b70-optimization-lab

Qwen3.6 35B Quark INT8 on B70

This folder is the durable result packet for the Qwen3.6 35B A3B Quark W8A8 INT8 lane on Intel Arc Pro B70. It exists to make the final state easy to reference before moving effort to another model.

Bottom Line

The Qwen3.6 35B 4x B70 lane is exhausted for now. The best strict-valid TP4 result is 93.55 tok/s corrected output on the PIECEWISE forced-comm graph baseline. A legacy LocalMaxxing-approved run reached 99.43 tok/s, but newer deep gates are stricter and should be used for current claims.

No attempted speculative decode path produced a valid >150 tok/s result for this model. The fastest numbers above the baseline were invalid, synthetic, or crashed before validity gates.

Start Here

• 4x B70 results: valid, legacy, smoke, invalid, and ceiling results for TP4.
• 2x B70 reference: TP2 numbers and caveats.
• Validity gates: what counts as a real result.
• Reproduce: commands for the best valid 4x and 2x references.
• Bugs and failed paths: MTP, ReplaySSM, DFlash, graph, and metadata shortcuts.
• Intel/vLLM suggestions: actionable upstream and platform feedback.
• Next model carryover: what to reuse when switching away from this model.

Best Results Summary

Scope	Result	Validity	Primary artifact
4x strict-valid current base	93.55 tok/s	JSON 128/128, color 256/256, quality pass; LocalMaxxing cmqq4mw4c00yfqo01gb2ucgxj	deep-gate-summary
4x legacy public approved	99.43 tok/s	LocalMaxxing approved, older gates	localmaxxing snapshot
2x reference smoke	85.87 tok/s	JSON 16/16, color 16/16, quality skipped; LocalMaxxing cmqq4mwgm00yiqo0133bj962q	tp2-smoke-summary
2x older raw reference	91.59 tok/s	smoke/reference only, not promoted	tp2-latency-truth
Fastest raw artifact	198.95 tok/s	invalid/reference only	ngram5 raw
Fastest synthetic ceiling	181.91 tok/s	canaries skipped, synthetic accept	eagle2 ceiling summary
Fastest MTP-ish invalid	107.77 tok/s	JSON and color failed immediately	mtp parity fix v2

Status Recommendation

Stop spending ad hoc benchmark time on Qwen3.6 35B Quark INT8 unless the task is one of these controlled follow-ups:

• upstream XPU kernel/runtime bakeoff against the strict 93.55 tok/s identity;
• deep graph-compatible speculative-state engineering, with full endpoint gates;
• extracting reusable DFlash/ReplaySSM lessons for a different model.

For new optimization work, switch to another model. Prior successful lanes include Qwen 27B and MiniMax M2.7; Gemma 4 12B also has strong TP4 production material. See next model carryover.

This site is open source. Improve this page.