All pull requests

Pull requests

• [NVIDIA][AgentX] Retune GB300 DSV4.1-Flash Dynamo + SGLang HiCache disagg points (stacked on #4595) / [NVIDIA][AgentX] 重新调整 GB300 DSV4.1-Flash Dynamo + SGLang HiCache 分离式点位（基于 #4595）

  #4673·weireweire opened Oct 30, 2026Collaborator·

  ·
  1

• [Klaud Cold] Update DeepSeek-V4.1-Flash FP4 MI355X ATOM AgentX DSpark image to nightly_202610291449 / [Klaud Cold] 将 DeepSeek-V4.1-Flash FP4 MI355X ATOM AgentX DSpark 镜像更新至 nightly_202610291449

  full-sweep-fail-fastklaud-parked
  #4672·Klaud-Cold opened Oct 30, 2026Collaborator·

  ·
  12

• Combine #4537 and #4463 for MiniMax-M3.1-Flash MI355X vLLM: K7 at C16/C32, CPU offload at C64 / 合并 #4537 与 #4463：MiniMax-M3.1-Flash MI355X vLLM C16/C32 使用 K7、C64 使用 CPU 卸载

  full-sweep-enabled
  #4671·Fangzhou-Ai opened Oct 30, 2026Collaborator·

  ·
  4

• [AgentX] Retune Muse Spark Open GB200 SGLang per concurrency / [AgentX] 按并发重新调优 Muse Spark Open GB200 SGLang

  full-sweep-fail-fast
  #4670·shyeh25 opened Oct 30, 2026Collaborator·4 of 11 tasks·

  ·
  1

• Add DeepSeek-V4.1-Flash MI355X vLLM TP8 c1 monokernel scan on latest nightly / 新增 DeepSeek-V4.1-Flash MI355X vLLM TP8 c1 monokernel 单点扫描（最新 nightly）

  full-sweep-fail-fast
  #4666·Fangzhou-Ai opened Oct 30, 2026Collaborator·

  ·
  5

• [AgentX] Qwen3.5-397B-A17B B200 vLLM: add NVFP4 TP2 points with MTP and host KV offload / [AgentX] Qwen3.5-397B-A17B B200 vLLM：新增 NVFP4 TP2 测试点（MTP + 主机 KV offload）

  #4665·aniskumar-nv opened Oct 29, 2026·4 of 6 tasks·

  ·
  2

• Select golden AL curves by verifier and add SPEED-Bench collectors / 按验证器选择黄金 AL 曲线并新增 SPEED-Bench 收集脚本

  #4663·Ankur-singh opened Oct 29, 2026Collaborator·

  ·
  7

• MiniMax-M3.1-Flash GB300 vLLM AgentX: update agentic config / MiniMax-M3.1-Flash GB300 vLLM AgentX：更新 agentic 配置

  full-sweep-enabled
  #4662·hshrivastava-droid opened Oct 29, 2026Collaborator·4 of 9 tasks·

  ·
  4

• Standardize golden AL curve names and selection / 规范黄金 AL 曲线命名与选择

  #4661·Ankur-singh opened Oct 29, 2026Collaborator·

  ·
  16

• [VideoGenX] verify GPU ownership and seal execution evidence / [VideoGenX] 校验 GPU 归属并封存执行证据

  #4659·edwingao28 opened Oct 29, 2026Collaborator·3/5·3 of 10 tasks·

  ·

• [VideoGenX] measure requests and validate generated media / [VideoGenX] 测量请求并校验生成媒体

  #4658·edwingao28 opened Oct 29, 2026Collaborator·2/5·3 of 10 tasks·

  ·

• [VideoGenX] define reproducible H3 workload sets / [VideoGenX] 定义可复现的 H3 负载集合

  #4657·edwingao28 opened Oct 29, 2026Collaborator·1/5·3 of 10 tasks·

  ·

• [PowerX] collect GLM-5.2 B200 Dynamo-SGLang AgentX power / [PowerX] 采集 GLM-5.2 B200 Dynamo-SGLang AgentX 功耗

  full-sweep-fail-fast
  #4656·edwingao28 opened Oct 29, 2026Collaborator·7 of 11 tasks·

  ·
  4

• feat(ci): report-only AgentX speculative acceptance gate / feat(ci)：仅作报告的 AgentX 投机解码接受长度门禁

  #4654·adibarra opened Oct 29, 2026Collaborator·7 of 11 tasks·

  ·
  15

• [Docs] Explain when to use append-only / [Docs] 说明何时使用 append-only

  #4653·edwingao28 opened Oct 29, 2026Collaborator·

  ·

• [power_model] calibrate x86 and Grace CPU state power from measurements / [power_model] 基于实测校准 x86 与 Grace CPU 状态功耗

  #4645·edwingao28 opened Oct 29, 2026Collaborator·3 of 11 tasks·

  ·
  4

• [VideoGenX] add GB200 H3 SGLang deployment recipes / [VideoGenX] 添加 GB200 H3 SGLang 部署配方

  #4644·edwingao28 opened Oct 29, 2026Collaborator·5/5·7 of 10 tasks·

  ·
  2

• [Docs] Clarify that the draft KV-cache dtype is not restricted / [Docs] 澄清 draft KV-cache dtype 不受限制

  #4640·nvpohanh opened Oct 29, 2026Collaborator·

  ·

• Move Kimi-K3 GB300 AgentX to the chosen Oct 26 points on the latest vLLM nightly / 将 Kimi-K3 GB300 AgentX 更新为 Oct 26 选定点位并使用最新 vLLM nightly

  non-canary-full-sweep-enabled
  #4638·wzhao18 opened Oct 29, 2026Collaborator·

  ·
  5

• [Klaud Cold] Update DeepSeek-V4.1-Flash FP4 B200 SGLang AgentX DSpark image to nightly-dev-cu13-20261029-c10e0e17 / [Klaud Cold] 将 DeepSeek-V4.1-Flash FP4 B200 SGLang AgentX DSpark 镜像更新至 nightly-dev-cu13-20261029-c10e0e17

  full-sweep-fail-fastklaud-finished
  #4635·Klaud-Cold opened Oct 29, 2026Collaborator·

  ·
  17

• [Klaud Cold] Update Qwen3.5 FP8 H100 SGLang AgentX (+MTP) image to v0.5.22-cu130 / [Klaud Cold] 将 Qwen3.5 FP8 H100 SGLang AgentX（含 MTP）镜像更新至 v0.5.22-cu130

  full-sweep-fail-fastklaud-parked
  #4634·Klaud-Cold opened Oct 29, 2026Collaborator·

  ·
  21

• Tune MiniMax M3.1 GB200 AgentX disagg / 调优 MiniMax M3.1 GB200 AgentX 分离式配置

  non-canary-full-sweep-enabled
  #4632·xinli-sw opened Oct 29, 2026Collaborator·

  ·
  4

• Tune MiniMax M3.1 GB300 AgentX disagg / 调优 MiniMax M3.1 GB300 AgentX 分离式配置

  non-canary-full-sweep-enabled
  #4631·xinli-sw opened Oct 29, 2026Collaborator·

  ·
  3

• perf(amd): cap GLM-5.2 FP4 MI355X SGLang TP8 c4 admission at 2 / perf(amd): 将 GLM-5.2 FP4 MI355X SGLang TP8 c4 的准入上限降到 2

  #4630·EricKing626 opened Oct 29, 2026Collaborator·

  ·
  6

• [AMD][MI355X] Add GLM-5.2 MXFP4 AgentX P/D behind the Infera router (infera-sglang) / [AMD][MI355X] 新增 GLM-5.2 MXFP4 AgentX PD 分离部署（Infera router，infera-sglang）

  #4629·llying-001 opened Oct 29, 2026Collaborator·

  ·
  2