Machine Learning

Submit, list and stop GPU training jobs on the Qizhi platform

Deploy and run ML training jobs on local, remote, Vast.ai or Modal GPUs

Run GPU training, fine-tuning and inference on Modal with no SSH or Docker setup

Watch training metrics for NaNs, loss divergence and idle GPUs while jobs run

Apply image classification practices to CIFAR, ImageNet and similar benchmarks

Apply object detection practices for COCO, VOC and YOLO or DETR models

Speed up data loaders so preprocessing never starves the GPU during training

Scale PyTorch training across GPUs with DDP and multi-GPU patterns

Speed up GPU training and cut memory use with FP16 and BF16 mixed precision

Apply LLM alignment practices for RLHF, DPO and instruction tuning

Apply language model pretraining and fine-tuning practices to training code

Build dependable PyTorch training loops and review existing training code

Apply reinforcement learning practices for PPO, SAC and reward design

Compress LLMs to 4-bit with AWQ for faster inference with minimal accuracy loss

Fine-tune 100+ LLMs with Axolotl YAML configs, LoRA and preference methods

Run image captioning, visual question answering and image-text retrieval with BLIP-2

Classify images zero-shot and search images by text with OpenAI CLIP

Align models for safety with self-critique, revision and RL from AI feedback

Scale model training with DeepSpeed ZeRO, pipeline parallelism and mixed precision

Pretrain large LLMs across many GPUs with torchtitan and 4D parallelism

Evaluate Cosmos robot manipulation policies in LIBERO and RoboCasa simulators

Track ML experiment runs and media logs in SwanLab with local or self-hosted dashboards

Fine-tune OpenVLA-OFT robot policies with LoRA and evaluate them on LIBERO

Fine-tune and serve OpenPI robot policy models with JAX or PyTorch backends

Fine-tune LLMs with TRL using SFT, DPO, PPO, GRPO and reward models

Quantize models to GGUF from 2 to 8 bits for llama.cpp on consumer hardware

Apply GPTQ 4-bit post-training quantization to large LLMs with little accuracy loss

Fine-tune reasoning models with GRPO reinforcement learning in TRL

Quantize LLMs to 2, 3 or 4 bits without calibration data using HQQ

Train and run fast Hugging Face tokenizers for BPE, WordPiece and Unigram models

Implement, train and fine-tune Llama, Gemma, Phi and Qwen models with LitGPT

Compress large models into smaller students with knowledge distillation and soft targets

Run LLM inference on CPUs, Apple Silicon and consumer GPUs with llama.cpp

Fine-tune 100+ LLMs with LLaMA-Factory WebUI, QLoRA and multimodal support

Run vision-language chat and visual question answering with LLaVA models

Extend transformer context windows to 128k tokens with RoPE, YaRN and ALiBi

Build and use Mamba-1 and Mamba-2 state-space models for long sequences

Run enterprise RL training for large MoE models with FP8 or INT4 precision

Track ML experiments, manage model versions and deploy models with MLflow

Merge fine-tuned models with mergekit SLERP, TIES and similar methods without retraining

Prune LLMs with Wanda and SparseGPT to reach high sparsity without retraining

Train mixture-of-experts models with DeepSpeed or Hugging Face for lower compute cost

Train a small GPT from scratch with the compact nanoGPT codebase for learning

Curate LLM training data on GPUs with deduplication, quality filters and PII redaction

Run interpretability experiments on model internals, including 70B+ models via NDIF

Run distributed RLHF training with PPO, GRPO and DPO on Ray and vLLM

Speed up transformer attention and cut memory with Flash Attention in PyTorch

Add PyTorch FSDP2 sharding, mixed precision and checkpointing to training scripts

Structure PyTorch training with Lightning Trainer, callbacks and automatic multi-GPU runs