[ Kernel ] FP8 Dynamic-Per-Token Quant Kernel #6511
Conversation
|
👋 Hi! Thank you for contributing to the vLLM project. Full CI run is still required to merge this PR so once the PR is ready to go, please make sure to run it. If you need all test signals in between PR commits, you can trigger full CI as well. To run full CI, you can do one of these:
🚀 |
| @@ -320,6 +320,17 @@ def scaled_fp8_quant( | |||
| return output, scale | |||
|
|
|||
|
|
|||
| def dynamic_per_token_scaled_fp8_quant( | |||
| input: torch.Tensor) -> Tuple[torch.Tensor, torch.Tensor]: | |||
There was a problem hiding this comment.
To fit the existing interface, where all quantization schemes go through scaled_fp8_quant, this should probably also be merged up with the scaled_fp8_quant. How do we want to handle this ?
I believe the dynamic-per-token vs dynamic-per-tensor model quantization scheme will come from the model itself and we can plumb it down into the scaled_fp8_quant function. I'd appreciate any references into will how to plumb this down. Thanks.
@robertgshaw2-neuralmagic @comaniac @tlrmchlsmth ?
There was a problem hiding this comment.
I think this looks good as is. I can handle the plumbing from here in a follow up with some end-to-end model tests
There was a problem hiding this comment.
Definitely prefer to have a unified API.
| @@ -3,6 +3,8 @@ | |||
|
|
|||
There was a problem hiding this comment.
Mostly refactors in this file!
robertgshaw2-neuralmagic
changed the title
|
/ready |
github-actions
bot
added
the
ready
csrc/quantization/fp8/common.cu
Outdated
| __syncthreads(); | ||
|
|
||
| float const inverted_scale = FP8_E4M3_MAX / block_absmax_val; | ||
| bool const vectorize_conversions = hidden_size % 4 == 0; |
There was a problem hiding this comment.
Why only vectorize conversions if the hidden_size is divisible by 4? scaled_fp8_conversion_vec has a loop to "Handle the remaining elements if num_elems is not divisible by 4"
There was a problem hiding this comment.
The input and output pointers need to be aligned at 8-byte and 4-byte addresses respectively. Each row of the input is processed by a single block.
FP8 output data-type requires that hidden-size (each token/row of the output has hidden-size elements) be a multiple of 4. When hidden-size is not a multiple of 4,
- The first row can be processed like you mentioned.
- But the second row's address will not be 4-byte aligned and throws a runtime error.
There was a problem hiding this comment.
There was a problem hiding this comment.
Why didn't we run into this with the original scaled_fp8_quant_kernel?
There was a problem hiding this comment.
the scaled_fp8_quant_kernel has a single scale that is applied to all of the input values and it doesn't have to process row-by-row. For quant purposes, that kernel can treat the inputs and output as one big row of values (the "first row" case from above).
There was a problem hiding this comment.
Approved to unblock.
- Same question @tlrmchlsmth that why not use vectorization for non-dividable problem size?
- Better to unify the API. Ok to move to the next PR.
robertgshaw2-neuralmagic
merged commit b5241e4
into
vllm-project:main
Co-authored-by: Varun Sundar Rabindranath <[email protected]>
Co-authored-by: Varun Sundar Rabindranath <[email protected]>
Co-authored-by: Varun Sundar Rabindranath <[email protected]>
Co-authored-by: Varun Sundar Rabindranath <[email protected]> Signed-off-by: Alvant <[email protected]>
Add support for dynamic-per-token fp8 quantization.
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!