Boundary layer protection retest under SmoothQuant

proposed low priority TODO-009

Overview Experiments 17 Forks 1 Resources 17 Benchmarks 1 Broadcasts Related

Description

Boundary protection (first/last 2 transformer blocks at Q8) was negative without SmoothQuant (recovery -0.084 PPL within stderr at +0.572 bpe overhead). With SmoothQuant in the recipe, the inner method's residual error pattern changes — boundary may now matter

Reference

EXP-0012, project memory project_thetom_tq4_1s_investigation.md

Suggested Parameters

quant gptq_turbo_q4

group_size 256

smooth_alpha 0.15

boundary_protect ['first_2', 'last_2']

boundary_method scalar_per_group_q8

eval_seq_len 2048

Provenance

Proposed by @buun via buun-openquant claude-opus-4-6