nvidia-nlp-finetune-mixtral

What problem does it solve? Evaluating Mixtral-8x7B-Instruct-v0.1 finetuning performance on NVIDIA GPUs requires coordinating Docker containers, distributed training with xtuner, log collection, and throughput calculation, which is error-prone when done manually. ## Core Features & Use Cases - Automated Finetune Execution: Launches single-node 8-GPU or multi-node (e.g., 2x8) full finetuning of Mixtral-8x7B using xtuner, DeepSpeed ZeRO-3, and the OpenAssistant Guanaco dataset via the bundled finetune.sh script. - Structured Metric Collection: Parses training logs with calc.py through calc.sh, validates the output as a single finite number, and writes throughput into result.json. - Multi-Host Support: Provides rank-aware distributed launch guidance in references/multi_host.md, including MASTER_ADDR/NODE_RANK injection and rank-suffixed logging. - Use Case: Ask the agent to run a Mixtral full finetune benchmark on one node with 8 NVIDIA GPUs; it starts the container, validates model shards and dataset, runs training, and produces a result.json containing the measured throughput. ## Quick Start Run a Mixtral-8x7B full finetuning performance evaluation on a single node with 8 NVIDIA GPUs and report the training throughput.