YAMY

### Checklist - [x] I searched related issues but found no solution. - [x] The bug persists in the latest version. - [x] Issues without environment info and a minimal...

[DeepSeekV3.2] Centralize NSA dispatch logic in NativeSparseAttnBackend

3

## Motivation NativeSparseAttnBackend currently spreads dispatch logic for NSA prefill/decode implementations and MHA vs. MLA selection across multiple places: - Global `NSA_PREFILL_IMPL` / `NSA_DECODE_IMPL` variables that are mutated in `__init__`....

Upgrade flashmla kernel for NSA tp support

1

## Motivation ## Modifications ## Accuracy Tests ## Benchmarking and Profiling ## Checklist - [ ] Format your code according to the [Format code with pre-commit](https://docs.sglang.ai/developer_guide/contribution_guide.html#format-code-with-pre-commit). - [ ] Add...

[DeepSeekV3.2] Enable pure TP & Partial DP Attention

2

## Motivation DeepSeekV3.2 NSA currently has rough edges when running in **pure TP mode** (`dp_size < tp_size`): - FlashMLA sparse can see an invalid `num_heads` per rank after TP sharding....

[Bug]: 下载知识库一半时出现403 Forbidden（考虑限流？）

2

### 问题描述 下载到一半的时候出现403 forbidden的情况： feishu2md dl --wiki -o /Users/yangminl/Documents/learn/FeishuConversion "https://yamy12344.feishu.cn/wiki/settings/XXXX" Captured document token: wikcnFo5OpfxTaC0r9SVlIexiQc Captured document token: FI8EwT2xairbEvkF7Q0cC11Anzg Captured document token: wikcnXOVourehpiWwSjZS9vOt5d Captured document token: wikcnOO3P4dwtFnV5rtUKKo81Kd Captured document token: wikcnfDhByEoIPgeHHNozhtAPcc...