GEMM / kernel fusion
-    SGLANG_OPT_FP8_WO_A_GEMM = EnvBool(False)
+    SGLANG_OPT_FP8_WO_A_GEMM = EnvBool(True)
     SGLANG_OPT_BF16_FP32_GEMM_ALGO = EnvStr("cublas")
     SGLANG_OPT_USE_JIT_EP_ACTIVATION = EnvBool(True)
     SGLANG_OPT_USE_JIT_NORM = EnvBool(False)

python/sglang/srt/models/deepseek_v4.py

@@ -1095,7 +1095,10 @@ class DeepseekV4ForCausalLM(nn.Module):
     def _setup_fp8_wo_a_scales(self, is_nextn: bool) -> None:
         from deep_gemm import transform_sf_into_required_layout
 
-        layers = self.model.layers
+        if is_nextn:
+            layers = [self.model.decoder]
+        else:
+            layers = self.model.layers
         for layer in layers:
             attn = layer.self_attn
             G = attn.n_local_groups

python/sglang/srt/server_args.py

@@ -4259,6 +4259,15 @@ class ServerArgs:
                     "Debug mode for CUDA graph is enabled via breakable CUDA graph. "
                     "All operations will run eagerly through the graph capture/replay path."
                 )
+        # FP8 W_o GEMM requires Blackwell (sm100+). Auto-disable on Hopper.
+        if is_cuda() and envs.SGLANG_OPT_FP8_WO_A_GEMM.get() and get_device_sm() < 100:
+            if envs.SGLANG_OPT_FP8_WO_A_GEMM.is_set():
+                logger.warning(
+                    "Disabling SGLANG_OPT_FP8_WO_A_GEMM: requires sm100+ (Blackwell), "
+                    "detected sm%d.",
+                    get_device_sm(),
+                )
+            envs.SGLANG_OPT_FP8_WO_A_GEMM.set(False)
 
     def _handle_cache_compatibility(self):
         if self.enable_hierarchical_cache and self.disable_radix_cache: