Chapter 26 β Model Optimization Overview
βThe best model is not the most accurate one β itβs the one that runs on your hardware within your latency budget with acceptable accuracy.β
Why Optimize?
Planned content: Research vs. deployment requirements. A ViT-H has 632M parameters β fine for GPU inference, impractical for a smartphone. Optimization bridges this gap. The three axes: accuracy, latency, model size. The tradeoff triangle.
π Planned diagram: The optimization tradeoff triangle β accuracy vs. latency vs. model size, with different techniquesβ positions labeled.
The Optimization Landscape
Planned content: Four main families: (1) Quantization β reduce numerical precision. (2) Pruning/Sparsification β remove weights or structure. (3) Knowledge Distillation β compress knowledge into smaller model. (4) Neural Architecture Search β find efficient architectures. (5) Inference optimization β graph compilation, operator fusion.
π Planned diagram: Optimization family tree β taxonomy with representative techniques per family.
Hardware Targets
Planned content: Server GPU (A100, H100): FP16/BF16, INT8, FP8. Edge GPU (Jetson Orin): INT8, INT4. Mobile GPU (Adreno, Apple Neural Engine): FP16, INT8. CPU (x86, ARM): INT8, sparse. FPGA: custom bitwidth. Each target has different optimal optimization strategy.
π Planned table: Hardware targets Γ optimization compatibility matrix β which techniques work well on which hardware.
Profiling and Bottleneck Analysis
Planned content: Before optimizing, profile! Latency breakdown: memory-bound vs. compute-bound ops. PyTorch Profiler. Roofline model for CV models. Memory bandwidth vs. arithmetic intensity. FLOPs vs. latency (why more FLOPs β more latency). Batch size effects. Layer-wise latency analysis.
π Planned diagram: Roofline model for a ResNet-50 layer β plotting arithmetic intensity vs. performance, roof = max compute or max bandwidth.
Optimization Pipeline
Planned content: Typical pipeline: train full model β profile β quantize PTQ β evaluate accuracy β if not enough: QAT β prune β distill smaller model β export to ONNX/TensorRT β benchmark on target hardware.
π Planned diagram (flowchart): Model optimization decision tree β start with PTQ, evaluate accuracy drop, if too large β try QAT β if still not enough β distillation or architecture replacement.
FLOP Counting
Planned content: FLOPs (Floating Point Operations): 2 Γ M Γ N Γ K for matmul
[M, K] Γ [K, N]. FLOPs for Conv2d:2 Γ C_in Γ C_out Γ kΒ² Γ H_out Γ W_out. Tools: thop, fvcore, torchinfo. MACs vs. FLOPs (1 MAC = 2 FLOPs). Why FLOPs are a poor proxy for latency.
1
2
3
4
5
6
7
8
9
from thop import profile
import torch
from torchvision.models import resnet50
model = resnet50()
x = torch.randn(1, 3, 224, 224) # [1, 3, 224, 224]
macs, params = profile(model, inputs=(x,))
print(f"MACs: {macs/1e9:.2f}G, Params: {params/1e6:.2f}M")
# MACs: 4.09G, Params: 25.56M
Next: Chapter 27 β Quantization β
Last updated: May 2026