Make vllm optional when using int8-sgl

[sitatec]

4 steps durations (after 4 warmups):

• int8-sgl+sglang | DiT cost 0.949300 seconds
• int8-vllm | DiT cost 0.925486 seconds
• int8-triton | DiT cost 1.361189 seconds
• int8-sgl+torchao | Run DiT cost 2.184162 seconds
• int8-sgl+torchao-compile | Run DiT cost 1.305696 seconds - Decorated def act_quant_int8_perchannel_sym_torchao(self, x) with @torch.compile(). Just an experiment, not included in this PR.
• int8-torchao | DiT cost 9.677256 seconds

Env:

• GPU: Nvidia H100
• Base docker image: nvidia/cuda:12.8.0-cudnn-devel-ubuntu24.04
• Pytorch: 2.9.1 (forked lightx2v and upgraded, it's working fine. Even 2.10.0 worked)
• FlashAttention 3

fp8 4steps	fp8 8stpes
int8 4stpes	int8 8steps

fp8 5stpes	int8 5steps
fp8 5stpes	int8 5steps
fp8 5stpes	int8 5steps

The quality of triton int8 quant was worse, that's why I put it last in the priority.

[gemini-code-assist]

Summary of Changes

Hello @sitatec, 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 enhances the flexibility and resilience of the quantization pipeline by making vLLM and sglang optional dependencies. It introduces a prioritized fallback system for INT8 activation quantization, ensuring that the system can still operate effectively by leveraging alternative backends like torchao or triton if the preferred ones are not available. This change improves the adaptability of the quantization process to different environments and available libraries.

Highlights

• Optional vLLM Dependency: The vLLM custom operations (_custom_ops) are now treated as an optional dependency, with their import aliased to vllm_ops and handled gracefully if not available.
• SGLang Integration: Introduced sglang.srt.layers.quantization.int8_kernel as an optional dependency for INT8 activation quantization, aliased as sglang_int8_act_quant.
• Quantization Backend Fallback: Implemented a robust fallback mechanism for int8-sgl activation quantization, prioritizing sglang, then vllm, followed by torchao, and finally triton as a last resort.
• FP32 Scale Casting: Added logic to optionally cast quantization scales to torch.float32 within the act_quant_int8_perchannel_sym_torchao function if self.scale_force_fp32 is enabled.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Changelog

• lightx2v/common/ops/mm/mm_weight.py
  • Renamed the import alias for vllm._custom_ops from ops to vllm_ops to avoid naming conflicts and clarify its origin.
  • Added a conditional import for sglang.srt.layers.quantization.int8_kernel as sglang_int8_act_quant, making sglang an optional dependency.
  • Modified act_quant_int8_perchannel_sym_torchao to include an optional casting of input_tensor_scale to torch.float32 based on self.scale_force_fp32.
  • Updated all references from the old ops alias to the new vllm_ops alias for vLLM-related function calls and availability checks.
  • Revised the docstring for MMWeightWint8channelAint8channeldynamicSglActVllm to accurately reflect the new priority order for activation dynamic quantization backends.
  • Implemented a new assignment logic for self.act_quant_func in MMWeightWint8channelAint8channeldynamicSglActVllm, establishing a priority chain: sglang_int8_act_quant > vllm_ops > torchao_int8_quant > int8_quantize_triton.

Activity

• No human activity (comments, reviews) has been recorded on this pull request yet.

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 by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

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 pull request 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 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. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

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

1. 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. ↩

[gemini-code-assist]

Code Review

This pull request successfully makes vllm an optional dependency for int8-sgl quantization by introducing a fallback mechanism for activation quantization. The changes are logical and well-implemented. I've provided a couple of suggestions to improve code readability and fix typos in the documentation. Overall, this is a good improvement.