launcher/specs/tech-architecture/upstream-watch-2026-10-26.md

Upstream watch — 2026-10-26

Pass across the sources named in the todo list and the 2026-10-23/24 night reports. Only 2026-10-26 changes are treated as "today"; older changes are flagged as context.

1. FLARE — HEAD moved, 12 commits today (the headline)

NVIDIA/flare-local-inference main, commits on 2026-10-26:

sha	message
de6f53e	Count single-token reasoning delimiters only when generated as that token (#94)
aa84dd3	Carry the checkpoint license across package promotion (#83)
3a6b46f	Return undeclared MiniMax tool calls under automatic tool choice (#49)
9e35eeb	Prepare an opt-in MXFP8 DeepSeek V4.1 draft vocabulary head (#98)
d592f1f	Overlap DeepSeek V4.1 query expansion with cache preparation (#97)
63a4008	Fuse DSpark draft scoring and sparse support restoration (#95)
230c5b4	Move the host prefix-cache tiers into a CUDA-free selene crate (#80)
2f2875a	Align preparation guidance and examples with CLI defaults
82cb15a	Add optional GPU clock locking during Ignition autotuning
c370a85	Default disk embedding scales to host RAM on discrete GPUs
9e19b3b	Pin FlashInfer to the supported tuning-tier API (#85)
328ddce	test: check sectioned unverified payload inspection output (#89)

Why it matters to our prod lane (dsv41-flash-1017-4x-flare-gb10):

• 63a4008 + 9e35eeb + d592f1f are exactly the DSpark/draft/prefill path our recipe uses (--mtp-draft-tokens 5 --mtp-adaptive-verification).
• c370a85 overlaps our manual --embedding-table-resident-scales choice: the new head makes disk embedding scales resident by default on discrete GPUs.
• 82cb15a overlaps our before_serve clock lock: Ignition autotune can lock the GPU clock instead of relying on the hook.
• 9e19b3b pins FlashInfer to the tuning-tier API — a compatibility pin, so a binary/package mismatch is a real risk.

Our pins today: specs/tech-architecture/flare-upstream.pin = 3b9e86a0 (2026-10-22 15:48); the prod binary was built at 26646d85 (2026-10-23 19:18). HEAD is de6f53e — a bump needs a binary rebuild and a package prepared at the matching revision; the "proxy ABI" rule in flare-bringup.md is why a binary-only bump must not be applied to prod.

2. tool-eval-bench — v3.0.0 released today

SeraphimSerapis/tool-eval-bench released v3.0.0 on 2026-10-26 09:22 (commit d85126e). Highlights:

• optional local web UI (tool-eval-bench serve, [web] extra);
• tool-eval-bench card — one SVG per session, optional PNG;
• --perf-preset long up to 128K tokens, --context-fill-tokens;
• TC-21 and TC-89 grading changed; TC-45 structured-output retries before exclusion; suite remains 69 standard / 97 Hard Mode.

The release explicitly says TC-21/TC-89 and probe corrections are not comparable with earlier versions. The harness pin was moved to v3.0.0 on pluto.lan on 2026-10-26 (verified: tool-eval-bench --version = 3.0.0; install receipt commit d85126e). Reports such as 91/100, 88/100, 95/100 are v2.8.0 and stay historical. The anchor lane must be re-baselined before any v3.0.0 comparison.

Forum topic 366903 ("Introducing Tool Eval Bench CLI") has today's announcement (serapis, 09:12) and a user run with a card --png overwrite bug (Digital_David, 13:30/14:23), plus a request to delete old runs from the UI.

3. vLLM — no new release today

Latest stable tag is still v0.32.0 (2026-10-17); v0.32.1rc0 is 2026-10-18. No new stable tag on 2026-10-26. Our vLLM 0.32 lanes stay on the existing pin; the v0.32.0 stable re-pin remains a supervised task.

4. GR8 — 7 commits today

kindlingai/gr8 main, 2026-10-26:

• 33b1aa7 — require the model by name (glm53, glm53f) and refuse anything else (#234/#245);
• 29abc14 — k8s extraVolumes/extraVolumeMounts for a local release as MODEL (#246);
• dc8284d — LeaderWorkerSet Helm chart that runs ./gr8 up on every box (#235/#243);
• 64816bb — ./gr8 kv health readiness check for KV node containers (#241/#242);
• eb14798 — ring shape with mentat: discover cable order instead of --peers (#185/#190);
• db8576d — ./gr8 <model> health one readiness check for compose/K8s probes (#233/#240);
• 3264cb1 — container entrypoint: one bin/ for every TP and its own ulimits (#238).

This touches our GLM-5.3 / GR8 lanes: model-name enforcement, readiness, ring discovery and container packaging. A bump should be bench-gated.

5. DGPP — 4 commits today (+ DFlash2 batch yesterday)

HawkBearPig/dgpp master, 2026-10-26:

• 1babfa6 — merge PR #144 template structural limits;
• bb5fdfe — complete template limit boundary, concurrency and HTTP coverage;
• 882c2ee — align reasoning-control test with the always-on model contract;
• d9aae09 — merge origin/master into the template-limits branch.

2026-10-25 batch (relevant to Qwen3.8 DFlash2):

• 7f8f6de — merge DFlash2 Flash-Next PR #145;
• 314401a — document DFlash2 performance choices and test acceptance lengths;
• afce06f — honor draft-head screening and recurrent replay for Flash-Next DFlash2;
• 9513ab3 — shared DFlash2 graph, cache and sampling integration;
• 3dbeddc — honor explicit DFlash disable regardless of flag order;
• 90d602a — paired speculative probe accounting and request attribution;
• b671677 — drop the 27B YaRN x4 template from the DFlash2 branch;
• d0fafab — make the paired spec probe runnable;
• e8ed6da — serve: make the block-draft graph gate earnable;
• fc7b2d3 — ship the DFlash2 drafter as a knob.

The 2026-10-24 DGPP commit 606a514 (DeepSeek-V4.1-Flash-1017 EXL3, 16 slots, 96-row decode batch, five-draft block) is the DSV41-EXL3 path.

6. Blackwell LLM Docker — no commits today

Latest commit 2026-10-20: 8bb0967 / 5f803f7 — "ds41-flash: serve top_p 1.0, the DeepSeek-V4.1-Flash-1017 recommendation". Not today, but a substantive upstream change we should carry into the DSV41 recipe/bench context.

No new commits today in local-inference-lab/blackwell-llm-docker; its latest is Oct 20.

7. SparkRing — no commits today

Latest commits 2026-10-19: dashboard endpoints, --checkpoint 1017 of deepseek-v41-flash-sparkcache-tp4, Install Builder profiles. The todo #14 tag (dev-20261016-sparkcache) still needs checking against these.

8. TensorFold — no commits today

x00byte/DeepSeek-V4.1-Flash-1017-Dual-Spark latest commits 2026-10-18 (double production KV pool, dashboard throughput); no 2026-10-26 change. local-inference-lab/b12x was archived on 2026-10-06; its development moved to FlashInfer.

9. NVIDIA forum — today's posts

• 384438 "New Deepseek 4.1 flash recipe" — huxiaofengtiger (14:17 UTC) upgraded a 3-node TP3 setup to 1025 and ran the single-stream suite against R6x TP3: 60.94 vs 59.73 tok/s (+2.0%), TPOT p50 16.20 ms, TTFT higher; long-prefill workloads next. rob01 (02:19) reports the crash fix and TP3 low-RAM context.
• 366903 "Introducing Tool Eval Bench CLI" — v3.0.0 announcement and the card-overwrite bug noted above.
• 386811 "Welcome Athena Logos & Engine" — created 2026-10-25; today's comments are user feedback, not a new lane.

10. HF model cards — no new 2026-10-26 revision

• RedHatAI/GLM-5.3-MXFP4xFP8 last modified 2026-10-22 14:05 (a5cf1930…).
• local-inference-lab/DeepSeek-V4.1-Flash-1017-FLARE last modified 2026-10-23 00:31 (6861e802…).

No new checkpoint today; the package/revision pins stay as they are.

What this means for the todo list

1. FLARE tuning is now upstream-ready: bump de6f53e is tempting because DSpark scoring/query expansion/MXFP8 draft head/embedding scales/clock lock all touch our lane, but prod requires a matching package rebuild. Treat as a supervised experiment on dgx5–8 first, not a prod patch.
2. tool-eval-bench v3.0.0 is the one launcher-side upstream change today; the pin is now v3.0.0 on pluto.lan. The anchor re-baseline is the open task; do not mix v2.8.0 report scores with v3.0.0 runs.
3. GR8 and DGPP moved today; both are bench-gated bumps, not overnight re-pins.
4. vLLM has no new release; the 0.32.0/0.32.1rc0 picture is unchanged.
5. No new checkpoint/package today — no download or staging action is justified by upstream alone.