Hugging Face's logo

Datasets:
prometheus04
/
apac-submission-build

Dataset card Files
xet
 Community
Dataset Viewer
Duplicate
The dataset viewer is not available for this split.
Cannot load the dataset split (in streaming mode) to extract the first rows.
Error code:   StreamingRowsError
Exception:    CastError
Message:      Couldn't cast
1024: struct<BLOCK_SIZE_M: int64, BLOCK_SIZE_N: int64, BLOCK_SIZE_K: int64, GROUP_SIZE_M: int64, num_warps (... 27 chars omitted)
  child 0, BLOCK_SIZE_M: int64
  child 1, BLOCK_SIZE_N: int64
  child 2, BLOCK_SIZE_K: int64
  child 3, GROUP_SIZE_M: int64
  child 4, num_warps: int64
  child 5, num_stages: int64
2048: struct<BLOCK_SIZE_M: int64, BLOCK_SIZE_N: int64, BLOCK_SIZE_K: int64, GROUP_SIZE_M: int64, num_warps (... 27 chars omitted)
  child 0, BLOCK_SIZE_M: int64
  child 1, BLOCK_SIZE_N: int64
  child 2, BLOCK_SIZE_K: int64
  child 3, GROUP_SIZE_M: int64
  child 4, num_warps: int64
  child 5, num_stages: int64
4096: struct<BLOCK_SIZE_M: int64, BLOCK_SIZE_N: int64, BLOCK_SIZE_K: int64, GROUP_SIZE_M: int64, num_warps (... 27 chars omitted)
  child 0, BLOCK_SIZE_M: int64
  child 1, BLOCK_SIZE_N: int64
  child 2, BLOCK_SIZE_K: int64
  child 3, GROUP_SIZE_M: int64
  child 4, num_warps: int64
  child 5, num_stages: int64
8192: struct<BLOCK_SIZE_M: int64, BLOCK_SIZE_N: int64, BLOCK_SIZE_K: int64, GROUP_SIZE_M: int64, num_warps (... 27 chars omitted)
  child 0, BLOCK_SIZE_M: int64
  child 1, BLOCK_SIZE_N: int64
  child 2, BLOCK_SIZE_K: int64
  child 3, GROUP_SIZE_M: int64
  child 4, num_warps: int64
  child 5, num_stages: int64
320: struct<BLOCK_SIZE_M: int64, BLOCK_SIZE_N: int64, BLOCK_SIZE_K: int64, GROUP_SIZE_M: int64, num_warps (... 27 chars omitted)
  child 0, BLOCK_SIZE_M: int64
  child 1, BLOCK_SIZE_N: int64
  child 2, BLOCK_SIZE_K: int64
  child 3, GROUP_SIZE_M: int64
  child 4, num_warps: int64
  child 5, num_stages: int64
256: struct<BLOCK_SIZE_M: int64, BLOCK_SIZE_N: int64, BLOCK_SIZE_K: int64, GROUP_SIZE_M: int64, num_warps (... 27 chars omitted)
  child 0, BLOCK_SIZE_M: int64
  child 1, BLOCK_SIZE_N: int64
  child 2, BLOCK_SIZE_K: int64
  child 3, GROUP_SIZE_M: int64
  child 4, num_warps: int64
  child 5, num_stages: int64
192: struct<BLOCK_SIZE_M: int64, BLOCK_SIZE_N: int64, BLOCK_SIZE_K: int64, GROUP_SIZE_M: int64, num_warps (... 27 chars omitted)
  child 0, BLOCK_SIZE_M: int64
  child 1, BLOCK_SIZE_N: int64
  child 2, BLOCK_SIZE_K: int64
  child 3, GROUP_SIZE_M: int64
  child 4, num_warps: int64
  child 5, num_stages: int64
384: struct<BLOCK_SIZE_M: int64, BLOCK_SIZE_N: int64, BLOCK_SIZE_K: int64, GROUP_SIZE_M: int64, num_warps (... 27 chars omitted)
  child 0, BLOCK_SIZE_M: int64
  child 1, BLOCK_SIZE_N: int64
  child 2, BLOCK_SIZE_K: int64
  child 3, GROUP_SIZE_M: int64
  child 4, num_warps: int64
  child 5, num_stages: int64
to
{'192': {'BLOCK_SIZE_M': Value('int64'), 'BLOCK_SIZE_N': Value('int64'), 'BLOCK_SIZE_K': Value('int64'), 'GROUP_SIZE_M': Value('int64'), 'num_warps': Value('int64'), 'num_stages': Value('int64')}, '256': {'BLOCK_SIZE_M': Value('int64'), 'BLOCK_SIZE_N': Value('int64'), 'BLOCK_SIZE_K': Value('int64'), 'GROUP_SIZE_M': Value('int64'), 'num_warps': Value('int64'), 'num_stages': Value('int64')}, '320': {'BLOCK_SIZE_M': Value('int64'), 'BLOCK_SIZE_N': Value('int64'), 'BLOCK_SIZE_K': Value('int64'), 'GROUP_SIZE_M': Value('int64'), 'num_warps': Value('int64'), 'num_stages': Value('int64')}, '384': {'BLOCK_SIZE_M': Value('int64'), 'BLOCK_SIZE_N': Value('int64'), 'BLOCK_SIZE_K': Value('int64'), 'GROUP_SIZE_M': Value('int64'), 'num_warps': Value('int64'), 'num_stages': Value('int64')}}
because column names don't match
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/utils.py", line 147, in get_rows_or_raise
                  return get_rows(
                      dataset=dataset,
                  ...<4 lines>...
                      column_names=column_names,
                  )
                File "/src/libs/libcommon/src/libcommon/utils.py", line 272, in decorator
                  return func(*args, **kwargs)
                File "/src/services/worker/src/worker/utils.py", line 127, in get_rows
                  rows_plus_one = list(itertools.islice(safe_iter(ds, dataset=dataset), rows_max_number + 1))
                File "/src/services/worker/src/worker/utils.py", line 483, in safe_iter
                  yield from ds.decode(False) if ds.features else ds
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2951, in __iter__
                  for key, example in ex_iterable:
                                      ^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2461, in __iter__
                  for key, pa_table in self._iter_arrow():
                                       ~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 2486, in _iter_arrow
                  for key, pa_table in self.ex_iterable._iter_arrow():
                                       ~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 547, in _iter_arrow
                  for key, pa_table in iterator:
                                       ^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/iterable_dataset.py", line 430, in _iter_arrow
                  for key, pa_table in self.generate_tables_fn(**gen_kwags):
                                       ~~~~~~~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 343, in _generate_tables
                  self._cast_table(pa_table, json_field_paths=json_field_paths),
                  ~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
                File "/usr/local/lib/python3.14/site-packages/datasets/packaged_modules/json/json.py", line 132, in _cast_table
                  pa_table = table_cast(pa_table, self.info.features.arrow_schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2378, in table_cast
                  return cast_table_to_schema(table, schema)
                File "/usr/local/lib/python3.14/site-packages/datasets/table.py", line 2306, in cast_table_to_schema
                  raise CastError(
                  ...<3 lines>...
                  )
              datasets.table.CastError: Couldn't cast
              1024: struct<BLOCK_SIZE_M: int64, BLOCK_SIZE_N: int64, BLOCK_SIZE_K: int64, GROUP_SIZE_M: int64, num_warps (... 27 chars omitted)
                child 0, BLOCK_SIZE_M: int64
                child 1, BLOCK_SIZE_N: int64
                child 2, BLOCK_SIZE_K: int64
                child 3, GROUP_SIZE_M: int64
                child 4, num_warps: int64
                child 5, num_stages: int64
              2048: struct<BLOCK_SIZE_M: int64, BLOCK_SIZE_N: int64, BLOCK_SIZE_K: int64, GROUP_SIZE_M: int64, num_warps (... 27 chars omitted)
                child 0, BLOCK_SIZE_M: int64
                child 1, BLOCK_SIZE_N: int64
                child 2, BLOCK_SIZE_K: int64
                child 3, GROUP_SIZE_M: int64
                child 4, num_warps: int64
                child 5, num_stages: int64
              4096: struct<BLOCK_SIZE_M: int64, BLOCK_SIZE_N: int64, BLOCK_SIZE_K: int64, GROUP_SIZE_M: int64, num_warps (... 27 chars omitted)
                child 0, BLOCK_SIZE_M: int64
                child 1, BLOCK_SIZE_N: int64
                child 2, BLOCK_SIZE_K: int64
                child 3, GROUP_SIZE_M: int64
                child 4, num_warps: int64
                child 5, num_stages: int64
              8192: struct<BLOCK_SIZE_M: int64, BLOCK_SIZE_N: int64, BLOCK_SIZE_K: int64, GROUP_SIZE_M: int64, num_warps (... 27 chars omitted)
                child 0, BLOCK_SIZE_M: int64
                child 1, BLOCK_SIZE_N: int64
                child 2, BLOCK_SIZE_K: int64
                child 3, GROUP_SIZE_M: int64
                child 4, num_warps: int64
                child 5, num_stages: int64
              320: struct<BLOCK_SIZE_M: int64, BLOCK_SIZE_N: int64, BLOCK_SIZE_K: int64, GROUP_SIZE_M: int64, num_warps (... 27 chars omitted)
                child 0, BLOCK_SIZE_M: int64
                child 1, BLOCK_SIZE_N: int64
                child 2, BLOCK_SIZE_K: int64
                child 3, GROUP_SIZE_M: int64
                child 4, num_warps: int64
                child 5, num_stages: int64
              256: struct<BLOCK_SIZE_M: int64, BLOCK_SIZE_N: int64, BLOCK_SIZE_K: int64, GROUP_SIZE_M: int64, num_warps (... 27 chars omitted)
                child 0, BLOCK_SIZE_M: int64
                child 1, BLOCK_SIZE_N: int64
                child 2, BLOCK_SIZE_K: int64
                child 3, GROUP_SIZE_M: int64
                child 4, num_warps: int64
                child 5, num_stages: int64
              192: struct<BLOCK_SIZE_M: int64, BLOCK_SIZE_N: int64, BLOCK_SIZE_K: int64, GROUP_SIZE_M: int64, num_warps (... 27 chars omitted)
                child 0, BLOCK_SIZE_M: int64
                child 1, BLOCK_SIZE_N: int64
                child 2, BLOCK_SIZE_K: int64
                child 3, GROUP_SIZE_M: int64
                child 4, num_warps: int64
                child 5, num_stages: int64
              384: struct<BLOCK_SIZE_M: int64, BLOCK_SIZE_N: int64, BLOCK_SIZE_K: int64, GROUP_SIZE_M: int64, num_warps (... 27 chars omitted)
                child 0, BLOCK_SIZE_M: int64
                child 1, BLOCK_SIZE_N: int64
                child 2, BLOCK_SIZE_K: int64
                child 3, GROUP_SIZE_M: int64
                child 4, num_warps: int64
                child 5, num_stages: int64
              to
              {'192': {'BLOCK_SIZE_M': Value('int64'), 'BLOCK_SIZE_N': Value('int64'), 'BLOCK_SIZE_K': Value('int64'), 'GROUP_SIZE_M': Value('int64'), 'num_warps': Value('int64'), 'num_stages': Value('int64')}, '256': {'BLOCK_SIZE_M': Value('int64'), 'BLOCK_SIZE_N': Value('int64'), 'BLOCK_SIZE_K': Value('int64'), 'GROUP_SIZE_M': Value('int64'), 'num_warps': Value('int64'), 'num_stages': Value('int64')}, '320': {'BLOCK_SIZE_M': Value('int64'), 'BLOCK_SIZE_N': Value('int64'), 'BLOCK_SIZE_K': Value('int64'), 'GROUP_SIZE_M': Value('int64'), 'num_warps': Value('int64'), 'num_stages': Value('int64')}, '384': {'BLOCK_SIZE_M': Value('int64'), 'BLOCK_SIZE_N': Value('int64'), 'BLOCK_SIZE_K': Value('int64'), 'GROUP_SIZE_M': Value('int64'), 'num_warps': Value('int64'), 'num_stages': Value('int64')}}
              because column names don't match

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

Qwen3-235B on Firmus AI Cloud (2× HGX H200) — PD serving reproduction guide

APAC HPC-AI Challenge 2026. Serving Qwen3-235B-A22B-Instruct-2507-FP8 with SGLang prefill/decode disaggregation on the Firmus AI Cloud cluster — two HGX H200 nodes (16 GPUs, 8× InfiniBand NDR), 17-minute walltime per job. Every number in this folder is read out of results/logs/profile_export_aiperf.json from a real full-protocol run.

Current result (job 22002, "bor10d"): goodput 24.20 req/s = 4,840 good output tokens/s at declared N=108, good_request_fraction 0.9648.

(Fractions in this guide are the scored-window values — the export's top-level good_request_fraction. Warmup-phase values are noted separately.)

The metric 

score = goodput (req/s) x OSL (200) = good output tokens/s

A request counts as good only if it satisfies every SLO:

gate limit
TTFT <= 3000 ms
ITL (per-request average) <= 20 ms
good_request_fraction >= 0.90 (below this the run scores zero)

ISL 4000, OSL 200, closed-loop at the declared concurrency N, official AIPerf command verbatim (warmup 60 / benchmark 120 / grace 30, seed 2026, --cache-bust first_turn_prefix, greedy).

Result — official protocol, job 22002 (bor10d, argmax-gather patch)

metric value
goodput 24.20 req/s -> 4,840 good tok/s
good_request_fraction (scored) 0.9648 (2,984 of 3,093 requests; 0 errors)
request throughput 25.08 req/s
TTFT avg / p50 / p99 734.0 / 587.8 / 3,539.8 ms
ITL avg / p50 / p99 17.68 / 17.76 / 20.27 ms
output token throughput 5,016 tok/s
input token throughput 99,978 tok/s
N declared 108

Failures: 109 misses in the scored window (3.5%), zero request errors — both tails improved vs the prior best (TTFT p99 3,654.6 -> 3,539.8 ms; ITL p99 20.60 -> 20.27 ms). The final lever is the argmax-gather patch (build/patches/03 + 04): the per-step full-vocab logits gather and the degenerate-penalty pass on the EAGLE3 greedy verify path are replaced by a packed candidate gather (per-rank top-4 + one small packed all-gather; penalties applied to candidates only, assert-guarded). Lossless under the official command: greedy-only gate, penalties asserted to be {0, -inf}, accept held at 2.00. Measured: goodput 23.49 -> 24.20 req/s at the same N (4,697.8 -> 4,839.7).

History at this N: spec-mode decode 4,681.5 (job 21544) / 4,697.8 (job 21698); the N-sweep closed at N=108 (N=116 invalid — frac 0.8618).

Artifacts: results/logs/ carries the job-22002 export; the 21544 and 21698 samples are results/confirmation-*-profile_export_aiperf.json; the verbatim scored scripts are run/champion-22002.sbatch (best) and run/champion-21698.sbatch (prior best). Two run_meta.json fields are harness-template defaults, not the run's actual env: mem_frac records 0.9 (the prefill value) and rail_affinity records true; the authoritative per-worker values are the export lines in run/champion-22002.sbatch — MEM_FRAC=0.9 prefill / 0.55 decode, RAIL_AFFINITY=0 on all seven workers, and SGLANG_OPT_ARGMAX_GATHER=1 on the decode worker.

Portability: every runnable script (build/setup.sh, run/submit.sbatch, both champion-*.sbatch) derives personal paths from SG_* env vars with generic $USER-based defaults — no account is hardcoded anywhere. The results/ logs, run_meta.json, and milestones/ are as-run records and therefore contain the team account's paths.

Config delta vs 20719: the tuned decode MoE config for E=128,N=384 (build/configs/), the declared concurrency (N=108, job 21257), the spec-mode decode flag (4,660.3 -> 4,697.8), and the argmax-gather patches (4,697.8 -> 4,839.7). The "Fallback to triton version 3.2.0" warning is absent from all these runs' worker logs; goodput went 22.50 -> 23.22 -> 23.30 -> 23.49 -> 24.20 req/s.

Configuration (job 22002 = 20938 stack + spec-mode decode + argmax-gather, declared N=108)

  • Topology 12+4: 6x prefill TP2/EP1 (12 GPUs) + 1x decode TP4/EP1 (4 GPUs)
  • KV transfer nixl; router round_robin; FP8 e4m3 KV; ctx 8192; mem-frac 0.9 prefill / 0.55 decode per-worker (run_meta.json's mem_frac field records the prefill value); rail-affinity off (all workers RAIL_AFFINITY=0)
  • Decode: --max-running-requests 256, chunked-prefill 8192, --disaggregation-decode-extra-slots 128, dense CUDA-graph ladder 1 2 4 8 12 16 ... 120 128
  • Prefill: chunked-prefill 8192, --max-running-requests 8, --disable-radix-cache (cache-bust workload never reuses prefixes)
  • Speculative: EAGLE3, draft nebius/EAGLE3-Qwen3-235B-A22B-Instruct-2507 (FP8), num-steps 2 / topk 1 / num-draft-tokens 3, triton draft attention, attention mode decode (4,660.3 -> 4,697.8 across two runs)
  • Decode MoE: tuned Triton config table (E=128,N=384, per-M 192-384) from build/configs/
  • Argmax-gather patch: SGLANG_OPT_ARGMAX_GATHER=1 on the decode worker; build/patches/03-argmax-gather.diff + 04-argmax-eagle.diff (EAGLE3 greedy verify path; loud asserts otherwise) — 4,697.8 -> 4,839.7
  • Stack overlay: sglang 0.5.19 + 4 source patches + 2 tuned MoE configs (see build/)

The tuning journey (three eras)

result (good tok/s) era key change evidence
132.00 1. baseline router defaults, upstream bugs job 14766
605.10 1. router round_robin for --cache-bust traffic 14994
808.76 1. DISAGG_QUEUE=16 THREADS=12 (default cap was 4) 15033
2,106.29 2. 8+8 topology full 8p8d rebuild + exact graph ladder 16505 x2
2,772.8 2. nixl + 0.5.19 overlay transfer mooncake->nixl, decode mem-frac 0.50 17860
2,960.0 2. EAGLE3 fine-tune (pilot) domain-adapted draft, +0.15 accept, validated 18918
3,000.0 2. EAGLE3 fine-tune (5K) data scaling closed by measurement 18963
3,602.7 3. engine era 8+8 -> 12+4, dense graph ladder, N=84 bracket 20520
4,227.7 3. draft extend-attention launch-config patch (+~0.35 accept-equivalent in step time) 20677
4,457.7 3. decode EP4 -> EP1 (kills expert-skew stalls; E=128) 20681
4,500.9 3. N=100 (patch + ep1 + dense ladder, stock MoE config) 20719
4,644.0 3. tuned decode MoE config live (per-M table; fallback warning gone) 20938
4,660.3 3. N=108 bracket — the ceiling run (frac 0.9404) 21257
4,681.5 3. speculative-attention-mode decode — first sample (frac 0.9528) 21544
4,697.8 3. confirmation repeat of spec-mode decode — prior best (frac 0.9545) 21698
4,839.7 3. argmax-gather patch — sampling collective + degenerate-penalty pass removed on the greedy path (frac 0.9648, accept 2.00) 22002
4,277.6 (INVALID) 3. N=116: broke the 0.90 gate (frac 0.8618); input tok/s fell 103.7k → 98.9k 21258

Fine-tune note: the team-trained EAGLE3 drafts (HF prometheus04/…Pilot-FP8, …5K-FP8) were validated on-cluster (2,960 / 3,000) and later superseded by the official Nebius release (accept 2.02 vs 1.65), which is what the final config uses. The fine-tune pipeline and its measured sim-to-real transfer (0.31 -> 0.11) are documented in milestones/.

Building / running

Stack: build/setup.sh (idempotent, login-node) rebuilds the overlay, applies the four patches, installs the two tuned configs, and stages the draft from HF. Details and md5s in build/build.log. Prerequisites: a pip venv (default ~/venv-sglang, or set SG_VENV), network access (pip + HF), and patch; every path is SG_*-env-driven end to end, and step [5/5] hard-fails if any patch or config did not land.

Job: run/submit.sbatch — a single Slurm script in which deploying the stack and invoking the official AIPerf command are consecutive steps. It reproduces the 4,839.7 run; the deltas along the chain were the tuned MoE config file that build/setup.sh installs, the declared concurrency, the spec-mode flag, and the argmax-gather patch (env-gated on the decode worker).

bash build/setup.sh                # overlay + patches + configs + slurmconf + draft
sbatch run/submit.sbatch           # deploy + official AIPerf, one job

Paths are parameterized (SG_REPO_ROOT, SG_OVERLAY, SG_DRAFT, SG_SLURMCONF); defaults are generic ($USER-based) and match the team account layout on this cluster.

Cold-account behavior (reproducing the score from scratch)

The job is self-contained except for the cluster-provided images/model and the persistent per-worker caches under ${ROOT}/jitcache/home/* (Triton/DeepGEMM JIT). The harness states its own coldness in the log:

DeepGEMM cache: 0 entries (0 = COLD, expect a slow first run)

Job 22002 ran with warm caches (ready at 510 s). On a brand-new account the first deploy pays the compiles during warmup and may not fit the 17-minute wall. Recommended sequence:

bash build/setup.sh                       # once
PREWARM=1 sbatch run/submit.sbatch        # deploy + smoke only; warms caches
sbatch run/submit.sbatch                  # scored run — reproduces 4,840 (N=108, spec-mode decode + argmax-gather)

PREWARM=1 exits right after the PD smoke gate (before the benchmark); the scored run never sets it. The same script, same official command — only the cache state differs.

Negative results (measured, keep the search space small)

lever result verdict
DeepGEMM MoE backend (decode + prefill) decode: step +27%, score 51.6; prefill: goodput 23.22 vs 23.30 req/s at N=108 (frac 0.9284) — no gain closed
cutlass decode MoE ITL flat (19.37 vs 19.39) closed
draft bf16 KV + FA3 backend accept collapsed 2.00 -> 1.23 closed (broken path)
chain 3/1/4 (more draft tokens) accept 2.00 -> 2.21, step +3.4 ms wash
num-continuous-decode-steps 2 ITL flat closed
tree drafting (eagle-topk >= 2) boot-blocked: tree path requires flashinfer >= 0.6.18; no public 0.6.18 cubin (overlay keeps the matched 0.6.15.post1 trio) closed
N > 108 (past the optimum) 112: 4,712.7, frac 0.950, input 99.7k tok/s; 116: frac 0.8618, input fell to 98.9k — gate broken closed (jobs 22020/21258)
prefill deep_gemm masked layout CUDA OOM in DeepGEMM JIT warmup at MEM_FRAC 0.9 — masked buffers don't fit closed
page-size 64 engine requires flashinfer >= 0.6.18; image ships 0.6.15.post1 closed
chunked-prefill-size 16384 (prefill) TTFT p50 588 -> 655 ms, input tok/s 100.0k -> 97.9k, score 4,839.7 -> 4,746.9 — shapes > 8k fall off the tuned MoE table and the start-of-window wave drain is unchanged reverted (job 22119)
DP8/EP8 + DeepEP prefill (8v8 rebuild) 58.6k input tok/s on 8 GPUs vs 79.4k for 4x TP2 pairs — 0.74x per-GPU; DeepEP a2a dispatch overhead is structural; the EAGLE3 draft also faults under DP8 closed (jobs 21955/21956)
prefill admission 4 + RAIL_AFFINITY=1 frac 0.9442 vs 0.9545, TTFT p99 3,420 vs 3,655 ms — wash; the tail is a ~1% multi-second stall class, not admission or rail closed (job 21970)
fused qk-norm-rope / moe-sum-AR flags qk-norm-rope: CUDA-graph capture crash; moe-sum-AR alone (decode): runs but ITL 17.68 -> 18.24 ms, score 4,839.7 -> 4,603.6 — the fused kernel is slower than separate moe_sum + custom-AR-v2 closed (job 22346)
MoE retune expectation decode MoE already ~HBM-bound; retune = single-digit % bounded

Submission logistics (per the organisers, Oct 2026)

  • Oct 9 deliverable = presentation slides, job scripts, and the logs that support the reported results. It is used to plan the online interviews — it is not the final package.
  • The full package (docs, code, scripts, logs, run instructions) is due after the interview, when result reproduction starts.
  • How to submit: a folder on Google Drive (or similar) shared with pengzhiz@hpcadvisorycouncil.com; upload plain files, not a zip; the folder may keep being updated after Oct 9, up to the interview. Mirror: huggingface.co/datasets/prometheus04/apac-submission-build.
  • Both tasks (HPC + AI) are scored separately; this folder covers the AI task.

Folder map 

submission/
├── CHECKSUMS.txt             # md5 of every file shipped here (build_inputs/ + node_modules excluded)
├── build/
│   ├── setup.sh              # stack reconstruction (self-built overlay)
│   ├── build.log             # exact history + artifact md5s
│   ├── patches/              # the four source diffs (applied by setup.sh)
│   └── configs/              # tuned Triton MoE configs (decode + prefill)
├── run/
│   ├── submit.sbatch         # champion job: deploy + official AIPerf (portable roots;
│   │                         #   reproduces 4,839.7 — N=108 + spec-mode + argmax-gather)
│   ├── champion-22002.sbatch # scored script — 4,839.7 (job 22002; SG_* defaults = as-run paths)
│   ├── champion-21698.sbatch # scored script — 4,697.8 (job 21698, prior best; SG_* defaults)
│   └── slurmconf/            # slurm.conf + cli_filter.lua + plugstack.conf (cluster workaround)
├── results/
│   ├── run_meta.json         # job 22002 config + protocol record
│   ├── confirmation-*.json   # 21544 + 21698 cross-check exports
│   └── logs/                 # profile_export_aiperf.{json,csv}, console, phases
├── analysis/                 # NVIDIA Nsight Systems kernel summaries (decode + prefill; recipe in analysis/README.md)
└── presentation/             # slides — delivered separately on the shared drive, not mirrored here
Downloads last month
265
Total file size:
32.7 MB