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

QLoRA with fp16=True crashes: trainers cast the adapters to bf16 unconditionally
huggingface/trl#7776 · opened on Oct 16, 2026
SFTTrainer: chat_template_path crashes with vision-language models
huggingface/trl#7775 · opened on Oct 16, 2026
VLM training with vLLM in server mode ignores the images for Gemma3
huggingface/trl#7772 · opened on Oct 16, 2026
LoRA on MoE models with shared experts does not train the dense and shared expert layers with transformers v5 and peft 0.19 to 0.21.2
huggingface/trl#7767 · opened on Oct 16, 2026
1
AsyncGRPO: TokenBudgetBatcher closes a micro-batch even when another row has free room
huggingface/trl#7766 · opened on Oct 16, 2026
LoRA on MoE models does not train the routed experts with transformers v5 and peft<0.19
huggingface/trl#7765 · opened on Oct 16, 2026
1
[Docs] Docs for Async GRPO and the Harbor example still use Qwen3.5, which will crash on startup
huggingface/trl#7761 · opened on Oct 16, 2026
1
Add an AgentEnv integration: trl.experimental.agentenv
huggingface/trl#7759 · opened on Oct 15, 2026
Proposal: audit zero-advantage group filtering under imperfect rewards
huggingface/trl#7747 · opened on Oct 15, 2026
Agree on when to add, drop and change the default Python version
huggingface/trl#7742 · opened on Oct 15, 2026
Add an AMD GPU runner to CI
huggingface/trl#7735 · opened on Oct 14, 2026
1
DiffusionGemma SFT example evicts sampled-block prefix from sliding-window KV cache
huggingface/trl#7734 · opened on Oct 14, 2026
5
GRPO/RLOO on Trainium: each new prompt length bucket recompiles generation for ~12 minutes
huggingface/trl#7733 · opened on Oct 14, 2026
[RFC] Document torchrun as the default launcher
huggingface/trl#7732 · opened on Oct 14, 2026
Agree on when to move the vLLM supported range
huggingface/trl#7722 · opened on Oct 14, 2026
Saving a checkpoint fails on torch 2.6 when generating the model card: UnicodeDecodeError: 'ascii' codec can't decode byte 0xc3
huggingface/trl#7720 · opened on Oct 14, 2026
3
import trl computes importlib.metadata.packages_distributions() eagerly (1-3 s with torch >= 2.14 on Python 3.12)
huggingface/trl#7716 · opened on Oct 14, 2026
3
CheckpointError and wrong gradients in later training tests after a training error on the same test worker
huggingface/trl#7714 · opened on Oct 14, 2026
1
Proposing to remove TPOTrainer
huggingface/trl#7709 · opened on Oct 13, 2026
Iterable datasets with at least as many shards as processes skip and duplicate prompts in multi-process GRPO/RLOO/Distillation/GOLD training
huggingface/trl#7690 · opened on Oct 13, 2026
1
KTO: a paired dataset is unpaired into single-label blocks, so no training batch mixes desirable and undesirable examples
huggingface/trl#7682 · opened on Oct 13, 2026
2
Use reasoning_content in the gpt-oss and LFM2.5 response schemas once transformers aliases reasoning fields
huggingface/trl#7673 · opened on Oct 13, 2026
GRPO/RLOO cache_implementation="static" fails with the default fp32 model: the static cache is fp32, the keys are bf16
huggingface/trl#7664 · opened on Oct 13, 2026
2
torch_compile=True fails on the second step on Trainium: the Neuron dynamo backend can't run the dynamic-shape recompile
huggingface/trl#7660 · opened on Oct 13, 2026
[Tracking] TRL on AWS Trainium (Neuron)
huggingface/trl#7658 · opened on Oct 13, 2026

Navigation Menu

Repository navigation

Issues

Issues

QLoRA with fp16=True crashes: trainers cast the adapters to bf16 unconditionally

SFTTrainer: chat_template_path crashes with vision-language models

VLM training with vLLM in server mode ignores the images for Gemma3

LoRA on MoE models with shared experts does not train the dense and shared expert layers with transformers v5 and peft 0.19 to 0.21.2

AsyncGRPO: TokenBudgetBatcher closes a micro-batch even when another row has free room

LoRA on MoE models does not train the routed experts with transformers v5 and peft<0.19

[Docs] Docs for Async GRPO and the Harbor example still use Qwen3.5, which will crash on startup

Add an AgentEnv integration: `trl.experimental.agentenv`

Proposal: audit zero-advantage group filtering under imperfect rewards

Agree on when to add, drop and change the default Python version

Add an AMD GPU runner to CI

DiffusionGemma SFT example evicts sampled-block prefix from sliding-window KV cache

GRPO/RLOO on Trainium: each new prompt length bucket recompiles generation for ~12 minutes

[RFC] Document torchrun as the default launcher

Agree on when to move the vLLM supported range

Saving a checkpoint fails on torch 2.6 when generating the model card: UnicodeDecodeError: 'ascii' codec can't decode byte 0xc3

import trl computes importlib.metadata.packages_distributions() eagerly (1-3 s with torch >= 2.14 on Python 3.12)

CheckpointError and wrong gradients in later training tests after a training error on the same test worker

Proposing to remove TPOTrainer

Iterable datasets with at least as many shards as processes skip and duplicate prompts in multi-process GRPO/RLOO/Distillation/GOLD training

KTO: a paired dataset is unpaired into single-label blocks, so no training batch mixes desirable and undesirable examples

Use `reasoning_content` in the gpt-oss and LFM2.5 response schemas once transformers aliases reasoning fields

GRPO/RLOO `cache_implementation="static"` fails with the default fp32 model: the static cache is fp32, the keys are bf16

`torch_compile=True` fails on the second step on Trainium: the Neuron dynamo backend can't run the dynamic-shape recompile

[Tracking] TRL on AWS Trainium (Neuron)