Bump FlashInfer to 0.6.15 and revert regressions (#31502)

Co-authored-by: Brayden Zhong <b8zhong@uwaterloo.ca>

docker/Dockerfile

@@ -19,7 +19,7 @@ ARG PIP_DEFAULT_INDEX
 ARG UBUNTU_MIRROR
 ARG GITHUB_ARTIFACTORY=github.com
 ARG INSTALL_FLASHINFER_JIT_CACHE=0
-ARG FLASHINFER_VERSION=0.6.14
+ARG FLASHINFER_VERSION=0.6.15
 ARG MOONCAKE_VERSION=0.3.11.post1
 ARG MSCCLPP_VERSION=sglang-v0.9.1
 #if need other arg please add in MOONCAKE_COMPILE_ARG

python/pyproject.toml

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

@@ -222,6 +222,17 @@ 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.
 
@@ -237,7 +248,7 @@ def ensure_cutedsl_wrapper(layer: torch.nn.Module) -> None:
         return
 
     try:
-        from flashinfer import CuteDslMoEWrapper
+        from flashinfer import ActivationType, CuteDslMoEWrapper
     except ImportError as e:
         raise ImportError(
             "flashinfer_cutedsl backend requires FlashInfer with CuteDSL support. "
@@ -284,7 +295,9 @@ 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=layer.moe_runner_config.activation,
+            activation_type=_cutedsl_wrapper_activation_type(
+                layer.moe_runner_config.activation, ActivationType
+            ),
         )
 
     w1_alpha, fc2_input_scale, w2_alpha, used_input_scale = (

python/sglang/srt/models/deepseek_v2.py

@@ -459,14 +459,6 @@ 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., "0.6.14")
+        min_version: Minimum version required (e.g., "0.6.15")
 
     Returns:
         True if package is installed and version >= min_version, False otherwise

test/registered/models_e2e/test_dsa_glm52_nvfp4_tp_mtp.py

@@ -21,7 +21,7 @@ class TestGLM52NVFP4TPMTP(
     model = "nvidia/GLM-5.2-NVFP4"
     tp_size = 4
     mem_fraction_static = 0.8
-    bs_1_speed_thres = 250
+    bs_1_speed_thres = 280
     extra_server_args = [
         "--moe-runner-backend",
         "flashinfer_trtllm",