Search results

Qwen3.5 FSDP GRPO with vLLM 0.18.0 / torch 2.10 requires flash-attn, but no compatible flash-attn wheel is available

Status: Open.

#6662 In verl-project/verl;

· nadphwr opened on Jun 9, 2026

Disable "use_mbridge" in version 0.8.x

Ascend

Status: Open.

#6658 In verl-project/verl;

· geolvr opened on Jun 8, 2026

triton 版本是多少？

bugSomething isn't workingSomething isn't working

Status: Open.

#6645 In verl-project/verl;

· cqray1990 opened on Jun 7, 2026

bug when training gemma-4

bugSomething isn't workingSomething isn't working

Status: Open.

#6613 In verl-project/verl;

· congdama opened on Jun 5, 2026

[Megatron][MoE] grad_norm explosion under context parallel (CP>1): loss_func must return 3-tuple when calculate_per_token_loss=True

Status: Open.

#6609 In verl-project/verl;

· insightLLM opened on Jun 4, 2026

hccl debug

bugSomething isn't workingSomething isn't working

Status: Open.

#6608 In verl-project/verl;

· code10244096 opened on Jun 4, 2026

[Roadmap] verl on AMD/ROCm — 26Q2/Q3

Status: Open.

#6590 In verl-project/verl;

· PeterYang12 opened on Jun 3, 2026

升级VLLM镜像后，reward model报错

bugSomething isn't workingSomething isn't working

Status: Open.

#6584 In verl-project/verl;

· jianhai0527 opened on Jun 3, 2026

[rollout, vllm] EngineCore "RPC call to sample_tokens timed out" persists in long agent-loop rollout

bugSomething isn't workingSomething isn't working

Status: Open.

#6583 In verl-project/verl;

· tongruiliu opened on Jun 3, 2026

Abnormal Rollout based on vllm_ascend with Ascend 910B when training Qwen3-30B-A3B-Instruct-2507

bugSomething isn't workingSomething isn't working

Status: Open.

#6580 In verl-project/verl;

· xkazm opened on Jun 3, 2026

[RFC]Dynamo KV-aware Routing, kv cache migration for Elastic scheduling in async RL, and KVBM Multi-tier KV Cache

Status: Open.

#6577 In verl-project/verl;

· sophiayyya opened on Jun 2, 2026

Add Megatron-LM precision options

Status: Open.

#6576 In verl-project/verl;

· kolehma8 opened on Jun 2, 2026