Update transformers to v5.x, unsloth, and add MoE LoRA conversion

[Kovbo]

Summary

• Update transformers from v4.57.x to v5.x (>=5.1.0)
• Update unsloth from 2025.12.9 to 2026.2.1 and unsloth-zoo from VCS fork to PyPI 2026.2.1
• Update peft from >=0.14.0 to >=0.18.0
• Add MoE LoRA adapter conversion utility for vLLM compatibility
• Fix multiple transformers v5 breaking changes (import paths, apply_chat_template return types, removed warnings_issued, deprecated TrainerArgs fields)

Changes

Dependencies (pyproject.toml, requirements/backend.vcs.txt)

• transformers>=5.1.0 (was >=4.55.2,<=4.57.3)
• unsloth==2026.2.1 (was 2025.12.9)
• unsloth-zoo==2026.2.1 via PyPI (was VCS pin to bradhilton fork)
• peft>=0.18.0 (was >=0.14.0)
• Added override-dependencies for transformers and trl to bypass unsloth's overly strict PyPI constraints (confirmed working per unsloth Feb 2026 release notes)

Code fixes

• src/art/unsloth/service.py: Fix import path (GenerationMixin, PreTrainedModel now directly from transformers instead of transformers.utils.dummy_pt_objects). Add warnings_issued compat shim. Integrate convert_checkpoint_if_needed calls.
• src/art/preprocessing/tokenize.py: Add return_dict=False to apply_chat_template calls (v5 default changed from list[int] to BatchEncoding)
• src/art/transformers/patches.py: Update return type for v5's _preprocess_mask_arguments (now returns 5 values instead of 4)
• src/art/dev/model.py: Remove deprecated TrainerArgs fields
• src/art/tinker/server.py: Add return_dict=False to apply_chat_template

New files

• src/art/utils/convert_moe_lora.py: Converts fused MoE LoRA adapters (produced by unsloth + transformers v5) to per-expert format for vLLM compatibility. Runs automatically after checkpoint save; no-op for non-MoE models.

Test results

Tested on H200 GPU cluster with:

• transformers: 5.2.0
• unsloth: 2026.2.1
• unsloth-zoo: 2026.2.1
• peft: 0.18.1
• vllm: 0.15.1
• torch: 2.9.1+cu128
• trl: 0.20.0

Test: 3-step yes-no-maybe RL training with Qwen2.5-7B-Instruct (LocalBackend)

=== Running 3 training steps ===

--- Step 1/3 ---
Skipping tuning (all same reward) → Advanced step 0→1

--- Step 2/3 ---
Packed 8 trajectories into 1 sequences of length 2048
train: loss=-0.00164, grad_norm=0.127, policy_loss=-0.00164, entropy=0.000413

--- Step 3/3 ---
Skipping tuning (all same reward) → Advanced step 2→3

=== All steps completed successfully! ===
Final step: 3

Full pipeline verified: model loading → inference (vLLM) → rollouts → tokenization → training (unsloth) → checkpoint save → LoRA swap → resume inference.

Test plan

• RL training with Qwen2.5-7B-Instruct on H200 (LocalBackend, 3 steps)
• CI tests (pending)

Closes #575

🤖 Generated with Claude Code