All pull requests

Pull requests

• [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

  klaud-parked
  #5055·Klaud-Cold opened Nov 13, 2026Collaborator·

  ·
  3

• 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·

  ·
  5

• [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

• perf(amd): Kimi-K3.1 MI355X vLLM AgentX: image kimi-k31-2ded1ca, env flags, c1/c4 specialized decode, c48 three draft tokens / perf(amd)：Kimi-K3.1 MI355X vLLM AgentX：镜像 kimi-k31-2ded1ca、环境变量、c1/c4 specialized decode、c48 三个草稿 token

  #5043·hyukjlee opened Nov 12, 2026Collaborator·

  ·
  4

• [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

  #5038·ZhengGong-amd opened Nov 12, 2026Contributor·5 of 11 tasks·

  ·
  1

• [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

  #5037·lishuoshuo-amd opened Nov 12, 2026Contributor·3 of 10 tasks·

  ·
  1

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

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

  ·
  2

• [AMD][AgentX] DeepSeek-V4.1-Flash MI355X SGLang: bump image to 20261111 and add DP attention c128/c256 points / [AMD][AgentX] DeepSeek-V4.1-Flash MI355X SGLang：升级镜像至 20261111 并新增 DP attention c128/c256 配置点

  full-sweep-fail-fast
  #5034·limou102 opened Nov 12, 2026Collaborator·5 of 11 tasks·

  ·
  4

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

  #5032·jiaqiang-dot-liu opened Nov 12, 2026·4 of 11 tasks·

  ·
  3

• [AMD][MI35X] 1111 DSV41 Flash sglang mtp agentic / [AMD][MI35X] 1111 DSV41 Flash sglang mtp agentic 配置更新

  agentxAMDnon-canary-full-sweep-enabled
  #5029·zijiecode opened Nov 12, 2026Collaborator·

  ·
  3

• [K3.1-Pro] GB200 Dynamo+SGLang: add DCP8 aggregated and PP2-prefill disaggregated points, drop superseded points / [K3.1-Pro] GB200 Dynamo+SGLang：新增 DCP8 聚合式与 PP2 prefill 分离式测试点并移除被取代的测试点

  full-sweep-fail-fast
  #5027·nvpohanh opened Nov 12, 2026Collaborator·5 of 11 tasks·

  ·
  5

• [K3.1-Pro] Add B300 Dynamo+SGLang TP8 aggregated and 1P1D/1P2D disaggregated curves at concurrency 1-56 / [K3.1-Pro] 新增 B300 Dynamo+SGLang 并发 1-56 的 TP8 聚合式与 1P1D/1P2D 分离式曲线

  full-sweep-fail-fast
  #5026·nvpohanh opened Nov 12, 2026Collaborator·5 of 11 tasks·

  ·
  9

• fix: fail SRT jobs on unhealthy Slurm nodes / fix: Slurm 节点异常时让 SRT 任务失败

  #5025·chunfangamd opened Nov 12, 2026Collaborator·6 of 11 tasks·

  ·

• Add GLM-5.2 FP4 MI355X SGLang disaggregated AgentX with MTP / 新增 GLM-5.2 FP4 MI355X SGLang 分离式 AgentX（MTP）

  #5024·ichbinblau opened Nov 12, 2026Collaborator·7 of 11 tasks·

  ·
  3

• [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

  #5023·lishuoshuo-amd opened Nov 12, 2026Contributor·2 of 4 tasks·

  ·
  1

• [AgentX] MiniMax-M3.1-Flash GB300 Dynamo SGLang PD: add DEP16 points at conc 500-950 / [AgentX] MiniMax-M3.1-Flash GB300 Dynamo SGLang PD：新增 conc 500-950 的 DEP16 点

  full-sweep-fail-fast
  #5021·zijiexia opened Nov 11, 2026Collaborator·

  ·
  4

• [feat] add H3 video on A100 and download its snapshot on a100-cw / [feat] 在 A100 上新增 H3 视频，并在 a100-cw 上下载其快照

  #5018·edwingao28 opened Nov 11, 2026Collaborator·6 of 11 tasks·

  ·
  1

• [AgentX] DeepSeek-V4.1-Flash GB300 Dynamo vLLM: vLLM main 01fba62, TP4 specialized decode and 5P/7P/12P disagg points / [AgentX] DeepSeek-V4.1-Flash GB300 Dynamo vLLM：vLLM main 01fba62，TP4 specialized decode 与 5P/7P/12P 分离式测点

  agentxnon-canary-full-sweep-enabled
  #5015·Juntian777 opened Nov 11, 2026Collaborator·

  ·
  5

• [Klaud Cold] Update DeepSeek-V4.1-Flash FP4 MI325X vLLM AgentX DSpark image to nightly-rocm100-f5d93296 / [Klaud Cold] 将 DeepSeek-V4.1-Flash FP4 MI325X vLLM AgentX DSpark 镜像更新至 nightly-rocm100-f5d93296

  full-sweep-fail-fastklaud-finished
  #5014·Klaud-Cold opened Nov 11, 2026Collaborator·

  ·
  10

• Sanitize stored CI output / 清理存储的 CI 输出

  #5013·adibarra opened Nov 11, 2026Collaborator·

  ·
  6

• perf(amd): DSv4.1-Flash ATOM AgentX TP4 c1 only on the 1111 nightly / DSv4.1-Flash ATOM AgentX 1111 nightly 只测 TP4 c1

  full-sweep-enabled
  #5012·yhl-amd opened Nov 11, 2026Collaborator·

  ·
  2

• operatorx: GEMM testlist cleanup, MoE data-parallel groups bugfix

  #5007·hbarclay opened Nov 11, 2026Collaborator·

  ·
  3

• Add Qwen3.6-35B FP8 vLLM AL collection / 新增 Qwen3.6-35B FP8 vLLM 接受长度收集

  #5004·functionstackx opened Nov 11, 2026Collaborator·

  ·
  3

• [Klaud Cold] Update Qwen3.8-Flash-Next FP4 MI355X SGLang AgentX image to v0.5.23-rocm10-mi35x-20261110 / [Klaud Cold] 将 Qwen3.8-Flash-Next FP4 MI355X SGLang AgentX 镜像更新至 v0.5.23-rocm10-mi35x-20261110

  full-sweep-fail-fastklaud-finished
  #5000·Klaud-Cold opened Nov 11, 2026Collaborator·

  ·
  12