Deploying Quantized LLMs for Efficient Inference
Chapter 1: Advanced LLM Quantization Fundamentals
Revisiting Quantization Principles for Large Models
Low-Bit Quantization Techniques (Below INT8)
Understanding Quantization Data Types and Formats
Post-Training Quantization (PTQ) Algorithms for LLMs
Quantization-Aware Training (QAT) Considerations
Mixed-Precision Quantization Strategies
Calibration Data Selection and Preparation
Hands-on Practical: Applying GPTQ to an LLM
Chapter 2: Implementing LLM Quantization with Toolkits
Overview of LLM Quantization Libraries
Using bitsandbytes for Low-Bit Operations
Quantization with Hugging Face Transformers and Accelerate
Applying GPTQ using AutoGPTQ
Applying AWQ using AutoAWQ
Comparing Toolkit Outputs and Performance
Handling Model Compatibility Issues
Practice: Quantizing Models with Multiple Toolkits
Chapter 3: Performance Evaluation of Quantized LLMs
Metrics for Quantized Model Evaluation
Measuring Inference Latency and Throughput
Assessing Memory Consumption (Disk and Runtime)
Evaluating Accuracy Degradation
Benchmarking Frameworks and Tools
Analyzing Performance on Target Hardware
Visualizing Performance Trade-offs
Hands-on Practical: Benchmarking a Quantized LLM
Chapter 4: Optimizing and Deploying Quantized LLMs
Inference Optimization Techniques Post-Quantization
Choosing the Right Deployment Framework
Deploying with Text Generation Inference (TGI)
Leveraging vLLM for High-Throughput Inference
GPU Optimization with NVIDIA TensorRT-LLM
Deployment using ONNX Runtime
Containerization and Scaling Strategies
Monitoring Deployed Quantized Models
Hands-on Practical: Deploying via an Inference Server
Chapter 5: Addressing Advanced Challenges
Mitigating Accuracy Loss in Low-Bit Regimes
Handling Activation and Weight Outliers
Quantizing Specific LLM Components (Attention, Normalization)
Hardware Constraints and Kernel Availability
Dynamic Quantization vs. Static Quantization Trade-offs
Debugging Quantization Issues
Integrating Quantized Models into Production Pipelines
Practice: Fine-tuning Quantization Parameters
Understanding Quantization Data Types and Formats
Was this section helpful?
- QLoRA: Efficient Finetuning of Quantized LLMs, Tim Dettmers, Artidoro Pagnoni, Ari Holtzman, Luke Zettlemoyer, 2023 arXiv preprint arXiv:2305.14314 DOI: 10.48550/arXiv.2305.14314 - Introduces QLoRA, a memory-efficient finetuning approach that uses 4-bit NormalFloat (NF4) quantization, along with Double Quantization.
- GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers, Elias Frantar, Saleh Ashkboos, Maximilian Seidl, Sebastian P. Stich, Torsten Hoefler, 2022 arXiv preprint arXiv:2210.01730 DOI: 10.48550/arXiv.2210.01730 - Presents GPTQ, a highly effective post-training quantization method for LLMs, primarily targeting 4-bit integer weights.
- AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration, Ji Lin, Jiaming Tang, Haotian Tang, Shang Yang, Wei-Ming Chen, Wei-Chen Wang, Guangxuan Xiao, Xingyu Dang, Chuang Gan, Song Han, 2023 MLSys 2024 DOI: 10.48550/arXiv.2306.00978 - Proposes AWQ, an activation-aware weight quantization method that selectively quantizes weights based on their impact on activation outliers, demonstrating good accuracy for low-bit LLM inference.
- Transformer Engine: A Unified Deep Learning Acceleration Library for Transformers, Oleg Kuchaiev, Jianyu Huang, Brian Catanzaro, 2022 arXiv preprint arXiv:2209.05193 DOI: 10.48550/arXiv.2209.05193 - Describes the NVIDIA Transformer Engine, which enables efficient large model training and inference, including the use of 8-bit floating-point (FP8) formats for activations and weights.