Repository navigation

Issues

New issue

[Tracking] Move off liger-kernel and make the memory-efficient loss the default
#7063 · qgallouedec opened on Sep 4, 2026
8
[Tracking] TRL on AWS Trainium (Neuron)
#7658 · qgallouedec opened on Oct 13, 2026
[Tracking] Promote AsyncGRPO to the stable API
#7841 · qgallouedec opened on Oct 20, 2026

from trl import GRPOTrainer fails without a visible GPU with bitsandbytes 0.45.1
huggingface/trl#7863 · opened on Oct 21, 2026
AsyncGRPO multi-turn tool trajectories can mix policy versions and falsely drop fresh groups
huggingface/trl#7861 · opened on Oct 21, 2026
Online DPO VLM training fails for Qwen2-VL and Qwen3-VL: ValueError: mm_token_type_ids is missing
huggingface/trl#7858 · opened on Oct 21, 2026
AsyncDistillation rollout records wrong policy version if a publish races with prompt preparation
huggingface/trl#7857 · opened on Oct 21, 2026
The minimum-versions CI job does not test deepspeed at its floor
huggingface/trl#7854 · opened on Oct 21, 2026
Test every dependency floor at that floor
huggingface/trl#7853 · opened on Oct 21, 2026
Bug: SFTTrainer unconditionally casts QLoRA trainable params to bfloat16, causing GradScaler crash when fp16=True
huggingface/trl#7851 · opened on Oct 21, 2026
3
Tests workflows run twice for PRs whose head branch starts with ci-
huggingface/trl#7847 · opened on Oct 21, 2026
[Tracking] Promote AsyncGRPO to the stable API
huggingface/trl#7841 · opened on Oct 20, 2026
AsyncGRPO: two files, async_grpo_config.py and async_grpo_trainer.py
huggingface/trl#7840 · opened on Oct 20, 2026
AsyncGRPO metrics: sample/ and the rollout queue
huggingface/trl#7839 · opened on Oct 20, 2026
AsyncGRPO metrics: tools/
huggingface/trl#7838 · opened on Oct 20, 2026
AsyncGRPO metrics: rollout/
huggingface/trl#7837 · opened on Oct 20, 2026
AsyncGRPO metrics: suffix conventions
huggingface/trl#7836 · opened on Oct 20, 2026
AsyncGRPO metrics: same names as GRPO for the same quantities
huggingface/trl#7835 · opened on Oct 20, 2026
AsyncGRPO: rename heartbeat_stale_after_s to rollout_worker_timeout
huggingface/trl#7834 · opened on Oct 20, 2026
AsyncGRPO: one timeout, vllm_server_timeout
huggingface/trl#7833 · opened on Oct 20, 2026
AsyncGRPO: remove fork_threshold_tokens
huggingface/trl#7832 · opened on Oct 20, 2026
AsyncGRPO: rename max_inflight_tasks to max_concurrent_rollouts
huggingface/trl#7831 · opened on Oct 20, 2026
AsyncGRPO: remove queue_maxsize
huggingface/trl#7830 · opened on Oct 20, 2026
AsyncGRPO: token_budget modes and what per_device_train_batch_size means
huggingface/trl#7829 · opened on Oct 20, 2026
AsyncGRPO: drop the weight_transfer argument
huggingface/trl#7828 · opened on Oct 20, 2026
AsyncGRPO: drop the rollout_worker argument
huggingface/trl#7827 · opened on Oct 20, 2026
AsyncGRPO: add quantization_config
huggingface/trl#7826 · opened on Oct 20, 2026
GRPO and RLOO continuous batching size the KV cache to the GPU instead of the batch, making training ~4x slower
huggingface/trl#7824 · opened on Oct 20, 2026
1

Navigation Menu

Repository navigation

Issues

Issues

`from trl import GRPOTrainer` fails without a visible GPU with bitsandbytes 0.45.1

AsyncGRPO multi-turn tool trajectories can mix policy versions and falsely drop fresh groups

Online DPO VLM training fails for Qwen2-VL and Qwen3-VL: ValueError: mm_token_type_ids is missing

AsyncDistillation rollout records wrong policy version if a publish races with prompt preparation

The minimum-versions CI job does not test deepspeed at its floor

Test every dependency floor at that floor

Bug: SFTTrainer unconditionally casts QLoRA trainable params to bfloat16, causing GradScaler crash when fp16=True

Tests workflows run twice for PRs whose head branch starts with `ci-`

[Tracking] Promote AsyncGRPO to the stable API

AsyncGRPO: two files, `async_grpo_config.py` and `async_grpo_trainer.py`

AsyncGRPO metrics: `sample/` and the rollout queue

AsyncGRPO metrics: `tools/`

AsyncGRPO metrics: `rollout/`

AsyncGRPO metrics: suffix conventions

AsyncGRPO metrics: same names as GRPO for the same quantities

AsyncGRPO: rename `heartbeat_stale_after_s` to `rollout_worker_timeout`

AsyncGRPO: one timeout, `vllm_server_timeout`

AsyncGRPO: remove `fork_threshold_tokens`

AsyncGRPO: rename `max_inflight_tasks` to `max_concurrent_rollouts`

AsyncGRPO: remove `queue_maxsize`

AsyncGRPO: `token_budget` modes and what `per_device_train_batch_size` means

AsyncGRPO: drop the `weight_transfer` argument

AsyncGRPO: drop the `rollout_worker` argument

AsyncGRPO: add `quantization_config`

GRPO and RLOO continuous batching size the KV cache to the GPU instead of the batch, making training ~4x slower