[Speculative decoding] [Multi-Step] decouple should_modify_greedy_probs_inplace #6971
Conversation
|
👋 Hi! Thank you for contributing to the vLLM project. Once the PR is approved and ready to go, please make sure to run full CI as it is required to merge (or just use auto-merge). To run full CI, you can do one of these:
🚀 |
|
Will take a look tomorrow. |
There was a problem hiding this comment.
LGTM.
- Can we add a test for this? e.g. something that runs
include_gpu_probs_tensorbut without modified greedy probs, and we verify that the gpu probs tensor is there but the greedy probs are not modified. - The modification of greedy probs should not incur a CPU sync 😄 . I'll take a look at that.
| (self.scorer_worker.model_runner.model.sampler. | ||
| should_modify_greedy_probs_inplace) = True |
There was a problem hiding this comment.
Can we put this behind an interface so refactors to the worker / model / sampler do spread everywhere?
There was a problem hiding this comment.
Not sure if there is a clean way to do this, as the line above sets include_gpu_probs_tensor before this PR also change the decoupled (this PR) should_modify_greedy_probs_inplace. Seems the scorer worker is passed around as WorkerBase so would need to add change that class which is not ideal. Am I missing something?
There was a problem hiding this comment.
Oh right.. sorry I must have misread.
|
By the way, what model is running in the profile ? pretty surprised by any CPU sync causing 17ms of overhead when the sampler already does a CPU sync at the beginning |
|
This is llama 8B on A10G, but the profile is with multi-step, which has removed all the other sources of CPU syncs (CPU prepare_input, GPU<>CPU transfer for sampled token, pythonization) |
| (self.scorer_worker.model_runner.model.sampler. | ||
| should_modify_greedy_probs_inplace) = True |
There was a problem hiding this comment.
Oh right.. sorry I must have misread.
| @@ -1067,6 +1067,10 @@ def org_vocab_size(self): | |||
| def include_gpu_probs_tensor(self): | |||
| return self.base_layer.include_gpu_probs_tensor | |||
|
|
|||
| @property | |||
| def should_modify_greedy_probs_inplace(self): | |||
There was a problem hiding this comment.
Confused why this is here, why does this layer need it?
There was a problem hiding this comment.
I'm not sure, I just greped for all the places that include_gpu_probs_tensor was used and added the change there to keep semantics identical. See the lines above.
|
Just added test for the inplace modification, let me know if there's anything else. |
github-actions
bot
added
the
ready
Preparation PR for multi step. Not a blocker, but will make MS PR smaller and improve perf.
Decouple
should_modify_greedy_probs_inplacefrominclude_gpu_probs_tensorso that multi-step can setinclude_gpu_probs_tensorwithout also settingshould_modify_greedy_probs_inplaceand incurring the overhead of the probs modification (causes a GPU<>CPU sync). Not a blocker for multi-step, but does add ~1ms of GPU bubble between each step on A10G, will be a much bigger slow down on H100.@cadedaniel This may be a perf bug for spec decode as well?
I also can't seem to get spec_decode tests to all consistently pass locally?Torch Profile with multi-step and without decoupling (
should_modify_greedy_probs_inplace == True)cc @WoosukKwon @zhuohan123 @Yard1 @comaniac @rkooo567
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,LLMclass, 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:
format.shto format your code.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-requiredand 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:
action-requiredlabel on the PR if there are changes required. The contributor should address the comments and ping the reviewer to re-review the PR.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!