[Bugfix][CI/Build][Hardware][AMD] Fix AMD tests, add HF cache, update CK FA, add partially supported model notes

[mawong-amd]

This PR attempts to fix existing failures on AMD tests.

• Metrics tests are currently failing because of [Core] draft_model_runner: Implement prepare_inputs on GPU for advance_step #6338, which accidentally added a hard requirement for vllm_flash_attn in vllm/spec_decode/draft_model_runner.py. This is not installed on ROCm, so the correct ROCm FA component is used instead.
• Basic correctness tests are currently failing because the test harness unconditionally installs flashinfer in a way that is not supported on ROCm. This installation is made optional: this shouldn't affect the esoteric case where flashinfer isn't correctly installed during CUDA tests, because 1. this will give an obvious error message at the beginning of the test and 2. flashinfer ModuleNotFoundError is also pretty obvious.
• Entrypoints tests are failing because they now depend on peft (due to [CORE] Adding support for insertion of soft-tuned prompts #4645) which is not currently installed in the AMD container. This in turns wants updated botocore and hence boto3 and awscli. These requirements are included in requirements-test.txt but are a small subset of it, so we do not install it for now. Once more AMD tests are enabled, we will begin using requirements-test.txt in earnest.
• Basic correctness tests were also failing because [ci][test] add correctness test for cpu offloading #6549 erroneously introduced a test involving compressed tensors quant on ROCm, where it is not supported.
• Model download sometimes takes an inordinately long time due to a flaky network configuration and contributes to tests (typically Distributed and Vision Language) hitting the 90 min timeout. This PR adds a HuggingFace models cache on AMD machines similar to what was done in Share HuggingFace downloads between test runs buildkite-ci#8. Note that this might not allow flaky tests to pass immediately, but the hope is that once each machine has had a chance to run all tests and incrementally build up its cache, the issue will eventually be resolved.
• Vision Language Models and Distributed tests are also failing because ROCm's Triton FA kernel hangs on a variety of cases, including PaliGemma full-precision tests (added in [Model] Add PaliGemma #5189) and multimodal broadcast with Phi-3-Vision. This is due to issues during Triton compilation which is solved by updating Triton.
• Even after updating Triton, Phi-3-Vision and PaliGemma are still problematic under certain settings because too much shared memory is reserved by the Triton kernel on some devices. As a temporary workaround, CK FA is updated and used instead for the relevant tests and the associated warnings for partial model support in ROCm are added.

---

PR Checklist (Click to Expand)

Thank you for your contribution to vLLM! Before submitting the pull request, please ensure the PR meets the following criteria. This helps vLLM maintain the code quality and improve the efficiency of the review process.

PR Title and Classification

Only specific types of PRs will be reviewed. The PR title is prefixed appropriately to indicate the type of change. Please use one of the following:

• [Bugfix] for bug fixes.
• [CI/Build] for build or continuous integration improvements.
• [Doc] for documentation fixes and improvements.
• [Model] for adding a new model or improving an existing model. Model name should appear in the title.
• [Frontend] For changes on the vLLM frontend (e.g., OpenAI API server, LLM class, etc.)
• [Kernel] for changes affecting CUDA kernels or other compute kernels.
• [Core] for changes in the core vLLM logic (e.g., LLMEngine, AsyncLLMEngine, Scheduler, etc.)
• [Hardware][Vendor] for hardware-specific changes. Vendor name should appear in the prefix (e.g., [Hardware][AMD]).
• [Misc] for PRs that do not fit the above categories. Please use this sparingly.

Note: If the PR spans more than one category, please include all relevant prefixes.

Code Quality

The PR need to meet the following code quality standards:

• We adhere to Google Python style guide and Google C++ style guide.
• Pass all linter checks. Please use format.sh to format your code.
• The code need to be well-documented to ensure future contributors can easily understand the code.
• Include sufficient tests to ensure the project to stay correct and robust. This includes both unit tests and integration tests.
• Please add documentation to docs/source/ if the PR modifies the user-facing behaviors of vLLM. It helps vLLM user understand and utilize the new features or changes.

Notes for Large Changes

Please keep the changes as concise as possible. For major architectural changes (>500 LOC excluding kernel/data/config/test), we would expect a GitHub issue (RFC) discussing the technical design and justification. Otherwise, we will tag it with rfc-required and might not go through the PR.

What to Expect for the Reviews

The goal of the vLLM team is to be a transparent reviewing machine. We would like to make the review process transparent and efficient and make sure no contributor feel confused or frustrated. However, the vLLM team is small, so we need to prioritize some PRs over others. Here is what you can expect from the review process:

• After the PR is submitted, the PR will be assigned to a reviewer. Every reviewer will pick up the PRs based on their expertise and availability.
• After the PR is assigned, the reviewer will provide status update every 2-3 days. If the PR is not reviewed within 7 days, please feel free to ping the reviewer or the vLLM team.
• After the review, the reviewer will put an action-required label on the PR if there are changes required. The contributor should address the comments and ping the reviewer to re-review the PR.
• Please respond to all comments within a reasonable time frame. If a comment isn't clear or you disagree with a suggestion, feel free to ask for clarification or discuss the suggestion.

Thank You

Finally, thank you for taking the time to read these guidelines and for your interest in contributing to vLLM. Your contributions make vLLM a great tool for everyone!

[mawong-amd]

/ready

[WoosukKwon]

This in turns wants updated botocore and hence boto3 and awscli

How is peft related to boto3 and awscli? Could you explain more?

[mawong-amd]

This in turns wants updated botocore and hence boto3 and awscli

How is peft related to boto3 and awscli? Could you explain more?

It's linked via urllib3 to botocore, which is linked to awscli and boto3. As far as I can tell, on the base image we use, awscli boto3 botocore peft is the smallest set to use to get the related dependencies working and reenable any tests we currently support.

[mawong-amd]

Rebased, fixed one more test regression, and is all green except LM Eval Small Models which seems to be broken in main. Ready to ship!