nunchaku-ai/cuda-challenge-flux-dump

---
license: apache-2.0
pretty_name: CUDA Challenge — FLUX.1-schnell dump (public)
tags:
- cuda
- quantization
- int4
- gemm
- flux
---

# 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

```python
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.