Repository navigation
Conversation
sufubao
marked this pull request as draft
November 12, 2026 04:45
sufubao
force-pushed
the
qwen38-flash-next-consolidated
branch
4 times, most recently
from
November 13, 2026 20:32
0b956d5 to
05fbb20
Compare
sufubao
force-pushed
the
qwen38-flash-next-consolidated
branch
from
November 14, 2026 13:43
05fbb20 to
c0de4ec
Compare
sufubao
changed the title
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
支持 Qwen3.8 Flash Next（
qwen4_exp/qwen4_exp_text），接入 QSA/GDN、n-gram/PLE、视觉编码器和 MTP，复用上游 ReplaySSM 状态管理。保留 BF16 专家并行、FP8 QSA、FlashInfer GDN prefill、CUDA Graph、MTP 接受行输出投影、TP 通信及小批量复制 decode 优化。硬件专用路径的适用条件集中在对应 decoder 中，并保留内核微基准和数值、缓存状态回归测试。
CPU 回归 1,001 通过、3,417 跳过；排除 7 个依赖 GPU 的 DSV4 夹具用例和 2 个 pinned-memory 用例，它们已在未修改的上游复现无 CUDA 时失败。181 个 Python 文件通过 Black、Flake8。当前提交未重跑 GPU 数值、完整模型生成或性能测试。