[FlashInfer] Bump FlashInfer version from 0.6.3 to 0.6.4 (#19005)

.github/workflows/release-docker-cu13-framework.yml

@@ -10,9 +10,9 @@ on:
         description: "Version to build (without v prefix, e.g., 0.5.8)"
         required: true
       flashinfer_version:
-        description: "FlashInfer version (default: 0.6.3)"
+        description: "FlashInfer version (default: 0.6.4)"
         required: false
-        default: "0.6.3"
+        default: "0.6.4"
 
 jobs:
   publish-x86:

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=0.6.3
+ARG FLASHINFER_VERSION=0.6.4
 ARG MOONCAKE_VERSION=0.3.9
 #if need other arg please add in MOONCAKE_COMPILE_ARG
 ARG MOONCAKE_COMPILE_ARG="-DUSE_HTTP=ON -DUSE_MNNVL=ON -DUSE_CUDA=ON -DWITH_EP=ON"

python/pyproject.toml

@@ -27,8 +27,8 @@ dependencies = [
   "datasets",
   "einops",
   "fastapi",
-  "flashinfer_python==0.6.3", # keep it aligned with jit-cache version in Dockerfile
+  "flashinfer_python==0.6.4", # keep it aligned with jit-cache version in Dockerfile
-  "flashinfer_cubin==0.6.3",
+  "flashinfer_cubin==0.6.4",
   "gguf",
   "hf_transfer",
   "huggingface_hub",

python/sglang/srt/entrypoints/engine.py

@@ -879,7 +879,7 @@ def _set_envs_and_config(server_args: ServerArgs):
         if server_args.attention_backend == "flashinfer":
             assert_pkg_version(
                 "flashinfer_python",
-                "0.6.3",
+                "0.6.4",
                 "Please uninstall the old version and "
                 "reinstall the latest version by following the instructions "
                 "at https://docs.flashinfer.ai/installation.html.",

python/sglang/srt/server_args.py

@@ -35,7 +35,6 @@ from sglang.srt.parser.reasoning_parser import ReasoningParser
 from sglang.srt.utils.common import (
     LORA_TARGET_ALL_MODULES,
     SUPPORTED_LORA_TARGET_MODULES,
-    check_pkg_version_at_least,
     configure_ipv6,
     cpu_has_amx_support,
     get_bool_env_var,
@@ -1745,12 +1744,10 @@ class ServerArgs:
                     and self.moe_a2a_backend == "none"
                     and self.moe_runner_backend == "auto"
                 ):
-                    # Only enable flashinfer_trtllm if flashinfer-python version is >= 0.6.3
+                    self.moe_runner_backend = "flashinfer_trtllm"
-                    if check_pkg_version_at_least("flashinfer-python", "0.6.3"):
+                    logger.info(
-                        self.moe_runner_backend = "flashinfer_trtllm"
+                        "Use flashinfer_trtllm as MoE runner backend on sm100 for Glm4MoeForCausalLM"
-                        logger.info(
+                    )
-                            "Use flashinfer_trtllm as MoE runner backend on sm100 for Glm4MoeForCausalLM"
-                        )
 
         elif model_arch in [
             "FalconH1ForCausalLM",

python/sglang/srt/utils/common.py

@@ -1114,7 +1114,7 @@ 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., "0.6.3")
+        min_version: Minimum version required (e.g., "0.6.4")
 
     Returns:
         True if package is installed and version >= min_version, False otherwise

scripts/ci/cuda/ci_install_dependency.sh

@@ -5,7 +5,7 @@ set -euxo pipefail
 # Set up environment variables
 IS_BLACKWELL=${IS_BLACKWELL:-0}
 CU_VERSION="cu129"
-FLASHINFER_VERSION=0.6.3
+FLASHINFER_VERSION=0.6.4
 OPTIONAL_DEPS="${1:-}"
 
 # Detect system architecture