Add challenge 81: INT4 Weight-Only Quantized MatMul (Medium)

[claude]

Summary

• Adds challenge 81: INT4 Weight-Only Quantized MatMul (W4A16), a medium-difficulty inference kernel challenge
• Solvers implement the fundamental dequantization + GEMM kernel powering modern LLM inference frameworks (AWQ, GPTQ, llama.cpp, vLLM)
• The challenge requires: (1) bit manipulation to unpack two INT4 values from each uint8 byte, (2) group-wise float16 scale application, and (3) mixed-precision matrix multiplication (INT4 weights × FP16 activations → FP16 output)

What makes this interesting

• Bit manipulation: high nibble (bits 7:4) holds one weight, low nibble (bits 3:0) holds the next — solvers must correctly unpack and apply the offset-8 encoding (signed = nibble - 8)
• Group quantization: each block of group_size weights shares a single float16 scale, requiring careful index arithmetic (k // group_size)
• Mixed precision: activations are FP16, weights are dequantized on-the-fly, accumulation should be FP32 for accuracy
• Genuinely distinct from existing challenge 32 (INT8×INT8→INT8, per-tensor scale) and challenge 64 (FP32 block-scale weight dequantization, no matmul)

Checklist

challenge.html

• Starts with <p> (problem description)
• Has <h2> sections for: Implementation Requirements, Example, Constraints
• First example matches generate_example_test() values
• Examples use LaTeX \begin{bmatrix} for matrices
• Constraints includes performance test size bullet
• SVG visualization included (dark theme, shows packing format and dataflow)

challenge.py

• class Challenge inherits ChallengeBase
• __init__ calls super().__init__() with all required fields
• reference_impl has assertions on shape, dtype, and device
• All 6 methods present
• generate_functional_test returns 10 cases covering edge, power-of-2, non-power-of-2, realistic, and zero inputs
• Performance test (M=4096, N=4096, K=4096, group_size=128) fits 5× in 16 GB VRAM (~360 MB total)

Starter files

• All 6 files present: .cu, .pytorch.py, .triton.py, .jax.py, .cute.py, .mojo
• Exactly 1 parameter description comment per file
• CUDA/Mojo use "device pointers" (no parenthetical — medium challenge)
• Python frameworks use "tensors on the GPU"; JAX has # return output tensor directly
• Starters compile/run but produce no correct output

General

• Directory follows 81_int4_matmul convention
• Linting passes: pre-commit run --all-files
• Validated with run_challenge.py --action run — solution passes

🤖 Generated with Claude Code