All pull requests

Pull requests

• Move CI priority scoring to the scheduler / 将 CI 优先级打分移至调度器

  #5085·adibarra opened Nov 15, 2026Collaborator·6 of 11 tasks·

  ·
  1

• [VideoGenX] sweep Wan2.2 B200 on 1, 2, 4 and 8 GPUs / [VideoGenX] 在 B200 上以 1、2、4、8 卡扫描 Wan2.2

  full-sweep-fail-fast
  #5084·edwingao28 opened Nov 15, 2026Collaborator·

  ·
  3

• [Fix] Link VideoGenX from the unofficial run visualizer comment / [Fix] 在非官方运行可视化评论中加入 VideoGenX 链接

  #5083·edwingao28 opened Nov 15, 2026Collaborator·

  ·

• [VideoGenX] add vLLM-Omni Wan2.2 on B300 with 1, 2, 4 and 8 GPUs / [VideoGenX] 在 B300 上新增 vLLM-Omni Wan2.2（1、2、4、8 卡）

  full-sweep-fail-fast
  #5082·edwingao28 opened Nov 14, 2026Collaborator·

  ·
  2

• [wip][NVIDIA][AgentX] MiniMax-M3.1-Flash GB200 Dynamo vLLM

  full-sweep-enabled
  #5076·xinli-sw opened Nov 14, 2026Collaborator·

  ·
  8

• [wip][NVIDIA][AgentX] MiniMax-M3.1-Flash GB300

  full-sweep-enabled
  #5075·xinli-sw opened Nov 14, 2026Collaborator·

  ·
  8

• [Klaud Cold] Update GLM-5.3-Flash FP8 MI325X SGLang AgentX MTP image to v0.5.23 / [Klaud Cold] 将 GLM-5.3-Flash FP8 MI325X SGLang AgentX MTP 镜像更新至 v0.5.23

  full-sweep-fail-fastklaud-finished
  #5074·Klaud-Cold opened Nov 14, 2026Collaborator·

  ·
  11

• fix(bench): honor the SSE end marker / fix(bench): 正确处理 SSE 结束标记

  #5072·alaydshah opened Nov 13, 2026·

  ·
  2

• [NV] Kimi-K3.1-Pro GB300 Dynamo+vLLM AgentX: new image, DEP32 c704/c960, 1P5D low-latency points / [NV] Kimi-K3.1-Pro GB300 Dynamo+vLLM AgentX：新镜像、DEP32 c704/c960、1P5D 低延迟点

  non-canary-full-sweep-enabled
  #5069·mispa-ms opened Nov 13, 2026Collaborator·6 of 11 tasks·

  ·
  4

• [NVIDIA][AgentX] B300 Kimi-K3.1-Pro FP4 Dynamo+vLLM disaggregated configs / [NVIDIA][AgentX] B300 Kimi-K3.1-Pro FP4 Dynamo+vLLM 分离式配置

  non-canary-full-sweep-enabled
  #5068·mispa-ms opened Nov 13, 2026Collaborator·

  ·
  6

• [AgentX] MiniMax-M3.1-Flash GB300 Dynamo TRT: add disaggregated points / [AgentX] MiniMax-M3.1-Flash GB300 Dynamo TRT：新增分离式运行点

  full-sweep-enabled
  #5066·Ankur-singh opened Nov 13, 2026Collaborator·5 of 11 tasks·

  ·
  6

• [NVIDIA][AgentX] GB300 DSV4.1-Flash Dynamo + SGLang: megakernel agg c4, c8 and c32 / [NVIDIA][AgentX] GB300 DSV4.1-Flash Dynamo + SGLang：agg c4、c8、c32 改用 megakernel

  #5065·weireweire opened Nov 13, 2026Collaborator·

  ·
  1

• [AMD][Qwen3.8-Flash-Next] MI325X FP8 SGLang AgentX on sglang-rocm v0.5.23-20261110 with sglang#46945 and no integer all-reduce / [AMD][Qwen3.8-Flash-Next] MI325X FP8 SGLang AgentX 升级至 sglang-rocm v0.5.23-20261110，应用 sglang#46945 并禁用整数 all-reduce

  #5064·ZhengGong-amd opened Nov 13, 2026Collaborator·5 of 11 tasks·

  ·
  4

• [NVIDIA][AgentX] Muse Spark Open GB300 SGLang megakernel point at concurrency 1 / [NVIDIA][AgentX] Muse Spark Open GB300 SGLang megakernel 并发 1 测试点

  full-sweep-fail-fast
  #5063·Fridge003 opened Nov 13, 2026Collaborator·

  ·
  9

• [AMD][AgentX] DeepSeek-V4.1-Flash MI355X SGLang TP2 c32: run the AMD monokernel with shortest-prefill-first / [AMD][AgentX] DeepSeek-V4.1-Flash MI355X SGLang TP2 c32：启用 AMD monokernel 和 shortest-prefill-first

  full-sweep-fail-fast
  #5062·lishuoshuo-amd opened Nov 13, 2026Collaborator·2 of 4 tasks·

  ·
  2

• [AMD][Qwen3.5] Bump MI355X FP4 AgentX image to sglang-rocm v0.5.23-20261109 / [AMD][Qwen3.5] 将 MI355X FP4 AgentX 镜像升级至 sglang-rocm v0.5.23-20261109

  full-sweep-fail-fast
  #5061·chuyeh opened Nov 13, 2026Collaborator·4 of 11 tasks·

  ·
  2

  

• [AMD][AgentX] MI300X Qwen3.5 FP8 SGLang MTP: run target verify on unified attention

  full-sweep-fail-fast
  #5060·jiaqiang-dot-liu opened Nov 13, 2026Collaborator·4 of 11 tasks·

  ·
  2

• [NVIDIA][AgentX] GB300 Kimi-K3.1-Pro FP4 Dynamo+SGLang: TP32 decode at c224, 1P3D c58 and trimmed megakernel options / [NVIDIA][AgentX] GB300 Kimi-K3.1-Pro FP4 Dynamo+SGLang：c224 TP32 decode、1P3D c58 与精简 megakernel 选项

  full-sweep-fail-fast
  #5059·nvpohanh opened Nov 13, 2026Collaborator·

  ·
  7

• [AMD][MI300X] DSv4.1 Flash vLLM: ROCm 10.0 nightly f5d93296 with SWA bounded replay / [AMD][MI300X] DSv4.1 Flash vLLM：升级至 ROCm 10.0 nightly f5d93296 并启用 SWA bounded replay

  full-sweep-fail-fast
  #5058·lishuoshuo-amd opened Nov 13, 2026Collaborator·3 of 10 tasks·

  ·
  7

• [AMD] DSv4.1-Flash MI355X vLLM DEP8 AgentX c256/c512/c1024 with tensor-scale FP8 KV cache / [AMD] DSv4.1-Flash MI355X vLLM DEP8 AgentX c256/c512/c1024 启用 tensor scale FP8 KV cache

  non-canary-full-sweep-enabled
  #5056·Fangzhou-Ai opened Nov 13, 2026Collaborator·

  ·
  5

• [Klaud Cold] Update Muse Spark Open MI355X SGLang image to v0.5.23-rocm10-mi35x-20261112 / 将 Muse Spark Open MI355X SGLang 镜像更新至 v0.5.23-rocm10-mi35x-20261112

  full-sweep-fail-fastklaud-finished
  #5055·Klaud-Cold opened Nov 13, 2026Collaborator·

  ·
  16

• Add Muse-Spark GB300 Dynamo TensorRT-LLM AgentX with DFlash2 / 新增 Muse-Spark GB300 Dynamo TensorRT-LLM DFlash2 AgentX 配置

  non-canary-full-sweep-enabled
  #5047·richardhuo-nv opened Nov 12, 2026Collaborator·6 of 11 tasks·

  ·
  7

• [AMD][MI300X] DSv4.1 Flash vLLM: move to the 2026-11-12 ROCm 10.0 nightly be47189f and enable SWA bounded replay

  #5046·maeehart opened Nov 12, 2026Collaborator·

  ·
  1

• [AMD][MI325X] DSv4.1 Flash vLLM: move to the 2026-11-12 ROCm 10.0 nightly be47189f

  #5045·maeehart opened Nov 12, 2026Collaborator·

  ·
  1

• [AMD][AgentX] DeepSeek-V4.1-Flash MI355X vLLM: add DP2 concurrency 128-384, with UMBP KV offloading at 256 and 384 / [AMD][AgentX] DeepSeek-V4.1-Flash MI355X vLLM：新增 DP2 并发 128–384 点，256、384 启用 UMBP KV offloading

  full-sweep-enabled
  #5036·shen-shanshan opened Nov 12, 2026Collaborator·4 of 11 tasks·

  ·
  6