Issues

Repository navigation

Issues
Assigned to me
Created by me
Mentioned
Recent activity
Views
Projects
Milestones
Labels
FeedbackPreview
Change the issues type filter
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
1
Search issues
Issues
Open
(160)
Closed
GRPO crashes on Qwen2-VL / Qwen2.5-VL with transformers < 5.3.0 when a completion ends with <|vision_start|>
huggingface/trl#8163 · opened on Nov 6, 2026
1
SFTTrainer keeps ShareGPT roles human/gpt: Qwen3 renders nothing and the train dataset ends up empty
huggingface/trl#8162 · opened on Nov 6, 2026
is_conversational result depends on PYTHONHASHSEED
huggingface/trl#8149 · opened on Nov 6, 2026
CI emits triton UserWarning: TRITON_DEFAULT_FP_FUSION is unset: defaulting to FP fusion ENABLED for backward compatibility in Triton 3.9
huggingface/trl#8136 · opened on Nov 5, 2026
2
CI emits transformers FutureWarning: gradient_checkpointing_enable is deprecated and will be removed in v6. Use activation_checkpointing_enable instead
huggingface/trl#8135 · opened on Nov 5, 2026
1
CI emits transformers FutureWarning: gradient_checkpointing is deprecated and will be removed in v6. Use activation_checkpointing instead
huggingface/trl#8134 · opened on Nov 5, 2026
1
CI fails for experimental tests with dev dependencies: ValueError: An error occurred while trying to load from 'kernels-community/flash-attn3': Failed to refresh TUF metadata.
huggingface/trl#8130 · opened on Nov 5, 2026
2
optimizer_cls_and_kwargs missing from DPO/GRPO/KTO/RLOO trainers (regression of #2302 for RLOO)
huggingface/trl#8127 · opened on Nov 5, 2026
8
Tiny GLM-5.3-Flash stores the mHC base/scale in bf16 while zai-org/GLM-5.3-Flash keeps them in float32
huggingface/trl#8123 · opened on Nov 5, 2026
2
SFT with assistant_only_loss trains the pre-filled thinking block on GLM-4.5 and GLM-4.6 with thinking off
huggingface/trl#8121 · opened on Nov 5, 2026
Training chat templates train tokens that the generation prompt already ends with
huggingface/trl#8109 · opened on Nov 4, 2026
AsyncGRPO fails on Gemma 4: FlashAttention forward only supports head dimension at most 256
huggingface/trl#8104 · opened on Nov 3, 2026
1
SFTTrainer saves output_router_logits=True in the config of the MoE models it trains
huggingface/trl#8102 · opened on Nov 3, 2026
Gemma 4 checkpoints are rejected for assistant-only loss: The chat template is not training-compatible
huggingface/trl#8095 · opened on Nov 3, 2026
GRPO/RLOO under tensor parallelism: TP ranks sample different completions, and the prompt sampler is sized for every process
huggingface/trl#8056 · opened on Nov 1, 2026
SFTTrainer silently skips ModernBERT Decoder's prediction head
huggingface/trl#8052 · opened on Oct 31, 2026
1
Docs: smol-course links in the SFT, DPO and ORPO trainer pages return 404
huggingface/trl#8051 · opened on Oct 31, 2026
1
parse_response puts Qwen3.5 tool calls into reasoning_content when the model skips the prefilled <think>
huggingface/trl#8050 · opened on Oct 30, 2026
2
Proposing to remove OnlineDPOTrainer
huggingface/trl#8042 · opened on Oct 30, 2026
Built-in reward functions crash with the standard dataset format
huggingface/trl#8023 · opened on Oct 29, 2026
1
Async trainers write lora_name into a private attribute that RolloutWorkerProtocol does not declare
huggingface/trl#8008 · opened on Oct 29, 2026
2
[GRPO] vllm_mode="server" + PEFT: last packed chunk is broadcast after unmerge_adapter(), so vLLM receives un-merged (stale) LoRA weights
huggingface/trl#8003 · opened on Oct 29, 2026
1
SFTTrainer.evaluate() repacks prepared BFD datasets and loses document boundaries
huggingface/trl#7999 · opened on Oct 28, 2026
4
FSDP2: reference model is zeros outside local rank 0's shard when cpu_ram_efficient_loading is on
huggingface/trl#7993 · opened on Oct 28, 2026
3
[BUG] GRPO/RLOO continuous batching casts the trained weights to bf16 in place
huggingface/trl#7965 · opened on Oct 28, 2026
2
1 2 3 4 5 6 7 Next

Navigation Menu

Repository navigation

Issues

GRPO crashes on Qwen2-VL / Qwen2.5-VL with transformers < 5.3.0 when a completion ends with `<|vision_start|>`

SFTTrainer keeps ShareGPT roles `human`/`gpt`: Qwen3 renders nothing and the train dataset ends up empty

`is_conversational` result depends on `PYTHONHASHSEED`

CI emits triton UserWarning: TRITON_DEFAULT_FP_FUSION is unset: defaulting to FP fusion ENABLED for backward compatibility in Triton 3.9

CI emits transformers FutureWarning: `gradient_checkpointing_enable` is deprecated and will be removed in v6. Use `activation_checkpointing_enable` instead

CI emits transformers FutureWarning: `gradient_checkpointing` is deprecated and will be removed in v6. Use `activation_checkpointing` instead

CI fails for experimental tests with dev dependencies: ValueError: An error occurred while trying to load from 'kernels-community/flash-attn3': Failed to refresh TUF metadata.

optimizer_cls_and_kwargs missing from DPO/GRPO/KTO/RLOO trainers (regression of #2302 for RLOO)

Tiny GLM-5.3-Flash stores the mHC `base`/`scale` in bf16 while `zai-org/GLM-5.3-Flash` keeps them in float32

SFT with assistant_only_loss trains the pre-filled thinking block on GLM-4.5 and GLM-4.6 with thinking off

Training chat templates train tokens that the generation prompt already ends with

AsyncGRPO fails on Gemma 4: FlashAttention forward only supports head dimension at most 256

SFTTrainer saves `output_router_logits=True` in the config of the MoE models it trains

Gemma 4 checkpoints are rejected for assistant-only loss: The chat template is not training-compatible

GRPO/RLOO under tensor parallelism: TP ranks sample different completions, and the prompt sampler is sized for every process

SFTTrainer silently skips ModernBERT Decoder's prediction head

Docs: smol-course links in the SFT, DPO and ORPO trainer pages return 404

`parse_response` puts Qwen3.5 tool calls into `reasoning_content` when the model skips the prefilled `<think>`

Proposing to remove OnlineDPOTrainer

Built-in reward functions crash with the standard dataset format

Async trainers write `lora_name` into a private attribute that `RolloutWorkerProtocol` does not declare

[GRPO] vllm_mode="server" + PEFT: last packed chunk is broadcast after unmerge_adapter(), so vLLM receives un-merged (stale) LoRA weights

SFTTrainer.evaluate() repacks prepared BFD datasets and loses document boundaries

FSDP2: reference model is zeros outside local rank 0's shard when cpu_ram_efficient_loading is on

[BUG] GRPO/RLOO continuous batching casts the trained weights to bf16 in place