CI / Layer 2 / Trusted PR GPU

Workflow runs for CI / Layer 2 / Trusted PR GPU

• [moe] sync and shard expert grads over moe_dp * sp with all_to_all sequence parallelism

  completed successfully:
  CI / Layer 2 / Trusted PR GPU#35:Pull request#6468ready_for_reviewbyJazyJiang
  fix/moe-sp-expert-grad-sync

  Oct 15, 20261h 0m 29s
  More actions for workflow run 35

• [zero] compute the grad norm over all gradient stores at once

  completed successfully:
  CI / Layer 2 / Trusted PR GPU#34:Pull request#6469ready_for_reviewbyJazyJiang
  fix/zero-ep-grad-norm

  Oct 15, 20261h 18m 45s
  More actions for workflow run 34

• [zero] compute the grad norm over all gradient stores at once

  cancelled:
  CI / Layer 2 / Trusted PR GPU#33:Pull request#6469openedbyJazyJiang
  fix/zero-ep-grad-norm

  Oct 15, 202613m 39s
  More actions for workflow run 33

• [moe] sync and shard expert grads over moe_dp * sp with all_to_all sequence parallelism

  cancelled:
  CI / Layer 2 / Trusted PR GPU#32:Pull request#6468openedbyJazyJiang
  fix/moe-sp-expert-grad-sync

  Oct 15, 202614m 46s
  More actions for workflow run 32

• [moe] remove DPGradScaler that over-scales expert grads under moe_dp

  completed successfully:
  CI / Layer 2 / Trusted PR GPU#31:Pull request#6467ready_for_reviewbyJazyJiang
  fix/moe-dp-grad-scale

  Oct 15, 202617m 49s
  More actions for workflow run 31

• [moe] remove DPGradScaler that over-scales expert grads under moe_dp

  completed successfully:
  CI / Layer 2 / Trusted PR GPU#30:Pull request#6467synchronizebypre-commit-ci[bot]Bot
  fix/moe-dp-grad-scale

  Oct 15, 202610m 28s
  More actions for workflow run 30

• [moe] remove DPGradScaler that over-scales expert grads under moe_dp

  cancelled:
  CI / Layer 2 / Trusted PR GPU#29:Pull request#6467openedbyJazyJiang
  fix/moe-dp-grad-scale

  Oct 15, 20261m 25s
  More actions for workflow run 29

• [example]: fix Pytorch_DDP/Pytorch_ZeRO plans crashing in gpt gemini demo

  requires action with the application:
  CI / Layer 2 / Trusted PR GPU#28:Pull request#6445synchronizebypre-commit-ci[bot]Bot
  LiRunGuo:fix/gpt-gemini-demo-ddp-tp-degree

  Oct 14, 2026Action required
  More actions for workflow run 28

• [shardformer]: fix qwen2/qwen3 causal mask with pipeline + sequence parallelism and no flash attention

  requires action with the application:
  CI / Layer 2 / Trusted PR GPU#27:Pull request#6451synchronizebypre-commit-ci[bot]Bot
  LiRunGuo:fix/qwen-pp-sp-attn-mask

  Oct 14, 2026Action required
  More actions for workflow run 27

• [gemini]: fix loading checkpoints of models with tied parameters

  requires action with the application:
  CI / Layer 2 / Trusted PR GPU#26:Pull request#6447synchronizebypre-commit-ci[bot]Bot
  LiRunGuo:fix/gemini-load-tied-weights

  Oct 14, 2026Action required
  More actions for workflow run 26

• [tensor]: fix saving checkpoints of padded parameters with tp_size=1 (keep detach/clone of padded tensors unpaddable)

  requires action with the application:
  CI / Layer 2 / Trusted PR GPU#25:Pull request#6448synchronizebypre-commit-ci[bot]Bot
  LiRunGuo:fix/padded-tensor-detach

  Oct 14, 2026Action required
  More actions for workflow run 25

• [shardformer]: support Qwen3-MoE with expert parallelism

  requires action with the application:
  CI / Layer 2 / Trusted PR GPU#24:Pull request#6450synchronizebypre-commit-ci[bot]Bot
  LiRunGuo:feat/qwen3-moe

  Oct 14, 2026Action required
  More actions for workflow run 24

• [checkpoint]: unpad padded parameters when saving sharded MoE checkpoints

  requires action with the application:
  CI / Layer 2 / Trusted PR GPU#23:Pull request#6449synchronizebypre-commit-ci[bot]Bot
  LiRunGuo:fix/moe-ckpt-unpad

  Oct 14, 2026Action required
  More actions for workflow run 23

• [shardformer]: fall back to native RMSNorm when apex is not installed

  requires action with the application:
  CI / Layer 2 / Trusted PR GPU#22:Pull request#6446synchronizebypre-commit-ci[bot]Bot
  LiRunGuo:fix/fused-rmsnorm-fallback-without-apex

  Oct 14, 2026Action required
  More actions for workflow run 22

• [kernel]: fix Q/K cross-wiring and out-of-bounds accesses in fused_rotary_emb triton kernel

  requires action with the application:
  CI / Layer 2 / Trusted PR GPU#21:Pull request#6444synchronizebypre-commit-ci[bot]Bot
  LiRunGuo:fix/fused-rotary-emb-qk-crosswiring

  Oct 14, 2026Action required
  More actions for workflow run 21

• [shardformer]: support Qwen3-MoE with expert parallelism

  requires action with the application:
  CI / Layer 2 / Trusted PR GPU#20:Pull request#6450synchronizebyLiRunGuo
  LiRunGuo:feat/qwen3-moe

  Oct 14, 2026Action required
  More actions for workflow run 20

• [kernel]: fix Q/K cross-wiring and out-of-bounds accesses in fused_rotary_emb triton kernel

  requires action with the application:
  CI / Layer 2 / Trusted PR GPU#19:Pull request#6444synchronizebyLiRunGuo
  LiRunGuo:fix/fused-rotary-emb-qk-crosswiring

  Oct 14, 2026Action required
  More actions for workflow run 19

• [shardformer]: fall back to native RMSNorm when apex is not installed

  requires action with the application:
  CI / Layer 2 / Trusted PR GPU#18:Pull request#6446synchronizebyLiRunGuo
  LiRunGuo:fix/fused-rmsnorm-fallback-without-apex

  Oct 14, 2026Action required
  More actions for workflow run 18

• [shardformer]: fix qwen2/qwen3 causal mask with pipeline + sequence parallelism and no flash attention

  requires action with the application:
  CI / Layer 2 / Trusted PR GPU#17:Pull request#6451synchronizebyLiRunGuo
  LiRunGuo:fix/qwen-pp-sp-attn-mask

  Oct 14, 2026Action required
  More actions for workflow run 17

• [gemini]: fix loading checkpoints of models with tied parameters

  requires action with the application:
  CI / Layer 2 / Trusted PR GPU#16:Pull request#6447synchronizebyLiRunGuo
  LiRunGuo:fix/gemini-load-tied-weights

  Oct 14, 2026Action required
  More actions for workflow run 16

• [tensor]: fix saving checkpoints of padded parameters with tp_size=1 (keep detach/clone of padded tensors unpaddable)

  requires action with the application:
  CI / Layer 2 / Trusted PR GPU#15:Pull request#6448synchronizebyLiRunGuo
  LiRunGuo:fix/padded-tensor-detach

  Oct 14, 2026Action required
  More actions for workflow run 15

• [checkpoint]: unpad padded parameters when saving sharded MoE checkpoints

  requires action with the application:
  CI / Layer 2 / Trusted PR GPU#14:Pull request#6449synchronizebyLiRunGuo
  LiRunGuo:fix/moe-ckpt-unpad

  Oct 14, 2026Action required
  More actions for workflow run 14

• [ci]: Run GPU CI in Torch 2.13/CUDA 13.0 containers

  completed successfully:
  CI / Layer 2 / Trusted PR GPU#13:Pull request#6463synchronizebykiseki-zj
  codex/torch213-ci-image

  Oct 13, 20261h 29m 2s
  More actions for workflow run 13

• [ci]: Run GPU CI in Torch 2.13/CUDA 13.0 containers

  cancelled:
  CI / Layer 2 / Trusted PR GPU#12:Pull request#6463ready_for_reviewbykiseki-zj
  codex/torch213-ci-image

  Oct 13, 20263m 24s
  More actions for workflow run 12

• CI / Layer 2 / Trusted PR GPU

  completed successfully:
  CI / Layer 2 / Trusted PR GPU#11:Manually runbykiseki-zj
  codex/torch213-ci-image

  Oct 13, 202618m 35s
  More actions for workflow run 11