perf: Avoid unnecessary data type conversions for DeepSeek-V3 on Blackwell (#9834)

Signed-off-by: Jinyang Yuan <154768711+jinyangyuan-nvidia@users.noreply.github.com>

python/sglang/srt/entrypoints/engine.py

@@ -655,7 +655,8 @@ def _set_envs_and_config(server_args: ServerArgs):
     os.environ["CUDA_DEVICE_MAX_CONNECTIONS"] = "4"
     os.environ["CUDA_MODULE_LOADING"] = "AUTO"
     # flashinfer uses this environment variable for various kernels from MoE to quant kernels
-    os.environ["TRTLLM_ENABLE_PDL"] = "1"
+    if os.environ.get("TRTLLM_ENABLE_PDL", "1") != "0":
+        os.environ["TRTLLM_ENABLE_PDL"] = "1"
 
     # Can also be passed as argument
     os.environ["SGLANG_RUN_ID"] = (

python/sglang/srt/models/deepseek_v2.py

@@ -67,7 +67,10 @@ from sglang.srt.layers.moe import (
     should_use_flashinfer_cutlass_moe_fp4_allgather,
 )
 from sglang.srt.layers.moe.ep_moe.layer import DeepEPMoE, get_moe_impl_class
-from sglang.srt.layers.moe.fused_moe_triton.layer import FusedMoE
+from sglang.srt.layers.moe.fused_moe_triton.layer import (
+    FusedMoE,
+    _is_fp4_quantization_enabled,
+)
 from sglang.srt.layers.moe.topk import TopK
 from sglang.srt.layers.quantization import deep_gemm_wrapper
 from sglang.srt.layers.quantization.base_config import QuantizationConfig
@@ -299,7 +302,9 @@ class MoEGate(nn.Module):
             and _device_sm >= 90
         ):
             # router gemm output float32
-            logits = dsv3_router_gemm(hidden_states, self.weight)
+            logits = dsv3_router_gemm(
+                hidden_states, self.weight, out_dtype=torch.float32
+            )
         elif _use_aiter_gfx95 and hidden_states.shape[0] <= 256:
             logits = aiter_dsv3_router_gemm(
                 hidden_states, self.weight, gemm_output_zero_allocator
@@ -364,6 +369,9 @@ class DeepseekV2MoE(nn.Module):
             prefix=add_prefix("experts", prefix),
         )
 
+        correction_bias = self.gate.e_score_correction_bias
+        if _is_fp4_quantization_enabled():
+            correction_bias = correction_bias.to(torch.bfloat16)
         self.topk = TopK(
             top_k=config.num_experts_per_tok + self.num_fused_shared_experts,
             renormalize=config.norm_topk_prob,
@@ -371,7 +379,7 @@ class DeepseekV2MoE(nn.Module):
             num_expert_group=config.n_group,
             num_fused_shared_experts=self.num_fused_shared_experts,
             topk_group=config.topk_group,
-            correction_bias=self.gate.e_score_correction_bias,
+            correction_bias=correction_bias,
             routed_scaling_factor=self.routed_scaling_factor,
             apply_routed_scaling_factor_on_output=self.experts.should_fuse_routed_scaling_factor_in_topk(),
             force_topk=quant_config is None,