Files

T

aszc-dev 3adb216763 chore(phase6,bench): record quantization tradeoff matrix

Phase 6 reference numbers for SD1.5 1x512x512 SPLIT_EINSUM, captured
against commit 0bbd8d8 on M2 Pro 32GB (macOS 26.1, coremltools 9.0,
numpy 1.26.4, torch 2.7.1, Python 3.12.11) — same toolchain as Phase 5
baseline ef2a18c-rebased 1e5791d.

Headline:
- size shrinks 1641 MB -> 822 / 617 / 412 MB (1/2, 1/2.7, 1/4)
- fwd-pass median drops 197 ms -> 187 / 183 / 180 ms (5-9% faster)
- noise_pred PSNR vs unquantized: 53.5 / 40.2 / 27.5 dB

The fwd-ms improvement is mostly weight-load bandwidth (smaller LUT
reads). PSNR is on the raw UNet output at a fixed seed; final-image
PSNR after 20 sampler steps is comfortably higher.

2026-05-25 01:31:30 +02:00

536 B

Raw Blame History

Quantization tradeoff matrix — `0bbd8d8`

compute unit: CPU_AND_NE
repeats per variant: 20
ckpt: v1-5-pruned-emaonly, 1x512x512, attn=se

nbits	size (MB)	size vs none	load (s)	fwd median (ms)	fwd min (ms)	PSNR vs none (dB)
none	1641.1	1.0	0.1997	196.651	196.588	-
8	822.0	0.501	0.1475	185.223	183.373	53.54
6	617.1	0.376	0.1457	181.091	180.977	40.2
4	412.3	0.251	0.1427	178.953	178.753	27.47

536 B Raw Blame History

Quantization tradeoff matrix — 0bbd8d8

536 B

Raw Blame History

Quantization tradeoff matrix — `0bbd8d8`