[WIP] Run eagle with full cudagraph #20190
Conversation
zixi-qi
force-pushed
the
eagle-full-cudagraph
branch
from
e9096c5 to
7ca19b0
Compare
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Summary of Changes
Hello @zixi-qi, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!
This pull request aims to enable the full CUDA graph for the Eagle speculative decoding in vLLM. It introduces a new command-line argument for specifying compilation configurations and modifies the dummy run methods to accommodate attention metadata. The PR is marked as a work in progress, as there's a numerical gap observed when the full CUDA graph is enabled.
Highlights
- Eagle Speculative Decoding: This PR introduces changes to support running v1 Eagle speculative decoding with full CUDA graph.
- Command-line Argument: Added a
--compilation_configargument toexamples/offline_inference/eagle.pyto allow specifying compilation configurations, including enabling full CUDA graph. - Dummy Run: Modified the
dummy_runmethod invllm/v1/spec_decode/eagle.pyto acceptattn_metadata. - Drafter Dummy Run: Modified the
_dummy_runmethod invllm/v1/worker/gpu_model_runner.pyto passattn_metadatatoself.drafter.dummy_run.
Using Gemini Code Assist
The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.
Invoking Gemini
You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.
| Feature | Command | Description |
|---|---|---|
| Code Review | /gemini review |
Performs a code review for the current pull request in its current state. |
| Pull Request Summary | /gemini summary |
Provides a summary of the current pull request in its current state. |
| Comment | @gemini-code-assist | Responds in comments when explicitly tagged, both in issue comments and review comments. |
| Help | /gemini help |
Displays a list of available commands. |
Customization
To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.
Limitations & Feedback
Gemini Code Assist is currently in preview and may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments to provide feedback.
You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.
Footnotes
-
Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩
There was a problem hiding this comment.
Code Review
This pull request enables full CUDA graph support for Eagle speculative decoding by plumbing attn_metadata to the drafter's dummy_run method during CUDA graph capture. This ensures the model's state is consistent between eager execution and a graphed run, resolving numerical discrepancies. The changes in vllm/v1/spec_decode/eagle.py and vllm/v1/worker/gpu_model_runner.py are well-targeted, and the modifications to the example script in examples/offline_inference/eagle.py are appropriate for testing this new functionality.
| compilation_config=( | ||
| json.loads(args.compilation_config) if args.compilation_config else None | ||
| ), |
There was a problem hiding this comment.
The direct call to json.loads can cause the script to crash with a json.JSONDecodeError if an invalid JSON string is passed to the --compilation_config argument. Consider adding a try-except block to handle potential parsing errors gracefully.
compilation_config = None
if args.compilation_config:
try:
compilation_config = json.loads(args.compilation_config)
except json.JSONDecodeError as e:
raise ValueError(f"Invalid JSON for --compilation_config: {e}") from e
|
👋 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 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 🚀 |
I think most probably there is something wrong with how draft model's attention metadata is being captured and updated during replay here. I have tried to do some manipulations but no luck so far |
zixi-qi
force-pushed
the
eagle-full-cudagraph
branch
from
7ca19b0 to
3dda3b3
Compare
|
This pull request has merge conflicts that must be resolved before it can be |
vllm/v1/spec_decode/eagle.py
Outdated
| @@ -169,7 +171,7 @@ def propose( | |||
| self.positions[:num_tokens] = target_positions | |||
| self.hidden_states[:num_tokens] = target_hidden_states | |||
|
|
|||
| with set_forward_context(per_layer_attn_metadata, | |||
| with set_forward_context(None, | |||
There was a problem hiding this comment.
AFAIK, in full cudagraph mode, the attention metdata passed into context manager should not have an impact on the results since no python code is executed. However changing the code here does have an impact:
- with set_forward_context(per_layer_attn_metadata...
--------------------------------------------------
mean acceptance length: 2.08
--------------------------------------------------
acceptance at token 0:0.63
acceptance at token 1:0.28
acceptance at token 2:0.11
acceptance at token 3:0.05
acceptance at token 4:0.02
acceptance at token 5:0.00
acceptance at token 6:0.00
- with set_forward_context(None...
--------------------------------------------------
mean acceptance length: 1.53
--------------------------------------------------
acceptance at token 0:0.40
acceptance at token 1:0.12
acceptance at token 2:0.01
acceptance at token 3:0.00
acceptance at token 4:0.00
acceptance at token 5:0.00
acceptance at token 6:0.00
Repro command:
VLLM_LOGGING_LEVEL=DEBUG VLLM_USE_V1=1 python examples/offline_inference/eagle.py --num_spec_tokens 7 --max_num_seqs 1 --num_prompts 1 --compilation_config '{"full_cuda_graph": true, "cudagraph_capture_sizes": [1]}'
This seems a bit unexpected, wondering if @zou3519 may have any insights?
There was a problem hiding this comment.
it does affect things when cudagraph is being captured, right?
There was a problem hiding this comment.
I think capture is supposed to happen inside the dummy_run function instead of here though?
There was a problem hiding this comment.
Some initial thoughts, if it helps. I'm still digging into this.
- full cuda graphs only applies to the nn.Module that is decorated with support_torch_compile. In eagle, the eagle head gets decorated with support_torch_compile (somewhere the self.model invokes a model, e.g. LlamaModel). The line of code we are commenting on always runs in Python, even when cudagraphs has been recorded.
There was a problem hiding this comment.
Also dug more and found that this line is the "prefill" section for the draft model, where num_tokens can be greater than max configured batch size for cudagraph capture. In this case, the model seems to be running in eager mode which utilizes the passed-in attn_metadata instead of the captured one, which is probably the expected behavior?
There was a problem hiding this comment.
Offline Zixi and I tried some things:
- turn off inductor
- turn off vllm_compile_cache
- check for dynamic shape issues
All of the above seemed to not do anything, so this is still probably a CUDAGraphs issue
| @@ -1982,7 +1982,7 @@ def _dummy_run( | |||
|
|
|||
| if self.speculative_config and self.speculative_config.use_eagle(): | |||
| assert isinstance(self.drafter, EagleProposer) | |||
| self.drafter.dummy_run(num_tokens) | |||
| self.drafter.dummy_run(num_tokens, attn_metadata) | |||
There was a problem hiding this comment.
Here's my hypothesis:
- the attn_metadata contains tensors
- cudagraphs is baking in the addresses of those tensors
- during runtime, the captured cudagraphs still read from these tensors.
Does the eagle forward pass use the tensors in the attn_metadata? If so, every time we invoke the eagle head, we may need to copy data into the tensors in the attn_metadata.
There was a problem hiding this comment.
You are right this is partially the reason for the numerical gap. As an experiment I copied over the attn_metadata constructed for eager mode into the captured attn_metadata in latest commit:
# copy attention metadata for full cudagraph mode
if self.draft_attn_metadata is not None:
self.draft_attn_metadata.seq_lens[:attn_metadata.seq_lens.shape[0]].copy_(attn_metadata.seq_lens.clone())
self.draft_attn_metadata.slot_mapping[:attn_metadata.slot_mapping.shape[0]].copy_(attn_metadata.slot_mapping.clone())
self.draft_attn_metadata.query_start_loc[:attn_metadata.query_start_loc.shape[0]].copy_(attn_metadata.query_start_loc.clone())
self.draft_attn_metadata.block_table[:attn_metadata.block_table.shape[0]].copy_(attn_metadata.block_table.clone())
As a result, I got better numerics but there is still a gap comparing with piecewise mode:
- VLLM_USE_V1=1 python examples/offline_inference/eagle.py --num_spec_tokens 7 --num_prompts 1 --compilation_config '{"full_cuda_graph": true, "cudagraph_capture_sizes": [1]}'
--------------------------------------------------
mean acceptance length: 2.46
--------------------------------------------------
acceptance at token 0:0.69
acceptance at token 1:0.38
acceptance at token 2:0.20
acceptance at token 3:0.12
acceptance at token 4:0.06
acceptance at token 5:0.00
acceptance at token 6:0.00
- VLLM_USE_V1=1 python examples/offline_inference/eagle.py --num_spec_tokens 7 --num_prompts 1 --compilation_config '{"full_cuda_graph": false, "cudagraph_capture_sizes": [1]}'
--------------------------------------------------
mean acceptance length: 2.82
--------------------------------------------------
acceptance at token 0:0.77
acceptance at token 1:0.51
acceptance at token 2:0.28
acceptance at token 3:0.13
acceptance at token 4:0.05
acceptance at token 5:0.03
acceptance at token 6:0.03
There was a problem hiding this comment.
So it seems there might still be some discrepancy in attention computation between eager mode and cudagraph mode. Will try to investigate more and would also appreciate if you have any suggestions to check from torch.compile perspective
zixi-qi
force-pushed
the
eagle-full-cudagraph
branch
from
3dda3b3 to
74ae072
Compare
Signed-off-by: qizixi <qizixi@meta.com>
zixi-qi
force-pushed
the
eagle-full-cudagraph
branch
from
74ae072 to
b5adae6
Compare
WIP change to support running v1 eagle speculative decoding with full cudagraph. Currently there is a numerical gap when full cudagraph is turned on: