SC FID eval (KID) + 9-GPU FP/W8A8 runners + MP calibration fix

[heroarmor]

Summary

SC FID-evaluation + multi-GPU run tooling from this session, plus a calibration bug fix. Stacked on feat/sc-halve so the diff is only this work.

Fix

• scripts/calibrate_mp_thresholds.py — the MLP chunked path in _run_mlp_linear_level called sc_matmul_enable_triton_mlp, a stale symbol that no longer exists (runtime NameError), with obsolete max/min positional args. Switched to sc_matmul(granularity="per_row", ...), matching the attention chunked branch and the non-chunked MLP branch. Unblocks MP threshold calibration in this repo (previously only runnable in scmp_diffusion_prev).

New tooling

• scripts/eval/kid_openai.{py,sh} — KID (polynomial-kernel unbiased MMD², 100×1000 subsets) on the same Inception pool3 features the ADM/OpenAI evaluator uses for FID, so KID is directly comparable to reported FID/sFID/IS. Re-derives FID from the same features as a consistency check.
• scripts/eval_openai.sh — ADM FID/sFID/IS/Precision/Recall wrapper (packs PNGs→npz, runs the evaluator in the tfeval env vs the ImageNet-256 reference).
• scripts/run_quant_9gpu.sh + _quant_node_worker.sh — generalized no-SC baseline runner, 3 nodes × 3 GPU. WBITS=16→FP16 baseline; WBITS=8→W8A8 (no stochastic computing). Generate → eval (FID/sFID/IS/P/R) + KID.
• scripts/run_fp_9gpu.sh + _fp_node_worker.sh — FP16 baseline 9-GPU runner.
• scripts/_mp_calib_worker.sh — builds fresh MP threshold tables (avg 64/96/128) via the fixed calibrate_mp_thresholds.py.

Results produced (DiT-XL/2 256×256, cfg=1.5, 50 DDIM steps, bitrev, 10k samples)

run	FID↓	sFID↓	IS↑	Prec↑	Rec↑	KID×10³↓
FP (w16a16)	4.79	19.78	164.5	0.802	0.675	0.32
W8A8 no-SC	7.55	27.41	140.5	0.731	0.704	1.84
W8A8+SC sl=64	9.07	22.70	130.5	0.701	0.710	3.04
W8A8+SC sl=96	7.66	22.72	138.4	0.722	0.700	2.09
W8A8+SC sl=128	6.08	21.05	148.0	0.751	0.693	1.06

FID/sFID use the ADM precomputed reference; KID + Precision/Recall use the 10k arr_0 reference images.

Notes

• Base is feat/sc-halve (depends on its per-row QK granularity) — retarget to main if preferred.
• Other untracked working-tree files (run_3gpu_*.sh, run_ngpu_uniform_nohalve_fanout.sh, tools/*) were intentionally left out; say the word to include them.
• Scripts use absolute cluster paths, consistent with the rest of the repo.

🤖 Generated with Claude Code