Repository Issue Activity (beta)

#3883 Opt out of upcoming SAC × saved_tensors_hooks change (pytorch/pytorch#190581)

#3973 Qwen3.5 monkeypatch `get_cu_seqlens` crashes under sample_packing: `.view(-1)` on a non-contiguous tensor

#3967 Expose FP8 Scaling Recipe (Rowwise) Instead of Hardcoded Tensorwise

#3981 Add support for new flash-attn (and kernel) interface in ring-flash-attention

#3972 Qwen3.5 monkeypatch reads `layer_type` but transformers 5.14.1 defines `block_type` (AttributeError in first forward pass)

#3971 Async GRPO fails with `num_tiles` when data producer and prefetch are enabled

#2650 Deprecated `flash_attn_fuse_qkv` toggle for the latest transformers (4.51.3 or lower)

#3955 [New model] Add Qwen4

#3952 nemotron_h sample-packing monkeypatch fails on both transformers 5.14.1 and 5.15.0 (the versions in the recent official images)

#3951 nemotron_h: 4-bit quantization packs `out_proj`; the mamba fused kernel reads `out_proj.weight` directly and crashes at step 0 (falcon_h1 already has this skip)

#3959 `vllm-serve` ignores `revision_of_model` and loads the default Hugging Face main branch

#3940 TokensPerSecondCallback captures resume_from_checkpoint before auto-resume detection resolves it, silently zeroing token metrics

#3203 OOM for causal lm evaluation and missing logging

#3836 check_batch_size_fields rejects configs that rely on the documented micro_batch_size/gradient_accumulation_steps defaults

#3754 `chat_template` Strategy Never Trains the Gemma Turn Terminator

#3890 model support system follow-ups

#3626 [Feature] Automatic LoRA rank recommendation based on dataset size

#3848 MultipackBatchSampler.generate_batches() stale-reference bug corrupts total_token_slots and can crash with IndexError

#3801 PEFT bump pending

#3908 MLflow/WandB/Comet/Trackio env vars not set on Ray Train worker (use_ray: true)

#3608 Ring Attention w/ document packing produces different results

#3856 kks

#3857 asj

#3787 Expert-granularity CPU offload for quantized MoE: 30B-A3B QLoRA in ~7 GB (single resident expert layer, extends the layer_offloading idea)

#2396 EXTREMELY SLOW (unusable) towards end of tokenization of dataset with long multi turn conversations