Skip to content

flashqla-megafusion-3500-spark1-report

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

flashqla-megafusion-3500-spark1-report

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
ledgers
ledgers
 
 
raw-logs
raw-logs
 
 
raw-patches
raw-patches
 
 
raw-results
raw-results
 
 
summary
summary
 
 
README.md
README.md
 
 

README.md

FlashQLA megafusion 3500 Spark 1 attempt — audit report

Snapshot report for the follow-on Spark 1 FlashQLA fusion loop that attempted to push the valid API/default PP2048 result above 3500 pp tok/s.

Verdict

No PASS artifact exists in this snapshot.

Required PASS contract:

  • latency_mode=api
  • prefix_caching_enabled=false
  • PP=2048
  • TG=32
  • CONCURRENCY=1
  • WARMUP_RUNS=2
  • RUNS=30
  • exactly 30 measured prefill values
  • mean prefill throughput >3500.0 pp tok/s

Best valid API/default N=30 artifact remains:

  • Artifact: raw-results/result-20260509-034541-alias-kpack2-nobaro-api-n30.json
  • Variant: alias-kpack2-nobaro
  • Mean: 3315.967641 pp tok/s
  • Median: 3315.498018 pp tok/s
  • Std: 10.305973

Best valid N=30 artifact from the 2026-05-10 >3500 attempt:

  • Artifact: raw-results/result-20260510-0920-abeta-api-n30.json
  • Variant: abeta
  • Mean: 3309.443457 pp tok/s
  • Median: 3308.529754 pp tok/s
  • Std: 10.815157

What this report contains

  • ledgers/GOAL.md — live goal text for the >3500 loop.
  • ledgers/PLAN.md — iteration plan/history.
  • ledgers/RESULTS.md — validated result notes and completion audits.
  • ledgers/FAILED_ATTEMPTS.md — failed candidate ledger.
  • ledgers/RESEARCH_CUTEDSL_FUSION_PLAYBOOK.md — CuTeDSL / next-gen fusion research playbook injected into the loop.
  • raw-results/ — all JSON result artifacts copied from the Spark 1 attempt workspace.
  • raw-logs/ — raw bench/server/build logs copied from the Spark 1 attempt workspace.
  • raw-patches/ — patch/source snapshots from the attempt workspace.
  • summary/all_results_summary.csv — parsed metrics for all copied JSON artifacts.
  • summary/valid_api_n30_summary.csv — strict API/default N=30 artifacts sorted by mean pp throughput.
  • summary/short_api_frontier_summary.csv — short API/default frontier artifacts, useful only as proxy evidence.
  • summary/summary.json — machine-readable headline summary.

Valid API/default N=30 scoreboard

Variant Artifact Mean pp tok/s Median Std N Contract
alias-kpack2-nobaro result-20260509-034541-alias-kpack2-nobaro-api-n30.json 3315.968 3315.498 10.306 30 api, prefix=False, pp=2048, tg=32, c=1
abeta result-20260510-0920-abeta-api-n30.json 3309.443 3308.530 10.815 30 api, prefix=False, pp=2048, tg=32, c=1
alias-kpack2-nographmem result-20260508-200800-alias-kpack2-nographmem-n30.json 3299.602 3300.958 14.149 30 api, prefix=False, pp=2048, tg=32, c=1
kktv1k2 result-20260510-0940-kktv1k2-api-n30.json 3278.103 3275.169 11.120 30 api, prefix=False, pp=2048, tg=32, c=1
alias-kpack2-nographmem-warm10 result-20260508-202200-alias-kpack2-nographmem-warm10-n30.json 3198.075 3201.343 32.991 30 api, prefix=False, pp=2048, tg=32, c=1
directfq result-20260510-1000-directfq-api-n30.json 3188.241 3225.526 115.044 30 api, prefix=False, pp=2048, tg=32, c=1
abetaempty result-20260510-0840-abetaempty-api-n30.json 3160.311 3201.557 127.401 30 api, prefix=False, pp=2048, tg=32, c=1
emptyout result-20260510-0820-emptyout-api-n30.json 3160.152 3194.928 124.775 30 api, prefix=False, pp=2048, tg=32, c=1
fastpack result-20260510-1040-fastpack-api-n30.json 3157.145 3193.309 108.104 30 api, prefix=False, pp=2048, tg=32, c=1
nooshared64 result-20260510-0800-nooshared64-api-n30.json 3154.558 3189.586 110.678 30 api, prefix=False, pp=2048, tg=32, c=1
alias-kpack2-nograph-n30 result-20260509-031112-alias-kpack2-nograph-api-n30-fresh.json 3153.125 3188.448 104.119 30 api, prefix=False, pp=2048, tg=32, c=1
kktalias result-20260510-1100-kktalias-api-n30.json 3151.261 3202.192 123.112 30 api, prefix=False, pp=2048, tg=32, c=1
nobaro2 result-20260510-1205-nobaro2-api-n30.json 3136.912 3170.312 109.476 30 api, prefix=False, pp=2048, tg=32, c=1
vdkpack2 result-20260510-1120-vdkpack2-api-n30.json 3133.116 3191.189 128.398 30 api, prefix=False, pp=2048, tg=32, c=1
bk128-bv128-noprefix-pp2048-tg32-c1-n30-20260506-161035 result-flashqla-v2-bk128-bv128-noprefix-pp2048-tg32-c1-n30-20260506-161035.json 3132.100 3131.063 6.722 30 api, prefix=False, pp=2048, tg=32, c=1
pgkpack2 result-20260510-1135-pgkpack2-api-n30.json 3130.526 3180.779 157.399 30 api, prefix=False, pp=2048, tg=32, c=1
norstd result-20260510-1020-norstd-api-n30.json 3128.772 3193.944 219.974 30 api, prefix=False, pp=2048, tg=32, c=1
alias-kpack2-nographmem result-20260508-201500-alias-kpack2-nographmem-n30.json 3126.937 3183.403 141.530 30 api, prefix=False, pp=2048, tg=32, c=1
fusedo-qkgemm-alias result-20260508-190600-fusedo-qkgemm-alias-n30.json 3110.206 3196.971 188.322 30 api, prefix=False, pp=2048, tg=32, c=1
metanosync result-20260510-0905-metanosync-api-n30.json 3108.613 3153.357 132.613 30 api, prefix=False, pp=2048, tg=32, c=1
noharg2 result-20260510-1220-noharg2-api-n30.json 3108.018 3167.666 180.764 30 api, prefix=False, pp=2048, tg=32, c=1
noinit result-20260510-1150-noinit-api-n30.json 3101.753 3161.217 164.826 30 api, prefix=False, pp=2048, tg=32, c=1
fusedo-qkgemm-alias-kpack2 result-20260508-192900-fusedo-qkgemm-alias-kpack2-n30.json 3101.555 3175.795 170.745 30 api, prefix=False, pp=2048, tg=32, c=1
alias-kpack2-preheated-nographmem result-20260508-205500-alias-kpack2-preheated-nographmem-n30.json 3094.709 3153.467 146.645 30 api, prefix=False, pp=2048, tg=32, c=1
pp2048-tg32-c1-n30-20260506-163233 result-flashqla-hkv-pp2048-tg32-c1-n30-20260506-163233.json 3030.626 3028.046 12.611 30 api, prefix=False, pp=2048, tg=32, c=1

2026-05-10 >3500-attempt N=30 promotions

These were valid API/default N=30 gates but all failed the >3500 criterion.

Variant Artifact Mean pp tok/s Median Std N Contract
abeta result-20260510-0920-abeta-api-n30.json 3309.443 3308.530 10.815 30 api, prefix=False, pp=2048, tg=32, c=1
kktv1k2 result-20260510-0940-kktv1k2-api-n30.json 3278.103 3275.169 11.120 30 api, prefix=False, pp=2048, tg=32, c=1
directfq result-20260510-1000-directfq-api-n30.json 3188.241 3225.526 115.044 30 api, prefix=False, pp=2048, tg=32, c=1
abetaempty result-20260510-0840-abetaempty-api-n30.json 3160.311 3201.557 127.401 30 api, prefix=False, pp=2048, tg=32, c=1
emptyout result-20260510-0820-emptyout-api-n30.json 3160.152 3194.928 124.775 30 api, prefix=False, pp=2048, tg=32, c=1
fastpack result-20260510-1040-fastpack-api-n30.json 3157.145 3193.309 108.104 30 api, prefix=False, pp=2048, tg=32, c=1
nooshared64 result-20260510-0800-nooshared64-api-n30.json 3154.558 3189.586 110.678 30 api, prefix=False, pp=2048, tg=32, c=1
kktalias result-20260510-1100-kktalias-api-n30.json 3151.261 3202.192 123.112 30 api, prefix=False, pp=2048, tg=32, c=1
nobaro2 result-20260510-1205-nobaro2-api-n30.json 3136.912 3170.312 109.476 30 api, prefix=False, pp=2048, tg=32, c=1
vdkpack2 result-20260510-1120-vdkpack2-api-n30.json 3133.116 3191.189 128.398 30 api, prefix=False, pp=2048, tg=32, c=1
pgkpack2 result-20260510-1135-pgkpack2-api-n30.json 3130.526 3180.779 157.399 30 api, prefix=False, pp=2048, tg=32, c=1
norstd result-20260510-1020-norstd-api-n30.json 3128.772 3193.944 219.974 30 api, prefix=False, pp=2048, tg=32, c=1
metanosync result-20260510-0905-metanosync-api-n30.json 3108.613 3153.357 132.613 30 api, prefix=False, pp=2048, tg=32, c=1
noharg2 result-20260510-1220-noharg2-api-n30.json 3108.018 3167.666 180.764 30 api, prefix=False, pp=2048, tg=32, c=1
noinit result-20260510-1150-noinit-api-n30.json 3101.753 3161.217 164.826 30 api, prefix=False, pp=2048, tg=32, c=1

Short-run frontier warning

Short API/default N=3/N=5 artifacts repeatedly over-predicted performance and collapsed at N=30. Do not promote a candidate on short-run evidence alone.

Variant Artifact Mean pp tok/s Median Std N Contract
nooshared64 result-20260510-0230-nooshared64-api-n3.json 3346.915 3343.872 4.703 3 api, prefix=False, pp=2048, tg=32, c=1
emptyout result-20260509-2140-emptyout-api-n3.json 3345.556 3355.935 14.821 3 api, prefix=False, pp=2048, tg=32, c=1
abetaempty result-20260510-0215-abetaempty-api-n3.json 3339.630 3333.235 11.773 3 api, prefix=False, pp=2048, tg=32, c=1
metanosync result-20260509-1220-metanosync-api-n3.json 3334.844 3334.947 12.088 3 api, prefix=False, pp=2048, tg=32, c=1
abeta result-20260509-1512-abeta-api-n3.json 3327.616 3328.807 9.124 3 api, prefix=False, pp=2048, tg=32, c=1
directfq result-20260509-2150-directfq-api-n3.json 3327.550 3322.460 8.216 3 api, prefix=False, pp=2048, tg=32, c=1
kktv1k2 result-20260509-1342-kktv1k2-api-n3.json 3321.043 3326.812 9.453 3 api, prefix=False, pp=2048, tg=32, c=1
norstd result-20260510-0530-norstd-api-n3.json 3320.875 3320.273 8.281 3 api, prefix=False, pp=2048, tg=32, c=1
fusedo-qkgemm-alias-kpack2 result-20260508-192400-fusedo-qkgemm-alias-kpack2-n3.json 3318.318 3313.225 7.479 3 api, prefix=False, pp=2048, tg=32, c=1
fastpack result-20260509-1212-fastpack-api-n3.json 3317.904 3306.532 17.419 3 api, prefix=False, pp=2048, tg=32, c=1
abetafold result-20260509-1522-abetafold-api-n3.json 3313.147 3310.742 4.237 3 api, prefix=False, pp=2048, tg=32, c=1
metafallback result-20260509-1240-metafallback-api-n3.json 3309.495 3316.311 13.653 3 api, prefix=False, pp=2048, tg=32, c=1
kktalias result-20260509-1330-kktalias-api-n3.json 3309.207 3306.908 6.954 3 api, prefix=False, pp=2048, tg=32, c=1
fusedo-qkgemm-alias result-20260508-190000-fusedo-qkgemm-alias-n3.json 3309.202 3311.364 6.927 3 api, prefix=False, pp=2048, tg=32, c=1
vdkpack2 result-20260509-1838-vdkpack2-api-n3.json 3306.525 3316.791 14.963 3 api, prefix=False, pp=2048, tg=32, c=1

Direction after this snapshot

The ledger conclusion is to stop promoting low-headroom proxy branches and return to actual fusion/elision work:

  • trace the FlashQLA/vLLM return path for real transpose, contiguous, allocation, wrapper, or launch boundaries;
  • redesign consumer-native output/final-state layout rather than changing store policy blindly;
  • remove a materialization boundary such as post-conv prep, raw g cumsum, local partial state, or wrapper-created tensor churn;
  • use the CuTeDSL/CUTLASS playbook for any new fixed-shape Blackwell candidate.

Contract guardrail

This report does not change the benchmark ruler. LATENCY_MODE=generation, prefix-cache changes, sample dropping, or post-hoc metric rewrites are not PASS artifacts for this goal.