feat(orchestrator): validate renderer auto-resolution at config time (#2537)

[hallerite]

Summary

• Adds an OrchestratorConfig @model_validator(mode="after") that rejects use_renderer=True + renderer.name='auto' when model.name isn't in MODEL_RENDERER_MAP, so --dry-run catches it (closes Surface unsupported renderer config errors earlier #2537).
• Sweeps the public configs to opt into the right renderer explicitly.
• Swaps PrimeIntellect/Qwen3-0.6B → Qwen/Qwen3-0.6B in the two public RL configs that referenced it (gsm8k, alphabet_sort) and sets renderer.preserve_all_thinking = true to match PI's template behavior under the auto-resolved qwen3 renderer.
• Includes submodule bumps to pick up the three companion PRs (now merged): renderers#48 (pre-flight overflow check with auto-discovery), renderers#50 (Qwen3-30B-A3B-{Instruct,Thinking}-2507 and GLM-5-FP8 added to MODEL_RENDERER_MAP), and verifiers#1408 (renderer-client OverlongPromptError translation).

Why config-time and not orchestrator startup?

The SFT path already rejects this at startup (src/prime_rl/trainer/sft/train.py:172-179). Putting the same guard at orchestrator startup wouldn't fire on --dry-run. Moving to a config validator catches it in cli(RLConfig) — i.e. on dry-run and on every config load — without touching the runtime code path.

What the error looks like

Value error, orchestrator.use_renderer=True with renderer.name='auto' but
'PrimeIntellect/Qwen3-0.6B-Reverse-Text-SFT' is not in
renderers.base.MODEL_RENDERER_MAP, so it would silently fall back to
DefaultRenderer. Pick one:
(a) [orchestrator.renderer] name='default' — for fine-tunes / vendored
    mirrors with custom chat templates (DefaultRenderer calls
    apply_chat_template); pair with tool_parser=<name> if the env uses tools.
(b) [orchestrator.renderer] name=<model-specific renderer> — if <model> is
    template-identical to a mapped family (and ideally also add it upstream
    to renderers.base.MODEL_RENDERER_MAP).
(c) orchestrator.use_renderer=false — opt out of the renderer client entirely.

Exits non-zero; Dry run complete never prints. The full list of mapped models is omitted from the error (40+ lines was too noisy); users can grep renderers.base.MODEL_RENDERER_MAP for it.

How I classified each unmapped config

After md5-hashing chat_template.jinja for each unmapped model against its closest mapped sibling, plus diffing PI mirror templates against upstream Qwen / MiniMax to confirm same-weights-only:

Action	Configs
Swap RL configs: PrimeIntellect/Qwen3-0.6B → Qwen/Qwen3-0.6B, add renderer.preserve_all_thinking = true	gsm8k/rl, ci/integration/alphabet_sort
Keep PI finetune id, set renderer.name = "qwen3" + preserve_all_thinking = true — chat template byte-identical (md5 c574d57…) to PrimeIntellect/Qwen3-0.6B base	examples/wordle/rl.toml (PrimeIntellect/Qwen3-1.7B-Wordle-SFT)
renderer.name = "default" — PI / Qwen finetune with custom chat template	9 configs — multi_reverse_text, ci/integration/reverse_text/{start,resume}, ci/integration/reverse_text_lora/{start,resume}, ci/integration/reverse_text_moe/start, ci/integration/reverse_text_multi_run, examples/{reverse_text, Intellect-3.1}
renderer.name = "default" + reasoning_parser = "think" — DeepSeek-R1 distill	8 configs — acereason_math/{stage1,stage2}, ci/nightly/acereason_math, deepscaler/{stage1,stage2,stage3}, hendrycks_math/sanity, examples/hendrycks_sanity
Model-specific renderer.name — template-identical (md5-confirmed) to a mapped sibling. After the submodule bump for renderers#50, four are auto-resolved and reverted; only the PI-mirror entry stays explicit.	1 config — examples/minimax_m2.5_swe (→ minimax-m2, PrimeIntellect/MiniMax-M2.5-bf16 deliberately not in upstream map)

Why we're not swapping all PI mirrors

PrimeIntellect/* re-uploads of upstream models mostly exist to carry chat-template patches that the hand-coded renderers ignore anyway. For the two RL configs that use the PI Qwen3 mirror, the swap to upstream + preserve_all_thinking = true reproduces PI's "always emit prior <think>" template patch under the qwen3 renderer.

SFT configs that reference PrimeIntellect/Qwen3-{0.6B,1.7B} are intentionally left untouched — SFT defaults use_renderer = false, so HF apply_chat_template runs directly against the model's template; swapping the model id would silently change SFT tokenization (PI emits empty <think></think> for every assistant message; upstream Qwen does not).

Genuine PI finetunes with custom chat templates (Qwen3-0.6B-Reverse-Text-SFT, Reverse-Text-SFT, INTELLECT-3*, INTELLECT-3-Base) keep the PI id and use renderer.name = "default" so DefaultRenderer runs the model's own (patched) apply_chat_template. Exception: Qwen3-1.7B-Wordle-SFT ships the standard PI template patch (byte-identical to PrimeIntellect/Qwen3-0.6B base), so it gets the same qwen3 renderer + preserve_all_thinking = true treatment as the two RL configs that swapped to upstream Qwen3.

The one exception is PrimeIntellect/MiniMax-M2.5-bf16: identical weights to MiniMaxAI/MiniMax-M2.5 modulo a bf16 dtype cast that prime-rl's trainer requires, and a byte-identical chat template. Keep the PI id (we need the bf16 weights), set renderer.name = "minimax-m2" explicitly. The corresponding PI-mirror entry was deliberately removed from PR #50 (no PI ids in the upstream renderer map).

Slim configs package

prime-rl-configs now lists renderers>=0.1.8.dev4 as a hard dependency. The validator's from renderers.base import MODEL_RENDERER_MAP lives inside the function body and only fires on renderer.name='auto' configs, so the slim install CI's "no heavy deps in sys.modules at import time" check still passes — renderers (and its transitive transformers) stay out of sys.modules until validation actually needs them. Verified locally against both slim CI steps and Bugbot's failure mode (default OrchestratorConfig() now resolves instead of raising ImportError).

Submodule bumps

• deps/renderers: 17d0584f → 8704f9d (renderers-v0.1.8.dev4)
  • Includes renderers#48 (overflow check) and renderers#50 (map additions).
• deps/verifiers: dd89b5e9 → 58b119fa
  • Includes verifiers#1408 (OverlongPromptError translation).

Tests

• 5 new unit tests in tests/unit/test_configs.py cover: reject unmapped, accept mapped, explicit renderer.name bypass, use_renderer=false bypass, explicit renderer.name="default" opt-in.
• All 78 parametrized test_load_configs cases pass after the sweep.
• All 13 tests/unit/train/test_runs.py + tests/unit/train/rl/test_packer.py cases pass after adding use_renderer = False to test fixtures that use placeholder "test-model" (now correctly rejected by the validator).
• Broad tests/unit/ sweep (excluding pre-existing unrelated CUDA/Laguna failures): 240 passed.

---

Note

Medium Risk
Medium risk: introduces a new config-time validation that can hard-fail previously working orchestrator configs that relied on renderer.name='auto' silently falling back to DefaultRenderer, and adds a new renderers dependency to prime-rl-configs.

Overview
Adds a new OrchestratorConfig validator that rejects use_renderer=true + renderer.name='auto' when the configured model isn’t in renderers.base.MODEL_RENDERER_MAP, surfacing misconfiguration during config load / --dry-run instead of silently falling back to DefaultRenderer.

Updates public configs/ and examples/ to explicitly opt into an appropriate renderer (name="default" or a model-specific renderer, plus reasoning_parser="think" / preserve_all_thinking where needed), and adjusts unit tests/fixtures to either use a mapped model or set use_renderer=false. Also adds renderers>=0.1.8.dev4 to prime-rl-configs dependencies and updates the lockfile accordingly.

^{Reviewed by Cursor Bugbot for commit 30653f4. Bugbot is set up for automated code reviews on this repo. Configure here.}

[cursor]

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

There are 2 total unresolved issues (including 1 from previous review).

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

^{Reviewed by Cursor Bugbot for commit aecf703. Configure here.}