qwentensor / README.md
Bobliu's picture
Upload folder using huggingface_hub
d525c16 verified
|
Raw History Blame Contribute Delete
1.62 kB
metadata
license: mit
base_model: Qwen/Qwen3.5-0.8B
tags:
  - qwen
  - qwen3.5
  - quantization
  - pytorch
  - research
pretty_name: qwentensor run logs

qwentensor: run logs (no model weights)

qwentensor compiles a small quantized PyTorch network into a prompt-written program that a Qwen3.5 (with the qwenTPU array) executes, bit-exact against a numpy golden. Code and video: https://github.com/Bobliuuu/qwentensor

No weights are released here. Base models are the unmodified Qwen/Qwen3.5-0.8B, -2B, -4B (Apache-2.0, Alibaba); the chip layers are compiled in memory at run time from the qwenTPU code. This is a dataset-style card for the A100 logs and the explainer video.

Contents

path what
results/a100_*.log A100 (bf16) runs: digits MLP, general programs, residual, paged and wide networks, vectorized machine, throughput, limits
results/readme_data.json the aggregate parsed from those logs, used by the README
qwentensor_explainer.mp4 explainer video with voice-over

Numbers (parsed from the logs)

  • Digits MLP on Qwen3.5 0.8B, 2B and 4B: 120 of 120 runs exact.
  • One compiled machine, four programs: 72 of 72 runs exact across the three sizes.
  • Paged and wide networks on 0.8B: 118 of 118 digit x readout runs exact, 3,454 page runs, 0 mismatches.

Limits: one machine holds at most 128 ops and 126 columns; larger networks run in pages with the host carrying activations; no loops or data-dependent branches; paged and wide runs were verified on 0.8B only.

Video: https://github.com/Bobliuuu/qwentensor/blob/main/video/qwentensor_explainer.mp4