Repository Issue Activity (beta)

#3865 [RL/DeepEP] Generator OOMs in vLLM cudagraph capture on memory-tight MoE (expandable_segments incompatible with DeepEP; gpu_memory_utilization is a no-op)

#613 Ability to train based on epoch

#2936 TODO Debt Report

#3780 [graph_trainer] Graph-PP RFC

#3775 transformers_modeling_backend: pretrained dense models produce near-random loss / crash on current PyTorch

#3699 [GraphTrainer] [AutoParallel] Tracker for outstanding issues

#3122 torch.compile + DTensor: backward crash on TP-parallelized sub-module with a param-shadowing pre-hook

#3816 GraphTrainer DSv3-16B: healthy single-node, MI355 FSDP EP=2 parity gap

#3439 HybridEP DispatchHandle not working with torch.compile

#3846 `sys.argv` frozen at torchtitan import time

#3830 DeepEP v2 doesn't work with per-layer compile

#3820 RFC CUDAGraphable GraphPP

#3817 TP + MXFP8 crashing with "mat2 must be col_major, got stride (4096, 1)" error

#2720 Change default attn backend to varlen/flex attention

#3374 [TensorParallel] NoParallel does not handle multi-output modules

#3434 RFC: Graph-Based EP Overlap For GraphTrainer

#3709 [rl] vLLM cudagraph: FULL corrupts generation, FULL_AND_PIECEWISE unsupported with the varlen attention backend

#3786 Clarify or revisit Qwen3 default seq_len / RoPE max_seq_len of 4096

#3790 Default `optimizer.implementation = "fused"` can OOM long FSDP2 runs due to upstream DTensorSpec leak

#3438 `LoRAConverter` silently leaves non-Linear parameters trainable

#3756 [RL] Support Qwen3.5 in the RL experiment

#3761 8 GPU Model Tests red on main: gpt-oss PP+FSDP+EP+SACOP MoE _grouped_mm index OOB (CUDA scatter assert == ROCm HSA 0x1016)

#3735 [RL] No mid-run checkpoint/resume in the RL training loop -- preemption loses all progress

#3755 [rl] StickySessionRoutingStrategy can overload a generator for long multi-turn rollouts (no KV-budget awareness / no rebalancing)

#3752 vllm v2 model runner does not work with VLLMGenerator DP routing