A high-throughput and memory-efficient inference and serving engine for LLMs
[Bugfix][CPU] Don't build triton-cpu on arm64 release image (#45401)
Signed-off-by: khluu <khluu000@gmail.com>
Kevin H. Luu committed
Jun 12, 2026 at 14:51 UTC
e3e31e54b05391d21a4b492d3bde612f47696975
1 file changed
+11
docker/Dockerfile.cpu
+11
@@ -168,6 +168,12 @@ RUN --mount=type=cache,target=/root/.cache/uv \
168
######################### TRITON-CPU BUILD IMAGE #########################
169
FROM base AS vllm-triton-cpu-build
170
171
+# Support for cross-compilation with x86 ISA including AVX2 and AVX512: docker build --build-arg VLLM_CPU_X86="true" ...
172
+# Re-declared here because this stage is `FROM base` (not `vllm-build`), so it
173
+# does not inherit the ARG/ENV defined there. Without it, the guard below would
174
+# see an empty value and build triton-cpu on non-x86 targets (e.g. arm64).
175
+ARG VLLM_CPU_X86=0
176
+
177
WORKDIR /vllm-workspace
178
179
RUN mkdir dist
@@ -269,6 +275,11 @@ ENV HF_HUB_DOWNLOAD_TIMEOUT 60
275
######################### RELEASE IMAGE #########################
276
FROM base AS vllm-openai
277
278
+# Re-declared here because this stage is `FROM base` (not `vllm-build`), so the
279
+# RUN below that gates the triton-cpu wheel install on $VLLM_CPU_X86 would
280
+# otherwise see an empty value and try to install it on non-x86 targets.
281
+ARG VLLM_CPU_X86=0
282
+
283
WORKDIR /vllm-workspace
284
285
RUN --mount=type=cache,target=/root/.cache/uv \