All pull requests

Pull requests

• Add DSv4 ATOM 1P1D c256 EP8 Mega + LMCache MP point / 新增 DSv4 ATOM 1P1D c256 EP8 Mega + LMCache MP 测试点

  full-sweep-enabled
  #3831·yhl-amd opened Oct 8, 2026Collaborator·

  ·
  1

• [AMD][Qwen3.5] Update MI355X FP8 AgentX image to v0.5.21 and add TP8

  full-sweep-fail-fast
  #3830·yichiche opened Oct 8, 2026Collaborator·

  ·
  2

• [Klaud Cold] Update qwen3.5-fp4-gb300-dynamo-trt-mtp TensorRT-LLM image to 1.3.0rc29 / [Klaud Cold] 将 qwen3.5-fp4-gb300-dynamo-trt-mtp 的 TensorRT-LLM 镜像更新为 1.3.0rc29

  #3827·Klaud-Cold opened Oct 8, 2026Collaborator·

  ·
  3

• [Klaud Cold] Update qwen3.5-fp4-gb300-dynamo-sglang SGLang image to v0.5.21-cu130 / [Klaud Cold] 将 qwen3.5-fp4-gb300-dynamo-sglang 的 SGLang 镜像更新至 v0.5.21-cu130

  full-sweep-fail-fast
  #3825·Klaud-Cold opened Oct 8, 2026Collaborator·

  ·
  3

• refactor(recipes): fold per-point recipe families into one bundle per config key

  #3824·adibarra opened Oct 8, 2026Collaborator·8/8·

  ·
  1

• refactor(srt): bind script-installed component versions from the master config

  #3821·adibarra opened Oct 8, 2026Collaborator·7/8·

  ·
  1

• refactor(srt): take cluster fabric values from runners.yaml via @fabric refs

  #3820·adibarra opened Oct 8, 2026Collaborator·6/8·

  ·
  1

• perf(amd): DSv4-Pro ATOM AgentX on MegaMoE DEP with LMCache MP on the 1007 nightly / DSv4-Pro ATOM AgentX 改用 MegaMoE DEP 并在 1007 nightly 启用 LMCache MP

  agentxAMDfull-sweep-fail-fast
  #3819·yhl-amd opened Oct 8, 2026Collaborator·

  ·
  4

• refactor(srt): compose AgentX recipes and select power from the master row

  #3818·adibarra opened Oct 8, 2026Collaborator·4/8·

  ·
  1

• Add B300 Dynamo+SGLang AgentX configs for DeepSeek V4 with NIXL KV transfer / 添加使用 NIXL KV 传输的 DeepSeek V4 B300 Dynamo+SGLang AgentX 配置

  full-sweep-fail-fast
  #3817·nvpohanh opened Oct 8, 2026Collaborator·7 of 11 tasks·

  ·
  3

• feat(config): add Qwen3.8 GB300 AgentX / feat(config): 新增 Qwen3.8 GB300 AgentX 配置

  full-sweep-enabled
  #3816·RohitNagraj opened Oct 8, 2026Collaborator·6 of 11 tasks·

  ·
  2

• [Klaud Cold] Update dsr1-fp8-mi325x-sglang-mtp SGLang image to v0.5.21-rocm720-mi30x / [Klaud Cold] 将 dsr1-fp8-mi325x-sglang-mtp 的 SGLang 镜像更新至 v0.5.21-rocm720-mi30x

  full-sweep-fail-fast
  #3814·Klaud-Cold opened Oct 8, 2026Collaborator·

  ·
  9

• [Klaud Cold] Update qwen3.5-fp4-b200-sglang SGLang image to v0.5.21-cu130 / [Klaud Cold] 将 qwen3.5-fp4-b200-sglang 的 SGLang 镜像更新至 v0.5.21-cu130

  full-sweep-fail-fast
  #3813·Klaud-Cold opened Oct 8, 2026Collaborator·

  ·
  7

• feat(matrix): fingerprint the concrete recipe each row runs and list affected config keys

  #3812·adibarra opened Oct 8, 2026Collaborator·5/8·

  ·
  1

• refactor(srt): compose fixed-sequence recipes from fragments and bind the matrix point

  #3809·adibarra opened Oct 8, 2026Collaborator·3/8·

  ·
  1

• feat(launch): record one job_event.json per benchmark job

  #3807·adibarra opened Oct 7, 2026Collaborator·

  ·

• feat(results): add versioned result row contract with collector quarantine

  #3806·adibarra opened Oct 7, 2026Collaborator·

  ·

• refactor(matrix): select srt-slurm recipes with srt-recipe-dir and srt-recipe

  #3805·adibarra opened Oct 7, 2026Collaborator·2/8·

  ·
  1

• fix(launch): cap single-node srun steps at their GPUs' share of node DRAM

  #3804·adibarra opened Oct 7, 2026Collaborator·

  ·
  1

• docs: add perf-changelog entries only when collecting results

  #3803·adibarra opened Oct 7, 2026Collaborator·

  ·
  1

• chore(recipes): delete unreferenced srt-slurm recipes

  #3802·adibarra opened Oct 7, 2026Collaborator·1/8·

  ·
  1

• Update B300 MiniMax-M3 vLLM and MSA decode / 更新 B300 MiniMax-M3 vLLM 和 MSA decode

  full-sweep-fail-fast
  #3801·xinli-sw opened Oct 7, 2026Collaborator·

  ·
  4

• Enable B200 MiniMax-M3 CUTLASS MSA / 启用 B200 MiniMax-M3 CUTLASS MSA

  full-sweep-fail-fast
  #3800·xinli-sw opened Oct 7, 2026Collaborator·

  ·
  2

• [PowerX] tidy the #3781 power consumers and trim their driver test / [PowerX] 整理 #3781 功耗消费端并精简其 driver 测试

  #3798·edwingao28 opened Oct 7, 2026Collaborator·4 of 11 tasks·

  ·

• feat(evals): add CRUXEval lm-eval tasks loadable through EVAL_TASKS_DIR

  #3790·Oseltamivir opened Oct 7, 2026Collaborator·

  ·
  6