Search results

[FSDP2] GRPO reproducibly crashes at step 11 — "ACL stream synchronize failed, error code:507035" (device-side AIV MTE out-of-range) on Atlas 800T A3; FSDP1 stable for 40 steps

Status: Open.

#6894 In verl-project/verl;

· TengTu0311 opened on Jun 29, 2026

Clarify parameter weight storage dtype in FP8 end-to-end training

Status: Open.

#6893 In verl-project/verl;

· Eisenhower opened on Jun 29, 2026

Crash in aclnnConvolutionBackwardGetWorkspaceSize during convolution backward

Ascend

bugSomething isn't workingSomething isn't working

Status: Open.

#6885 In verl-project/verl;

· llllchao opened on Jun 29, 2026

[Feature request] Per-request abort for rollout LLMServerClient

Status: Open.

#6866 In verl-project/verl;

· cr-gao opened on Jun 27, 2026

[docker] Recommended image / Dockerfile for DeepSeek-V4 + DSA on the Megatron-Lite backend (B200 / H200)?

Status: Open.

#6864 In verl-project/verl;

· spacegoing opened on Jun 27, 2026

[Bug] vLLM rollout DP>1 fails when a RolloutReplica spans multiple nodes

bugSomething isn't workingSomething isn't working

Status: Open.

#6856 In verl-project/verl;

· MakeKJ opened on Jun 26, 2026

Multi-turn GRPO with thinking models: how do you handle the chat template stripping prior-turn <think> from history?

Status: Open.

#6854 In verl-project/verl;

· dipta007 opened on Jun 26, 2026

GRPO training wtih lora using qwen2.5-7b-vl tain_batch_size=1,mini_batch_size=1,and micro_mini_batch_size_per_gpu=1

bugSomething isn't workingSomething isn't working

Status: Open.

#6851 In verl-project/verl;

· cqray1990 opened on Jun 25, 2026

[Bug] FlashInfer FP16 MoE produces corrupted rollouts with Qwen3-Coder-30B-A3B in multi-node colocate-async

bugSomething isn't workingSomething isn't working

Status: Open.

#6847 In verl-project/verl;

· HanlinDu opened on Jun 25, 2026

process_validation_metrics() crashes on None-filled sparse reward_extra_info keys (metrics emitted for only some samples)

bugSomething isn't workingSomething isn't working

help wantedExtra attention is neededExtra attention is needed

Status: Open.

#6830 In verl-project/verl;

· giladfrid009 opened on Jun 24, 2026

Qwen3.5-4b ppo_kl != 0

bugSomething isn't workingSomething isn't working

Status: Open.

#6829 In verl-project/verl;

· dipta007 opened on Jun 23, 2026

[Feature] On-Policy Self-Distillation (OPSD): privileged-context self-teacher

Status: Open.

#6827 In verl-project/verl;

· HaozheZhang6 opened on Jun 23, 2026