[Flashinfer] Upgrade Flashinfer version to 0.6.18 (#54313)
Signed-off-by: wzhao18 <wzhao18.sz@gmail.com> Signed-off-by: Wei Zhao <51183510+wzhao18@users.noreply.github.com> Co-authored-by: Jee Jee Li <pandaleefree@gmail.com> Co-authored-by: Yongye Zhu <zyy1102000@gmail.com> Co-authored-by: Jiangyun Zhu <riverclouds.zhu@qq.com>
Wei Zhao committed
Aug 30, 2026 at 19:01 UTC
7ab29234890b29b005e46b53037309e597425095
5 files changed
+7
-5
docker/Dockerfile
+1
-1
@@ -943,7 +943,7 @@ RUN --mount=type=cache,target=/opt/uv/cache \
943
# Install FlashInfer JIT cache (requires CUDA-version-specific index URL)
944
# https://docs.flashinfer.ai/installation.html
945
# From versions.json: .flashinfer.version
946
-ARG FLASHINFER_VERSION=0.6.17
946
+ARG FLASHINFER_VERSION=0.6.18
947
RUN --mount=type=cache,target=/opt/uv/cache \
948
if [ "${INSTALL_RUBIN_PRERELEASE}" != "true" ]; then \
949
uv pip install --system flashinfer-jit-cache==${FLASHINFER_VERSION} \
docker/versions.json
+1
-1
@@ -71,7 +71,7 @@
71
"default": "true"
72
},
73
"FLASHINFER_VERSION": {
74
- "default": "0.6.17"
74
+ "default": "0.6.18"
75
},
76
"GDRCOPY_CUDA_VERSION": {
77
"default": "12.8"
requirements/cuda.txt
+2
-2
@@ -14,8 +14,8 @@ PyNvVideoCodec==2.0.4
14
# flashinfer-cubin is not on PyPI since 0.6.14; setup.py excludes it from
15
# install_requires so the published wheel does not carry an unresolvable pin
16
--extra-index-url https://flashinfer.ai/whl/
17
-flashinfer-python==0.6.17
18
-flashinfer-cubin==0.6.17
17
+flashinfer-python==0.6.18
18
+flashinfer-cubin==0.6.18
19
apache-tvm-ffi==0.1.11
20
tilelang==0.1.12
21
nvidia-cudnn-frontend>=1.19.1
tests/evals/gpt_oss/configs/gpt-oss-20b-flashinfer-mxfp4-bf16-cutlass.yaml
+1
-1
@@ -3,4 +3,4 @@
3
model_name: "openai/gpt-oss-20b"
4
metric_threshold: 0.568
5
reasoning_effort: "low"
6
-server_args: "--tensor-parallel-size 2 --moe-backend flashinfer_cutlass"
6
+server_args: "--tensor-parallel-size 2 --moe-backend flashinfer_cutlass --gpu-memory-utilization 0.85"
vllm/model_executor/layers/mamba/gdn/qwen_gdn_linear_attn.py
+2
@@ -197,6 +197,8 @@ def fi_chunk_gated_delta_rule(
197
fi_state = initial_state.to(torch.float32)
198
fi_g = g.to(torch.float32)
199
fi_beta = beta.to(torch.float32)
200
+ if cu_seqlens is not None:
201
+ cu_seqlens = cu_seqlens.to(torch.int64)
202
result = chunk_gated_delta_rule_fi(
203
q=q,
204
k=k,