[torch nightlies] Use main Dockerfile with flags for nightly torch tests

[orionr]

Use standard Docker image instead of torch_nightly image for PyTorch nightlies testing and CI runs.

Moving this from #239 to a branch on upstream for testing purposes outlined at https://github.com/vllm-project/ci-infra?tab=readme-ov-file#how-to-test-changes-in-this-repo

Tests to confirm:

1. Baseline (my vllm fork matching HEAD, no ci-infra changes) at https://buildkite.com/vllm/ci/builds/42874/steps/canvas. Allowed 5 test runs to move forward. -> Seems like PT nightlies build itself failed on installing flashinfer so all tests failed afterwards.
2. New (my vllm changes at [CI][torch nightlies] Use main Dockerfile with flags for nightly torch tests vllm#30443, my ci-infra changes at [torch nightlies] Use main Dockerfile with flags for nightly torch tests #244) with a successful build at https://buildkite.com/vllm/ci/builds/45736/steps/canvas?sid=019b9459-43ce-46d3-99c2-c10a1a8ce96c. One downstream test is failing, but that looks real and something we will investigate.

We will remove https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile.nightly_torch in a separate commit.

Looking for review and landing with help from @khluu, @amrmahdi, @atalman, @huydhn . Thanks!

[orionr]

@khluu I might need your help on this one and/or have you point me to an expert on Buildkite configs.

I'm trying to use the standard Docker builds here for PyTorch nightly testing, but need to also run uv pip install torch torchvision torchaudio --pre --extra-index-url https://download.pytorch.org/whl/nightly/cu128 (or create a Docker image layer) before each test runs. I thought I'd figured this out by adding an extra commands section, but looks like that might need to get propagated down to and through render_cuda_config. Is that the right way to do this or should I go a different path?

Current status is that the main Docker image is used (which is good), but tests are all running on release PyTorch versions (not good) without the latest changes.

Latest failing run is at https://buildkite.com/vllm/ci/builds/42927/steps/canvas?sid=019b0a30-bee1-4b6b-8393-7f85b537d2ef with the error

[2025-12-10T22:21:19Z] public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:2dcbac9077ecadff0aa78b7c282f9e147a260e86
--
Error: Can't use both a step level command and the command parameter of the plugin

because of e596c0d#diff-b5c060fa4acd68fd48a2b3cdcd4069bd9eae5b0ee8512e1b25d8f8e2526834e5R480

Any thoughts? cc @atalman as well and I'll keep digging.

[huydhn]

@khluu I might need your help on this one and/or have you point me to an expert on Buildkite configs.

I'm trying to use the standard Docker builds here for PyTorch nightly testing, but need to also run uv pip install torch torchvision torchaudio --pre --extra-index-url https://download.pytorch.org/whl/nightly/cu128 (or create a Docker image layer) before each test runs. I thought I'd figured this out by adding an extra commands section, but looks like that might need to get propagated down to and through render_cuda_config. Is that the right way to do this or should I go a different path?

Current status is that the main Docker image is used (which is good), but tests are all running on release PyTorch versions (not good) without the latest changes.

Latest failing run is at https://buildkite.com/vllm/ci/builds/42927/steps/canvas?sid=019b0a30-bee1-4b6b-8393-7f85b537d2ef with the error

[2025-12-10T22:21:19Z] public.ecr.aws/q9t5s3a7/vllm-ci-test-repo:2dcbac9077ecadff0aa78b7c282f9e147a260e86
--
Error: Can't use both a step level command and the command parameter of the plugin

because of e596c0d#diff-b5c060fa4acd68fd48a2b3cdcd4069bd9eae5b0ee8512e1b25d8f8e2526834e5R480

Any thoughts? cc @atalman as well and I'll keep digging.

I think the uv pip install torch --pre --extra-index-url https://download.pytorch.org/whl/nightly/cu12x could only be done as a Docker layer inside https://github.com/vllm-project/vllm/blob/main/docker/Dockerfile#L139-L143. Something likes:

if NIGHTLY == 1:
   uv pip install torch --pre --extra-index-url ${PYTORCH_CUDA_NIGHTLY_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')
   python use_existing_torch
else:
    uv pip install -r requirements/cuda.txt --extra-index-url ${PYTORCH_CUDA_INDEX_BASE_URL}/cu$(echo $CUDA_VERSION | cut -d. -f1,2 | tr -d '.')

[orionr]

Good call on needing build as well as test signal. Let me see what I can do to modify the base Dockerfile.

[orionr]

Code changes should be ready for review after the final Buildkite test runs. And now done.

[orionr]

Note that this matches standard Dockerfile build flags both here and down below. The key is that vLLM builds with PT nightlies and standard vLLM builds should be identical here minus the --build-arg PYTORCH_NIGHTLY=1 flag. Unfortunately we can't unify further yet, but we can do that in some additional commits.

[orionr]

For the main Nvidia GPU build this has now moved to buildkite/scripts/ci-bake.sh, but we should still keep these incremental changes towards what the main build was doing.

[orionr]

@simon-mo and @khluu thank you for merging vllm-project/vllm#30443! This is the other part of it. After this one we can land vllm-project/vllm#32426