Qwen3.8-27B-heretic-ara · 496 Hessian/YAQA-correctable real targets · auto-refresh 8s
2026-09-30 07:45:16 Running for 9h 47m
Overall progress
496 Hessian/YAQA-correctable real targets — of 497 total language-model tensors (112 stay full precision, 1 handled separately, see the tensor taxonomy below)
Estimated completion
Complete
Real output model already saved -- nothing left to estimate.
Total time running9h 47m
System memory: 41.2 GB free · swap: total = 3072.00M used = 1975.12M free = 1096.88M (encrypted)
Live process
COMPLETE
real output model saved
Live log
[MTP-YAQA] mlp.gate_proj: {'label': 'mtp.layers.0.mlp.gate_proj', 'effective_rank_in': 1.150061244185548, 'effective_rank_out': 1.2112867836212802, 'trials': [{'finite': True, 'weighted_err': 0.00012025704199913889, 'naive_weighted_err': 0.16949787735939026, 'beats_weighted': True, 'frob': 11.043423652648926, 'naive_frob': 10.9600830078125, 'within_frob': True, 'max_abs_hatW': 0.3469604551792145, 'max_abs_W': 0.34765625, 'within_magnitude': True, 'safe': True, 'sigma_O': 1.0}], 'path': 'default'}
mtp.layers.0.mlp.up_proj: real effective rank -- H_I=1.3/5120, H_O=1.8/17408
mtp.layers.0.mlp.up_proj/Hin: block_LDL succeeded on attempt 1, final sigma_reg=1.0000
mtp.layers.0.mlp.up_proj/Hout(sigma_O=1.0): block_LDL succeeded on attempt 1, final sigma_reg=1.0000
mtp.layers.0.mlp.up_proj: sigma_O=1e+00 weighted=0.0004x naive frob=1.011x naive SAFE=True
[MTP-YAQA] mlp.up_proj: {'label': 'mtp.layers.0.mlp.up_proj', 'effective_rank_in': 1.2619153522066242, 'effective_rank_out': 1.816212852172597, 'trials': [{'finite': True, 'weighted_err': 0.00020498040248639882, 'naive_weighted_err': 0.49700331687927246, 'beats_weighted': True, 'frob': 12.054214477539062, 'naive_frob': 11.917745590209961, 'within_frob': True, 'max_abs_hatW': 0.36344078183174133, 'max_abs_W': 0.36328125, 'within_magnitude': True, 'safe': True, 'sigma_O': 1.0}], 'path': 'default'}
mtp.layers.0.mlp.down_proj: real effective rank -- H_I=1.3/17408, H_O=1.5/5120
mtp.layers.0.mlp.down_proj/Hin: block_LDL succeeded on attempt 1, final sigma_reg=1.0000
mtp.layers.0.mlp.down_proj/Hout(sigma_O=1.0): block_LDL succeeded on attempt 1, final sigma_reg=1.0000
mtp.layers.0.mlp.down_proj: sigma_O=1e+00 weighted=0.0003x naive frob=1.014x naive SAFE=True
[MTP-YAQA] mlp.down_proj: {'label': 'mtp.layers.0.mlp.down_proj', 'effective_rank_in': 1.3327763110607582, 'effective_rank_out': 1.5435623522595716, 'trials': [{'finite': True, 'weighted_err': 7.94425795902498e-05, 'naive_weighted_err': 0.26578760147094727, 'beats_weighted': True, 'frob': 11.539873123168945, 'naive_frob': 11.385348320007324, 'within_frob': True, 'max_abs_hatW': 0.8824055790901184, 'max_abs_W': 0.8828125, 'within_magnitude': True, 'safe': True, 'sigma_O': 1.0}], 'path': 'default'}
[MTP-YAQA] Wrote real YAQA-corrected sidecar to /Users/hghelab/.mtplx/models/Qwen3.8-27B-heretic-ara-YAQA-5bpw-v2-HESSIAN-PROBE/optiq/mtp.safetensors (29 tensors, 7 real corrected).
Part 5c PASS: real output model saved to /Users/hghelab/.mtplx/models/Qwen3.8-27B-heretic-ara-YAQA-5bpw-v2-HESSIAN-PROBE.
[hessian-scores] 128 real tensor(s) scored -- wrote /Users/hghelab/.mtplx/models/Qwen3.8-27B-heretic-ara-YAQA-5bpw-v2-HESSIAN-PROBE.yaqa_resume/hessian_scores.json
Real completion velocity — tensors per 15-minute window
Real tensor taxonomy — what every number below is actually counted out of: this plan covers 497 real language-model tensors in total (vision/audio/MTP sidecars are reattached separately, not counted here). Of those: 112 (23%) stay at full precision (bits=16, never quantized at all); 385 (77%) are real quantization targets, split into 385 tensors that go through Hessian/YAQA two-sided correction (the number the panel below tracks) and exactly 1 (n/a) handled separately because its own real output-side Hessian would need 246.7GB.
Real quality result, scoped to the 496 Hessian/YAQA-correctable real targets only — how much error YAQA actually removed vs. naive rounding
496/496
of 496 Hessian/YAQA-correctable tensors: reduces real weighted error
0
tensors where it does not (flagged, not hidden)
99.54%
avg error reduction · range 96.49% – 99.99%
350/496
raw Frobenius got worse (expected, see below)
Distribution of real per-tensor error reduction (zoomed into the real 90-100% band -- nearly everything real is in it)
99.6655%4-bit (255 tensors)
99.5328%5-bit (94 tensors)
99.8196%6-bit (1 tensors)
99.3097%8-bit (146 tensors)
Live interpretation — computed fresh from this exact refresh's real data
Right now, 496 of 496 completed tensors (100%) show a real reduction in Hessian-weighted error vs. naive rounding -- a strong, consistent real win. Most of them (496 tensors, 100%) cluster in the 90-100% reduction band. The single best real result so far is language_model.model.layers.16.linear_attn.in_proj_z, at 99.9855% error reduction.
err_yaqa / err_naive (Metric A, what's plotted above): the real Hessian‑weighted
reconstruction error — how much the rounding error matters once weighted by real sensitivity
from a real backward pass. Lower is better; this is what YAQA actually optimizes. frob_yaqa / frob_naive (Metric B, "Frobenius got worse" above): plain, unweighted distance
from the original weight. YAQA can legitimately make this worse while making the weighted error
better — it deliberately trades raw magnitude for lower‑impact error, not a bug.
Reveal the real build sequence — final model assembly
✓Loading cached tensors
✓Verifying correction results
✓Packing the final model
✓Computing final bits-per-weight
✓Saving model to disk
✓Reattaching vision/audio sidecars
✓YAQA-correcting MTP sidecar
✓Complete
Trunk bits per weight: 5.970 -- language-model tensors only (vision sidecars stay at original precision, untouched; MTP is corrected separately with no bpw of its own logged)
Real output model saved to: /Users/hghelab/.mtplx/models/Qwen3.8-27B-heretic-ara-YAQA-5bpw-v2-HESSIAN-PROBE