Features

Even more training chat templates

Three more model families gain training-compatible templates with {% generation %} markers (so assistant_only_loss=True just works):

• Phi-3.5 by @DagaBhai in #5746
• Qwen3-VL by @aazizyan in #5764
• Qwen3.5 Think / NoThink by @aazizyan in #5824

Final logits softcapping for async GRPO

The chunked LM-head path used by AsyncGRPOTrainer now supports models that use final_logit_softcapping (notably Gemma 2). _ChunkedLogProbFunction applies logit_scale, optional tanh-based softcapping, and temperature consistently in both forward and backward — softcapped models are no longer rejected.

by @mlarnouhet in #5691

KTO ↔ DPO alignment continues

Two more cycles closer to KTO graduation:

• Align compute_loss flow by @albertvillanova in #5810
• Align _compute_loss_liger flow by @albertvillanova in #5816

Trainer telemetry (opt-out)

_BaseTrainer.__init__ now emits a single anonymous huggingface_hub.send_telemetry ping per trainer instantiation, so we can finally see which trainers / model families / distributed backends are actually being used in practice and prioritize accordingly.

The payload is intentionally minimal — TRL version, trainer class name, model architecture, PEFT yes/no, distributed backend (deepspeed/fsdp/ddp/none), bucketed world size, device type, GPU model when available. No user data, no dataset names, no model paths, no hyperparameter values, never sent in CI / offline / HF_HUB_DISABLE_TELEMETRY mode.

See usage_stats.md for what's collected and how to opt out.

by @qgallouedec in #5758

Other

• OpenRewardSpec: fix omitting task-scoped tools during rollout binding (fixes #5727) by @rycerzes in #5729
• Add OpenReward example to the list of examples by @sergiopaniego in #5752
• Add DDP-2 members to invariant test suite by @qgallouedec in #5736
• Align and simplify the stable training scripts by @qgallouedec in #5812
• Replace uv installation script with setup action by @qgallouedec in #5735

Fixes

• Fix exponential backtracking in qwen3 / qwen3_5 / glm4moe response parsing — GRPOTrainer was hanging indefinitely on truncated <tool_call> blocks (a degenerate case that happens naturally when generation hits max_completion_length mid-tool-call). Rewrote the regex to be non-backtracking — worst case goes from O(2ⁿ) to O(n). By @xodn348 in #5798
• CUDA memory leak: release BNB dequantization buffers & stale state in OffloadActivations — follow-up to v1.4's activation-offloading leak fix. By @butterwecksolutions in #5730
• Invalidate ZeRO-3 param coordinator trace in add_hooks by @roycho96 in #4693
• Fix nested vocab_size for DistillationTrainer and GOLDTrainer by @Beichen-Ma in #5592
• Fix MPS support in experimental empty_cache() by @jamie-peterson-ml in #5799
• Fix metric_for_best_model for trainer-specific eval metrics by @qgallouedec in #5811
• Fix generate_batch: inference tensors blocking inplace ops in background thread by @albertvillanova in #5818
• Replace deprecated torch_dtype with dtype across examples, docs, notebooks, tests, and experimental distillation / gold trainers by @qgallouedec in #5717

Documentation and Examples

• docs(grpo): align model to Qwen2.5 and add GRPO OOM tab in quickstart by @xodn348 in #5740

CI

• Migrate tests to Qwen3.5 Think/NoThink fixtures + tiny-model generation scripts by @aazizyan in #5819 and #5821
• Align tiny Glm4MoeForCausalLM / Cohere / Cohere2 / Qwen2.5-VL configs with their reference models by @qgallouedec in #5638, #5706, #5707 and #5739
• Fix tiny Qwen3-VL deepstack_visual_indexes and drop the test skip by @qgallouedec in #5779
• Fix tiny Qwen2.5-VL fullatt_block_indexes out of range for depth=2 by @albertvillanova in #5805
• Remove non-existent params from tiny Qwen2-VL model by @albertvillanova in #5795
• Fix vision config num_heads key in Qwen VL tiny model scripts by @matdou in #5792
• Drop unjustified model.visual. skip in GRPO/RLOO Qwen2.5-VL tests by @qgallouedec in #5780
• Make the LLaVA / LLaVA-Next test guard explicit by @qgallouedec in #5778
• Remove obsolete Gemma 3 vision-head guard from VLM training tests by @qgallouedec in #5772
• Fix OOM in CI: reduce batch size in VLM SFT / GRPO/RLOO VLM / toolcall tests by @albertvillanova in #5687, #5767, #5801
• Fix OOM in CI by clearing chained exception tracebacks by @albertvillanova in #5776
• Fix OOM in CI by reducing intermediate_size and image token budget for tiny Gemma 4 by @albertvillanova in #5760
• Fix CI errors in response parsing for gpt-oss/llama with transformers v5 by @albertvillanova in #5755
• Fix CI AttributeError: 'GptOssConfig' object has no attribute 'num_experts' by @albertvillanova in #5756
• Fix CI apply_model_revisions by removing _commit_hash kwarg by @albertvillanova in #5762
• Fix CI test to avoid skipping model.visual params by @albertvillanova in #5806
• Fix transformers min version for tiny gemma 4 as 5.5.0 by @albertvillanova in #5763
• Hotfix CI: pin torch < 2.12.0 (later reverted) by @albertvillanova in #5769
• Fix catch-all empty string in Makefile pytest --only-rerun by @albertvillanova in #5784
• chore: update tests_latest.yml by @hf-security-analysis[bot] in #5733

New Contributors

• @hf-security-analysis[bot] made their first contribution in #5733
• @Beichen-Ma made their first contribution in #5592
• @DagaBhai made their first contribution in #5746
• @xodn348 made their first contribution in #5740
• @mlarnouhet made their first contribution in #5691
• @matdou made their first contribution in #5792
• @jamie-peterson-ml made their first contribution in #5799
• @rycerzes made their first contribution in #5729

What's Changed

• ⬆️ Bump dev version by @qgallouedec in #5734
• chore: update tests_latest.yml by @hf-security-analysis[bot] in #5733
• fix: CUDA memory leak / release BNB dequantization buffers & stale state in OffloadActivations by @butterwecksolutions in #5730
• fix: invalidate ZeRO-3 param coordinator trace in add_hooks by @roycho96 in #4693
• Fix nested vocab_size for DistillationTrainer and GOLDTrainer by @Beichen-Ma in #5592
• feat: add Phi-3.5 training chat templates with generation markers by @DagaBhai in #5746
• docs(grpo): align model to Qwen2.5 and add GRPO OOM tab in quickstart by @xodn348 in #5740
• torch_dtype -> dtype by @qgallouedec in #5717
• Add OpenReward example to the list of examples by @sergiopaniego in #5752
• Fix CI errors in response parsing for gptoss/llama with transformers v5 by @albertvillanova in #5755
• Add DDP-2 members to invariant test suite by @qgallouedec in #5736
• Hotfix CI param not updated AssertionError: Pin torch < 2.12.0 by @albertvillanova in #5769
• Align tiny-Glm4MoeForCausalLM with GLM-4.5 reference config by @qgallouedec in #5638
• Align tiny Cohere config with aya-expanse-8b by @qgallouedec in #5706
• Align tiny Cohere2 config with tiny-aya-earth by @qgallouedec in #5707
• Fix OOM in CI by reducing intermediate_size and image token budget for tiny Gemma4 by @albertvillanova in #5760
• Fix CI AttributeError: 'GptOssConfig' object has no attribute 'num_experts' by @albertvillanova in #5756
• Fix CI apply_model_revisions by removing _commit_hash kwarg by @albertvillanova in #5762
• Remove obsolete Gemma3 vision-head guard from VLM training tests by @qgallouedec in #5772
• Replace uv installation script with setup action by @qgallouedec in #5735
• Fix OOM in CI by clearing chained exception tracebacks by @albertvillanova in #5776
• Fix transformers min version for tiny gemma4 as 5.5.0 by @albertvillanova in #5763
• Final logits softcapping support for async GRPO Trainer by @mlarnouhet in #5691
• Fix vision config num_heads key in Qwen VL tiny model scripts and revert torch pin by @matdou in #5792
• Drop unjustified model.visual. skip in GRPO / RLOO Qwen2.5-VL tests by @qgallouedec in #5780
• Fix OOM in CI by reducing batch size and sequence length for toolcall tests by @albertvillanova in #5801
• Fix exponential backtracking in qwen3 / qwen3_5 / glm4moe response parsing by @xodn348 in #5798
• Add telemetry to trainers by @qgallouedec in #5758
• Add Qwen3-VL training chat template with generation markers by @aazizyan in #5764
• Align tiny Qwen2.5-VL with Qwen/Qwen2.5-VL-3B-Instruct by @qgallouedec in #5739
• Fix tiny Qwen3-VL deepstack_visual_indexes and drop the test skip by @qgallouedec in #5779
• Fix OOM in CI by reducing batch size in GRPO/RLOO VLM tests by @albertvillanova in #5767
• Fix catch-all empty string in Makefile pytest --only-rerun by @albertvillanova in #5784
• Remove non-existent params from tiny Qwen2-VL model by @albertvillanova in #5795
• Fix tiny Qwen2.5-VL fullatt_block_indexes out of range for depth=2 by @albertvillanova in #5805
• Make the LLaVA / LLaVA-Next test guard explicit by @qgallouedec in #5778
• Fix MPS support in experimental empty_cache() by @jamie-peterson-ml in #5799
• Fix CI test to avoid skipping model.visual params by @albertvillanova in #5806
• Align KTO with DPO: Align compute_loss flow by @albertvillanova in #5810
• Fix generate_batch: inference tensors block inplace ops in background thread by @albertvillanova in #5818
• Fix metric_for_best_model for trainer-specific eval metrics by @qgallouedec in #5811
• Align and simplify the stable training scripts by @qgallouedec in #5812
• Align KTO with DPO: Align _compute_loss_liger flow by @albertvillanova in #5816
• Add tiny Qwen3.5 Think/NoThink fixture generation scripts by @aazizyan in #5819
• Migrate tests to Qwen3.5 Think/NoThink fixtures by @aazizyan in #5821
• Fix OpenRewardSpec omitting task‑scoped tools during rollout binding (fixes #5727) by @rycerzes in #5729
• Add Qwen3.5 Think/NoThink training chat templates with generation markers by @aazizyan in #5824
• Release: v1.5 by @qgallouedec in #5835

Full Changelog: v1.4.0...v1.5.0