Revert "Bump FlashInfer to 0.6.15 and revert regressions" (#31625)

python/pyproject.toml

@@ -31,7 +31,7 @@ dependencies = [
   "einops",
   "fastapi",
   "flash-attn-4==4.0.0b15",
-  "flashinfer_python[cu13]==0.6.15", # keep it aligned with jit-cache version in Dockerfile
+  "flashinfer_python[cu13]==0.6.14", # keep it aligned with jit-cache version in Dockerfile
   "gguf",
   "humming-kernels[cu13]==0.1.10",
   "interegular",

python/sglang/srt/entrypoints/engine.py

@@ -1304,7 +1304,7 @@ def _set_envs_and_config(server_args: ServerArgs):
         if server_args.attention_backend == "flashinfer":
             assert_pkg_version(
                 "flashinfer_python",
-                "0.6.15",
+                "0.6.14",
                 "Please uninstall the old version and "
                 "reinstall the latest version by following the instructions "
                 "at https://docs.flashinfer.ai/installation.html.",

python/sglang/srt/layers/moe/moe_runner/flashinfer_cutedsl.py

@@ -222,17 +222,6 @@ def resolve_cutedsl_standard_scales(
     return w1_alpha, fc2_input_scale, w2_alpha, used_input_scale
 
 
-def _cutedsl_wrapper_activation_type(activation: str, activation_type_cls: Any) -> Any:
-    if activation == "silu":
-        return activation_type_cls.Swiglu
-    if activation == "relu2":
-        return activation_type_cls.Relu2
-    raise ValueError(
-        f"CuteDSL MoE wrapper supports 'silu' (gated) or 'relu2' (non-gated) "
-        f"activation, got {activation!r}."
-    )
-
-
 def ensure_cutedsl_wrapper(layer: torch.nn.Module) -> None:
     """Lazily create CuteDslMoEWrapper and resolve scales on first forward.
 
@@ -248,7 +237,7 @@ def ensure_cutedsl_wrapper(layer: torch.nn.Module) -> None:
         return
 
     try:
-        from flashinfer import ActivationType, CuteDslMoEWrapper
+        from flashinfer import CuteDslMoEWrapper
     except ImportError as e:
         raise ImportError(
             "flashinfer_cutedsl backend requires FlashInfer with CuteDSL support. "
@@ -295,9 +284,7 @@ def ensure_cutedsl_wrapper(layer: torch.nn.Module) -> None:
             local_expert_offset=layer.moe_ep_rank * layer.num_local_experts,
             output_dtype=layer.moe_runner_config.params_dtype,
             device=str(layer.w13_weight.device),
-            activation_type=_cutedsl_wrapper_activation_type(
-                layer.moe_runner_config.activation, ActivationType
-            ),
+            activation=layer.moe_runner_config.activation,
         )
 
     w1_alpha, fc2_input_scale, w2_alpha, used_input_scale = (

python/sglang/srt/models/deepseek_v2.py

@@ -459,6 +459,14 @@ class MoEGate(nn.Module):
                     "quark",
                 ):
                     correction_bias_dtype = torch.bfloat16
+                # NOTE(kpham-sgl): flashinfer trtllm routing requires a bf16
+                # routing_bias; an fp32 bias yields NaN routing on exact ties.
+                # Mirror the fp8 path's cast.
+                if (
+                    quant_config.get_name() == "modelopt_fp4"
+                    and get_moe_runner_backend().is_flashinfer_trtllm()
+                ):
+                    correction_bias_dtype = torch.bfloat16
             self.e_score_correction_bias = nn.Parameter(
                 torch.empty((config.n_routed_experts), dtype=correction_bias_dtype)
             )

python/sglang/srt/utils/common.py

@@ -1907,7 +1907,7 @@ def check_pkg_version_at_least(pkg: str, min_version: str) -> bool:
 
     Args:
         pkg: Package name (distribution name, e.g., "flashinfer-python")
-        min_version: Minimum version required (e.g., "0.6.15")
+        min_version: Minimum version required (e.g., "0.6.14")
 
     Returns:
         True if package is installed and version >= min_version, False otherwise