What
The recipes.vllm.ai sweep (#605–#610) resolved recipe models to architectures by
reading each model's HF config.json. 20 of the 157 could not be read — gated
repos returning 401, or 404. Their architectures are therefore UNVERIFIED, not
confirmed absent, and every count that sweep produced is a floor rather than a
total.
Unreadable at the time of the sweep (vllm-project/recipes @ 86c7777a):
| Reason |
Models |
| 401 gated |
meta-llama/Llama-3.1-8B-Instruct, meta-llama/Llama-3.3-70B-Instruct, meta-llama/Llama-4-Scout-17B-16E-Instruct, google/translategemma-27b-it, inclusionAI/Ring-1T-FP8, pfnet/plamo-3-nict-31b-base, stabilityai/stable-audio-open-1.0, stabilityai/stable-diffusion-3.5-medium, zai-org/GLM-GA |
| 404 |
IndexTeam/IndexTTS-2.5, mistralai/Mistral-Large-3-675B-Instruct-2512, mistralai/Voxtral-4B-TTS-2603, zai-org/GLM-TTS |
no top-level architectures (diffusion pipelines / omni) |
MiniMaxAI/MiniMax-H3, Qwen/Qwen-Image, Wan-AI/Wan2.2-T2V-A14B-Diffusers, zai-org/GLM-Image, meituan-longcat/LongCat-Image-Edit, openbmb/VoxCPM2, fishaudio/s2-pro |
Several are almost certainly already rowed (the Llama ones resolve to
LlamaForCausalLM), but "almost certainly" is not the standard for an inventory
that claims exhaustiveness.
The second, larger hole
A recipe-driven sweep can only see architectures some recipe references. Any
vllm-omni architecture with no recipe is invisible to it by construction —
vllm-omni's own docs/models/supported_models.md lists TTS/audio architectures
beyond the six rowed in #610.
Why it matters
.agents/model-matrix.md presents itself as the exhaustive architecture
inventory, and its rollup is CI-enforced — which makes the count look
authoritative regardless of how it was derived. A floor presented in a surface
that reads as a total is the failure mode worth closing.
Scope
Resolve the 20 by a route that does not depend on anonymous HF reads — the
transformers/vllm config machinery in the pinned oracle venv, an authenticated
fetch, or the recipe guide: prose where it names the architecture. Then sweep
vllm-omni's supported-models doc directly rather than through recipes, and row
whatever is missing. Records-only unless something surprising turns up.
Not urgent — but the sweep's own numbers should not be quoted as totals until it
is done.
Found while auditing recipes.vllm.ai coverage; flagged by the implementer on
#625, who declined to infer coverage it could not verify.
What
The
recipes.vllm.aisweep (#605–#610) resolved recipe models to architectures byreading each model's HF
config.json. 20 of the 157 could not be read — gatedrepos returning 401, or 404. Their architectures are therefore UNVERIFIED, not
confirmed absent, and every count that sweep produced is a floor rather than a
total.
Unreadable at the time of the sweep (
vllm-project/recipes@86c7777a):meta-llama/Llama-3.1-8B-Instruct,meta-llama/Llama-3.3-70B-Instruct,meta-llama/Llama-4-Scout-17B-16E-Instruct,google/translategemma-27b-it,inclusionAI/Ring-1T-FP8,pfnet/plamo-3-nict-31b-base,stabilityai/stable-audio-open-1.0,stabilityai/stable-diffusion-3.5-medium,zai-org/GLM-GAIndexTeam/IndexTTS-2.5,mistralai/Mistral-Large-3-675B-Instruct-2512,mistralai/Voxtral-4B-TTS-2603,zai-org/GLM-TTSarchitectures(diffusion pipelines / omni)MiniMaxAI/MiniMax-H3,Qwen/Qwen-Image,Wan-AI/Wan2.2-T2V-A14B-Diffusers,zai-org/GLM-Image,meituan-longcat/LongCat-Image-Edit,openbmb/VoxCPM2,fishaudio/s2-proSeveral are almost certainly already rowed (the Llama ones resolve to
LlamaForCausalLM), but "almost certainly" is not the standard for an inventorythat claims exhaustiveness.
The second, larger hole
A recipe-driven sweep can only see architectures some recipe references. Any
vllm-omniarchitecture with no recipe is invisible to it by construction —vllm-omni's owndocs/models/supported_models.mdlists TTS/audio architecturesbeyond the six rowed in #610.
Why it matters
.agents/model-matrix.mdpresents itself as the exhaustive architectureinventory, and its rollup is CI-enforced — which makes the count look
authoritative regardless of how it was derived. A floor presented in a surface
that reads as a total is the failure mode worth closing.
Scope
Resolve the 20 by a route that does not depend on anonymous HF reads — the
transformers/vllmconfig machinery in the pinned oracle venv, an authenticatedfetch, or the recipe
guide:prose where it names the architecture. Then sweepvllm-omni's supported-models doc directly rather than through recipes, and rowwhatever is missing. Records-only unless something surprising turns up.
Not urgent — but the sweep's own numbers should not be quoted as totals until it
is done.
Found while auditing
recipes.vllm.aicoverage; flagged by the implementer on#625, who declined to infer coverage it could not verify.