Night shift 2026-09-12 (local) — maximum shift: eugr 2x, sparkring re-run, Aiden DCP2 A/B

Operator plan (before start): "максимальная смена" — first update-check of all sources, then 2x window (eugr + sparkring), morning: Aiden image re-check, r31 back to production, dgx5 power-off consideration.

Phase 0 — source update-check (22:15 UTC)

Aiden published production-1.1 at 16:54 UTC (Docker Hub) — the event of the shift. Pulled to all 4 nodes + dgx5 (23.4 GB, digest ede55d46).
vLLM main is 338 commits ahead of our pin e47aa780 (09-10, "[Model] Support DeepSeek-V4.1-Flash", still not on main) and DSV4 fixes started flowing into main (2 today: indexer gather kernel, dflash attention padding) → next-rebase campaign material.
FlashInfer: today's feat(sm120): runtime KV row stride + canonical 528B GLM53_NOPE rows + NaN-safe masked gather — candidate for the GLM lanes.
DeepGEMM: new "Public Release 26/09" (09-10).
SparkRing upstream: new r33-arm64-c8646b0 tag (R33 TP4/DCP4 SparkCache profile) + the memory-guard rebuild we needed tonight: glm53-shared-8b32ef2 = digest 4774a7c5 (confirmed via ghcr digest lookup; glm53-shared-74275757 = f25cdb6 = our current pin).
tonyd2wild: new EXL3 TP4 context lane (3.3M-token KV pool on four Sparks), now their default serving config — note for the next DSV4.1 campaign.
eugr: no fresh nightly (20260910 still the newest, already on the nodes).
LMCache: quiet since 09-04.

eugr 2x — nightly-20260910 (report `2026-09-11_223443_eugr-2x-0910`)

4x DSV4.1 stopped at 22:18 UTC; eugr 2x on dgx1+dgx2.
166/176 = 94/100 (81 pass / 4 partial TC-51/57/62/63 / 3 fail TC-45/68/80) vs r28 2x baseline 158/176 (90) — their b12x + humming-MTP + 1M-ctx stack is +8 points on 2 nodes.
Perf d0 c1 (pp/tg/TTFT ms): 1,198 / 29.0 / 2,113 vs r28 1,475 / 27.2 / 1,362 — slower prefill and a 55% higher TTFT (1M-ctx prefill buffer, the same effect we measured on DSV4.1 at 1M).
⚠ tool-eval-bench drifted to v2.6.1.dev65+g6be685f0e this shift (all earlier runs: dev46+g9d412464a) — cross-run comparisons spanning the drift are approximate.

sparkring 2x re-run on the memory-guard image (report `2026-09-11_235045_...`)

Pulled ghcr.io/fujitsupolycom/sparkring:glm53-shared-8b32ef2 (= 4774a7c5, the 2 GB memory-guard rebuild) to dgx1+dgx2.
Boot-1 bug (fixed in recipe, commit 13720ae): the new image's b12x loader JIT-compiles its native module at first boot and the linker cannot find libcudart (the old image prebaked it). Verified the fix in a scratch container: LIBRARY_PATH=/usr/local/cuda/lib64 (+ CUDA_HOME), plus a persistent B12X_CACHE_DIR mount so the compile survives reboots.
Reliability verdict: FAIL, again. The engine died at scenario ~62 (TC-62 ran 123 s, then all 26 remaining scenarios infra-failed: timeout/connection/5xx). 112/124, 70.5% completion — the "90/100" header is a partial score, not comparable to full runs. No system OOM (114 GB free at inspection) — the death is container-internal (their liveness/memory guard, per the 09-08 forensics). The 2 GB guard fix did not fix the sustained-load death.

Aiden production-1.1 — DCP2 A/B (report `2026-09-12_003640_aiden11-dcp2-0912`)

Image internals (verified on dgx5): vllm 0.28.1rc1.dev388+g8a728663c (base vllm/vllm-openai:nightly-8a728663, not our e47aa780 pin) + b12x 1.3.0 + FI 0.7.0rc1 + torch 2.13.0+cu130; the 24-file DCP2 patch set is baked (manifest at /opt/dsv41-dcp2-mounts.txt) + FI source with trtllm_dcp_alltoall.cu / cake_dcp.py. DCP env defaults are OFF (DSV41_DCP_STAGE=3, DCP_COMBINE=ag_rs, DCP_PREFILL_GATHER=0, ENGRAM_MMAP/PREFETCH=0); CLI: --decode-context-parallel-size 2 (-dcp), --dcp-comm-backend {ag_rs|a2a}.

Recipe dsv41-flash-4x-aiden11 = their DCP2 core (TP4 + dcp=2 + a2a + 8 GiB pinned KV, their validated 300K row) on our serving config (DSpark k=5 exact capture, SEQS 8, thinking+max, engram-local, deepseek-v4-flash alias).

Boot: clean, Worker_TP0_DCP0, weights 48/48 in 92 s (warm NFS), "DSA indexer: DCP2 with compress_ratio=2 (state-sharded K cache)", CUDA graphs 11 s, KV 2,518,386 tokens @ 300K (better than their claimed 1.89M), 8.39× 300K concurrency.
Result: 158/176 = 90/100 (74 pass / 10 partial / 4 fail TC-43/51/68/84) vs our b12x baseline 167/176 (95) @ 512K.
Perf d0 c1 (pp/tg/TTFT ms): 1,517 / 41.8 / 2,446 vs ours 1,642 / 47.0 / 1,738 — we win prefill, decode, and TTFT.
Verdict: the DCP2 A/B is ours to keep. DCP2 buys KV capacity (2.5M tokens) / high-concurrency production, not quality or single-stream speed. TC-43/84 (fixed by our b12x numerics) fail on their stack again.
Option left open: the 24-file DCP2 set is extractable from the image (docker cp of the manifest's targets) if the operator ever wants DCP2 on our clone-at-pin tree — no need to wait for Aiden's file dump.

End state

Production: glm53-flash-r31-4x-dcp4 (4nodes-switch) — restarted as morning prod per operator plan.
DSV4.1 lane fully staged: dsv41-flash-4x-b12x recipe + images on all nodes + engram-local copies; Aiden 1.1 image on all 4 nodes for re-runs.
dgx5: powered off (nothing pending on it; power on via HA when a build is needed).
Gist (published daytime): https://gist.github.com/kilork/1ba224bfe28779c571d096cf954fff31 — base recipe card + live b12x overlay instructions in the README.

Production state after the daytime window

Production: glm53-flash-r35-4x-dcp4 (4nodes-switch, promoted 09-12 per operator decision — r35 stays; r31 recipe remains as the rollback lane, image tagged on all nodes).
Follow-ups added: (a) r31 re-bench on dev65 at the next 4x window for a harness-clean delta (and to settle TC-51 variance-vs-regression); (b) 32K-prefill probe if the split-MoE prefill gain on GB10 ever matters (outside the current bench matrix).

Morning follow-ups (ranked)

Aiden production-1.2 on Hub (DCP2-1.2 with the launcher env values) — re-check; the 1.1 A/B already answers the direction.
vLLM rebase campaign: DSV4.1 support is landing in main (338 commits ahead of our pin) — re-pin + re-patch when the branch stabilizes.
FlashInfer sm120 update (today's GLM53_NOPE rows commit) for the GLM lanes.
SparkRing r33-arm64-c8646b0 (R33 profile) — if the reliability story ever changes.
tonyd2wild EXL3 TP4 lane (3.3M KV pool) — next DSV4.1 quant option.
gpuflip.py GB10 slow-state probe on our 4 nodes (read-only, not done).

Daytime follow-up: R35 adoption + A/B (09-12, operator window)

Adopted R35 (latest lab rev, qualified 2026-09-11, swiglu-reviewed- composition-r35.source.lock): full 3-fork cut — vLLM de982a50 (tree 6d5cc349), b12x 98086604 (tree d963f030), LMCache 29bc5a2e (tree 5a88a1ea). All pins verified fetchable + tree-matched; build-dep set r31↔r35 identical (no fetch-deps resync). profiles/glm53-r35.env + recipe glm53-flash-r35-4x-dcp4.
Image built on dgx5 (71 m 18 s, checksum 8bb87348), boot gates green (vllm string 0.1.dev0+glm53.r35.de982a50, b12x/lmcache/flashkda OK); pre-transferred dgx5→dgx1 (docker save|ssh, ~11 min) before the window.
A/B (4nodes-switch, r31 stopped, dev65 harness, seed 42, hardmode, c1/c2): | | r31 (09-10, dev46) | r35 (09-12, dev65) | |---|---|---| | score | 159/176 (90) | 163/176 (93) | | pp/tg/TTFT d0 c1 (ms) | 1,519 / 50.0 / 1,524 | 1,554 / 49.6 / 1,544 | | KV | 16 GiB pin | 8,060,928 tokens @512K (15.4x) | Scenario delta: r31's TC-33/TC-80 fails and TC-46/61/67 partials are GONE on r35 (TC-68 partial→pass); NEW on r35: TC-51 fail (step-ordering: sent email before observing create_calendar_event result) + TC-35/TC-40 partials. Cold inductor recompile cost ~25 min first boot (JIT warm since).
Interpretation: perf FLAT on GB10 (their +8.8% prefill gain was a 32K-prefill cell on RTX PRO 6000; our matrix maxes at 8192 prefix — unmeasured here), quality +4 but harness-confounded (dev46→dev65 credit fixes lean up). r35's real value: SwiGLU-clamp-safe MoE split path (the R33/R34 runaway- generation defect does not exist in r35), LMCache fs-eviction accounting, DFlash2 K7 qualified, latest lab line. TC-51 = variance vs regression unresolved (needs an r31 dev65 re-bench or a second r35 run).

Night shift (22:00–00:10 UTC+2 local, unattended ops + DFlash2 A/B)

DFlash2 4x A/B — WINNER on latency, +2 points

Recipe glm53-flash-r35-4x-dflash2 (clone of the r35 MTP3 prod, ONLY the speculative-config swapped: MTP3 → {"method":"dflash","model":"...DFlash2-MXFP8", "num_speculative_tokens":7,"kv_cache_dtype":"auto"}). No build, no download: the r35 image ships the DFlash2 machinery (DFlash2DraftModel, dedicated dflash_attn graph-safe split-KV backend, FlashInferCutlassMxfp8LinearKernel for the fused draft K/V projection) and the 1.2 GB MXFP8 draft was already on the shared NAS.

Boot gates green: Resolved architecture: DFlash2DraftModel, method=dflash K7, and — the 08-29 blocker is DEAD — the r35 tree pads the mamba page size (+0.70%) so draft/target page geometry unifies instead of the old 4096 % 660 != 0 abort. KV: 6,333,399 tokens @512K (12.08x) vs MTP3's 8,060,928 — the draft's 5 sliding-attention layers + BF16 draft pages eat part of the 16 GiB pin.

	r35 MTP3 (prod, 11:31)	r35 DFlash2 K7 (22:20)
score (dev65, seed 42)	163/176 (93)	165/176 (94)
pp / tg / TTFT, d0 c1	1,554 / 49.6 / 1,544 ms	1,977 / 47.3 / 1,071 ms
pp / tg, d0 c2	1,489 / 53.0	1,475 / 55.8
KV	8.06M tokens	6.33M tokens

Same harness both runs (clean A/B). Prefill +27%, TTFT −31%, decode −4.6% @c1 / +5.3% @c2, score +2 (TC-40/50/63 now pass; new partials TC-60 + one more; shared fails TC-43/45/51/68). The run package (manifest + recipe copy, the 081a0e6 feature) landed on its first live run; found and fixed a utc_now_iso bug in the same commit (civil_date got seconds, produced year 4900778).

Verdict (operator, 09-12 night): the DFlash2 lane is the better 4x config on this bench — but the operator keeps r35 MTP3 as prod for now (the day's decision) and logs DFlash2 as the 4x candidate: +2 points, big TTFT/prefill gain, −22% KV (6.3M vs 8.06M — matters for the 512K concurrency story). Restore path: sm121 stop + serve the MTP3 recipe.

DFlash2 2x Spark — BLOCKED on the engine side (definitive)

glm53-flash-r35-2x-dflash2 (r31-2x shape, r35 image, NVFP4-Spark target, same draft). First attempt (bare TP2, no DCP): both ranks abort 5 min into boot with ValueError: GLM MLA parent-page stride must be an exact number of C4 pages — the r35-tree successor of the 08-29 page-unification blocker. Second attempt (+DCP2, interleave 2): ValueError: GLM C4 DCP requires cp_kv_cache_interleave_size to be divisible by 4, got 2. The interleave must be a multiple of 4 → the only working DFlash2 shapes are the lab's certified TP4/DCP1 and TP4/DCP4 (4x only). This matches the lab's qualification exactly (their 2x Spark profile is MTP3-only). 2x stays MTP3; DFlash2 is a 4x-only lane on this tree. The recipe is kept as a regression canary (BLOCKED notes in the file).

Ops note: sm121's docker run carries --rm by default — a crashed rank-0 container vanished with its logs. sm121 serve --keep suppresses --rm and was used for the second attempt (logs survived, root cause captured).

Power-off (operator instruction)

Cluster 4nodes-switch powered off via HA after the A/Bs (prod state: down until the next serve; MTP3 r35 = prod of record, DFlash2 = candidate).

Day follow-up 09-13: DeepSeek-V4.1 lab r36 — built, works, LOSES on GB10

The lab published their FIRST DS4.1 lock (09-12): image jovian-judgement-community-20260912-r36 (vLLM 202a11a98e / JJ b40673cd + #738/#740, b12x 323107ff with DS4.1 native kernels, LMCache 29bc5a2e, FlashKDA 3b225bf2, DeepGEMM 8b1392b9, FI 0.6.17, + liburing 2.5 for the Engram disk-mode io_uring reads). DS4.1 profile: TP4/DCP1, native DSpark K7 greedy + standard rejection + adaptive verification (draft inside the checkpoint), B12X attn/MoE/linear, 131K ctx, 4 seqs / 4096 batched, block 256, FULL_AND_PIECEWISE ≤128, Engram RAM(188.83 GiB)/disk. Their RTX quartet numbers: 32K prefill ~14K t/s, C1 244–261 t/s.

We built the same source composition for GB10 (the r35 build dir + new profile jovian-r36-ds41, docker-images/glm53-jovian-r27):

Build 1: failed at our final gate — the b12x marker B12X_PCIE_ONESHOT_PDL no longer exists in b12x 323107ff (gate updated to _PREPARED_ONESHOT_LAUNCHERS, same file).
Build 2: image built, but first serve hard-errored at boot: RuntimeError: io_uring PLE support is unavailable: install liburing development headers and pkg-config, then rebuild the b12x native loader. The b12x loader is C sources (_ple_reader.c et al.) compiled at RUNTIME by b12x/loader/_native.py (cc, -Werror, liburing optional via pkg-config probe) — the image needs liburing-dev (headers), not just the liburing2 runtime lib. Build 3 (13m47s) with both: OK.
Boot gates on 4nodes-switch (warm cache transfers ~4 min/node, label 49b368a0): DSpark K7 SpeculativeConfig(method='dspark', num_spec_tokens=7), EngramConfig(table_memory='disk') (no io_uring error), KV 3,377,679 tokens (25.77x @131K), Application startup complete, smoke "четыре" ✅. One engine warning: "torch.compile is turned on, but the model ... does not support it" — the DS4.1 model class lacks @support_torch_compile in the r36 tree (true for their RTX runs too; piecewise cudagraphs, not inductor fusion).

A/B vs the current dsv41 prod (eugr image, DSpark K5 probabilistic, same dev65 harness/seed, 4nodes-switch):

	dsv41 prod (eugr)	lab r36 (this run)
score	167/176 (95)	158/176 (90)
pp / tg / TTFT @d0 c1	1,642 / 47.0 / 1,738 ms	401 / 33.7 / 5,061 ms
tg @d4096 c2	(b12x lane 1,642/47 anchor)	12.0
KV	512K pin	3.38M @131K

Verdict: r36 is a clear loser on GB10 — prefill 4.1x slower, TTFT 3x worse, decode −28%, quality −9. Prime suspect for the prefill collapse: the Engram PLE reader's io_uring row reads over our NFS (their RTX runs use a local NVMe checkpoint; 510 GB won't fit our 132 GB local free space), compounded by RTX-tuned scheduler/compile settings (131K ctx, 4 seqs, block 256, DCP1). The lab's stack is optimized for the RTX PRO 6000 (1.8 TB/s GDDR7); on GB10 the eugr lane stays production. Follow-ups (in order of value): (1) diagnose the Engram read path (log PLE read latency; try the r36 stack with our prod scheduler flags — 512K, DCP4, block 128, seqs 8 — to separate storage from tuning); (2) wait for the lab's sm_121 port if they ever tune for Spark; (3) the built image jovian-r36-ds41:local (label 49b6...) + recipe dsv41-flash-jovian-r36-4x stay in the repo as a known-working (losing) lane.

Day follow-up 2 (09-13): Engram on local NVMe — the NFS suspect CONFIRMED

Operator: "use engram equivalently from NVMe — NAS is a guaranteed loss." The eugr/aiden dsv41 lanes already stage per-rank sparse Engram copies on node NVMe (/var/tmp/engram-local/DeepSeek-V4.1-Flash, 48 GB real per node — two full-size 101.5 GB files, only the rank's ceil-row TP shard materialized, verified per-node row starts 0 / 96,000,564 / 192,001,740 / 288,003,654). The r36 engine's io_uring PLE reader reads (file, offset+rows) from the model dir — so it works unchanged against those files: same filename (the engine looks shards up by name via the index.json), per-node individual content.

Launcher feature: composed_model.local_files (commit eb63d88) — the standard composed-dir mechanism extended: entries are HARDLINKED (not copied; sparse stays sparse, zero disk) from a per-node host path into the per-node /tmp composed dir, after the repo symlink pass (local wins at the same name). Recipe dsv41-flash-jovian-r36-4x now uses [composed_model]: whole DSV4.1 snapshot symlinked into the HF cache + the two Engram shards hardlinked from the eugr lane's sparse copies. (A hand-rolled ~/models farm first failed in the container — host-side /mnt/nas symlink targets dangle inside the container; composed-dir symlinks use container-side paths by design, and hardlinks are path-independent — the failure is the feature validation.)

Result (4nodes-switch, dev65/seed 42, vs NFS run 158/176 and prod 167/176):

	prod (eugr K5)	r36 + Engram NFS	r36 + Engram NVMe
score	167/176 (95)	158/176 (90)	149/176 (85)
pp / tg / TTFT @d0 c1	1,642 / 47.0 / 1,738	401 / 33.7 / 5,061	3,654 / 43.9 / 826
pp / tg @d4096 c2	—	12.0 tg	1,711 / 34.9
KV	512K pin	3.38M @131K	3.34M @131K

The storage hypothesis is CONFIRMED: prefill 401 → 3,654 tok/s (9.1x, now 2.2x PROD), TTFT 5,061 → 826 ms (6.1x, 2.1x better than prod), d4096c2 decode 12.0 → 34.9 tg. The io_uring row reads over NFS were the entire prefill collapse — with the table on NVMe the r36 stack (K7 DSpark + B12X + V2 runner) is the fastest prefill lane we have.

Quality: 149 vs 158 (NFS) is run variance, not storage — the 11 differing scenarios are tool-discipline wobbles (skipped steps, empty query, extra calls; two improved: TC-35, TC-50), plus 2 new safety warnings on long-horizon P scenarios (TC-82/85). The r36-vs-prod gap (−9…−18) is the engine-stack delta (K7 greedy+adaptive vs K5 probabilistic), present in both r36 runs. r36 + local Engram is now a legitimate CANDIDATE lane: best prefill/TTFT we have, −18 quality vs prod. Next levers (if pursued): r36 with our prod scheduler flags (512K/DCP4/block 128/seqs 8) to separate stack from tuning; or K7 speculative-config tuning.

Day follow-up 3 (09-13): r36 @512K (our floor) + forum 381350 post #521

Operator: "128K is not usable — we need ≥512K." Recipe: --max-model-len 524288 (rest of the lab profile unchanged: DCP1, 4 seqs, 4096 batched, block 256, GMU 0.88, B12X, DSpark K7). Boot green, KV pool GREW to 9,616,957 tokens (18.34x concurrency @512K) — the larger max-len changes the engine's pool computation (vs 3.34M @131K), a free win.

dev65 (seed 42, same bench):

lane (4x)	pp / tg / TTFT @d0c1	score
dsv41 prod (eugr K5, 512K)	1,642 / 47.0 / 1,738	167/176 (95)
r36 + Engram NFS (131K)	401 / 33.7 / 5,061	158/176
r36 + Engram NVMe (131K)	3,654 / 43.9 / 826	149/176
r36 + Engram NVMe (512K)	3,774 / 45.7 / 822	154/176 (88)

d4096c2: 2,554 pp / 58.8 tg (vs 1,711/34.9 at 131K — the bigger pool shows under load). Quality: 154 — the recurring failures are the same long-horizon cluster across every r36 run (TC-82/85 fail-or-partial every time, TC-60/74 flapping) — systematic K7-stack behavior on multi-turn P scenarios, not storage, not variance.

Verdict: r36 + local Engram @512K is a serious candidate — best prefill (2.3x prod) and TTFT (2.1x better) in the fleet, decode within 3% of prod, −13 quality (the K7 stack's long-horizon gap). The storage story is settled: Engram must live on node-local NVMe (the composed_model local_files mechanism + the eugr lane's per-rank sparse copies).

Forum 381350 (#521, tea24864, 09-13, verified): --load-format matters. On 4x Spark GB10 (same fleet), GLM-5.3-Flash 320B/18B NVFP4-Spark, TP4: single-stream prefill ~2,950 tok/s with --load-format instanttensor vs ~1,900 with --load-format b12x (−35%, flat ratio 0.63–0.65 across 8k–128k ctx — a constant per-token cost; no memory difference). The b12x loader targets 2x-Spark memory pressure (O_DIRECT into CPU-addressable CUDA storage) — irrelevant at TP4, cost remains. ALSO: the b12x loader cannot load a separate DFlash2 drafter (checkpoint destination was not allocated as a weight: candidate_selector.predecessor_codebook). All our lanes use instanttensor or the default loader (dsv41 prod has no --load-format) — no action needed, but the b12x loader is now a known anti-pattern at TP4.

320B/18B verified from the checkpoint itself (local-inference-lab/GLM-5.3-Flash-NVFP4-Spark config.json): 45 layers, hidden 4096, 288 routed experts + 1 shared, top-k 8, moe_inter 2048 → ~315–320B total / ~18–19B active. The thread title is real.

2026-09-12.md 20 KB Geçmiş Ham