[Misc][Model][Refactor] Pass the prefix into Linear layers #28259
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Signed-off-by: MengqingCao <cmq0113@163.com>
mergify
bot
added
deepseek
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Code Review
This pull request is a large-scale refactoring to pass a prefix argument to various linear layers across multiple models. This is a good improvement for code consistency and modularity, especially for weight loading and quantization. The changes are mostly correct, but I've identified two critical copy-paste errors that would break model loading. Please see the detailed comments for fixes.
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
ℹ️ About Codex in GitHub
Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
Contributor
Author
|
CI failed due to llama3.1 + ngram accept rate doesn't reach 66%, seems no related to this pr, is it a known issue on CI? <html>
<body>
<!--StartFragment-->
[2025-11-07T04:19:06Z] =================================== FAILURES ===================================
| [2025-11-07T04:19:06Z] ____________ test_ngram_and_suffix_correctness[speculative_config1] ____________
| [2025-11-07T04:19:06Z]
| [2025-11-07T04:19:06Z] speculative_config = {'method': 'suffix', 'suffix_decoding_max_spec_factor': 2.0, 'target_model_config': ModelConfig(model='meta-llama/Llam...rank=0, _data_parallel_master_port_list=[], decode_context_parallel_size=1, _api_process_count=1, _api_process_rank=0)}
| [2025-11-07T04:19:06Z] monkeypatch = <_pytest.monkeypatch.MonkeyPatch object at 0x7f6ce5646ab0>
| [2025-11-07T04:19:06Z] sampling_config = SamplingParams(n=1, presence_penalty=0.0, frequency_penalty=0.0, repetition_penalty=1.0, temperature=0, top_p=1.0, top...tokens=True, spaces_between_special_tokens=True, truncate_prompt_tokens=None, structured_outputs=None, extra_args=None)
| [2025-11-07T04:19:06Z] model_name = 'meta-llama/Llama-3.1-8B-Instruct'
| [2025-11-07T04:19:06Z]
| [2025-11-07T04:19:06Z] @pytest.mark.parametrize(
| [2025-11-07T04:19:06Z] "speculative_config",
| [2025-11-07T04:19:06Z] [
| [2025-11-07T04:19:06Z] {
| [2025-11-07T04:19:06Z] "method": "ngram",
| [2025-11-07T04:19:06Z] "prompt_lookup_max": 5,
| [2025-11-07T04:19:06Z] "prompt_lookup_min": 3,
| [2025-11-07T04:19:06Z] "num_speculative_tokens": 3,
| [2025-11-07T04:19:06Z] },
| [2025-11-07T04:19:06Z] {
| [2025-11-07T04:19:06Z] "method": "suffix",
| [2025-11-07T04:19:06Z] "suffix_decoding_max_spec_factor": 2.0,
| [2025-11-07T04:19:06Z] },
| [2025-11-07T04:19:06Z] ],
| [2025-11-07T04:19:06Z] )
| [2025-11-07T04:19:06Z] def test_ngram_and_suffix_correctness(
| [2025-11-07T04:19:06Z] speculative_config: dict,
| [2025-11-07T04:19:06Z] monkeypatch: pytest.MonkeyPatch,
| [2025-11-07T04:19:06Z] sampling_config: SamplingParams,
| [2025-11-07T04:19:06Z] model_name: str,
| [2025-11-07T04:19:06Z] ):
| [2025-11-07T04:19:06Z] """
| [2025-11-07T04:19:06Z] Compare the outputs of an original LLM and a speculative LLM
| [2025-11-07T04:19:06Z] should be the same when using ngram speculative decoding.
| [2025-11-07T04:19:06Z] """
| [2025-11-07T04:19:06Z] test_prompts = get_test_prompts(mm_enabled=False)
| [2025-11-07T04:19:06Z]
| [2025-11-07T04:19:06Z] ref_llm = LLM(model=model_name, max_model_len=1024)
| [2025-11-07T04:19:06Z] ref_outputs = ref_llm.chat(test_prompts, sampling_config)
| [2025-11-07T04:19:06Z] del ref_llm
| [2025-11-07T04:19:06Z] torch.cuda.empty_cache()
| [2025-11-07T04:19:06Z] cleanup_dist_env_and_memory()
| [2025-11-07T04:19:06Z]
|
<br class="Apple-interchange-newline"><!--EndFragment-->
</body>
</html> |
Collaborator
|
I also see this failure in other PRs, also cc @DarkLight1337 |
Contributor
Author
|
CI passed now, It seems like that was an occasional issue. @jeejeelee could you help merge this, thx! |
ZhengHongming888
pushed a commit
to ZhengHongming888/vllm
that referenced
this pull request
Nov 8, 2025
…ect#28259) Signed-off-by: MengqingCao <cmq0113@163.com>
xuebwang-amd
pushed a commit
to xuebwang-amd/vllm
that referenced
this pull request
Nov 13, 2025
…ect#28259) Signed-off-by: MengqingCao <cmq0113@163.com> Signed-off-by: xuebwang-amd <xuebwang@amd.com>
2 tasks
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Purpose
Refactor the modeling code: pass the prefix into Linear layers
Essential Elements of an Effective PR Description Checklist
supported_models.mdandexamplesfor a new model.