Skip to content

Create New Base Docker Image for vLLM - #204

Open
AryanAhadinia wants to merge 3 commits into
mainfrom
AA/vllm-new-image
Open

Create New Base Docker Image for vLLM#204
AryanAhadinia wants to merge 3 commits into
mainfrom
AA/vllm-new-image

Conversation

@AryanAhadinia

Copy link
Copy Markdown
Member

No description provided.

@AryanAhadinia AryanAhadinia self-assigned this Aug 11, 2026
@AryanAhadinia AryanAhadinia added the images Docker images used for model launch label Aug 12, 2026
@AryanAhadinia AryanAhadinia linked an issue Aug 12, 2026 that may be closed by this pull request
3 tasks

# libibverbs/libnuma back the RDMA and NUMA paths on Grace nodes; ffmpeg and
# libsndfile1 decode audio/video inputs; curl serves the router health checks.
# hadolint ignore=DL3008

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should be fixed.

Comment on lines +33 to +49
# JIT and compile caches. Left unset they default under $HOME, which is NFS --
# the worst filesystem here for the many small concurrent writes that
# torch.compile and Triton produce. These live on capstor instead, which is
# already mounted above and persists between jobs; node-local /tmp is tmpfs, so
# it would consume node RAM and be discarded at job end, recompiling on every
# launch.
#
# The paths are literal because variable references are not expanded reliably
# on the launch path, and an unexpanded ${SCRATCH} collapses to an unwritable
# root path. The cache is therefore shared across the infra01 group rather than
# per user; the directories carry setgid and group write so members can
# populate them. Entries are keyed by GPU architecture and library versions, so
# different users, models and images coexist without colliding.
#
# TORCHINDUCTOR_CACHE_DIR is deliberately absent: vLLM sets it itself, to a
# subdirectory of VLLM_CACHE_ROOT, so setting it here is silently discarded.
# TRITON_CACHE_DIR and CUDA_CACHE_PATH are not covered by VLLM_CACHE_ROOT.

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

So long comment.

@@ -0,0 +1,61 @@
image = "/capstor/store/cscs/swissai/infra01/container-images/ci/pr-204/vllm_base-{arch}.sqsh"

Copy link
Copy Markdown
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Don't forget to fix the path before merge.

@sonarqubecloud

Copy link
Copy Markdown

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

images Docker images used for model launch

Projects

None yet

Development

Successfully merging this pull request may close these issues.

vLLM Launch Crashes for Multi-node Models

1 participant