Bump to v0.9.1 build (#15)

Author: Apsu

.github/workflows/build-vllm.yaml

@@ -6,8 +6,8 @@ env:
   VLLM_FA_CMAKE_GPU_ARCHES: 90a-real
   FLASHINFER_REF: v0.2.6.post1
   FLASHINFER_BUILD_SUFFIX: cu128
-  VLLM_REF: v0.9.0.1
-  VLLM_BUILD_VERSION: 0.9.0.1
+  VLLM_REF: v0.9.1
+  VLLM_BUILD_VERSION: 0.9.1
   XFORMERS_REF: v0.0.30
   XFORMERS_BUILD_VERSION: 0.0.30+cu128
 

Dockerfile

@@ -85,8 +85,8 @@ RUN cd flashinfer && \
     python -m build -v --wheel --no-isolation -o /wheels
 
 FROM build-base AS build-vllm
-ARG VLLM_REF=v0.9.0.1
-ARG VLLM_BUILD_VERSION=0.9.0.1
+ARG VLLM_REF=v0.9.1
+ARG VLLM_BUILD_VERSION=0.9.1
 ENV BUILD_VERSION=${VLLM_BUILD_VERSION:-${VLLM_REF#v}}
 ENV SETUPTOOLS_SCM_PRETEND_VERSION=${BUILD_VERSION:-:}
 RUN git clone https://github.com/vllm-project/vllm.git