AI-Hypercomputer / AI-Hypercomputer/maxtext

Qwen3-Next: sequential GDN paths (lax.scan recurrence, solve_triangular) bound TPU step time

未關閉
#4,347 1 則留言 0 個 reaction 已指派 1 人 已被 @parambole 認領 在 GitHub 檢視
主要語言
Python
星號
2.4k
分支
608
平均合併
2 天 19 小時
30 天內合併 PR
158

描述

### Problem

In the Qwen3-Next Gated Delta Net (`jax_chunk_gated_delta_rule`), the two sequential computations dominate TPU step time:

- the inter-chunk recurrence runs as `lax.scan`, so the recurrent state round-trips HBM on every chunk step, and
- the UT-transform inverse `A = (I+S)^-1` uses `solve_triangular`, which substitutes row by row and barely uses the MXU.

Neither overlaps with other work, so they bound the step directly.

### Measurement

v5e-256, Qwen3-Next dense-8B config, seq 4096, bs 1, remat full (details in #4348):

| build | step time | tokens/s/device |
|---|---|---|
| main | 3.373s | 1,214 |
| #4348 (Pallas kernels) | **2.571s** | **1,593 (+31%)** |

Loss trajectories match to bf16 rounding level.

貢獻指南

開啟貢獻指南

研究方向

Start at the Qwen3-Next Gated Delta Net entry point jax_chunk_gated_delta_rule and inspect the lax.scan recurrence and solve_triangular inverse described in the issue. Compare the Pallas-kernel results referenced in #4348, then benchmark the v5e-256, dense-8B, sequence-4096, batch-1 configuration; done means lower step time with loss matching to bf16 rounding.

由索引模型根據 Issue 內容生成。

評估

技術堆疊
python
領域
machine-learning, performance
Issue 類型
重構
難度
4/5
預估耗時
3-5 天
活躍度
活躍
描述清晰度
基本清楚
新手友好度
35/100

把新 issue 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。