feat(llm): add QAT/Pruning/Sparsity model config support (#115)

[m96-chan]

Summary

Add support for loading models optimized with QAT, pruning, and sparsity techniques.

Closes #115

New Config Classes

• QATQuantConfig: Parse QAT/QAD configs from TensorRT Model Optimizer and HuggingFace formats
• PruningConfig: Detect structured/unstructured pruning configs
• SparsityConfig: Support 2:4 structured sparsity patterns for Ampere+ TensorCores
• ModelOptimizationInfo: Aggregate all optimization info

Supported Formats

Format	Source	Detection
NVFP4/FP8/INT8	TensorRT Model Optimizer	producer + quantization fields
AWQ/GPTQ/BNB	HuggingFace	quantization_config.quant_method
Pruned heads	HuggingFace	pruned_heads field
2:4 sparsity	Various	sparsity_config.pattern

Usage Example

from pygpukit.llm import ModelOptimizationInfo

# Parse all optimizations from config.json
opt_info = ModelOptimizationInfo.from_config(config)

if opt_info.has_any_optimization():
    print(opt_info.summary())  # e.g., "FP8(e4m3), Pruned(structured)"

# Check specific optimizations
if opt_info.sparsity_config and opt_info.sparsity_config.is_2_4_sparse():
    print("Model uses 2:4 structured sparsity")

Test plan

• Unit test config parsing for all formats
• Pre-commit checks pass (ruff, mypy)
• Build succeeds

Future Work (out of scope for this PR)

• Sparse GEMM kernels for 2:4 sparsity
• QAT weight loading with learned scales
• TensorRT-LLM checkpoint format support

🤖 Generated with Claude Code