[ BugFix ] Prompt Logprobs Detokenization

[robertgshaw2-redhat]

SUMMARY:

• Issue introduced by 603ad84#diff-9d1cd5050a7ec1e588cae646f3f95ca134cffbd8b41d6655c6a14de70117b869, where in rare cases detokenization of prompt logprobs can become broken
• Fixes issues LI was having with runtime on ARC in lm-eval-harness
• We had a hard to track down bug where prompt_logprob detokenization could be performed incorrectly due to an off-by-one-ish error in the creation of prompt_logprobs in sampler
• This has impacted vLLM v0.4.2-v0.5.1

BEFORE FIX:

from vllm import LLM, SamplingParams

PROMPT_TOKENS = [[128000, 128006, 882, 128007, 271, 14924, 25, 16299, 961, 315, 459, 19670, 76854, 279]]
PARAMS = SamplingParams(temperature=0, prompt_logprobs=1, max_tokens=1)

model = LLM("meta-llama/Meta-Llama-3-8B-Instruct", enforce_eager=True)
outputs = model.generate(prompt_token_ids=PROMPT_TOKENS, sampling_params=PARAMS)

print(outputs[0].prompt_logprobs[-1])
# >> {279: Logprob(logprob=-0.09461420774459839, rank=1, decoded_token='<|start_header_id|><|start_header_id|><|start_header_id|><|start_header_id|><|start_header_id|><|start_header_id|><|start_header_id|><|start_header_id|><|start_header_id|><|start_header_id|> the')}

AFTER FIX:

from vllm import LLM, SamplingParams

PROMPT_TOKENS = [[128000, 128006, 882, 128007, 271, 14924, 25, 16299, 961, 315, 459, 19670, 76854, 279]]
PARAMS = SamplingParams(temperature=0, prompt_logprobs=1, max_tokens=1)

model = LLM("meta-llama/Meta-Llama-3-8B-Instruct", enforce_eager=True)
outputs = model.generate(prompt_token_ids=PROMPT_TOKENS, sampling_params=PARAMS)

print(outputs[0].prompt_logprobs[-1])
# >> {279: Logprob(logprob=-0.09461420774459839, rank=1, decoded_token=' the')}

FIX #4904
FIX #4772
FIX #5334
FIX #5872

BEFORE SUBMITTING, PLEASE READ THE CHECKLIST BELOW AND FILL IN THE DESCRIPTION ABOVE

---

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!

[robertgshaw2-redhat]

The first token should have prompt_logprobs=None

[rkooo567]

I think this is not going to work with chunked prefill because the second chunk won't have None for the first prompt logprob

[njhill]

Nice find @robertgshaw2-neuralmagic

[Yard1]

Can we add a test to prevent regressions?

[zifeitong]

Can you try some of the test cases in #5846, #5872, w/ and w/o chunked prefill ?

Additionally, you should be able mark #4904, #4772, #5334, #5872 as fixed.

[robertgshaw2-redhat]

Can you try some of the test cases in #5846, #5872, w/ and w/o chunked prefill ?

Additionally, you should be able mark #4904, #4772, #5334, #5872 as fixed.

Thanks for bringing these to my attention. I will certainly try out the test cases

[simon-mo]

sampler test broke

[robertgshaw2-redhat]

sampler test broke

Yup, its due to chunked_prefill. Im fixing it. @simon-mo @njhill @Yard1 will need re-review

[rkooo567]

Can we have a regression test? Also I have impression the current fix won't work with chunked prefill (mainly because second chunk won't have None for the first prompt logprob)

[robertgshaw2-redhat]

Can we have a regression test? Also I have impression the current fix won't work with chunked prefill (mainly because second chunk won't have None for the first prompt logprob)

Yup, im fixing this right now

[robertgshaw2-redhat]

Okay, chunked prefill needed more fixes than I expected. I had to back my changes out the sampler because it required poking around too much in the sequence_data to detect if we were at the start of a sequence or not which was breaking the abstractions

• I think that prepending [None] makes sense to do in the process_prompt_logprobs function.
• I also had to update the decode_prompt_logprobs_inplace to handle detokenization of the second+ prompt

I added a regression test which tests the reconstructed prompts from the detokenized logprobs matches the original prompts with and without chunked prefill

[Yard1]

Suggested change

	def test_decode_logprobs_regression(
	def test_decode_logprobs_chunked_prefill(

[mgoin]

Really tricky to track down, nice job isolating this

[zifeitong]

Thanks for the fix!

About the CI OOM issue, I am not sure if it's the best workaround, but wait_for_gpu_memory_to_clear has been helpful.