Home / 512 GB

512 GB · 4× DGX Spark (GB10)

Bizuayeu's GLM-5.3-Flash XSFP4 on four DGX Sparks (TensorFold 1.2.9)

Weights: XSFP4: 4-bit NVFP4 routed experts, mostly 8-bit elsewhere, 165.7 GiB (requantised by Bizuayeu from NVIDIA's GLM-5.3-Flash-FP4-Local, revision 88736f91)

4x NVIDIA DGX Spark GB10 (128 GB each), linked by ConnectX-7 RDMA cables in a switchless ring (two links per cable, about 109 Gb/s each), GPU clock capped at 2200 MHz, kernel 7.0.0-1019-nvidia, driver 580.178.04

Weights (XSFP4 checkpoint, revision 88736f91) and serving settings (GLM-5.3-Flash BIZ 3.15.0, v3/config/tp4.env)BizuayeuGitHubHugging Face
TensorFold 1.2.9 official release package (linux-aarch64, host binary, Zig 0.17.0), NCCL 2.30.7 from the image of Mia's recipe 1104 via LD_LIBRARY_PATH, TP=4 ring, TF_GLM_PREFILL_REDUCE=gather, MTP draftscommit release v1.2.9 (2026-11-15 01:22 UTC), tarball sha256 91e90c489bef3d217c2d3d51ce34d2a9a160ffc46b36e8b25026016b90f32f04; tag object 539dc9d6; checkpoint rev 88736f916f8939607fde8a4b881dcae749f7551d4 GB10 boxes (CX7 Ring, TP4, switchless)measured 2026-11-16

Not ranked yet. The rank needs a finished, verified Gauntlet race (the game test), and this setup has not played it yet. It shows Speed v2, Quality v2 and the memory test, every number with its raw files and exact command. How the game test works.

Speed, the author's method

The checkpoint author's own test first

Bizuayeu's own tool (glm53_tf), his settings, temperature 0, MTP drafts on, one request at a time. His figures come from the setup README of GLM-5.3-Flash BIZ (v3/README.md, section Measured on the Release, table column TP=4, XSFP4, lines 188 to 196): release 3.12.0 of 12 Nov 2026, engine 1.2.7, his reference ring of four MSI EdgeXpert boxes. Ours ran on 16 Nov 2026 on the same weights (revision 88736f91) with TensorFold 1.2.9 on our four boxes. The model card of the checkpoint lists speeds for TP=2 and TP=3 only. Not repeated here: the decode check for the prose and code prompts, the 499,254 and 1,035,295-token prompts and the cut-prompt test. Our standard test, Speed v2, follows below: sampled at temperature 1.0 with thinking off, which accepts fewer drafts, so its decode figure is lower.

MeasurementAuthor (his README)This run (his tool)DetailRaw files
Decode, bench --kinds decode (512 tokens, median of 3)74.14 tok/s86.1 tok/sruns 87.13, 86.09, 86.10result · command
Edit reply, bench --kinds edit (median of 3)239.69 tok/s236.77 tok/sruns 229.72, 236.88, 236.77result · command
Prefill of about 39K tokens, bench --kinds prefill (median of 3; his prompt 38,960 tokens, ours 38,961 to 38,964)1,868.0 tok/s1,811.3 tok/sruns 1,811.3, 1,824.6, 1,802.9result · command
Decode check, prompt kind count (median of 3)58.31 tok/s66.6 tok/sruns 66.6, 65.29, 67.74; acceptance length 3.406 against his 3.483result · command
A prompt of about 200K tokens with one passphrase in the middle, first token (his 199,307 tokens, ours 199,652)107.9 s110.0 sone run, the passphrase was foundresult · command
Speed v2

How fast it runs

Decode
52.0tok/s
Decode, code
52.4tok/s
Decode, prose
51.6tok/s
500 words
≈ 13.1s
Prefill at 8K
1605.8tok/s
First token at 2K
1.3s
MeasurementResult
Decode52.0 tok/s
Words per second38.3 words/s
Characters per second242.2 chars/s
Decode, code prompt52.4 tok/s
Decode, prose prompt51.6 tok/s
Spread of the decode runs (CV)5.8 %
Spread of the code runs (CV)5.3 %
Spread of the prose runs (CV)5.8 %
Decode, window 151.9 tok/s
Prefill at 2K1603.8 tok/s
Prefill at 8K1605.8 tok/s
Prefill at 32K1734.7 tok/s
Time to first token at 2K1.3 s
Time to first token at 8K5.1 s
Time to first token at 32K18.9 s
Draft acceptance rate49.4 %
Mean accepted length2.2 tokens
1 request at once (total)50.3 tok/s
2 at once (total)63.6 tok/s
4 at once (total)76.5 tok/s
8 at once (total)89.6 tok/s
1 request at once (per request)50.3 tok/s
2 at once (per request)31.9 tok/s
4 at once (per request)19.5 tok/s
8 at once (per request)11.5 tok/s
Server slots (8: --parallel 8, the author's setting)8 slots

Long prompts

Prompt lengthPrefillTime to first tokenDecode, code promptDecode, prose prompt
16K tokens1705.8 tok/s9.6 s51.9 tok/s52.1 tok/s
32K tokens1752.8 tok/s18.7 s50.4 tok/s50.9 tok/s
64K tokens1784.5 tok/s36.7 s49.4 tok/s50.1 tok/s
128K tokens1788.9 tok/s73.3 s51.4 tok/s49.7 tok/s

Method: code+prose; 2 run(s) each; 512 tokens; 1 warm-up(s); median; requests at once (later supplementary run of 2026-11-16, same setup): C1,C2,C4,C8 at once; 2 rounds each; 512 tokens per request; prose prompt; total = all tokens / wall time, per request = tokens / end-to-end time; median over the rounds; long prompts (later supplementary run of 2026-11-16, same setup): prompts of exactly 16K, 32K, 64K and 128K tokens (a filler log first, the code or prose task prompt last), 2 runs per step, 512 output tokens; decode = tokens after the first / time after the first token; prefill = prompt tokens / time to first token of the same request, median over its requests. Measured 2026-11-16. Raw files: speed result · speed contract · log · speed result · speed contract · log · speed result · speed contract · log.

Requests at once: the server has 8 slots (from the author's recipe). What ran. Ash's TensorFold 1.2.9, the official release package for Linux arm64 (tar checksum 91e90c48, signature checked; run as the host binary, with the NCCL 2.30.7 library, the version the author names, taken from the container image of Mia's TensorFold GLM recipe of 4 Nov 2026 and loaded through LD_LIBRARY_PATH as the package's RUNTIME.md describes), serving Bizuayeu's GLM-5.3-Flash XSFP4 (Bizuayeu/GLM-5.3-Flash-XSFP4, revision 88736f91, 165.7 GiB; 42 large files checked against the Hugging Face SHA-256 values) on four DGX Spark boxes with tensor parallelism 4 in a switchless ring: each box is cabled to its two neighbours only, there is no link between rank 0 and rank 2 or between rank 1 and rank 3. The four boxes were checked to be identical (engine, library, settings) before the start, and the 8 ring links measured 109.25 to 109.34 Gb/s. The settings are the author's for TP=4 (v3/config/tp4.env and serve.sh of GLM-5.3-Flash BIZ 3.15.0) and the lines of the TensorFold recipe for a four-box ring. Where this run differs from the author's. (1) Engine: the release 1.2.9 as a host binary, where the author's package runs release 1.2.7 in his own container image. (2) The bootstrap connection between the ranks runs over the local network, where the author forwards between the two cable ports; it carries no tokens. (3) The image libraries libturbojpeg and libwebpdecoder were missing, so only JPEG and WebP input would fail to decode; the tests here send text only. The GPU clock cap of 2200 MHz is the same as the author's. Against the author's own figures. The model card of the checkpoint lists speeds for TP=2 and TP=3 only (decode 54.15 and 66.23 tok/s, prefill of 38,960 tokens 1,415.1 and 1,717.3 tok/s, with the author's own tool on his reference boxes); it has no figure for four boxes. For TP=4 the author's setup README (GLM-5.3-Flash BIZ v3/README.md, section Measured on the Release, release 3.12.0 of 12 Nov 2026, engine 1.2.7, reference ring of four MSI EdgeXpert boxes) gives: decode 74.14 tok/s, an edit reply 239.69 tok/s, a prefill of 38,960 tokens 1,868.0 tok/s, a 199,307-token prompt in 107.9 s. Those four were measured here first, with the author's own tool (glm53_tf bench, temperature 0, drafts on, the same prompts) on the same weights: see the table on this page, which sits before our standard test. Our standard test below is a different method: sampling at temperature 1.0 (the checkpoint's setting), thinking off, 512 output tokens, where MTP drafts are accepted less often (acceptance 49.4 %, 2.24 tokens per round; with greedy decoding the author's tool shows 3.4 tokens per round and more), so its decode speed is lower than the greedy figure of the author's method. The first run of the speed test stopped at the 8-at-once step: the guard saw a foreign kernel process (kcompactd0, memory compaction) above 25 % of one core for two samples in a row, up to 78 %, after 3 repeats. Its values for 1, 2 and 4 at once (48.5, 64.0, 76.4 tok/s) are not shown; the whole 1, 2, 4 and 8 series on this page is from a repeat on a freshly started server, and the context steps 16K to 128K are from a second repeat on another fresh start. Between the first run and these repeats the local-network port of one box dropped its link repeatedly for 14 minutes and ended a context run; that run is not part of this page, and the cause of both events is not found.

Exact command we measured withfor reproducing↓

You do not need this to run the model. It is the exact command of our measurement, so anyone can repeat the result.

Speed v2 (run glm-xsfp4-tf129-tp4-1611-speed-r0)
python3 speed2.py --base-url http://127.0.0.1:44727/v1 --label glm-xsfp4-tf129-tp4-1611-speed-r0 --prompts $BENCH/repos/rigmark-v2/prompts-v2.json --output $RUN/glm-xsfp4-tf129-tp4-1611-speed-r0/glm-xsfp4-tf129-tp4-1611-speed-r0/speed2 --contract $RUN/glm-xsfp4-tf129-tp4-1611-speed-r0/glm-xsfp4-tf129-tp4-1611-speed-r0/speed2/speed-contract.json --parts decode,prefill,conc,l1,ctx --prefill-depths 2048,8192,32768 --prompt-ids code,prose --runs 2 --max-tokens 512 --warmups 1 --prefill-runs 3 --conc-levels 1,2,4,8 --conc-rounds 2 --conc-max-tokens 512 --conc-prompt prose --ctx-depths 16384,32768,65536,131072 --ctx-runs 2 --ctx-max-tokens 512 --tokenizer-json $HOME/models/glm53-mia-4bpw-1104/tokenizer.json --tokenizer-python $HOME/not published --sampling {"temperature": 1.0, "top_p": 0.95} --sampling-source Checkpoint generation_config.json:9-10 (temperature 1.0, top_p 0.95; Bizuayeu/GLM-5.3-Flash-XSFP4 rev 88736f91, box-3 $HOME/hf-tp4-2510/xsfp4-glm53-1611/generation_config.json). Bizuayeu's serve.sh (v3/scripts/serve.sh:74-93) sets no top_k and no temperature: the server default applies to requests that give none. Speed v2 runs with thinking off (enable_thinking false per request), seed 42. --model GLM-5.3-Flash --no-cache-prompt --server-log $RUN/server-A.log --engine-version-file $RUN/engine-version.txt --server-slots 8 --server-slots-source Bizuayeu/GLM-5.3-Flash-DGX-Sparks-BIZ@542aacc4 v3/config/tp4.env:8 (PARALLEL=8); rank 0's start log shows 8 streams at once --server-ctx 1048576 --server-ctx-source the card of this setup, field ctx

# Requests at once: supplementary run of 2026-11-16 (raw folder glm-xsfp4-tf129-tp4-1611-conc-speed-r0-conc)
python3 speed2.py --base-url http://127.0.0.1:34349/v1 --label glm-xsfp4-tf129-tp4-1611-conc-speed-r0 --prompts $BENCH/repos/rigmark-v2/prompts-v2.json --output $RUN/glm-xsfp4-tf129-tp4-1611-conc-speed-r0/glm-xsfp4-tf129-tp4-1611-conc-speed-r0/speed2 --contract $RUN/glm-xsfp4-tf129-tp4-1611-conc-speed-r0/glm-xsfp4-tf129-tp4-1611-conc-speed-r0/speed2/speed-contract.json --parts conc --prefill-depths 2048,8192,32768 --prompt-ids code,prose --runs 2 --max-tokens 512 --warmups 1 --prefill-runs 3 --conc-levels 1,2,4,8 --conc-rounds 2 --conc-max-tokens 512 --conc-prompt prose --ctx-depths 16384,32768,65536,131072 --ctx-runs 2 --ctx-max-tokens 512 --tokenizer-json $HOME/models/glm53-mia-4bpw-1104/tokenizer.json --tokenizer-python $HOME/not published --sampling {"temperature": 1.0, "top_p": 0.95} --sampling-source Checkpoint generation_config.json:9-10 (temperature 1.0, top_p 0.95; Bizuayeu/GLM-5.3-Flash-XSFP4 rev 88736f91, box-3 $HOME/hf-tp4-2510/xsfp4-glm53-1611/generation_config.json). Bizuayeu's serve.sh (v3/scripts/serve.sh:74-93) sets no top_k and no temperature: the server default applies to requests that give none. Speed v2 runs with thinking off (enable_thinking false per request), seed 42. --model GLM-5.3-Flash --no-cache-prompt --server-log $RUN/server-A.log --engine-version-file $RUN/engine-version.txt --server-slots 8 --server-slots-source Bizuayeu/GLM-5.3-Flash-DGX-Sparks-BIZ@542aacc4 v3/config/tp4.env:8 (PARALLEL=8); rank 0's start log shows 8 streams at once --server-ctx 1048576 --server-ctx-source the card of this setup, field ctx --supplement-of glm-xsfp4-tf129-tp4-1611-speed-r0

# Long prompts: supplementary run of 2026-11-16 (raw folder glm-xsfp4-tf129-tp4-1611-ctx3-speed-r0-ctx)
python3 speed2.py --base-url http://127.0.0.1:36933/v1 --label glm-xsfp4-tf129-tp4-1611-ctx3-speed-r0 --prompts $BENCH/repos/rigmark-v2/prompts-v2.json --output $RUN/glm-xsfp4-tf129-tp4-1611-ctx3-speed-r0/glm-xsfp4-tf129-tp4-1611-ctx3-speed-r0/speed2 --contract $RUN/glm-xsfp4-tf129-tp4-1611-ctx3-speed-r0/glm-xsfp4-tf129-tp4-1611-ctx3-speed-r0/speed2/speed-contract.json --parts ctx --prefill-depths 2048,8192,32768 --prompt-ids code,prose --runs 2 --max-tokens 512 --warmups 1 --prefill-runs 3 --conc-levels 1,2,4,8 --conc-rounds 2 --conc-max-tokens 512 --conc-prompt prose --ctx-depths 16384,32768,65536,131072 --ctx-runs 2 --ctx-max-tokens 512 --tokenizer-json $HOME/models/glm53-mia-4bpw-1104/tokenizer.json --tokenizer-python $HOME/not published --sampling {"temperature": 1.0, "top_p": 0.95} --sampling-source Checkpoint generation_config.json:9-10 (temperature 1.0, top_p 0.95; Bizuayeu/GLM-5.3-Flash-XSFP4 rev 88736f91, box-3 $HOME/hf-tp4-2510/xsfp4-glm53-1611/generation_config.json). Bizuayeu's serve.sh (v3/scripts/serve.sh:74-93) sets no top_k and no temperature: the server default applies to requests that give none. Speed v2 runs with thinking off (enable_thinking false per request), seed 42. --model GLM-5.3-Flash --no-cache-prompt --server-log $RUN/server-A.log --engine-version-file $RUN/engine-version.txt --server-slots 8 --server-slots-source Bizuayeu/GLM-5.3-Flash-DGX-Sparks-BIZ@542aacc4 v3/config/tp4.env:8 (PARALLEL=8); rank 0's start log shows 8 streams at once --server-ctx 1048576 --server-ctx-source the card of this setup, field ctx --supplement-of glm-xsfp4-tf129-tp4-1611-speed-r0
Quality v2

How well it answers

Thinking off for the first six tests (per request), seed 42, 8 requests at once, sampling of the checkpoint (temperature 1.0, top-p 0.95). Answer limits: 1,024 tokens for MMLU-Pro and GPQA (letter answer), 4,096 for IFEval, 32,768 for LiveCodeBench, 128 for the two long-text tests, 16,384 for GPQA with thinking. Answers cut off at the limit count as wrong: in the first job 33 answers were cut off (MMLU-Pro 3, GPQA 0, IFEval 5, long text 8K 13, long text 32K 12). LiveCodeBench ran in its own job after the others, with the generated code run in a sandbox: 12 of the 40 answers were cut off at the limit. GPQA 30 with thinking on uses the checkpoint's thinking sampling: 5 of its 30 answers were cut off. The GPQA files show only counts: the text of the answers is left out because it quotes the questions. Raw numbers only, no overall score.

TestResult95 % marginRightCut offcount as wrongRaw files
MMLU-Pro
400 fixed questions, letter only
59.3 % ± 4.8 pts237 / 4003summary · log · contractcommand
GPQA Diamond
198 questions, letter only
44.4 % ± 6.9 pts88 / 1980answers · summary · log · contractcommand
IFEval
200 fixed prompts, strict
65.0 % ± 6.6 pts130 / 2005summary · log · contractcommand
LiveCodeBench
40 newest tasks, the code runs
65.0 % ± 14.2 pts26 / 4012summary · log · contractcommand
Long text 8K
4 tasks x 10 at 8192 tokens
62.5 % ± 14.4 pts25 / 4013log · contract · answers · summarycommand
Long text 32K
4 tasks x 10 at 32768 tokens
42.5 % ± 14.6 pts17 / 4012log · contract · answers · summarycommand
GPQA 30, thinking on
30 questions, limit 16,384 tokens
70.0 % ± 15.6 pts
median 846 tokens, 52.2 s per question
21 / 305log · answers · summarycommand

Thinking off, except the last test. Seed 42. Sampling from the model card. A cut-off answer counts as wrong. Measured 2026-11-16.

Exact commands we measured withfor reproducing↓

You do not need this to run the model. It is the exact command of our measurement, so anyone can repeat the result.

MMLU-Pro (run glm-xsfp4-tf129-tp4-1611-inst-r1)
bash tools/quality2/measure_all.sh tools/quality2/cards/glm53-flash-xsfp4-tf129-official-4x-tp4-ring-1611.json --parts mmlupro --job glm-xsfp4-tf129-tp4-1611
GPQA Diamond (run glm-xsfp4-tf129-tp4-1611-inst-r1)
bash tools/quality2/measure_all.sh tools/quality2/cards/glm53-flash-xsfp4-tf129-official-4x-tp4-ring-1611.json --parts gpqa --job glm-xsfp4-tf129-tp4-1611
IFEval (run glm-xsfp4-tf129-tp4-1611-inst-r1)
bash tools/quality2/measure_all.sh tools/quality2/cards/glm53-flash-xsfp4-tf129-official-4x-tp4-ring-1611.json --parts ifeval --job glm-xsfp4-tf129-tp4-1611
LiveCodeBench (run glm-xsfp4-tf129-tp4-1611-lcb-inst-r0)
bash tools/quality2/measure_all.sh tools/quality2/cards/glm53-flash-xsfp4-tf129-official-4x-tp4-ring-1611.json --parts lcb --job glm-xsfp4-tf129-tp4-1611-lcb
Long text 8K (run glm-xsfp4-tf129-tp4-1611-inst-r1)
bash tools/quality2/measure_all.sh tools/quality2/cards/glm53-flash-xsfp4-tf129-official-4x-tp4-ring-1611.json --parts ruler --job glm-xsfp4-tf129-tp4-1611
Long text 32K (run glm-xsfp4-tf129-tp4-1611-inst-r1)
bash tools/quality2/measure_all.sh tools/quality2/cards/glm53-flash-xsfp4-tf129-official-4x-tp4-ring-1611.json --parts ruler --job glm-xsfp4-tf129-tp4-1611
GPQA 30, thinking on (run glm-xsfp4-tf129-tp4-1611-think-r2)
bash tools/quality2/measure_all.sh tools/quality2/cards/glm53-flash-xsfp4-tf129-official-4x-tp4-ring-1611.json --parts think --job glm-xsfp4-tf129-tp4-1611
Credits

Also made this possible

Setup

Run it yourself

What you need and how to start it, step by step. The results below come from exactly this setup.

What you need

Hardware
4x NVIDIA DGX Spark GB10 (128 GB each), linked by ConnectX-7 RDMA cables in a switchless ring (two links per cable, about 109 Gb/s each), GPU clock capped at 2200 MHz, kernel 7.0.0-1019-nvidia, driver 580.178.04
Model
Bizuayeu GLM-5.3-Flash XSFP4 (rev 88736f91, 165.7 GiB; derivative of nvidia/GLM-5.3-Flash-FP4-Local rev 76348ddb, experts requantized from zai-org/GLM-5.3-Flash)
Weights (XSFP4 checkpoint, revision 88736f91) and serving settings (GLM-5.3-Flash BIZ 3.15.0, v3/config/tp4.env)
huggingface.co/Bizuayeu/GLM-5.3-Flash-XSFP4
Source checkpoint of the weights (GLM-5.3-Flash-FP4-Local, revision 76348ddb)
huggingface.co/nvidia/GLM-5.3-Flash-FP4-Local
Base model (GLM-5.3-Flash, BF16 weights the experts were requantized from)
huggingface.co/zai-org/GLM-5.3-Flash
Engine
TensorFold 1.2.9 official release, tensor parallel on four boxes in a switchless ring (TP=4) — the program that loads the model and answers requests · github.com/ashhart/TensorFold/releases/tag/v1.2.9

Steps

  1. Get the engine and the model files from the list above.

  2. Start the server with this command. It loads the model into memory. When the model has loaded, the server waits for requests on port 8888, only from this computer.

    env TF_GLM_RANK=<0..3> \
      TF_GLM_WORLD=4 \
      TF_GLM_PREFILL_REDUCE=gather \
      TF_GLM_ROCE_MULTI=0 \
      TF_KV_SHARE=32 \
      TENSORFOLD_PREFILL_SHARE=0.7 \
      TENSORFOLD_MEMORY_RESERVE_GIB=14 \
      TF_HEAT_HIGH=92 \
      TF_HEAT_LOW=88 \
      NCCL_CUMEM_ENABLE=1 \
      NCCL_RUNTIME_CONNECT=1 \
      NCCL_ALGO=Ring \
      NCCL_GRAPH_FILE=<one graph file per rank> (further NCCL lines as in the author's examples/tp4-rank0.env) taskset -c 0-4,10-14 tensorfold-native serve <XSFP4 model folder> --name GLM-5.3-Flash --host 127.0.0.1 --port 8888 --context 1048576 --max-tokens 32768 --no-update-check --reasoning-effort high --parallel 8 --mtp-confidence 0.5 --vision

    What the parts mean

    • --parallel 8how many requests it answers at the same time
    • --host 127.0.0.1who can connect: only this computer
    • --port 8888the network port the server listens on
    • --name GLM-5.3-Flashthe model ID the server advertises to clients
    • --reasoning-effort highthe reasoning effort given to the chat template: how much the model thinks before it answers
Start text for this setupfor reproducing↓

You do not need this to run the model. It is the start text of this setup, based on the setup card of this run, with local paths, machine names and addresses shortened or replaced by placeholders, notes added on some pages, and the engine build in front, so anyone can repeat the result.

Start, based on the setup card of this run
TensorFold 1.2.9 official release package (linux-aarch64, host binary, Zig 0.17.0), NCCL 2.30.7 from the image of Mia's recipe 1104 via LD_LIBRARY_PATH, TP=4 ring, TF_GLM_PREFILL_REDUCE=gather, MTP drafts; env TF_GLM_RANK=<0..3> TF_GLM_WORLD=4 TF_GLM_PREFILL_REDUCE=gather TF_GLM_ROCE_MULTI=0 TF_KV_SHARE=32 TENSORFOLD_PREFILL_SHARE=0.7 TENSORFOLD_MEMORY_RESERVE_GIB=14 TF_HEAT_HIGH=92 TF_HEAT_LOW=88 NCCL_CUMEM_ENABLE=1 NCCL_RUNTIME_CONNECT=1 NCCL_ALGO=Ring NCCL_GRAPH_FILE=<one graph file per rank> (further NCCL lines as in the author's examples/tp4-rank0.env) taskset -c 0-4,10-14 tensorfold-native serve <XSFP4 model folder> --name GLM-5.3-Flash --host 127.0.0.1 --port 8888 --context 1048576 --max-tokens 32768 --no-update-check --reasoning-effort high --parallel 8 --mtp-confidence 0.5 --vision (the same arguments on all four boxes, ranks 1 to 3 started first and rank 0 last; settings from the author's v3/config/tp4.env lines 6 to 14 and v3/scripts/serve.sh lines 55 and 74 to 85 at commit 542aacc4, ring lines from the TensorFold recipe docs/recipes/glm-5.3-flash-cuda.md lines 176 to 186; the official release package tensorfold-1.2.9-linux-aarch64, not built by us)

Context and sampling: 1,048,576 tokens (the model's limit and the author's setting), 8 server slots (--parallel 8), MTP drafts with --mtp-confidence 0.5, vision tower on, BF16 KV cache (engine default), memory left to the system 14 GiB, per-request thinking off in the speed and quality tests, seed 42, sampling of the checkpoint's generation_config.json (temperature 1.0, top-p 0.95). The window and the number of KV pages are read from the start log of rank 0: 6,874 pages, 8 streams.. Hardware: 4x NVIDIA DGX Spark GB10 (128 GB each), linked by ConnectX-7 RDMA cables in a switchless ring (two links per cable, about 109 Gb/s each), GPU clock capped at 2200 MHz, kernel 7.0.0-1019-nvidia, driver 580.178.04. All raw files of this setup: raw/v2/glm53-flash-4sparks-xsfp4-tensorfold-129-official-bizuayeu.