feat(map): add Qwen3-30B-A3B-{Instruct,Thinking}-2507, GLM-5-FP8

[hallerite]

Summary

Adds three entries to MODEL_RENDERER_MAP. Each has a chat_template.jinja that is byte-identical (md5-confirmed) to a model already in the map, so they can route to the same hand-coded renderer instead of falling back to DefaultRenderer:

New entry	Routes to	Identical to
Qwen/Qwen3-30B-A3B-Instruct-2507	qwen3	Qwen/Qwen3-4B-Instruct-2507
Qwen/Qwen3-30B-A3B-Thinking-2507	qwen3	Qwen/Qwen3-4B-Thinking-2507
zai-org/GLM-5-FP8	glm-5	zai-org/GLM-5

Why

Unblocks four prime-rl example configs (multinode, qwen30b_math, qwen30b_swe, glm5_pd_disag) that were silently falling back to DefaultRenderer because their model.name wasn't an exact match in the map. With the new config-time validator in prime-rl (PrimeIntellect-ai/prime-rl#2540), these would otherwise need explicit [orchestrator.renderer] overrides — and adding to the map is the right answer because the templates are confirmed identical.

Policy: no PrimeIntellect/* mirrors in the map

An earlier revision also added PrimeIntellect/MiniMax-M2.5-bf16 → minimax-m2. That entry has been removed.

PrimeIntellect/* re-uploads of upstream models mostly exist to carry chat-template patches. The renderers in this map ignore those patches — every hand-coded renderer applies its own hardcoded template, not the model's. So adding PI mirrors to the map gives them no benefit, and inviting them in makes the map a moving target as PI ships and retires patched mirrors.

Prime-rl configs should reference the upstream model id directly (e.g. Qwen/Qwen3-0.6B, not PrimeIntellect/Qwen3-0.6B) so auto-resolution picks the right hand-coded renderer. The rare exception — PrimeIntellect/MiniMax-M2.5-bf16 is needed because we have to train in bf16 and upstream ships fp8 — keeps the PI id but sets renderer.name = "minimax-m2" explicitly in the prime-rl config.

Verification

For each candidate, fetched chat_template.jinja (or extracted chat_template from tokenizer_config.json) from the HF repo and md5-hashed against the already-mapped sibling:

5795f12e815e  Qwen/Qwen3-4B-Instruct-2507/tokenizer_config.json chat_template (2630 B)
5795f12e815e  Qwen/Qwen3-30B-A3B-Instruct-2507/tokenizer_config.json chat_template (2630 B)

(4168 B)  Qwen/Qwen3-4B-Thinking-2507/tokenizer_config.json chat_template
(4168 B)  Qwen/Qwen3-30B-A3B-Thinking-2507/tokenizer_config.json chat_template

c39fa970531b98cb889b16714f8c5512  zai-org/GLM-5/chat_template.jinja
c39fa970531b98cb889b16714f8c5512  zai-org/GLM-5-FP8/chat_template.jinja

test_modality_registry_models_route_to_renderer passes locally.

Note

Add Qwen3-30B-A3B-Instruct-2507, Qwen3-30B-A3B-Thinking-2507, and GLM-5-FP8 to renderer map

Adds three new model ID entries to MODEL_RENDERER_MAP in renderers/base.py, routing Qwen/Qwen3-30B-A3B-Instruct-2507 and Qwen/Qwen3-30B-A3B-Thinking-2507 to the qwen3 renderer and zai-org/GLM-5-FP8 to the glm-5 renderer instead of falling back to the default renderer.

^{Macroscope summarized b3d3f2e.}

[macroscopeapp]

Approvability

Verdict: Approved

Purely additive changes to MODEL_RENDERER_MAP, mapping three new model identifiers to existing renderers. No new logic or behavioral changes - just extends existing configuration patterns.

^{You can customize Macroscope's approvability policy. Learn more.}