Add nvidia provider with 15 models

[TF0rd]

Type of change

• Add a model
• Add or fix parameters

Summary

Adds the nvidia provider with 15 model entries sourced from the official NVIDIA NIM API reference at docs.api.nvidia.com/nim/reference/llm-apis#nvidia.

Each model's parameters were scraped from its -infer endpoint documentation (e.g. nemotron-3-ultra-550b-a55b-infer).

Models added

Model	Notable params
nemotron-3-ultra-550b-a55b	reasoning_effort (none/medium/high), reasoning_budget
nemotron-3-super-120b-a12b	reasoning_effort (none/low/high), reasoning_budget
nemotron-3-nano-30b-a3b	temp, top_p, max_tokens, seed, stop
nemotron-mini-4b-instruct	tools, freq/presence penalties
nemotron-content-safety-reasoning-4b	temp, top_p, max_tokens, seed, stop
llama-3.1-nemotron-nano-8b-v1	full OpenAI-compatible params + seed
llama-3.1-nemotron-ultra-253b-v1	full OpenAI-compatible params + seed
llama-3.3-nemotron-super-49b-v1	full OpenAI-compatible params + seed
llama-3.3-nemotron-super-49b-v1.5	max_tokens up to 65536
llama-3.1-nemotron-safety-guard-8b-v3	temperature only (0-1, default 0)
llama-3.1-nemoguard-8b-topic-control	temp 0-2, top_p, freq/presence penalties
riva-translate-4b-instruct-v1.1	translation model, default temp 0
usdcode-llama-3.1-70b-instruct	expert_type enum (auto/code/knowledge/helperfunction)
gliner-pii	labels, threshold, chunk_length, overlap, flat_ner
nemoguard-jailbreak-detect	input param (security endpoint)

Models omitted

Two nvidia/ models were not added because their infer pages document only stream (a reserved MPS path that cannot be a parameter):

• llama-3.1-nemoguard-8b-content-safety — only stream, accept, model, messages
• nvidia-nemotron-nano-9b-v2 — infer page appears buggy (serves content-safety model data)

Checks

• npm run validate — OK (189 models)
• npm test — 101/101 passed
• npm run guard:params — no removals

Docs

• NVIDIA LLM APIs
• nemotron-3-ultra-550b-a55b-infer (example infer page)

[vercel]

@TF0rd is attempting to deploy a commit to the Manifest Team on Vercel.

A member of the Team first needs to authorize it.

[guillaumegay13]

Pushed a small maintainer follow-up in 3c8d451 after checking the NVIDIA docs and live API behavior:

• removed gliner-pii labels, because the API requires an array and current MPS only represents scalar parameter values
• removed nemotron-mini-4b-instruct tools, because the API requires a tools array/object payload, not a scalar model parameter
• removed nemoguard-jailbreak-detect input, because it is the request payload being classified rather than a saved/configurable model parameter

Local validation on the updated branch passed:

• npm run validate
• npm test
• npm run build
• npm run guard:params

[guillaumegay13]

Follow-up pushed in b8a740b to fix the CI package-codegen failure.

Root cause: the PR had NVIDIA catalog entries, but packages/modelparams/src/generated/* had not been regenerated after the branch was updated from main. I ran npm run codegen --workspace=modelparams and committed the generated package data.

One small generator fix is included too: models with no configurable params now emit Record<string, never> in ParamsById instead of {}, so generated output passes the repo lint rules.

Verified locally:

• npm run lint
• npm run typecheck
• npm run validate
• npm run build --workspace=modelparams
• npm test --workspace=modelparams
• npm run build
• npm test
• npm run guard:params

GitHub Actions are now passing for CI and Param guard. Vercel is still blocked by external authorization.