[Performance] Optimize NSA Indexer K/S Buffer Access with Fused Triton Kernels

[Johnsonms]

Motivation

#13811

Modifications

Implement fused Triton kernels that:

1. Retrieve both K and S data in a single kernel call
2. Implement GetK, GetS and GetKandS triton kernels,
3. Switch from torch_fast to optimized Triton implementations

Accuracy Tests

Before:

After:

Benchmarking and Profiling

1. Reduced CPU time from 166 µs to 63 µs per layer.
2. Reduced kernel latency from 63 µs to 9 µs per layer.
3. Reduced kernel launches from 11 times to 1 time per layer.

Before:

After:

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.
• Work with maintainers to merge your PR. See the PR Merge Process

[gemini-code-assist]

Summary of Changes

Hello @Johnsonms, 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 enhances the performance of the NSA Indexer's K/S buffer access within the SGLang framework. It achieves this by migrating from existing torch_fast implementations to highly optimized, fused Triton kernels. The primary goal is to reduce computational overhead and improve efficiency when retrieving key and scale data, leading to substantial gains in processing speed and resource utilization.

Highlights

• Performance Optimization: Achieved significant speedups by replacing existing torch_fast implementations with highly optimized Triton kernels for K/S buffer access.
• Fused K/S Data Retrieval: Introduced a new fused Triton kernel (GetKAndS) that retrieves both Key (K) and Scale (S) data in a single, more efficient operation.
• Dedicated Triton Kernels: Implemented specific Triton kernels for GetK, GetS, and the new fused GetKAndS operations, enhancing data gathering from paged buffers.
• Reduced Kernel Launches: The number of kernel launches per layer was drastically reduced from 11 to 1, contributing to overall efficiency.
• Improved Latency and CPU Time: Benchmarking shows kernel latency dropped from 63 µs to 9 µs, and CPU time per layer decreased from 166 µs to 63 µs.

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 introduces significant performance optimizations by implementing fused Triton kernels for accessing K/S buffers in the NSA Indexer. The changes are well-structured and effectively reduce kernel launch overhead. My review includes a few minor suggestions to enhance code clarity and maintainability, such as removing commented-out debug statements and refactoring a small part of a Triton kernel to avoid redundant computations.

[Fridge003]

Please add correctness tests for the get_k_and_s kernel (can put them under test/manual folder)

[Fridge003]

/tag-and-rerun-ci

[Fridge003]

@Johnsonms Can you verify the correctness of this PR on AIME/GPQA?
Just make sure it doesn't hit any IMA error related to triton kernel (https://docs.sglang.io/basic_usage/deepseek_v32.html#benchmarking-results)

[Johnsonms]

FP4 model gpqa result
python3 -m sglang.test.run_eval --port 54321 --eval-name gpqa --num-examples 198 --max-tokens 120000 --repeat 8 --thinking-mode deepseek-v3

aime 2025 result

#! /bin/bash
export NEMO_SKILLS_DISABLE_UNCOMMITTED_CHANGES_CHECK=1

ns prepare_data aime25

PORT=30000
BACKEND=sglang
MODEL="deepseek-ai/DeepSeek-V3.2-Exp"
MODEL_NAME="dsv32-fp4"

echo "Starting AIME25 evaluation with model $MODEL on port $PORT using backend $BACKEND..."
ns eval \
  --benchmarks=aime25:4 \
  --server_type=$BACKEND \
  --model=$MODEL \
  --server_address=http://localhost:${PORT}/v1 \
  --output_dir=nemo_skills_aime25_${MODEL_NAME}_output_${BACKEND}_$(date +%Y%m%d_%H%M%S) \
  ++max_concurrent_requests=512 \
  ++server.api_key=dummy \
  ++inference.tokens_to_generate=64000

[Johnsonms]

FP8
python3 -m sglang.test.run_eval --port 54321 --eval-name gpqa --num-examples 198 --max-tokens 120000 --repeat 8 --thinking-mode deepseek-v3

FP8
aime 2025 result