[Core] Implement sharded state loader #4690
Conversation
|
We should add a test to ensure it's working correctly (loaded weights and outputs are the same as with the default loader) |
examples/save_sharded_state.py
Outdated
| def main(args): | ||
| engine_args = EngineArgs.from_cli_args(args) | ||
| model_path = engine_args.model | ||
| if not Path(model_path).is_dir(): | ||
| raise ValueError("model path must be a local directory") | ||
| # Create LLM instance from arguments | ||
| llm = LLM(**dataclasses.asdict(engine_args)) | ||
| # Prepare output directory | ||
| Path(args.output).mkdir(exist_ok=True) | ||
| # Dump worker states to output directory | ||
| model_executor = llm.llm_engine.model_executor | ||
| model_executor.save_sharded_state(path=args.output, | ||
| pattern=args.pattern, | ||
| max_size=5 * 1024**3) | ||
| # Copy metadata files to output directory | ||
| for file in os.listdir(model_path): | ||
| if not any( | ||
| file.endswith(ext) for ext in (".bin", ".pt", ".safetensors")): | ||
| shutil.copy(f"{model_path}/{file}", args.output) |
There was a problem hiding this comment.
shouldn't this set the config.json or quant_config.json next to the model weights to inform vLLM loading the model what type of quantization the model checkpoint is in?
There was a problem hiding this comment.
Currently, it's just copying the config.json and quant_config.json from the input checkpoint that's being converted, which works for the use cases we've tested. Actually, I am not sure if it's correct to override these configs (or add a quant_config.json where it wasn't there previously) because then the config may mismatch the final loaded states?
Sure, that makes sense. What's the spec of the test runner machine we should target? OK to assume cuda device is present? |
|
@Yard1 added test, please take a look |
|
One question I have is that can this be implemented using safetensor's partial read? safetensors have all the metadata in headers so you can access the tensors partially |
Conceptually, I think so. Though for larger models like Arctic we prefer this implementation for a few reasons:
|
|
Ah, I hadn't seen that PR, thanks for bringing it up! From what I understand, both PRs address the problem of model loading speed, particularly for larger models, but the approaches seem pretty different. #3729 modifies the existing model loading path so each worker loads a different set of tensors, then broadcasts the shards to each other. This PR loads the model once using the default path and dumps each worker state to disk, then subsequent model loads can direct read these states, skipping the default model loading path altogether. Some first thoughts on the pros/cons:
|
There was a problem hiding this comment.
LGTM! The code looks clean and mergable. Thanks for addressing the comments from @Yard1.
Co-authored-by: Woosuk Kwon <[email protected]>
Co-authored-by: Woosuk Kwon <[email protected]>
Co-authored-by: Woosuk Kwon <[email protected]>
This PR implements a new model loader that directly loads the sharded states of each worker when using
DistributedGPUExecutor. When using tensor parallelism, this avoids each worker reading the full checkpoint just to load a small shard of it. Our tests using Arctic (#4652) showed a 10x improvement in model loading speed from NVMe when using 8x tensor parallelism.For quantization methods like DeepSpeed's FP_Quantize (also used in #4652) that quantize after loading, this PR allows easy creation of a quantized checkpoint that is directly loaded into each worker, further speeding up model loading.
This PR is separated out from #4652.
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,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!