[CORE] Quantized lm-head Framework

[Qubitium]

MOTIVATION

• lm-head and be quantized with minimal loss in accuracy
• models like llama-3 have very large vocab sizes

SUMMARY:

• support quantized lm head (for models without tied word embeddings)
• update autogptq integration to enable quantized lm-head
• added tests for ~all models that can fit in our L4 CI to confirm this does not break anything

IMPLEMENTATION:

• refactored VocabParallelEmbedding to use create_weights to create the parameters
• refactored ParallelLMHead to use apply() to generate output
• plumb changes through all models
• plumb quant_config into ParallelLMHead
• updated AutoGPTQ integration to use lm-head if detected in config

FOLLOW UPS:

• support for fp8 and int8
• look into quantized embeddings

TEST MODEL: (quantized by auto-round and load tested with autogptq):

• https://huggingface.co/LnL-AI/TinyLlama-1.1B-intermediate-step-1341k-3T-autoround-lm_head-symFalse

https://github.com/intel/auto-round/blob/8a3da144423322dfedb0b3fa702ae35d242496d8/docs/Meta-Llama-3-8B-Instruct-acc.md?plain=1#L3

Metric	BF16	w4g128 w/o lm-head	w4g128 with quantized lm-head
Avg.	0.6352	0.6312	0.6303
mmlu	0.6386	0.6306	0.6318
winogrande	0.7143	0.7238	0.7269
truthfulqa_mc1	0.3623	0.3537	0.3525
rte	0.6751	0.6859	0.6679
piqa	0.7867	0.7797	0.7802
openbookqa	0.3400	0.3300	0.3320
lambada_openai	0.7182	0.7200	0.7173
hellaswag	0.5769	0.5699	0.5701
boolq	0.8297	0.8309	0.8284
arc_easy	0.8152	0.8089	0.8106
arc_challenge	0.5299	0.5102	0.5154

---

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-neuralmagic]

@comaniac ready for re-review

(ignore spec decode test failure @njhill making a new model that avoids OOM at fp32)

[comaniac]

Overall LGTM

[comaniac]

Why we have to re-download Falcon?

[robertgshaw2-neuralmagic]

trust_remote_code did not work for falcon, not sure why

[comaniac]

How long does it take to run all models listed here? Can some of them be removed to reduce the CI time?

[robertgshaw2-neuralmagic]

I can just remove them. I just wanted to prove the accuracy was right

[njhill]

• @Yard1 for VocabParallelEmbedding weight_loading --> anything to be aware of for LoRA?

@robertgshaw2-neuralmagic @Yard1 it looks like there was some impact from this ... not sure if it actually exposed a latent bug where the lm_head for gpt_bigcode (and similar) was not previously adaptable: #6314

[robertgshaw2-neuralmagic]

• @Yard1 for VocabParallelEmbedding weight_loading --> anything to be aware of for LoRA?

@robertgshaw2-neuralmagic @Yard1 it looks like there was some impact from this ... not sure if it actually exposed a latent bug where the lm_head for gpt_bigcode (and similar) was not previously adaptable: #6314

@njhill do you know if this was working before?

[RanchiZhao]

hi！is 'look into quantized embeddings' available now? looking forward to this!