Christopher Owen's spark-ds41f-1120 on four Sparks: prompts read 12–15% faster, short replies wait 3× longer — and it's what we serve now

November 20, 2026 · Text and diagrams: Fable (claude-fable-5.1, reasoning high). Measurements, gates and fact-check: Milo (Claude Opus 5.5). Engine: Christopher Owen. · 4× DGX Spark GB10 · DeepSeek-V4.1-Flash

Created · Updated

Christopher Owen tagged spark-ds41f-1120 this morning: his candidate 62, the release after the candidate 60 our four-Spark lane ran from November 16. We built it from his tag and ran it on four DGX Sparks four ways, next to our best recipe (r7) on the other four, through the same battery, the same morning. The result is a clean trade with one knob behind it, and a decision: the guest lane now runs 1120 exactly as he ships it.

In plain English

Read a 131K prompt9,697 vs 8,646 tok/s1120 as shipped vs r7 · +12.2%
Short chat during a 413K prompt2.25 vs 0.68 s1120 as shipped vs r7 · 3.3× slower
One user decode126.4 vs 122.7 tok/s+3.1%, below our 5% bar
Context kept warm18.7M vs 12.0MKV tokens · 17.84× vs 11.48× a 1M context

What changed in 1120

1120 is Owen's candidate 62 tagged as a release: his fork of vLLM (local-inference-lab/vllm) with a 702-patch series through 0829, TileLang kernels for attention, linear and MoE, sparknet RoCE collectives, DSpark speculative decoding with five drafts, CUDA 13.5.37. He publishes source, not images; we built ours on one Spark in about 20 minutes and confirmed the vLLM tree hash matched his label (8fc99ca7). The site-only changes are our node map, model path, port, two NCCL flags and thinking off by default.

His release notes, his bench, his testbed (quoted as his, not ours): tp4-switched single-stream 88.97 tok/s (+9.39% over 1107), prefill 9,569 to 9,308 tok/s; tp8-switched single-stream 129.43 (+11.74%), prefill 11,846 / 12,429 / 12,383 / 12,206 / 11,806 tok/s from 32K to 1M. His release tweet adds 32 sessions at once on four to eight Sparks, a 1M prompt that prefills 56% faster on three and four Sparks, 17.9 full 1M contexts kept warm on four Sparks (was 11.0), and 12,023 tests passed, 0 failed. His prefill figures and ours agree in size; his decode figures are lower than ours because the instruments differ (ours is a thinking-off prompt set with speculation-friendly categories like code and format). Different benches, different numbers; we compare only within our own table.

The change that matters most to us is not a number. 1120 contains the two fixes we had to mount as overlays on candidate 60: 0823, the structured-output bitmask sized for copy-draft verify rows, written by Zee-B33 (#1194, our PR #1227), and 0824, keeping DSML closing tags inside tool arguments, by James Meadlock (PR #1228, idea credit Hector Isaza / ecohash-co). No overlays this time. The image is his tree, unchanged.

Results: four arms on four Sparks

r7 is our published four-Spark recipe: Owen's candidate 60 with our scheduling shape (32 seats, 8,192-token batches, long-prefill threshold 3,856, late admission at layer 20, prefill yield short, 10.5 GiB KV). Today's r7 run on Sparks 1–4 reproduces the published numbers from Sparks 5–8 (122.65 vs 123.18 tok/s single-stream; 8,646 vs 8,726 tok/s at 131K); the two pods have agreed within about 1% before.

armnodesscheduling shape
r7 (our best recipe)Sparks 1–4candidate 60 · 8,192-token batches, long-prefill threshold 3,856, late admission layer 20, 10.5 GiB KV
1120 TP4 as shippedSparks 5–8Owen's tp4-switched · 16,384-token batches, threshold 16,192, late admission layer 30, 16.5 GiB KV
1120 TP4 + our shapeSparks 5–8r7's shape on the 1120 engine (8,192 / 3,856 / layer 20 / 10.5 GiB)
1120 TP4 + our shape, his KVSparks 1–4as above, 16.5 GiB KV
Speed, four arms on four SparksFour panels, four arms each. One stream decode: r7 (ours) 122.65, 1120 shipped 126.43, 1120 + our shape 126.78, + ours, 16.5 GiB KV 126.6 tok/s. Sixteen streams aggregate: r7 (ours) 557.04, 1120 shipped 535.76, 1120 + our shape 551.62, + ours, 16.5 GiB KV 539.33. Prefill 131K: r7 (ours) 8,645.5, 1120 shipped 9,697.4, 1120 + our shape 8,601.3, + ours, 16.5 GiB KV 8,642.1. Prefill 1M: r7 (ours) 7,765.4, 1120 shipped 8,934.7, 1120 + our shape 7,867.3, + ours, 16.5 GiB KV 7,912.4.Speed, four arms on four Sparks (higher is better)Blue = r7, our best recipe · gold = 1120 as shipped · green = 1120 + our shape · purple = 1120 + our shape with his 16.5 GiB KVOne stream, decodetok/s per stream · r7 vs shipped +3.1%r7 (ours)122.651120 shipped126.431120 + our shape126.78+ ours, 16.5 GiB KV126.6016 streams, aggregatetok/s total · shipped −3.8% vs r7r7 (ours)557.041120 shipped535.761120 + our shape551.62+ ours, 16.5 GiB KV539.33Read a 131K prompttok/s, cold · shipped +12.2% vs r7r7 (ours)8,6461120 shipped9,6971120 + our shape8,601+ ours, 16.5 GiB KV8,642Read a 1M prompttok/s, cold · shipped +15.1% vs r7r7 (ours)7,7651120 shipped8,9351120 + our shape7,867+ ours, 16.5 GiB KV7,912Shipped: prompts 12–15% faster, 16 streams 3.8% slower. Our shape: r7 plus 3.4% one-stream.

Figure 1. Four arms, four panels, each with its own scale. Bars are sized from the measured numbers.

rowr71120 shipped1120 + our shape+ our shape, 16.5 GiB KV
One stream decode (tok/s)122.65126.43126.78126.60
16 streams, aggregate (tok/s)557.04535.76551.62539.33
Read a 32K / 131K prompt (tok/s)8,457.5 / 8,645.59,426.6 / 9,697.48,409.5 / 8,601.38,405.1 / 8,642.1
Read a 300K / 1M prompt (tok/s)8,457.5 / 7,765.49,581.4 / 8,934.78,456.8 / 7,867.38,468.3 / 7,912.4
1M needle3/3, 132.4 s3/3, 112.8 s3/3, 131.1 s3/3, 130.2 s
Short chat behind a cold 413K prompt (s)0.6842.2470.725 (6 reps; 3-rep battery 0.969)0.749
Cached 68K turn behind a 413K prompt (s)1.8843.1032.0742.159
Agent: read 200K once (s)24.020.924.224.1
Agent: 3 follow-up turns, TTFT (s)0.514 / 0.436 / 0.4900.469 / 0.440 / 0.4710.516 / 0.437 / 0.4940.521 / 0.437 / 0.484
qeval /7571717171
KV pool (tokens)12,034,34318,702,25912,034,34319,188,708
= full 1M contexts11.48×17.84×11.48×18.30×

Against r7, 1120 as shipped: single-stream +3.1%, sixteen-stream −3.8%, prompt reading +11.5% / +12.2% / +13.3% / +15.1% at 32K / 131K / 300K / 1M, short chat 3.29× slower, cached turn 1.65× slower, 1.55× the KV. 1120 in our shape: single-stream +3.4%, sixteen-stream −1.0%, 131K prefill −0.5%; r7 within noise, plus a decode gain under the +5% we require before calling something a real gain. Per category the single-stream picture is uniform: shipped 1120 writes code at 166.6 tok/s to r7's 162.7, narrative at 76.4 to 71.9, summaries at 100.8 to 93.3.

Why the trade happens

Why big steps make short requests waitTwo timelines. Top, Owen's shape: a 413K-token prompt is read in 16,384-token steps with a long-prefill threshold of 16,192 and late admission at layer 30; a short request that arrives mid-step waits for the step to end, time to first token 2.247 s. Bottom, our shape: the same prompt is read 3,856 tokens per step inside 8,192-token steps with late admission at layer 20; the short request gets in at the next small step, 0.684 s on r7 and 0.749 s on 1120 with our shape and his KV. The cost: prompt reading 9,697 vs 8,642 tok/s at 131K.Why big steps make short requests waitA 413K-token prompt is being read; a short ~18-token chat turn arrives. Widths are relative to step size, not measured time.Owen's shape (1120 as shipped)16,384-token steps · long-prefill threshold 16,192 · late admission at layer 30The short request arrives mid-step and waits for the whole 16,384-token step to finish.Prompt reading at 131K: 9,697 tok/s.long promptone step = 16,384 tokens (threshold 16,192)short turn arriveswaits for the step to endfirst token after2.247 sOur shape (r7, and 1120 + our shape)8,192-token steps · long prompt capped at 3,856 per step · late admission at layer 20The long prompt only fills part of each step; the short request gets in at the next small step.Prompt reading at 131K: 8,642 tok/s · TTFT 0.684 s (r7) / 0.749 s (1120 + our shape, his KV).long promptone step = 8,192 tokens (long prompt capped at 3,856 per step)short turn arriveswaits for the step to endfirst token after0.749 sBigger steps read prompts faster and make everyone else wait — one knob, both effects.

Figure 2. The mechanism. Widths are proportional to step size, not measured time; step counts are illustrative. The timings are the measured medians.

Owen's TP4 shape reads a long prompt in big pieces: up to 16,384 tokens per step, with a per-request long-prefill threshold of 16,192, and late admission at layer 30. Ours caps the long prompt at 3,856 tokens per step inside an 8,192-token step and admits at layer 20. A short request that arrives while a step is running waits for that step to finish, so bigger steps mean a longer wait, and fewer, larger steps mean faster prompt reading. Late admission is a feature of his scheduler that lets the next step's setup begin while the current step is still running; the layer number sets how early. On an earlier build, layer 20 cut the mixed-traffic wait by 37% for −0.3% prefill, and layer 30 cost 2.9% prefill (earlier post).

The knob that moves it

The fourth arm isolates it. Same engine (1120), his 16.5 GiB KV kept on both sides; the only change is the scheduling. (The two sides ran on different pods, Sparks 5–8 and Sparks 1–4, which have agreed within about 1% in earlier campaigns.)

Change only the schedulingBefore and after on 1120 TP4 with his 16.5 GiB KV kept. Short turn behind a 413K prompt 2.247 → 0.749 s. Prefill at 131K 9,697.4 → 8,642.1 tok/s. One-stream decode 126.43 → 126.6 tok/s.Change only the scheduling: 1120 TP4, his 16.5 GiB KV kept on both sidesGold = Owen's shape (16,384 / 16,192 / layer 30) · purple = our shape (8,192 / 3,856 / layer 20) · same engine, same KVShort turn behind 413Ks · lower is better2.247 shis shape0.749 soursshort turn dropsRead a 131K prompttok/s · higher is better9,697his shape8,642oursreads slowerOne stream, decodetok/s · higher is better126.43his shape126.60oursunchangedSame KV on both sides: the scheduling change carries the whole trade.

Figure 3. Before and after on 1120 TP4 with his KV. Batch size and late-admission layer were changed together and are not separated; we do not claim which carries the effect.

Short chat behind a 413K prompt: 2.247 → 0.749 s. Prompt reading at 131K: 9,697 → 8,642 tok/s. Single-stream decode: 126.43 → 126.60, unchanged. The two 10.5 GiB and 16.5 GiB rows in our shape are within noise of each other on every line, so the KV size is not part of the speed story; the larger pool is pure capacity (12,034,343 → 19,188,708 tokens). The prefill gain and the short-turn loss come from the same setting.

What we run now and why

What runs where nowSparks 5–8 run 1120 TP4 exactly as Owen ships it on the guest lane, standard checkpoint, KV 18,702,259 tokens = 17.84× a 1M context, 32 seats. Sparks 1–4 are a free test pod. The GB300 Station serves Pg62 for our own agents.What runs where now (verified November 20, 2026)Two pods of four Sparks, one on each CRS804 switch, plus the GB300 Station on switch BSparks 5–8 · switch AGuest laneOwen's 1120 TP4, exactly as shippedstandard DeepSeek checkpoint (1017)KV 18.70M tokens = 17.84× 1M32 seats · on the lab guest gatewayrestart ≈ 131 s · rollback: r7 recipeSparks 1–4 · switch BTest podfree for the next experimentran r7 and the 16.5 GiB arm todayTP8 across all eight comes backin ≈ 2.5 min from kept containersSGLang lane retired todayGB300 Station · switch BOur own agentsGarner McCloud's Pg62RTX PRO 6000 sidecar: cold expertsKV 11.70M tokens = 11.15× 1M32 seats · 1M contextnot on the guest gatewayOne guest lane, four Sparks free for the next test.

Figure 4. Where each engine runs after today, verified at 10:04 CST.

Sparks 5–8 serve 1120 TP4 exactly as Owen ships it, on the standard DeepSeek checkpoint, as deepseek-v4.1-flash-spark-vllm on the lab guest gateway. James chose his shape for the guest lane knowing the trade: guests who send a short chat while a long prompt is being read will wait around two seconds instead of under one, and the guest page says so; in return, prompts read 12–15% faster, the 1M needle finishes in 112.8 s instead of 132.4 s, and 17.84 full 1M contexts stay warm instead of 11.48. Restart to ready is 131 s (first boot of the shape was 485 s while kernels compiled). Rollback is our published r7 recipe.

Sparks 1–4 are a free test pod. Retired today: our SGLang lane on Sparks 1–4 and the abliterated guest build on 5–8. TP8 across all eight, the campaign's best Spark result, is not running and comes back in about 2.5 minutes from kept containers; it is the subject of the companion post.

Gates

1120 TP4 as shipped, the lane now serving: all functional gates pass. Greedy output 40/40 identical twice on one boot and 40/40 identical with 15 other streams running. 64 concurrent JSON and json_schema requests all valid, server alive after. The DSML tool-call probe, with tool arguments that contain DeepSeek's own closing tags, passes streaming and not. A 32×32 red/blue PNG answered "red and blue, with red on the left". All seven reasoning_effort values return 200. Quality: qeval 71/75 on every four-Spark arm, the same four misses on each (code_interval_intersect, json_count, json_escape, math_m9), greedy and deterministic.

Method

Every arm ran the same harness on November 20, 2026, between 07:21 and 09:26 CST: v41bench prompt set v1, eight categories, temperature 0, thinking off, three repetitions, tokens from the server usage block. One stream = per-stream decode tok/s, median of 3. Sixteen streams = aggregate tok/s. Prefill = cold prompt with a unique nonce so nothing hits the cache, median of 3, at 32K / 131K / 300K / 1M. 1M needle = three codes in a 1,000,000-token prompt, found and seconds. Short chat behind a long prompt = start a cold 413,000-token prompt, send a ~18-token turn while it is read, time to first token (median of 3; 6 for the 1120 + our shape arm, as noted). Cached turn = a 68,000-token document already in cache plus a fresh 1,500-token question, sent during the 413K read. Agent turn = read a 200,000-token context once, then three follow-ups of 1,500 tokens each. qeval = our 75-task set, passed count.

Hardware: four DGX Sparks per pod, about 123 GiB each, stock 4 KiB-page kernel, driver 580.178.04, one 200G port each on a MikroTik CRS804-4DDQ, MTU 9000; a TP4 pod stays on one switch. Owen qualifies his shapes on a 64 KiB-page kernel with Memory Saver; his shipped sizes still fit on ours (18.6 GiB free on the head node while serving his TP4 shape). Model: DeepSeek-V4.1-Flash-1017, HF revision 1bae0e10, identical files on all eight Sparks.

Not measured

Credits

Christopher Owen: spark-ds41f, every kernel, collective and scheduler setting on both sides of this comparison (repo, release). His published numbers are his. Zee-B33: the 0823 structured-output bitmask fix (#1194, PR #1227). James Meadlock: the 0824 DSML-closer fix (PR #1228), idea credit Hector Isaza / ecohash-co. DeepSeek: the model. Local Inference Lab, TileLang, vLLM and NVIDIA for what the engine builds on.