nunchaku-ai/cuda-challenge-flux-dump

FLUX.1-schnell activation & weight dump (public)

Benchmark tensors for the W4A4 Quantized GEMM CUDA Challenge. Download these, drop them in flux_dump/, and ./benchmark.sh will run correctness + TOPs measurement against them.

Extracted from FLUX.1-schnell at 1024×1024, seed 0, 4 inference steps, prompt "A cat holding a sign that says hello world".

Files

• weights.pt — {layer_name: [N, K] fp16} for the 4 target layers
• activations_1024x1024.pt — {layer_name: [M, K] fp16} input activations captured on the first forward pre-hook fire

Target GEMM shapes

layer	M	N	K
attn_to_qkv	4096	9216	3072
attn_to_out	4096	3072	3072
ff_up	4096	12288	3072
ff_down	4096	3072	12288

Usage

from huggingface_hub import snapshot_download
snapshot_download(
    repo_id="samuelt0207/cuda-challenge-flux-dump",
    repo_type="dataset",
    local_dir="flux_dump",
)

Or use download_data.py in the challenge repo. See dump_data.py for the full generation procedure.