Dataset Viewer
Auto-converted to Parquet Duplicate
models
list
models_long
list
programs
list
paged
list
paged_networks
list
throughput
list
vec
list
totals
dict
[ { "key": "m08", "model": "Qwen3.5-0.8B", "digits_exact": 40, "digits_total": 40, "general_exact": 24, "general_total": 24, "hidden": 3072, "bad": 0 }, { "key": "m2", "model": "Qwen3.5-2B", "digits_exact": 40, "digits_total": 40, "general_exact": 24, "gener...
[ { "model": "Qwen3.5-0.8B", "test": "digits MLP", "exact": 40 }, { "model": "Qwen3.5-0.8B", "test": "general (4 programs)", "exact": 24 }, { "model": "Qwen3.5-2B", "test": "digits MLP", "exact": 40 }, { "model": "Qwen3.5-2B", "test": "general (4 programs)", ...
[ { "program": "digits MLP 15-15-10", "exact": 6, "total": 6, "prompt": 459, "emitted": 50 }, { "program": "random MLP 20-8-4 (K=20 spans 2 tiles)", "exact": 6, "total": 6, "prompt": 282, "emitted": 24 }, { "program": "random 4-layer 8-5-5-5-4", "exact": 6, ...
[ { "network": "stream residual", "readout": "code", "source": "a100_stream.log", "digits_exact": 10, "digits_total": 10, "pages_run": 70, "page_mismatches": 0, "prompt_min": 161, "prompt_max": 574, "emitted": 1400 }, { "network": "stream residual", "readout": "lm_h...
[ { "network": "stream residual", "digit_runs": 20, "exact": 20, "pages_run": 140, "emitted": 2800, "mismatches": 0 }, { "network": "stream deep", "digit_runs": 20, "exact": 20, "pages_run": 200, "emitted": 5680, "mismatches": 0 }, { "network": "stream cnn",...
[ { "path": "eager", "exact": 20, "total": 20, "prefill_ms": 24.1, "loop_ms": 657.1, "tok_s": 76.1, "ms_per_token": 13.14, "e2e_ms": 681.2, "inf_s": 1.47 }, { "path": "CUDA-graph", "exact": 20, "total": 20, "prefill_ms": 102.1, "loop_ms": 162.7, "tok_s":...
[ { "machine": "scalar", "path": "eager", "exact": 20, "emitted": 50, "layers": 9, "tok_s": 76.2, "e2e_ms": 680.3, "series": "scalar eager" }, { "machine": "scalar", "path": "CUDA-graph", "exact": 20, "emitted": 50, "layers": 9, "tok_s": 307.7, "e2e_ms":...
{ "digits_total_runs": 120, "digits_runs": 120, "digits_exact": 120, "general_runs": 72, "general_exact": 72, "page_runs": 3454, "page_mismatches": 0, "paged_digits_exact": 118, "paged_digits_total": 118 }

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

Downloads last month
59