Support DeepSeekV3-style block FP8 quantization with CT #20279
Conversation
Signed-off-by: mgoin <mgoin64@gmail.com>
|
👋 Hi! Thank you for contributing to the vLLM project. 💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels. Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging. To run CI, PR reviewers can either: Add 🚀 |
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Summary of Changes
Hello @mgoin, I'm Gemini Code Assist1! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!
This pull request significantly extends vLLM's quantization capabilities by integrating support for DeepSeekV3-style block FP8 quantization. It involves substantial changes to how quantized weights are handled, from their creation and loading to their application during inference, ensuring compatibility with new model types and optimizing performance through specialized kernel utilization.
Highlights
- DeepSeekV3-style Block FP8 Support: This pull request introduces comprehensive support for DeepSeekV3-style block FP8 quantization, enabling vLLM to load and utilize models quantized with this specific scheme.
- Enhanced Weight Management: New logic has been added to handle block-quantized weights, including the introduction of
BlockQuantScaleParameterfor managing block-wise scales, and specific processing steps like FNUZ normalization and ROCm padding during weight loading. - Optimized Kernel Dispatch: The system now dispatches to a specialized
torch.ops.vllm.apply_w8a8_block_fp8_linearoperation for block FP8 linear layers, leveraging optimized kernels for efficient execution based on the detected quantization strategy. - Refactored Quantization Logic: Internal quantization parameter management has been streamlined by storing
weight_block_sizedirectly on the layer object and consistently passingQuantizationArgsobjects, improving type safety and code clarity.
Using Gemini Code Assist
The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.
Invoking Gemini
You can request assistance from Gemini at any point in your pull request via creating an issue comment (i.e. comment on the pull request page) using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands.
| Feature | Command | Description |
|---|---|---|
| Code Review | /gemini review |
Performs a code review for the current pull request in its current state. |
| Pull Request Summary | /gemini summary |
Provides a summary of the current pull request in its current state. |
| Comment | @gemini-code-assist | Responds in comments when explicitly tagged, both in issue comments and review comments. |
| Help | /gemini help |
Displays a list of available commands. |
Customization
To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.
Limitations & Feedback
Gemini Code Assist is currently in preview and may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments to provide feedback.
You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.
Footnotes
-
Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩
There was a problem hiding this comment.
Code Review
This pull request introduces support for DeepSeekV3-style block FP8 quantization, primarily by integrating the BLOCK quantization strategy into the compressed-tensors framework. Key changes include refactoring how weight_block_size is handled across linear layers and quantization methods, implementing new weight creation and processing logic for block-quantized weights, and ensuring the correct kernel dispatch for block FP8 linear operations. The changes are well-structured and address the core requirements for this new quantization scheme.
| def create_weights(self, layer: torch.nn.Module, | ||
| input_size_per_partition: int, | ||
| output_partition_sizes: list[int], input_size: int, | ||
| output_size: int, params_dtype: torch.dtype, | ||
| weight_loader: Callable, **kwargs): | ||
| maybe_create_device_identity() | ||
|
|
||
| output_size_per_partition = sum(output_partition_sizes) | ||
| layer.logical_widths = output_partition_sizes | ||
| layer.weight_block_size = None | ||
|
|
||
| if self.strategy == QuantizationStrategy.BLOCK: | ||
| tp_size = get_tensor_model_parallel_world_size() | ||
| assert self.weight_block_size is not None | ||
| layer.weight_block_size = self.weight_block_size | ||
| block_n, block_k = ( | ||
| layer.weight_block_size[0], | ||
| layer.weight_block_size[1], | ||
| ) | ||
| # Required by row parallel | ||
| if (tp_size > 1 | ||
| and input_size // input_size_per_partition == tp_size | ||
| and input_size_per_partition % block_k != 0): | ||
| raise ValueError( | ||
| f"Weight input_size_per_partition = " | ||
| f"{input_size_per_partition} is not divisible by " | ||
| f"weight quantization block_k = {block_k}.") | ||
| # Required by column parallel or enabling merged weights | ||
| if (tp_size > 1 and output_size // output_size_per_partition | ||
| == tp_size) or len(output_partition_sizes) > 1: | ||
| for output_partition_size in output_partition_sizes: | ||
| if output_partition_size % block_n != 0: | ||
| raise ValueError( | ||
| f"Weight output_partition_size = " | ||
| f"{output_partition_size} is not divisible by " | ||
| f"weight quantization block_n = {block_n}.") | ||
|
|
||
| # WEIGHT | ||
| weight = ModelWeightParameter(data=torch.empty( | ||
| output_size_per_partition, | ||
| input_size_per_partition, | ||
| dtype=torch.float8_e4m3fn), | ||
| input_dim=1, | ||
| output_dim=0, | ||
| weight_loader=weight_loader) | ||
| layer.register_parameter("weight", weight) | ||
|
|
||
| # WEIGHT SCALE | ||
| # TODO: update create_xxx_parameter functions to return | ||
| # the newly added parameters | ||
| if self.strategy == QuantizationStrategy.CHANNEL: | ||
| weight_scale = ChannelQuantScaleParameter( | ||
| data=torch.empty((sum(output_partition_sizes), 1), | ||
| dtype=torch.float32), | ||
| output_dim=0, | ||
| weight_loader=weight_loader) | ||
| elif self.strategy == QuantizationStrategy.BLOCK: | ||
| assert self.is_static_input_scheme is False | ||
| weight_scale = BlockQuantScaleParameter( | ||
| data=torch.empty( | ||
| (output_size_per_partition + block_n - 1) // block_n, | ||
| (input_size_per_partition + block_k - 1) // block_k, | ||
| dtype=torch.float32, | ||
| ), | ||
| input_dim=1, | ||
| output_dim=0, | ||
| weight_loader=weight_loader, | ||
| ) | ||
| else: | ||
| assert self.strategy == QuantizationStrategy.TENSOR | ||
| weight_scale = PerTensorScaleParameter(data=torch.empty( | ||
| len(output_partition_sizes), dtype=torch.float32), | ||
| weight_loader=weight_loader) | ||
|
|
||
| # min requirement for fp8 kernels | ||
| weight_scale[:] = torch.finfo(torch.float32).min | ||
| layer.register_parameter("weight_scale", weight_scale) | ||
|
|
||
| # INPUT SCALE | ||
| if self.is_static_input_scheme: | ||
| input_scale = PerTensorScaleParameter(data=torch.empty( | ||
| len(output_partition_sizes), dtype=torch.float32), | ||
| weight_loader=weight_loader) | ||
| input_scale[:] = torch.finfo(torch.float32).min | ||
| layer.register_parameter("input_scale", input_scale) |
There was a problem hiding this comment.
The new create_weights method is a substantial addition that correctly handles the creation of weight and scale parameters for block quantization. The divisibility checks for block_n and block_k are critical for ensuring proper alignment with tensor parallelism and merged weights, preventing potential errors during computation. The use of BlockQuantScaleParameter for block strategy is also correct.
...model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_fp8.py
Show resolved
Hide resolved
vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors.py
Show resolved
Hide resolved
...model_executor/layers/quantization/compressed_tensors/schemes/compressed_tensors_w8a8_fp8.py
Show resolved
Hide resolved
| if layer.weight_block_size is not None: | ||
| return torch.ops.vllm.apply_w8a8_block_fp8_linear( | ||
| input=x, | ||
| weight=layer.weight, | ||
| block_size=layer.weight_block_size, | ||
| weight_scale=layer.weight_scale, | ||
| input_scale=layer.input_scale, | ||
| bias=bias, | ||
| cutlass_block_fp8_supported=self.cutlass_block_fp8_supported, | ||
| use_aiter_and_is_supported=self.use_aiter_and_is_supported, | ||
| ) |
There was a problem hiding this comment.
This if block correctly dispatches to torch.ops.vllm.apply_w8a8_block_fp8_linear when layer.weight_block_size is set, indicating that block quantization is active. This ensures that the appropriate kernel is used for the forward pass, leveraging the new block quantization capabilities.
Signed-off-by: mgoin <mgoin64@gmail.com>
|
This pull request has merge conflicts that must be resolved before it can be |
Purpose
Relies on recent support in compressed-tensors (neuralmagic/compressed-tensors#372) and llm-compressor (vllm-project/llm-compressor#1607) to produce the models.
This PR implements W8A8 FP8 block quantization support for compressed-tensors models. This is focused on supporting the DeepSeekV3-style format, which has 128x128 block weights and 1x128 block activations (really per-token-group).
Most of the logic is ported directly from
fp8.pyand I hope to refactor the utilities to be shared eventually.Test Plan
Manual testing with newly produced models. I'll add lm-eval in another PR
Test Result
Dense
CT result:
Ref:
MoE
CT result:
Ref: