[Core] Set linear_weights directly on the layer
#3977
Conversation
Yard1
changed the title
linear_weights attribute into a dynamic propertylinear_weights directly on the layer
Yard1
changed the title
linear_weights directly on the layerlinear_weights directly on the layer
| params_dtype: torch.dtype) -> Dict[str, Any]: | ||
| """Create weights for a linear layer.""" | ||
| output_size: int, params_dtype: torch.dtype, | ||
| **extra_weight_attrs): |
There was a problem hiding this comment.
Could this just be an optional weight_loader?
I see the value of enabling kwargs here for future extensibility, but I don't see a case that exists yet other than weight_loader so far, perhaps making the argument explicit is better until we have a reason to allow kwargs
There was a problem hiding this comment.
Actually, making it a kwarg will be easier as we don't need extra handling to not set the weight_loader if it's left unspecified.
| layer.register_parameter("scales", scales) | ||
| set_weight_attrs(scales, extra_weight_attrs) | ||
|
|
||
| layer.exllama_state = exllama_state |
There was a problem hiding this comment.
The LinearLayer does not need to know this info, so I think it should be GPTQLinearMethod.exllama_state ... It probably should not have been in the weights dict before
This will avoid having this dangling member
There was a problem hiding this comment.
ah actually I don't think this is possible since GPTQLinearMethod is a singleton across all layers
There was a problem hiding this comment.
Sounds good
|
LGTM, left a few minor comments |
This reverts commit e20cdc1.
Currently, all vLLM linear layers use a
linear_weightsattribute to store a dictionary of tensors to be used for actually calculating the matmuls. Those same tensors are also set on the layer itself, leading to the tensors being referenced both in the dictionary and on the layer. If the references to the tensors on the layer are replaced (eg. by model loading code), then the tensors will stop being equal leading to incorrect matmul results. This PR removes the double reference by avoiding the storing of thelinear_weightsdictionary and instead setting the tensors directly as attributes on the layer, meaning we are left with only a single source of truth.See #3476 (comment) for more context
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!