b70-optimization-lab

MTP1 sub-gate portfolio policy and first crossover

Date: 2026-07-16

Status: method validated; first bundle positive but not promoted

Decision

The 0.50 ms/cycle integration threshold now applies to a predeclared, compatible optimization portfolio, not necessarily to every exact component in isolation. This is the supported way to evaluate small improvements without lowering the gate or benchmark-shopping.

The first portfolio combined:

  1. M=2 fused QNorm/RoPE/direct-FP8-KV insertion; and
  2. the ordered N128 small-M MXFP4 verifier policy.

The same-binary B-A-B crossover was positive, but neither bundle run exceeded the qualified 63.349928 tok/s record. Keep the record recipe at M=1 for the QNorm selector and N64 for MXFP4. No LocalMaxxing submission was made.

Portfolio admission rule

A bundle may receive a TP4 service load only when all of the following hold:

A bundle can establish that sub-gate changes add up, but promotion still requires a new qualified record, exact canaries, and the normal reproducibility and quality gates. One favorable row is not sufficient.

Frozen portfolio identity

The clean rebuild produced package-identical artifacts:

Hardware gates

The prior M=2 QNorm gate passed 40/40 changed eager inputs and 8/8 graph replays on each B70. It projected 1.68-1.90 ms/cycle saved over 43 layers.

The portfolio rebuild repeated the N128 gate on all four cards. Every card passed 48/48 bitwise-exact route cases. Projected savings versus N64 were:

Physical B70 Saved ms / 43 layers
0 0.3111
1 0.2782
2 0.2961
3 0.2677

The measured combined range is therefore approximately 1.95-2.21 ms/cycle. This clears the portfolio gate even though N128 does not clear it alone.

Raw N128 summaries are in ../../../data/deepseek-v4-reap-m2-portfolio-20260716/.

Same-binary service crossover

All three suites used the fixed 12-prompt realistic suite once per prompt, returned token IDs, generated 128 tokens, and reported zero cached prompt tokens.

Order Variant Median tok/s p10 tok/s Mean tok/s
B1 QNorm M2 + N128 62.446116 59.707219 62.918715
A QNorm M1 + N64 61.895036 58.898587 62.193886
B2 QNorm M2 + N128 62.767570 59.935732 63.137313

The two bundle medians average 62.606843 tok/s, +0.711806 tok/s over the same-binary control. Their p10 average is 59.821475, +0.922888 over control. The direction is positive, but both bundle medians remain below the 63.349928 record. This is useful evidence for the portfolio method, not a promotable speed record.

Raw service evidence:

After B2, 10/10 ordered exact capture suites passed: 60/60 expected answers, all cached-zero.

What this changes

Do not discard exact micro-wins solely because each is below 0.50 ms. Keep a portfolio ledger and combine only compatible candidates whose conservative, non-overlapping lower bounds clear the threshold. Do not revive candidates with negative service evidence merely to inflate the projected sum.

The next useful portfolio should add a newly measured exact component to this method, then repeat a frozen same-binary crossover. The present two-component bundle is closed below the record and should not be rerun until another non-overlapping component materially raises its conservative ceiling.