chore: bump transformers to 5.14.1

[hallerite]

Summary

Branch is merged with main (256809b). The merge kept package versions identical to main; only the transformers→5.14.1 pin + the exclude-newer-package exemption differ in pyproject.toml/uv.lock.

Bumps transformers 5.6.2 → 5.14.1 (the newest release, pulled ahead of the workspace 7-day cooldown via exclude-newer-package). The only transitive change from the bump itself is safetensors 0.7.0 → 0.8.0; all dependents' constraints are satisfied (tightest: vllm>=5.5.3, verifiers>=4.56.2).

The internal-API surface prime-rl vendors (checkpoint conversion mappings, Qwen3.5 internals, auto_factory, rope, hub_kernels, the _compat.py shims) is intact across the jump. Three fixes were needed for genuine API/behavior changes.

Fixes

• afmoe — create_causal_mask() no longer accepts cache_position. After merging main's seq_lens refactor (feat(trainer): use seq_lens as the packed-sample boundary contract #3032) the call site in the model is gone entirely, so the fix now only lives in the vendored HF test reference (tests/unit/train/models/afmoe_hf_modeling).
• nemotron_h — the shared Mamba2 mixer (via Zamba2MambaMixer) now reads config.layer_types, which our custom NemotronHConfig didn't expose. Added a read-only layer_types property returning the canonical-remapped schedule derived from layers_block_type (a plain attribute_map alias corrupts layers_block_type because the base config remaps legacy names and writes them back). The remap_legacy_layer_types import is lazy so importing the models package doesn't hard-require a newer transformers.
• custom-impl registration (perf) — transformers 5.13 made _LazyAutoMapping.register() a silent no-op for configs defined under transformers.* (a guard against remote code hijacking native configs). That silently dropped prime-rl's overrides of Llama / Qwen3 / GPT-OSS, so supports_custom_impl() returned False and training fell back to the slower HF implementation — a ~17% throughput regression (and higher multi-GPU peak memory) that benchmark_regression caught. Fixed by registering native-config overrides via the mapping's _extra_content directly (exactly what register() does past the guard).
  • Follow-up (4a52ee9, also flagged by Bugbot): the registration rewrite had accidentally dropped the Qwen3_5TextConfig → Qwen3_5ForCausalLM entry, silently falling back to the HF implementation for dense Qwen3.5 — restored. The AutoConfig.register("qwen3_5_text", ...) call is intentionally not restored: it's a native transformers config, already registered (same rationale as GPT-OSS).

Verification

Post-merge, on a 2× RTX PRO 6000 (workstation Blackwell, sm_120) box:

• CPU unit suite: 409 passed (the test_load_configs examples that need Hub taskset packages fail identically on main in this environment — not bump-related).

• GPU model suite (pytest tests/unit/train/models -m gpu, excluding test_nemotron_h.py): 40 passed, 3 skipped. Two test_nemotron_h cases hard-crash the process on this GPU arch — verified byte-for-byte identical crash point on a main worktree with transformers 5.6.2, so environmental and pre-existing, not from this bump.

• supports_custom_impl() verified True for Llama / Qwen3 / Qwen3.5-dense / GPT-OSS configs on 5.14.1.

• Benchmark root-caused + fixed locally (Qwen3-0.6B RL, 1-GPU, seq 65536), bisected to transformers 5.13.0:

  transformers	throughput	step time	custom impl
  5.6.2 (baseline)	7878 tok/s	16.65s	✅
  5.14.1 (before fix)	6503 tok/s	20.18s	❌ HF fallback
  5.14.1 (after fix)	7946 tok/s	16.51s	✅

Notes / out of scope

• gpt_oss grouped-MM single-process forward fails on a torchtitan.expert_parallel arg-count mismatch — transformers-independent (torchtitan is git-pinned), pre-existing, untested path.
• mamba-ssm is a non-default dependency group dropped by uv sync --all-extras; unrelated to this bump.
• test_qwen3_moe grad check flaked once at the tolerance margin (max grad diff 1024 vs atol=1000, ~2 bf16 ulps at that magnitude) during a full-suite run; passes deterministically standalone and on re-run, on both main and this branch.

🤖 Generated with Claude Code

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 0dc2539. Configure here.}