Add test case for compiling multiple graphs

[sarckk]

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

Purpose

In some cases, we might want to compile multiple subgraphs, e.g. if some submodules in the model are not traceable with torch dynamo. See #14913 and #19719 for motivating scenarios.

This PR adds a test case showcasing how this can be supported in vLLM for piecewise compilation. This requires several components:

• Each subgraph in the model being compiled needs to be initialized with set_model_tag(tag) to ensure compile cache is not shared between these subgraphs (initially added in [torch.compile] reorganize the cache directory to support compiling multiple models #19064 to support compiling multiple models)
• Pre-allocate static buffers to be used as arg inputs to each compiled subgraph as CUDA Graph expects static tensor addresses between runs
• Added ignore_torch_compile helper decorator in scenarios where a class inherits a parent module that has support_torch_compile decorator applied but this class does not want to compile its forward(). Added test_ignore_torch_compile_decorator that tests this.

Test Plan

pytest tests/compile/piecewise/test_multiple_graphs.py

Test Result

Both newly added unit tests (test_ignore_torch_compile_decorator and test_multi_graph_piecewise_compile) pass

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[sarckk]

cc: @zou3519 @youkaichao

[gemini-code-assist]

Code Review

This pull request introduces a new ignore_torch_compile decorator and adds test cases for compiling multiple graphs. There is a potential issue with the inheritance of the IGNORE_COMPILE_KEY that needs to be addressed.

[zou3519]

Looks reasonable to me. One minor comment

[zou3519]

cc @BoyuanFeng

[sarckk]

@zou3519 I noticed that when I remove this and enable compile caching, the compile for CompiledAttentionTwo fails when running pytest tests/compile/piecewise/test_multiple_graphs.py::test_multi_graph_piecewise_compile. The error is:

E               torch._dynamo.exc.BackendCompilerFailed: backend='<vllm.compilation.backends.VllmBackend object at 0x7f8addaf6cf0>' raised:
E               RuntimeError: vLLM failed to compile the model. The most likely reason for this is that a previous compilation failed, leading to a corrupted compilation artifact. We recommend trying to remove ~/.cache/vllm/torch_compile_cache and try again to see the real issue. 
E               
E               Set TORCHDYNAMO_VERBOSE=1 for the internal stack trace (please do this especially if you're reporting a bug to PyTorch). For even more developer context, set TORCH_LOGS="+dynamo"

I found that when I keep the prefix in inductor compile cache dir here, the test passes. Before this change, inductor cache dir would be /home/yhshin/.cache/vllm/torch_compile_cache/bbbe1d91c9/rank_0_0 for both attn_one and attn_two, but after the change these are separate (i.e. /home/yhshin/.cache/vllm/torch_compile_cache/bbbe1d91c9/rank_0_0/{prefix}).

For compiling multiple graphs in a single model, should we be setting inductor cache dir to separate directories?

[zou3519]

When we're compiling multiple graphs in a single model, each graph should be tagged with a different model tag. That will make it so that their intermediate artifacts are stored in a separate subfolder of the cache directory.

Are you saying that even with the set_model_tags, you need to set VLLM_DISABLE_COMPILE_CACHE=1?

[sarckk]

Are you saying that even with the set_model_tags, you need to set VLLM_DISABLE_COMPILE_CACHE=1?

Yes, and it seems to be due to shared inductor cache dir. If I force it to be stored in separate subfolder according to the model tag, it runs fine.

Having them share inductor cache seems to be a conscious decision in #19064:

This PR re-organizes the cache directory structure, so that the same vLLM instances will use the same TORCHINDUCTOR_CACHE_DIR and TRITON_CACHE_DIR, but just different storage for vllm_compile_cache.py etc.

[BoyuanFeng]

looks good! a few minor comments

[BoyuanFeng]

For test, should we random initialize inputs and check bitwise equivalence?

[sarckk]

Updated to randomize test input each test run. By bitwise equivalence do you mean torch.equal() on final output or something else? Using torch.equal (rtol=0,atol=0) fails currently, but that might be expected?