tune triton-fp8w8a8g64 and triton-fp8w8a8g128 in 4090D#1216
Conversation
Summary of ChangesHello, 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 integrates new autotuned Triton kernel configurations into the LightLLM project. These additions are crucial for optimizing the performance of Highlights
🧠 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
Activity
Using Gemini Code AssistThe 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
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 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
|
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
There was a problem hiding this comment.
Code Review
This pull request adds autotuned Triton kernel configurations for scaled_mm_act_per_group_w_perchannel on NVIDIA GeForce RTX 4090D and RTX 5090 GPUs. The changes consist of adding several JSON configuration files for various problem sizes. While the addition of tuned configurations is beneficial for performance, I've noticed a minor maintainability issue in the new JSON files. The keys, which correspond to matrix dimensions, are not sorted, making the files difficult to read and compare. I've left a comment with a suggestion to sort the keys numerically. This should be applied to all new configuration files. Also, the pull request title only mentions the 4090D, but configurations for the 5090 are also included. It would be good to update the title to accurately reflect the scope of the changes.
| { | ||
| "1": { | ||
| "BLOCK_K": 128, | ||
| "BLOCK_M": 8, | ||
| "BLOCK_N": 64, | ||
| "GROUP_M": 8, | ||
| "num_stages": 6, | ||
| "num_warps": 4 | ||
| }, | ||
| "100": { | ||
| "BLOCK_K": 128, | ||
| "BLOCK_M": 64, | ||
| "BLOCK_N": 128, | ||
| "GROUP_M": 8, | ||
| "num_stages": 4, | ||
| "num_warps": 8 | ||
| }, | ||
| "1024": { | ||
| "BLOCK_K": 128, | ||
| "BLOCK_M": 64, | ||
| "BLOCK_N": 64, | ||
| "GROUP_M": 8, | ||
| "num_stages": 3, | ||
| "num_warps": 4 | ||
| }, | ||
| "128": { | ||
| "BLOCK_K": 128, | ||
| "BLOCK_M": 64, | ||
| "BLOCK_N": 64, | ||
| "GROUP_M": 8, | ||
| "num_stages": 4, | ||
| "num_warps": 4 | ||
| }, | ||
| "16": { | ||
| "BLOCK_K": 128, | ||
| "BLOCK_M": 16, | ||
| "BLOCK_N": 64, | ||
| "GROUP_M": 8, | ||
| "num_stages": 4, | ||
| "num_warps": 4 | ||
| }, | ||
| "256": { | ||
| "BLOCK_K": 128, | ||
| "BLOCK_M": 64, | ||
| "BLOCK_N": 64, | ||
| "GROUP_M": 8, | ||
| "num_stages": 3, | ||
| "num_warps": 4 | ||
| }, | ||
| "32": { | ||
| "BLOCK_K": 128, | ||
| "BLOCK_M": 32, | ||
| "BLOCK_N": 64, | ||
| "GROUP_M": 8, | ||
| "num_stages": 4, | ||
| "num_warps": 8 | ||
| }, | ||
| "64": { | ||
| "BLOCK_K": 128, | ||
| "BLOCK_M": 64, | ||
| "BLOCK_N": 64, | ||
| "GROUP_M": 8, | ||
| "num_stages": 3, | ||
| "num_warps": 4 | ||
| }, | ||
| "8": { | ||
| "BLOCK_K": 128, | ||
| "BLOCK_M": 8, | ||
| "BLOCK_N": 64, | ||
| "GROUP_M": 8, | ||
| "num_stages": 5, | ||
| "num_warps": 4 | ||
| } | ||
| } No newline at end of file |
There was a problem hiding this comment.
The keys in this JSON file, which represent the 'M' dimension for tuning, are not sorted. This makes the file difficult to read and maintain. Please sort the keys numerically to improve readability. This should be applied to all new JSON configuration files in this pull request.
{
"1": {
"BLOCK_K": 128,
"BLOCK_M": 8,
"BLOCK_N": 64,
"GROUP_M": 8,
"num_stages": 6,
"num_warps": 4
},
"8": {
"BLOCK_K": 128,
"BLOCK_M": 8,
"BLOCK_N": 64,
"GROUP_M": 8,
"num_stages": 5,
"num_warps": 4
},
"16": {
"BLOCK_K": 128,
"BLOCK_M": 16,
"BLOCK_N": 64,
"GROUP_M": 8,
"num_stages": 4,
"num_warps": 4
},
"32": {
"BLOCK_K": 128,
"BLOCK_M": 32,
"BLOCK_N": 64,
"GROUP_M": 8,
"num_stages": 4,
"num_warps": 8
},
"64": {
"BLOCK_K": 128,
"BLOCK_M": 64,
"BLOCK_N": 64,
"GROUP_M": 8,
"num_stages": 3,
"num_warps": 4
},
"100": {
"BLOCK_K": 128,
"BLOCK_M": 64,
"BLOCK_N": 128,
"GROUP_M": 8,
"num_stages": 4,
"num_warps": 8
},
"128": {
"BLOCK_K": 128,
"BLOCK_M": 64,
"BLOCK_N": 64,
"GROUP_M": 8,
"num_stages": 4,
"num_warps": 4
},
"256": {
"BLOCK_K": 128,
"BLOCK_M": 64,
"BLOCK_N": 64,
"GROUP_M": 8,
"num_stages": 3,
"num_warps": 4
},
"1024": {
"BLOCK_K": 128,
"BLOCK_M": 64,
"BLOCK_N": 64,
"GROUP_M": 8,
"num_stages": 3,
"num_warps": 4
}
}2 participants
No description provided.