Configure EAGLE3 offline pipeline YAML files for new model checkpoints.
npx skills add https://github.com/NVIDIA/Model-Optimizer --skill eagle3-new-model
Please help me install this Agent Skill. Skill: eagle3-new-model Source: https://github.com/NVIDIA/Model-Optimizer/tree/main/.agents/skills/eagle3-new-model Command: npx skills add https://github.com/NVIDIA/Model-Optimizer --skill eagle3-new-model
This Skill helps users configure new EAGLE3 model checkpoints for offline pipeline execution without manually designing launcher YAML files or selecting incompatible settings.
Ask the skill to configure an EAGLE3 offline pipeline YAML for a new model checkpoint by selecting the correct backend and adapting an existing example.
High-intent search queries and answers about installing and using this skill.
To configure a new EAGLE3 model checkpoint, you adapt the closest existing hf_offline_eagle3.yaml example. This process adjusts model-specific parameters like hidden-state dump backend selection and GPU settings to prepare optimized training and benchmark workflows.
Backend selection for hidden-state dumping chooses between vLLM, HF, or TRT-LLM based on model characteristics. It evaluates multimodal support, attention patterns, and deployment compatibility to determine the optimal backend for the pipeline.
Yes, EAGLE3 onboarding supports multimodal models by selecting the appropriate backend. The configuration process checks model characteristics to ensure compatibility, adjusting settings like trust-remote-code requirements and tokenizer environments for new checkpoints.
Configuring MoE expert dimensions requires adjusting model-specific parameters in the offline pipeline YAML. The skill handles these adjustments automatically by adapting reference configurations to fit the new checkpoint's architecture.
The best way to set up GPU configurations is by adapting an existing reference YAML file. This approach ensures launcher setup and GPU adjustments match the model's specific requirements without manually designing incompatible settings.
Trust-remote-code settings are required when onboarding new checkpoints that use custom model architectures. The configuration process identifies this necessity and applies the correct environment parameters to ensure the offline pipeline executes properly.
NVIDIADiagnose failed EAGLE3 pipeline runs by identifying broken workflow stages.
ovachieverEvaluate language models across 60+ academic benchmarks with lm-evaluation-harness.
yo-stevenDeploy HF LLM checkpoints as OpenAI-compatible endpoints via vLLM, SGLang, or TRT-LLM.
qcmuuEvaluate LLM endpoints across standardized benchmarks using YAML-configured multi-backend setups.
