Update Dockerfile for uv and pyproject.toml migration - #260

Merged

natolambert merged 21 commits into

mainfrom

fix/dockerfile-uv-migration

Jan 31, 2026

Commits on Jan 30, 2026

Update Dockerfile for uv and pyproject.toml migration

Nathan Lambert
and
claude
committed
169b955
Add google-generativeai to generative extra

Nathan Lambert
and
claude
committed
2a0b7e9
Migrate generative.py to google-genai SDK

Nathan Lambert
and
claude
committed
4d7e495
Revert "Migrate generative.py to google-genai SDK"
Nathan Lambert
committed
48de371
Pin torch<=2.8 and use prebuilt flash-attn wheel

Nathan Lambert
and
claude
committed
760a269
Revert torch pin, build flash-attn from source

Nathan Lambert
and
claude
committed
6c2b7d4
Remove flash-attn from Docker build

Nathan Lambert
and
claude
committed
1a52d10
Restore flash-attn build from source

Nathan Lambert
and
claude
committed
8943428
Add Linux platform marker to vllm for macOS compatibility

Nathan Lambert
and
claude
committed
ebb9331
Pin vllm on all platforms, add MAX_JOBS to flash-attn build

Nathan Lambert
and
claude
committed
95f957d
Remove MAX_JOBS limit from flash-attn build

Nathan Lambert
and
claude
committed
ada1326
Refactor Dockerfile to use uv sync for better layer caching

Nathan Lambert
and
claude
committed
26314e5
Move flash-attn before source copy + add GHA layer caching

Nathan Lambert
and
claude
committed
66661bc
Split into two images: base (fast) and vllm (slow)

Nathan Lambert
and
claude
committed
2f7f6fa
Update docs for two-image setup and new extras

Nathan Lambert
and
claude
committed
d624a96
Fix uv pip install with --system flag and code formatting

Nathan Lambert
and
claude
committed
90d4a8f
Remove flash-attn from vllm image (vLLM bundles its own)

Nathan Lambert
and
claude
committed
9ac21b8
Update Dockerfile.vllm header comment

Nathan Lambert
and
claude
committed
e8b4cd4
Fix base image: downgrade torch after uv sync, use flash-attn 2.8.3

Nathan Lambert
and
claude
committed
6ae1d61
Remove unused packages: openjdk, git-lfs, GNU parallel

Nathan Lambert
and
claude
committed
ffba9ae
Remove GHA layer caching from Docker builds

Nathan Lambert
and
claude
committed
d9ef779

Navigation Menu

Uh oh!

Update Dockerfile for uv and pyproject.toml migration - #260

Update Dockerfile for uv and pyproject.toml migration#260
natolambert merged 21 commits into
mainfrom
fix/dockerfile-uv-migration

Commits on Jan 30, 2026

Update Dockerfile for uv and pyproject.toml migration

Add google-generativeai to generative extra

Migrate generative.py to google-genai SDK

Revert "Migrate generative.py to google-genai SDK"

Pin torch<=2.8 and use prebuilt flash-attn wheel

Revert torch pin, build flash-attn from source

Remove flash-attn from Docker build

Restore flash-attn build from source

Add Linux platform marker to vllm for macOS compatibility

Pin vllm on all platforms, add MAX_JOBS to flash-attn build

Remove MAX_JOBS limit from flash-attn build

Refactor Dockerfile to use uv sync for better layer caching

Move flash-attn before source copy + add GHA layer caching

Split into two images: base (fast) and vllm (slow)

Update docs for two-image setup and new extras

Fix uv pip install with --system flag and code formatting

Remove flash-attn from vllm image (vLLM bundles its own)

Update Dockerfile.vllm header comment

Fix base image: downgrade torch after uv sync, use flash-attn 2.8.3

Remove unused packages: openjdk, git-lfs, GNU parallel

Remove GHA layer caching from Docker builds