Summary
RocketRide has dedicated LLM nodes for OpenAI, Anthropic, Gemini, DeepSeek, Qwen, MiniMax, Kimi, and Baidu Qianfan — but nothing for NVIDIA's Nemotron family (open-weight hybrid Mamba-Transformer reasoning models, up to 1M context). Nemotron is served through NVIDIA's OpenAI-compatible cloud API (build.nvidia.com, nvapi- keys) and is self-hostable via NIM containers / vLLM / SGLang using the same model IDs.
Proposal
Add a dedicated llm_nemotron node following the llm_kimi / llm_minimax pattern:
- Node package with Nemotron 3 profiles (Super 120B default, Ultra 550B, Nano 30B), a Custom profile, and self-hosted "(Local)" NIM profiles
sync_models provider (ROCKETRIDE_NVIDIA_KEY) with a tight nvidia/*nemotron* model filter — the NVIDIA /v1/models catalog is multi-vendor (Llama, GLM, Kimi, ...), and embed/vision/safety/reward variants must be excluded- CI sync-workflow key wiring, test-framework mock credential, live profile-existence test, catalog docs
Reasoning models: responses may carry reasoning_content / <think> blocks; the node should strip these so downstream nodes only see the final answer .
Summary
RocketRide has dedicated LLM nodes for OpenAI, Anthropic, Gemini, DeepSeek, Qwen, MiniMax, Kimi, and Baidu Qianfan — but nothing for NVIDIA's Nemotron family (open-weight hybrid Mamba-Transformer reasoning models, up to 1M context). Nemotron is served through NVIDIA's OpenAI-compatible cloud API (build.nvidia.com,
nvapi-keys) and is self-hostable via NIM containers / vLLM / SGLang using the same model IDs.Proposal
Add a dedicated
llm_nemotronnode following thellm_kimi/llm_minimaxpattern:sync_modelsprovider (ROCKETRIDE_NVIDIA_KEY) with a tightnvidia/*nemotron*model filter — the NVIDIA/v1/modelscatalog is multi-vendor (Llama, GLM, Kimi, ...), and embed/vision/safety/reward variants must be excludedReasoning models: responses may carry
reasoning_content/<think>blocks; the node should strip these so downstream nodes only see the final answer .