[Others]【Hackathon 10th Spring No.49】【RFC】并行 N-gram 匹配 GPU Kernel 设计文档 (V2.0)

[cloudforge1]

Motivation

RFC V2.0 design document for Hackathon 10th Spring Task No.49: 为 FastDeploy 支持投机解码功能 — 并行 N-gram 匹配 GPU Kernel。

This RFC supersedes the design section (§三) of the prior RFC #1213 (NKNaN, 2026-02-07) by providing a complete parallel architecture specification that was absent from the original. The prior RFC's §三 totaled 12 lines with no thread model, no parallelism strategy, and no extreme-config performance analysis.

What this RFC adds over #1213:

Gap in #1213	This RFC
No parallel strategy	Two-phase kernel: Phase 1 <<<bsz, 1024>>> + Phase 2 <<<1, 1024>>>
No thread model	1024 threads/block parallel sliding-window search + atomicMin64 CAS
No threshold GPU-ization	CUB BlockScan prefix-sum, O(bsz) parallel
No extreme config analysis	Tested bsz=256, seq=128K: GPU 162µs vs CPU 275ms (1,700×)
No template specialization	parallel_ngram_search_specialized<1/2/3> with register caching
No early-exit mechanism	A2 early-exit: skip positions beyond current best match

Implementation PRs (already submitted, CI-verified on SM90 H20/H100):

• PR #6960 — Standard two-phase parallel kernel
• PR #7136 — Template-specialized optimization variant

Context: @freeliuzc clarified that the kernel must use parallel strategies for acceleration at bsz=256, seq_len=128K — requirements not present in the original task description or #1213. This RFC formally documents the parallel architecture that meets those requirements.

Modifications

• Added RFC design document: rfcs/FastDeploy/20260404_parallel_ngram_match_gpu_kernel_for_fastdeploy.md

Usage or Command

N/A — documentation only.

Accuracy Tests

N/A — design document. Implementation accuracy verified in PRs #6960 and #7136 (11/11 tests passed, GPU output bit-identical to CPU reference).

Checklist

• Add at least a tag in the PR title.
• Format your code, run pre-commit before commit.
• Add unit tests. N/A — RFC only.
• Provide accuracy results. N/A — RFC only.
• If submitting to release branch, cherry-pick from develop. N/A — targeting master.

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请检查PR提交格式和内容是否完备，具体请参考示例和模版。
Your PR has been submitted. Thanks for your contribution!
Please check its format and content. For this, you can refer to Template and Demo.

[cloudforge1]

Closing: will resubmit as inline edit to the existing merged RFC file per Baidu convention.