b70-optimization-lab

Qwen3.6 Research Map

Use this page as the first stop for the Qwen3.6-35B-A3B / Arc Pro B70 research lane. It is a navigation map, not a benchmark claim.

Current Status

As of 2026-06-22, this lane is a reference/archive lane rather than the primary optimization target. The best strict-valid TP4 result is the PIECEWISE forced-comm graph baseline at about 93.55 tok/s; no attempted speculative decode path produced a valid >150 tok/s result. Move new optimization effort to another model unless doing a controlled upstream/runtime bakeoff or a deep graph-compatible speculative-state engineering project.

The organized result packet is:

It contains best valid 4x and 2x results, reproduction commands, validity rules, invalid fast lanes, DFlash/ReplaySSM findings, and carryover notes for other models.

Baseline Identity

Known baseline identity to keep separate from graph-none results:

Never compare this lane to a run that is missing COMPILATION_CONFIG='{"cudagraph_mode":"PIECEWISE"}' or any of the graph/GDN identity fields listed in ../AGENTS.md.

Read Order

  1. ../results/qwen36-35b-quark-int8-b70/README.md for the final organized result packet.
  2. ../results/qwen36-35b-quark-int8-b70/validity-gates.md before comparing or promoting any result.
  3. ../notes/2026-06-20-master-plan.md for the baseline-first plan that guided the late lane.
  4. ../notes/2026-06-16-qwen36-current-handoff.md for detailed run history and the latest full-layerlet gate interpretation.
  5. ../notes/2026-06-21-qwen36-phase3-copy-skip.md before trying metadata-copy or num_computed_tokens shortcuts again.
  6. ../suggestions/findings/README.md for sourced upstream/fork/runtime leads.
  7. ../notes/2026-06-20-research-plan-replayssm-and-speed.md before touching ReplaySSM, DFlash, MTP, or EAGLE verifier state.

Current Decisions

Artifact Locations

Next Safe Work If This Lane Is Reopened

  1. Run an upstream kernel-package delta bakeoff on a clean branch, with the same canaries and full benchmark identity.
  2. Run a TP2 vs TP4 single-request A/B only after identity capture is explicit.
  3. Make speculative rollback/commit graph-compatible before chasing another >150 tok/s result.
  4. For driver/runtime changes, read local-ops.md first so sudo usage and credential hygiene stay consistent.
  5. Revisit ReplaySSM or DFlash only with endpoint canaries, not synthetic-only validation.