b70-optimization-lab

Exact-Q4 Native DFlash Trace Closeout

Date: 2026-07-13

Status: native hook and ABI smoke complete; real capture and adaptation screen not run because the Qwen optimization lane closed before the protected runtime tree could be frozen.

Preserved implementation

• Runtime helper: /home/steve/src/llama.cpp/common/dflash-target-trace.{h,cpp}.
• Narrow plumbing: common/CMakeLists.txt, common/speculative.{h,cpp}, and tools/server/server-context.cpp.
• Native two-request smoke: experiments/qwen27-dflash-sycl-b70/tests/dflash-q4-target-trace-native-smoke.cpp.
• Contract/parser blueprint: 2026-07-13-dflash-q4-target-trace-hook-blueprint.md.
• Protected-tree snapshot: patches/qwen27-dflash-exact-q4-native-trace-protected-stack-20260713.patch, SHA-256 ef865ff8397bba95a98c33e7e92bfe537ac05c84134a0031795891be9c4b8970.

The patch is deliberately labeled protected-stack: the shared speculative.cpp and server-context.cpp were already dirty with the active Q6 target-top1 experiment, so the snapshot contains that concurrent context as well as the trace plumbing. Do not blindly apply it to a different tree; use the isolated trace helper and named hook points from the blueprint when rebasing.

Verified

• Separate CPU llama-server build completed with both trace compile-time identities set; speculative.cpp, the helper, and server-context.cpp all compiled and linked.
• Warning-enabled standalone helper build passed.
• One trace object atomically captured and published two sequential requests, proving successful-finalize plus release-callback idempotency and reuse.
• Each native payload had the exact 256-byte header, three 51,216-byte rows, BF16 [3,5,5120] features, contiguous positions, and delayed labels.
• The native payload parsed into qwen36_eagle_sequence_v2 with two training anchors. Native identity, payload-byte, position, and label corruptions were rejected.
• The existing synthetic session, ABI, parser, truncation, identity, prompt-reuse, and incomplete-payload suite remained fully green.

Remaining risks

• No real Q4 target/DFlash model request exercised the gather hook. A real one-prompt capture is still required before trusting a corpus.
• The synchronous BF16 conversion and final write are diagnostic overhead; no performance number from a capture-enabled server is valid.
• Capture request gating checks cache disable and greedy temperature at the server boundary, but reopening should audit any newly added sampler transforms before collection.
• The compile-time dirty-patch SHA in the smoke was the earlier plan identity, not the final moving protected tree. It was only an ABI test identity.
• No acceptance training or screen was attempted. The hard continuation gate remains at least 4.0 accepted drafts and 5.0 emitted tokens per favorable width-six cycle.

Reopen procedure

1. Rebase the isolated helper and lifecycle hooks onto a clean/frozen runtime.
2. Compute the full dirty patch SHA-256, update capture-plan.json, and rebuild with LLAMA_DFLASH_TRACE_RUNTIME_COMMIT and LLAMA_DFLASH_TRACE_DIRTY_PATCH_SHA256 set to that exact identity.
3. Launch TP1/parallel-one DFlash with draft K/V F16, n_max=n_min=0, p_min=0, checkpoints off, reasoning off, and the two deep model hashes.
4. Run one collector prompt, parse it, and repeat the four corruption checks.
5. Only then collect train/heldout traces and run the bounded layer-position-bias, draft-token-five screen. Stop unless it clears the 4 accepted / 5 emitted hard gate.

This site is open source. Improve this page.