[novkien]

Add the selected SM70 small-batch Q6_K paths for the measured W4 shapes and W1 draft head, with backend fixtures.

Validation reused the native/RPC builds, 100 CPU-reference cases, and actual Qwen/Flash route proofs. The two-pair-per-model screen measured Qwen +3.51% to +3.81%; Flash ranged from -0.97% to +4.55%. These descriptive results have no confidence intervals and Flash was variable; they are not a universal speedup claim. The owner selected this candidate for deployment while accepting the positive Qwen gain below the earlier 5% threshold.

Addresses #9

AI disclosure: Prepared with OpenAI Codex assistance.

[novkien]

Pre-merge handoff and head validation:

• PR ggml: add SM70 small-batch Q6_K kernels #23 is based on master at 5423bddf352adc41a8cc59ed02d38a511b333d30; reviewed head is 46871624dfd0b8b28fcc91683b82838d392dedef.
• The exact base-to-head binary diff is 51b9b224b2ff6d3168b5f6cf91720d5a6018b7c77844bd1cabdb92a018b4f7d5. It contains only ggml/src/ggml-cuda/mmvq.cu and tests/test-backend-ops.cpp.
• Root reviewed the full diff and surrounding kernel/security logic; EditorConfig 3.0.3, 152 model names, and git diff --check passed. Independent review returned GO on the same patch. Reused evidence includes native/RPC builds, 100/100 CPU-reference cases, both model-route proofs, and the two-pair-per-model descriptive screen.
• Screen results remain descriptive: Qwen +3.51% to +3.81%; Flash -0.97% to +4.55%, with variable outputs. No confidence interval or universal speedup claim is made.
• GitHub reports MERGEABLE. master has no branch protection and the repository has no rulesets; optional CI jobs are queued/in progress, with no enforced gate to wait for.

Codex /root retains responsibility for the exact-merge native/RPC build, deployment, and bounded live acceptance on both models through proxy :8082. Keep issue #9 open until that delivery is verified.

[novkien]

Exact-merge delivery preparation

Root verified merge c2d2a20df48b261d4ce94a3b14374ce4bfed9282 has parents exact5423 and reviewed head 46871624dfd0b8b28fcc91683b82838d392dedef, the same tree as that head, and the unchanged pure patch SHA-256 51b9b224b2ff6d3168b5f6cf91720d5a6018b7c77844bd1cabdb92a018b4f7d5. The existing native/RPC numerical and model-route evidence remains applicable to this source tree. The owner has selected deployment of the modest positive Qwen result; no further 5% gate or repeated benchmark campaign is being imposed.

The prepared delivery interval starts with normal baseline wake and identity verification, unloads before build/install, converges both clean physical checkouts to the exact merge, builds and packages native/RPC from its git archive with an empty source-diff identity, and runs allocator checks. It then installs the matched immutable pair and sends one frozen structured cold8192+512 request through proxy8082 for each model, checking running hashes, model hashes, route/startup invariants, request completion, draft counters and idle slots. These are deployment checks, not additional comparative speed claims.

Success retains the newly verified pair; failures restore the previous exact5423 pair. The previous packages are retained for rollback. Cleanup unloads owned models and the outer wrapper immediately requests graceful shutdown and checks offline plus physical standby power. No idle inhibition, Block Load write, route retuning, alarm reset or proxy restart is used. Source publication is complete; build, activation and live acceptance are the remaining steps owned by root.

[novkien]

Exact-merge build complete; recover the live-check invocation

Both physical checkouts converged cleanly to c2d2a20df48b261d4ce94a3b14374ce4bfed9282. Native and RPC immutable packages final-q6-c2d2a20df4 built from that exact merge, with empty source-diff identity and passing allocator checks. The complete paired build receipt SHA-256 is 8d20af37547d0e031192f882cb9c10008d9dd406d8d1ff043ed2ddbb8a343ffb.

The first deployment interval installed the pair and loaded Qwen with matching running identity, but the request runner rejected a redundant --seed argument because the frozen request file already contains seed731. It exited during CLI validation before sending an inference request. This is a coordinator invocation defect, not an observed Q6 runtime failure. Root removed the redundant flag and exercised the real runner's argument validation offline, stopping before any network activity; the request's sampling fields remain unchanged.

Normal cleanup restored exact5423 on both hosts with no forced cleanup, then verified shutdown offline at 12:58:12 UTC with fresh6W, no hold and zero requests. The next interval will reuse the pinned completed merge-build receipt and immutable packages, recheck physical source and install hashes, and perform only the two planned live requests. No rebuild, extra comparative measurements, route changes or source repair is needed. Issue9 remains open until that deployment/acceptance completes.

[novkien]

Delivered and accepted: exact merged Q6 pair

PR23 is now deployed as matched native/RPC packages from merge c2d2a20df48b261d4ce94a3b14374ce4bfed9282, source identity c2d2a20df48b261d4ce94a3b14374ce4bfed9282:e3b0c44298fc1c149afbf4c8996fb92427ae41e4649b934ca495991b7852b855. Both physical source checkouts are clean at that merge. The reviewed tree and patch remain unchanged; the new package label is final-q6-c2d2a20df4. Both allocator checks passed. Existing 100 CPU-reference rows and both actual-route proofs bind the same source tree.

Root completed the corrected live check through proxy8082, using one frozen structured cold8192+512 request per model:

Model	Result	MTP drafted / accepted	Runtime invariants
Qwen 27B	HTTP200, 8192 new prompt tokens, 512 output, cache0, valid sample	481 / 350	Four native V100s; 66/66 offload; pipeline enabled; configured C400000 rounds to400128; two slots
Flash Next	HTTP200, 8192 new prompt tokens, 512 output, cache0, valid sample	587 / 315	Four V100s plus RPC GPU; target49/49 and draft50/50 offload; pipeline enabled; C266240/ub768; two slots

Both reported build b11406-c2d2a20df. Pre-/post-request executable and library hashes matched the exact new native/RPC manifests. Raw SSE hashes matched, native slots returned idle, model-file identities and persisted routes were unchanged, and normal unload retired both model/RPC processes. The final installed manifests still identify the NEW Q6 pair; cleanup did not revert the successful deployment. Previous exact5423 packages and atomic-install backups remain available for rollback.

The first attempt's CLI-only --seed error and successful rollback are preserved in the prior report. The completed build was reused; no additional comparative benchmark campaign ran. These post-deployment requests establish bounded running behavior, not new statistical speed or broad answer-quality claims. Owner selection accepts the earlier descriptive Qwen +3.51%-3.81% gain below the former5% threshold, with the previously reported variable Flash results unchanged.

Final coordinator receipt SHA-256: 696a0f83199e181b40764bd0316f5eb629dc092297d05b6f1dddfd383a08511e. Root's artifact readback report SHA-256: ed4c39875bea4c06c5ce4a4462f92cb28bc06a586b47562de59b0f88c39954bf. The wrapper exited0 with final target verified, no forced cleanup and no idle inhibition. Graceful shutdown was verified offline at 13:10:01 UTC on2026-10-04, with fresh 6W at13:10:00, no hold and zero active requests. No route write, Block Load write, alarm reset or proxy restart occurred.

Source publication, exact-source build, matched installation, normal activation, both-model live checks and shutdown are complete. Remaining work for the owner-selected Q6 delivery: none. Issue9 can close against this acceptance; unselected optimization families remain unselected.