[PR #50] [MERGED] fix: regressions (scheduler hang, cuda graph TypeError, MXFP4 cache, rsf double-apply) #50

New issue

Closed

opened 2026-10-29 21:06:45 +00:00 by bsakaguchi · 0 comments

bsakaguchi commented

2026-10-29 21:06:45 +00:00

Owner

📋 Pull Request Information

Original PR: https://github.com/kvcache-ai/sglang/pull/50
Author: @yyj6666667
Created: 5/11/2026
Status: ✅ Merged
Merged: 5/14/2026
Merged by: @yyj6666667

Base: main ← Head: fix/gpu-prefill-mxfp4-cache-and-rsf

📝 Commits (5)

bedcff3 fix(v4-flash): remove broken MXFP4 weight cache + fix rsf double-apply
de9d7bf fix(scheduler): revert PR #38 req_pool changes that break TP-only mode
2763727 fix(cuda_graph): use out-of-band _replay_forward_batch for non-DSV4 backends
91277ff Merge pull request #1 from yyj6666667/fix/scheduler-req-pool-regression
9fa1de6 fix: remove undefined _GraphBucket reference in cuda graph replay

📊 Changes

8 files changed (+34 additions, -35 deletions)

View changed files

📝 python/sglang/srt/layers/attention/deepseek_v4_backend_radix.py (+9 -3)
📝 python/sglang/srt/layers/moe/kt_ep_wrapper.py (+7 -21)
📝 python/sglang/srt/layers/quantization/mxfp4_deepseek.py (+6 -1)
📝 python/sglang/srt/layers/quantization/v4_triton_kernels_moe.py (+3 -2)
📝 python/sglang/srt/managers/scheduler.py (+2 -1)
📝 python/sglang/srt/managers/scheduler_runtime_checker_mixin.py (+1 -4)
📝 python/sglang/srt/mem_cache/memory_pool.py (+2 -2)
📝 python/sglang/srt/model_executor/cuda_graph_runner.py (+4 -1)

📄 Description

No description provided

_{🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.}

## 📋 Pull Request Information **Original PR:** https://github.com/kvcache-ai/sglang/pull/50 **Author:** [@yyj6666667](https://github.com/yyj6666667) **Created:** 5/11/2026 **Status:** ✅ Merged **Merged:** 5/14/2026 **Merged by:** [@yyj6666667](https://github.com/yyj6666667) **Base:** `main` ← **Head:** `fix/gpu-prefill-mxfp4-cache-and-rsf` --- ### 📝 Commits (5) - [`bedcff3`](https://github.com/kvcache-ai/sglang/commit/bedcff3786b8bb33b12f416867f580abe05d143a) fix(v4-flash): remove broken MXFP4 weight cache + fix rsf double-apply - [`de9d7bf`](https://github.com/kvcache-ai/sglang/commit/de9d7bf83aa62a0acf3974a386885a091f501fba) fix(scheduler): revert PR #38 req_pool changes that break TP-only mode - [`2763727`](https://github.com/kvcache-ai/sglang/commit/2763727f30a704faafb303fbf8230989c1e23897) fix(cuda_graph): use out-of-band _replay_forward_batch for non-DSV4 backends - [`91277ff`](https://github.com/kvcache-ai/sglang/commit/91277ffd6ee693e57419cf19463b8a5cac7d4d8f) Merge pull request #1 from yyj6666667/fix/scheduler-req-pool-regression - [`9fa1de6`](https://github.com/kvcache-ai/sglang/commit/9fa1de6bce255807add65ac14bc78c5f7412287a) fix: remove undefined _GraphBucket reference in cuda graph replay ### 📊 Changes **8 files changed** (+34 additions, -35 deletions) <details> <summary>View changed files</summary> 📝 `python/sglang/srt/layers/attention/deepseek_v4_backend_radix.py` (+9 -3) 📝 `python/sglang/srt/layers/moe/kt_ep_wrapper.py` (+7 -21) 📝 `python/sglang/srt/layers/quantization/mxfp4_deepseek.py` (+6 -1) 📝 `python/sglang/srt/layers/quantization/v4_triton_kernels_moe.py` (+3 -2) 📝 `python/sglang/srt/managers/scheduler.py` (+2 -1) 📝 `python/sglang/srt/managers/scheduler_runtime_checker_mixin.py` (+1 -4) 📝 `python/sglang/srt/mem_cache/memory_pool.py` (+2 -2) 📝 `python/sglang/srt/model_executor/cuda_graph_runner.py` (+4 -1) </details> ### 📄 Description _No description provided_ --- <sub>🔄 This issue represents a GitHub Pull Request. It cannot be merged through Gitea due to API limitations.</sub>

bsakaguchi

2026-10-29 21:06:45 +00:00

closed this issue
added the
pull-request
label

No labels

pull-request

No milestone

No project

No assignees

1 participant

Notifications

Due date

The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set.

Reference

kvcache-ai/sglang#50

No description provided.

Rows
Columns