ART · Issues· 143 open

Investigate output and gradient variability between matched same-source Qwen TrainerRank actors

TrainerRank reusable-cache admission can leave insufficient physical CUDA memory for backward library allocations

TrainerRank MoE memory admission: backward-recompute OOM and profile calibration failures

Support combined expert and data parallelism in TrainerRank

Support pipeline parallelism in TrainerRank without exposing stage scheduling to callers

Investigate large hidden-state differences when independent requests are co-packed (Qwen3.6, no-grad)

Retained sampled suffix can carry logprobs under a different causal prefix

Determine whether a large dynamics RL batch refusal is necessary or avoidable by execution layout

Feature: Dr. GRPO Support

Qwen3.5-35B-A3B at EP1 with CP>1 segfaults in Transformer Engine's grouped GEMM on real-data routing (likely empty local experts)

Dedicated LocalBackend vLLM server can become unreachable during LoRA adapter reload

Regression: forked GRPO run can produce malformed checkpoint outputs on newer ART ref

LoRA adapters are not properly syncing when using LocalBackend - stale adapters are being used in rollouts

LocalBackend fork_checkpoint doesn't update vLLM's initial LoRA

Add from_entity parameter to _experimental_fork_checkpoint