SLAI-AITP commited on
Commit
37fb47a
·
verified ·
1 Parent(s): a76318c

Remove website and publishing tools from the dataset package

Browse files

Keep the technical report and README logos under docs/. Update documentation, release metadata and checksums, and remove the unused GitHub Pages workflow. Task indexes, frozen benchmark inputs and evaluation code are unchanged.

This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +1 -3
  2. .github/workflows/pages.yml +0 -57
  3. CHECKSUMS.sha256 +7 -179
  4. README.md +5 -7
  5. README_PROJECT.md +5 -6
  6. {website/public → docs}/AscendGYM_TechReport.pdf +0 -0
  7. {website/public → docs}/assets/logo/mark-dark.svg +0 -0
  8. {website/public → docs}/assets/logo/mark-primary.svg +0 -0
  9. hf/DATASET_CARD.md +0 -216
  10. hf/LICENSES.md +0 -30
  11. hf/README.md +0 -80
  12. hf/prepare_release.py +0 -302
  13. metadata/release.json +17 -3
  14. website/.gitattributes +0 -5
  15. website/.gitignore +0 -15
  16. website/README.md +0 -81
  17. website/docs/api-cost-estimates-2026-10-26.csv +0 -22
  18. website/docs/perf-cost-analysis-2026-10-26.md +0 -69
  19. website/docs/shape-results-guide.zh.md +0 -45
  20. website/docs/shape-results-kit.zip +0 -3
  21. website/evaluate/OPERATOR_EVAL_SPEC.md +0 -467
  22. website/evaluate/Operator_eval.md +0 -567
  23. website/evaluate/evaluate_operator_runs.py +0 -0
  24. website/evaluate/visualize_operator_metrics.py +0 -136
  25. website/index.html +0 -15
  26. website/metrics_master.md +0 -389
  27. website/package-lock.json +0 -1388
  28. website/package.json +0 -36
  29. website/public/assets/harnesses/SOURCES.md +0 -9
  30. website/public/assets/harnesses/claude-code.svg +0 -1
  31. website/public/assets/harnesses/opencode.svg +0 -7
  32. website/public/assets/logo/favicon.svg +0 -6
  33. website/public/assets/logo/lockup-primary.svg +0 -11
  34. website/public/assets/logo/mark-monochrome.svg +0 -5
  35. website/public/assets/resources/SOURCES.md +0 -9
  36. website/public/assets/resources/arxiv.jpg +0 -3
  37. website/public/assets/resources/github.svg +0 -1
  38. website/public/assets/resources/huggingface.svg +0 -8
  39. website/public/assets/vendors/SOURCES.md +0 -16
  40. website/public/assets/vendors/alibabacloud.svg +0 -4
  41. website/public/assets/vendors/anthropic.svg +0 -4
  42. website/public/assets/vendors/baidu.svg +0 -1
  43. website/public/assets/vendors/bytedance.svg +0 -1
  44. website/public/assets/vendors/deepseek.svg +0 -4
  45. website/public/assets/vendors/google.svg +0 -4
  46. website/public/assets/vendors/kimi.svg +0 -5
  47. website/public/assets/vendors/meta.svg +0 -1
  48. website/public/assets/vendors/minimax.svg +0 -1
  49. website/public/assets/vendors/mistral.svg +0 -1
  50. website/public/assets/vendors/openai.svg +0 -4
.gitattributes CHANGED
@@ -18,8 +18,6 @@
18
 
19
  # Hugging Face binary storage; preserve frozen tensor bytes.
20
  *.pt filter=lfs diff=lfs merge=lfs -text
21
- website/docs/shape-results-kit.zip filter=lfs diff=lfs merge=lfs -text
22
- website/public/AscendGYM_TechReport.pdf filter=lfs diff=lfs merge=lfs -text
23
- website/public/assets/resources/arxiv.jpg filter=lfs diff=lfs merge=lfs -text
24
  docs/assets/leaderboard/leaderboard-performance-cost.png filter=lfs diff=lfs merge=lfs -text
25
  docs/assets/leaderboard/leaderboard-sts.png filter=lfs diff=lfs merge=lfs -text
 
 
18
 
19
  # Hugging Face binary storage; preserve frozen tensor bytes.
20
  *.pt filter=lfs diff=lfs merge=lfs -text
 
 
 
21
  docs/assets/leaderboard/leaderboard-performance-cost.png filter=lfs diff=lfs merge=lfs -text
22
  docs/assets/leaderboard/leaderboard-sts.png filter=lfs diff=lfs merge=lfs -text
23
+ docs/AscendGYM_TechReport.pdf filter=lfs diff=lfs merge=lfs -text
.github/workflows/pages.yml DELETED
@@ -1,57 +0,0 @@
1
- name: Deploy project website
2
-
3
- # Enable Pages in Settings before manually running this workflow.
4
- on:
5
- workflow_dispatch:
6
-
7
- permissions:
8
- contents: read
9
-
10
- concurrency:
11
- group: github-pages
12
- cancel-in-progress: false
13
-
14
- jobs:
15
- build:
16
- runs-on: ubuntu-latest
17
- permissions:
18
- contents: read
19
- pages: read
20
- steps:
21
- - uses: actions/checkout@v6
22
- - uses: actions/setup-node@v6
23
- with:
24
- node-version: '22'
25
- cache: npm
26
- cache-dependency-path: website/package-lock.json
27
- - uses: actions/setup-python@v6
28
- with:
29
- python-version: '3.12'
30
- - uses: actions/configure-pages@v5
31
- id: pages
32
- - name: Install website dependencies
33
- working-directory: website
34
- run: npm ci
35
- - name: Build website for Pages
36
- working-directory: website
37
- env:
38
- PAGES_BASE_PATH: ${{ steps.pages.outputs.base_path }}
39
- run: npm run build -- --base "${PAGES_BASE_PATH%/}/"
40
- - uses: actions/upload-pages-artifact@v4
41
- with:
42
- path: website/dist
43
-
44
- deploy:
45
- if: github.ref == 'refs/heads/main'
46
- needs: build
47
- runs-on: ubuntu-latest
48
- permissions:
49
- pages: write
50
- id-token: write
51
- environment:
52
- name: github-pages
53
- url: ${{ steps.deployment.outputs.page_url }}
54
- steps:
55
- - name: Deploy to GitHub Pages
56
- id: deployment
57
- uses: actions/deploy-pages@v4
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
CHECKSUMS.sha256 CHANGED
@@ -1,11 +1,10 @@
1
- 011f1c0bddb0afd0079df6d98021f9d295c67910fccd1f81825d284ea4fc4b69 .gitattributes
2
- becd6e44a3d211b1da2f8806cc94dc574f9a86a38e475a839c80662adbab0915 .github/workflows/pages.yml
3
  b40089560df62b980d4767263f8393bcdf38cf414044fd94bd0a18abf989ead7 .gitignore
4
  c755d20abf6031b09024f4b6b5594a727111075bd97ef9b844ca4405fba98936 LICENSE
5
  43070e2d4e532684de521b885f385d0841030efa2b1a20bafb76133a5e1379c1 LICENSE-APACHE-2.0
6
  37dc4945c5ebcedfdcd78a6c266e177526ebbe4b8e0a127e139a3bd8ab14cc8a LICENSES.md
7
- db582ee6864600fd2fba86b8b579e428ffecc232489208e97dca4dd88a57d1f5 README.md
8
- 9e039b589d54771c0a178db8685340a8dcaf8eb2962f9fc49dca85873a99a836 README_PROJECT.md
9
  5aaba4fd7cc4e4399e9792716cee7d67e05f651969c1a3369065d158d14f745a bridge/bare_bench/__init__.py
10
  5e4d6bdce75c36a5cafb1c874bb14488a5c938fb4c2c889c044c8b8165873fa1 bridge/bare_bench/__main__.py
11
  8fd7e8cebce8d008988cd2f4a23470065ebc350c1de7a618964ec6002aa42778 bridge/bare_bench/runner.py
@@ -1355,6 +1354,7 @@ c4c400462ba7be6cccdc98a064a8c5aa03c8d39a4973cc7c9138aaecfc459928 datasets/tasks
1355
  1991f6e7e17f4acb990f27f7eda5f7a710c08225579cdb88c22c18aa01ec2165 datasets/tasks/preserved-golden/transformer__gather_pa_kv_cache/task.json
1356
  bbf5fc282264b5cc2869c35ec1884a106c17ab29e3e815aba56c91060f3b0c05 datasets/tasks/preserved-golden/transformer__kv_rms_norm_rope_cache/README.md
1357
  2d486840ce20d43cf636dae71e88f77af8ce0563ce0ce3c2f69a5f3689058584 datasets/tasks/preserved-golden/transformer__kv_rms_norm_rope_cache/task.json
 
1358
  5b3beff07299df1018d1f8e20de2f10761faeca147f9700ef0f538bf1b232f66 docs/BRIDGE_CONFIG.md
1359
  1057d30bc4d89026341b4d7329744cc719767902a720ef43a82bd61550b7fe21 docs/BRIDGE_TOOLS.md
1360
  fd33d8922bd5573d750ff7e51db9f7da41941fe33ceac0c48300b130f2768fc7 docs/BRIDGE_USAGE.md
@@ -1367,6 +1367,8 @@ c6e41c60fb9e296eeee56afb5b603c509b3c64e84a5e4124f82b3630191ee4a2 docs/DATA_LAYO
1367
  2da03251246b591cbcd4360ef9d65b73cf2b68b4db47e9e00148c8b5c3bff612 docs/SELF_HOSTED_ENVIRONMENT.md
1368
  80eaa4ebe9373b62d5b56878ce85b5447142edfe9406589a7172e9d27c50ce21 docs/assets/leaderboard/leaderboard-performance-cost.png
1369
  434e519414fc72e73eee0c6c116699c08c9b8ec7f008de2ab5216b2f5cedd52e docs/assets/leaderboard/leaderboard-sts.png
 
 
1370
  f3edfb722789c407fc8538c4195889391f798d84147f111a4fe960bbc7ca9230 environment/public-dependencies.json
1371
  171fc2add32c705faf032b55225108143e67bc603c79f87f7896a2e11855b3d0 environment/reference-stack.json
1372
  8b0d0b526e961ae3d74a0729d7c2d516f7cc52ee4fbabcc70a655cac3d14faab environment/requirements-tests.txt
@@ -1446,181 +1448,7 @@ d0f44f8c30363c6c9dfe3a729ec25c4f22f55ac88ca21f3484d135cd504650ba examples/01_na
1446
  14442bf8facae08ddc7cf2e9af28db7b53d5523e042f708c325c903e36197ffa examples/README.md
1447
  845a604a94796afe16b636144b33b1f9949a915a406b2bd87fd19aa0ba7fdeac examples/env.sh
1448
  ff5de6d2da9b39acaaf840464751f0ed10b260be1ebefa3fb5534ced1a17e684 examples/run_all_serial.sh
1449
- b27af4951a495d6885d848ab371ccee6b1366d55f6914dc162e48bb80302e69c hf/DATASET_CARD.md
1450
- 37dc4945c5ebcedfdcd78a6c266e177526ebbe4b8e0a127e139a3bd8ab14cc8a hf/LICENSES.md
1451
- 6ac94f80e08b054af4c68e03cc690bbdf9d17e99e39f0136090d4f94a5d5795a hf/README.md
1452
- 81a4a35c4a1a6e72812a62c49cc213c288d15d089ada2c35a7eac6187561f69a hf/prepare_release.py
1453
  5e9348deec725c24e52017aaddb70e579d445d619ff8dc117df1893e6274977e metadata/batch1.jsonl
1454
  e208f0486ba91fe11c35adf88c378d6eb09cfb3de8277917fefd0f4460064002 metadata/preserved_golden.jsonl
1455
- dc48609158c17fd0decdc3e009d54360816cad39d0a46b86a6773ea790092f1c metadata/release.json
1456
  1e7f6a17ca2aad79e3ee13f55b3b9695a19a046b9d91aee7d056e6c136ea2444 requirements.txt
1457
- 5fa6c278fda137a35d92bdb0bcc9d38983fb98578c3e499625e8f16dfe6c5a4b website/.gitattributes
1458
- eeae9c8cfa7fcfed458d5fede33d39bae840fbb58f35372830317c90960b14fa website/.gitignore
1459
- 74d4894b40e58989c758e878e171dcda8cc02d74888c4fce8e37081ad467da20 website/README.md
1460
- a3e9fb29d249fb2363dc4d0ab0682ae5d72ddf4f050e81e17efe3231f714c88b website/docs/api-cost-estimates-2026-10-26.csv
1461
- bd216f87b71a473aa2db132b1f239ace4f265dfe6089d366e7864a885d370515 website/docs/perf-cost-analysis-2026-10-26.md
1462
- f774233e87bc85d811ccf1ed4f98e5db2dbdd2a783f84ed630525e768d796548 website/docs/shape-results-guide.zh.md
1463
- 796fd0e5349d447644c2dac9a4a0db7e72822dab60d12a7f9ea9967a65da1883 website/docs/shape-results-kit.zip
1464
- b4102f092544fe53ab3406baf07dbc77d4a9f6f879354217b53e89233ea06d84 website/evaluate/OPERATOR_EVAL_SPEC.md
1465
- 0e0db0ed0e5c02dfe1e7ac26d72f5c3238f67c8c51ed0acb20bdfbf9874a6bbe website/evaluate/Operator_eval.md
1466
- 57aab2efe69b8c1d4598e44b9f8ea1115c1089ed953e50950377435886b4aec5 website/evaluate/evaluate_operator_runs.py
1467
- ed6038d31e8622f3fb4038b7cd16a90a768aa83d59335eced359d0905f5614fc website/evaluate/visualize_operator_metrics.py
1468
- 9682a1c78bde1ab228548bd9893e42c682af21fbfd90cd85b3c91e12068dc024 website/index.html
1469
- 91938a3246aa806496374812e1ae66961a89cc153b778ee9fac29d3670f44d74 website/metrics_master.md
1470
- 1f13b8a6596ad13a424991640417e3e35f9e5df66c0ba1c554bd3ef340f44f41 website/package-lock.json
1471
- 80915e2671969a931af9f19708d62727ed2dc56a6b09d6edc6fd9214d0571dc3 website/package.json
1472
- 009a2553e149622eeda439b9c67f7a90b4cd6cd3bea8cf8a4a1d530807cae75b website/public/AscendGYM_TechReport.pdf
1473
- a0eaf4379a27e332d8cce73311c2a095d503ae75e30b930fb7e61b43087ab570 website/public/assets/harnesses/SOURCES.md
1474
- 645bc8fb1b0bb4b59a808249e81a02d9ad4b3e670e6a1c6ae39e220bcabb7ed9 website/public/assets/harnesses/claude-code.svg
1475
- e29bbe33380ad1c1ada9134b52f229d30e9776d60481512c9d81f2bb6f37def9 website/public/assets/harnesses/opencode.svg
1476
- abc6b3bb0f31d105f8160b48110a22429a3a4b72e29868ec61572859ce53b677 website/public/assets/logo/favicon.svg
1477
- c4f406b2423e29e3d9485dd5c764f27b5d9e18ba82b1de6ced024bec96f42d9a website/public/assets/logo/lockup-primary.svg
1478
- 20dc01f4d47bc75a7c4a15066c787e5ab07b31634044ebc8e351f39ec7e0ea98 website/public/assets/logo/mark-dark.svg
1479
- 02602e61d6113ec698e86781119fb1e4852ad309db69dc944b66f40a2c171556 website/public/assets/logo/mark-monochrome.svg
1480
- 9afbc997275d345728f7b7430b55fae9973ef92458758da02d50f09fa6f3f497 website/public/assets/logo/mark-primary.svg
1481
- c672754ecbf10e71967e6e44b97bed081cca16b53c7c9195021f19c8687a5287 website/public/assets/resources/SOURCES.md
1482
- deee1e08ef17db9f06342e1188b07ec1adfe2d5e574c1712efa313b118631f98 website/public/assets/resources/arxiv.jpg
1483
- bce494189797623c34e39a41c2c38a132bdca23ce4e6ac06b70116ed7e91ce26 website/public/assets/resources/github.svg
1484
- 942cad1ccda905ac5a659dfd2d78b344fccfb84a8a3ac3721e08f488205638a0 website/public/assets/resources/huggingface.svg
1485
- e6668815a390eaf52b3ed071b455c0ae59fc2ceb7680383a9bd0dae68dc17d35 website/public/assets/vendors/SOURCES.md
1486
- 3b4245cab4540f9c4efdd7bd7b8281b9a214100e3086e33b005026ec520e43e4 website/public/assets/vendors/alibabacloud.svg
1487
- b7a81f5d7a6af08579df88aec19641ba693f3304732266de11267d62ee37e0ea website/public/assets/vendors/anthropic.svg
1488
- e2eeaca4c5273ecff775ee4b357091678a7587b2d606190f8e5b7dce4d93a781 website/public/assets/vendors/baidu.svg
1489
- 03a322949b34209cb0446382bbaf1c458eeffd75a6930f6bf37ca910578a6429 website/public/assets/vendors/bytedance.svg
1490
- fdad254344f6b9acb4942744cad847809728bbb2e38c5fe37a89d2f235da9f7e website/public/assets/vendors/deepseek.svg
1491
- 2fb5ea89caf1fe28b7155d709b9cf2165a6af13b6fa9f2438094b5a72b4c5cb2 website/public/assets/vendors/google.svg
1492
- 27fc9a0960219e473892ac455a3ce431b1f45b41124927f33231c186d7799745 website/public/assets/vendors/kimi.svg
1493
- e3e9db62f59dc477e5917687349b8aa9685b841d40300c1afad60ad85075cd2e website/public/assets/vendors/meta.svg
1494
- ef95e28b290e5954a5c0be338925d81f13b3d06a0a32af29a4bc64583fbcc368 website/public/assets/vendors/minimax.svg
1495
- 88f00cf2e82e4d36bb4fdd0e913df200c6e7d6162ec627cc58cfdd6a4eb17d18 website/public/assets/vendors/mistral.svg
1496
- 98480cb4cf46ff374c58d095deadd21d34629786c84f58e9d6fb2dfd0c8d36c6 website/public/assets/vendors/openai.svg
1497
- 693e68863eceb8dc9f72e2acd386ab9c20a10858dab2c076212f7084cb7a32fe website/public/assets/vendors/xai.svg
1498
- 12be0c13838a87441ef267dc7aa206574828037a5f8719eb76e3939d9ca515a8 website/public/assets/vendors/zai.svg
1499
- 2fe9b28dd3d9c9ea6aa0d8fedc5a7ec38de9c78778613e5e03363c9089ce168a website/public/templates/shape-collection-index.csv
1500
- cb82f94f1ddc7ee8c0ed24cf74724cdc80ff4f75a8fca483a76dd38d22db1b04 website/public/templates/shape-details-design.html
1501
- 4811576dc238eae3459bd23958db8c1c2abab95728f2273144467ce25cf1f5bc website/public/templates/shape-results.all-kernels.json
1502
- 49a5cea2fada58f21ef01d2446f3a6ae94986144fa2a6c201a317aee25b85c8c website/public/templates/shape-results.blank.csv
1503
- f7d6b1b8c64e9ddd0c7345f28000d50cc6c75d07e6beebfc37b86159b09191c4 website/public/templates/shape-results.csv
1504
- 71f7b364c13f4fd23eb38c31a17f3a0e42cf500cdba892b28f0d7932cedd46e1 website/public/templates/shape-results.example.csv
1505
- 3104ff6a140f3828d117ad14259bac9b68521d46d89847eaac78cd0956458282 website/public/templates/shape-results.template.json
1506
- 80b9dcd36eefb02aa447f98e9794656e3eedab6ee86ea437786d7fdf1ce6fc3e website/scripts/audit-field-mapping.mjs
1507
- 34bf9292849ff2907ccdc13275d8d1f835e299547447ad0f345f7ea5775cef3b website/scripts/audit-precision.mjs
1508
- bb97dba4a4d2f77eabe947b0df4b64065f4adf2464fe88022615c4e6f0aec63e website/scripts/build-bundle.mjs
1509
- 0301d7c15c79e04dea3e54d0b17c4b29c518045a708d9e9ecceda0afc9672ea6 website/scripts/build-operator-bundle.mjs
1510
- 9b1e980e45f6c169edf46e1437ca2b4c4658c7c692a63a6d2c63dba97c3bc8e0 website/scripts/build-shape-bundle.py
1511
- 25f5879c174f13e29f5314d48e133dd6517c5abe7fcaf09372b8e86ab6742836 website/scripts/compact-scale.mjs
1512
- 86dbef60a487cde2a503fd54b68881ec0514a2bcaba768019df27b1051f85986 website/scripts/export-api-cost-estimates.mjs
1513
- 4c106bb42530f0d5721bcb8298bb64a9386a534f4b18156ba2cdc94525689d6a website/scripts/mock-api.mjs
1514
- b0c7f1af16c52781af8ab1aa30debf32791aee7d95a19f8bc3a14736599a45dc website/scripts/sync-supplement.py
1515
- 9479b6665c96ac914ef6d9223082830d090babb46761b531922bda0bae968662 website/scripts/validate-shape-results.py
1516
- 87b955fc850b661b13891bb99923e2ed9e5c18124d07cfa91b19d7a3a3559484 website/scripts/verify-anchors.mjs
1517
- 3269428e773a25e041695e26854423725035bd7944cec2abd19a5e544a4e2c44 website/scripts/verify-api-pricing.mjs
1518
- da0ce7027ece73afd01b14ab11b14959cd16ec30d86bf02b9757316c20fa153e website/scripts/verify-pagination.mjs
1519
- befca52b5586dea85cfb703d1d956915c44725206a96a9f05379bd05b010c340 website/scripts/verify-shape-results.mjs
1520
- 71dbcc0c70aea8248bdcc5b01665e9b1998f9ffc09874b476449da5450b12ae5 website/scripts/verify-submission.mjs
1521
- f46d8923148fcd0358660093a1e0efe76db6355d66615eede299edf359ab2969 website/scripts/verify-supplement.py
1522
- 81cffc50fabf3de7161ac16c81bab417d9983aaf3a316703d31aac52584b9a12 website/src/App.vue
1523
- 44b729ec6bfd20f765ebbebabb4866b8ad9c7c274569a43dedabc258537e086e website/src/assets/design-system.css
1524
- 57b80ed8b78fdcfada474d60407c1961014673a2054c97c5a9ebba5f13b74396 website/src/assets/desktop-typography.css
1525
- a8fde88286f62c55424ad6695a67f42409152740dc0b05141e579bf10d3cd8be website/src/assets/display-scale.css
1526
- abc6b3bb0f31d105f8160b48110a22429a3a4b72e29868ec61572859ce53b677 website/src/assets/logo/favicon.svg
1527
- c4f406b2423e29e3d9485dd5c764f27b5d9e18ba82b1de6ced024bec96f42d9a website/src/assets/logo/lockup-primary.svg
1528
- 20dc01f4d47bc75a7c4a15066c787e5ab07b31634044ebc8e351f39ec7e0ea98 website/src/assets/logo/mark-dark.svg
1529
- 02602e61d6113ec698e86781119fb1e4852ad309db69dc944b66f40a2c171556 website/src/assets/logo/mark-monochrome.svg
1530
- 9afbc997275d345728f7b7430b55fae9973ef92458758da02d50f09fa6f3f497 website/src/assets/logo/mark-primary.svg
1531
- f8a19f4ee8e57eb8dddc52d29afb0f7ad72b4de171a741ea5bca6dc1626a8a15 website/src/assets/readability.css
1532
- 044b6532213b541c5902f27627d669ca7334e07b66004869245e189fa4e3c20b website/src/assets/styles-v2.css
1533
- 1e2ecc862434e97f457b43fa7c53ac732068f659d344efec25bcfa2a53cb5aae website/src/assets/styles.css
1534
- 90af35c84a2deddd8e2f1cd76ce58233857eccc185443ea0c611eb81e2def018 website/src/components/AppToast.vue
1535
- e1a0f341d82ae50674ba55982302286d52d681686ba3293d003b89f993c30f3c website/src/components/BenchmarkSelect.vue
1536
- 48f3540870d47d1a57d6f6d5b610deaa3a20c45f133cc926a4e7bbb853715019 website/src/components/BestByOperator.vue
1537
- e543dbac1c4243cb1121c03966dea083933d3d36abad70101a4713f9e56f2ed1 website/src/components/ConfigurationFilter.vue
1538
- f7d4cb69feb893776082f160c0bcd5355060193a77ba510388ad9c86adbcb7ad website/src/components/DataTooltip.vue
1539
- e42d1619dd426191a391f7445ededcf92bb75155c42271d8ec705edcb61448f2 website/src/components/DurationSection.vue
1540
- 399d06776af6c91ad849c2e4728dd01496bce5f960e98f8449c11f3a46adc652 website/src/components/EntryDrawer.vue
1541
- 143b7617579d9fb0fd38ebf3c78bb270ebba214ba1a18fecd71f6a8a12e98669 website/src/components/EntryLabel.vue
1542
- f8f133ca7b45ed6e3e9bc41797c20ed1400024ca2bd2dc7c6340947f48a191de website/src/components/EntryRankingSection.vue
1543
- 9c78072323fd8db9699b7b8f1fc6cd4ef980ce1ca657bcc27e9e72af6226fcc8 website/src/components/FlagBadges.vue
1544
- 36c67e69a96057c547e97fe7d377e4238aa564b269a36e3c09d200ed4cad36cb website/src/components/GainChart.vue
1545
- 2c0881c225001ecbdc18ffb292b3ef67993620a9f3e2683d76d231c3daab358e website/src/components/HardwareSection.vue
1546
- 936983267bb1617ac0f762a007a5ce1a4db7d698a3caaf263848324a13948819 website/src/components/HarnessChip.vue
1547
- 36875c82aba5a79218c333a9a69c2361bc8f08e1dd38d42f70a5e0513364477d website/src/components/ImportSection.vue
1548
- caf71178335427f1cefa465b038b3efe57b4f5fff40551b28feb2e3f3a186764 website/src/components/KernelSelect.vue
1549
- fffbfbf448117241be77794d0ed4b260117b3e5a0fdef61989cce6e501bece7b website/src/components/KernelShapeDetails.vue
1550
- 3afbfefd1afcde4faa03fa53746e226dc31bff73b4774342f7159b44318229d2 website/src/components/LeaderTable.vue
1551
- cb5ada576e3f391a099bb5c297d3d4fa97b8a5677060b36553b729db1a1946f1 website/src/components/MathText.vue
1552
- 44f3140c17d059bd3d5ac020dc65554bde8e1e89d460927ed1de10a7a3d1f571 website/src/components/MethodologySection.vue
1553
- c6462472c90eb145ae4d270f9aabd91ab0041d45db9525a6596403bea6bc5d78 website/src/components/MetricsDetailSection.vue
1554
- ed399e48e2d116dbcde3ce8dd0f124c8fd931c559f8e31e648e8aad95345b5e8 website/src/components/OperatorProcessSection.vue
1555
- 68d2498ac1e93df73f552e844e70f64e7db66ab2dfce0e160feb4d9a6063fe94 website/src/components/PerformanceCostTable.vue
1556
- 0d013aed091c0955df8a0608358b32ebe7f5f743bbff04ea688efabb75e640cc website/src/components/PerformanceGainSection.vue
1557
- ee8d0c037bcf5c4d83cda51ccadf5c241a10c5ef4d8bc8c89277e28cd1082e08 website/src/components/QualityCostSection.vue
1558
- 65fa48053aef25a7f334439383b4b0cfb9275af1a6774b199a16254b0dfabfef website/src/components/RaceTracks.vue
1559
- 0e827401e2e0ac7a1c232ab91290c93d3a1e2a85dffa6dbf38a6ea41d739af1f website/src/components/RankChart.vue
1560
- 62506a75541fade643ccd3dcb68116fb0df2d6cb0a4a5fe7dd9475688f21b4d6 website/src/components/RatioFingerprint.vue
1561
- 35d258318834213140e5da8e132009a6683c055d215e90cea717a96cc0770a44 website/src/components/RatioLanes.vue
1562
- 6b59cde7e21d99fe814d942321e1761c0a02ccb0fcf185e2c840b9a03be9926b website/src/components/ScatterChart.vue
1563
- 9b3ca00e3af1a62d08db400ef98a071cc39b5adf5cecc78cbd8b5ad2324b159c website/src/components/ShapeRings.vue
1564
- 2725db473d573d928ce20269550467b57d7867cfabe06b6525cd3b539fc0e26a website/src/components/SiteFooter.vue
1565
- 8bb3c90cc2d20055914363150d4a014cccc99b8c469d171c2df5e5d354808691 website/src/components/SiteHeader.vue
1566
- 7e658522ee2fa0cfa52eead55fb7083ace03adc50c359838800243a6b55a0dae website/src/components/Sparkline.vue
1567
- d09bd18f6705640136232716a48abd8e8b4deb1c1057b2b8b4ac713bec30639c website/src/components/TablePagination.vue
1568
- ab8d1d657c2689e1506bf5edc8164838cdd7b15be06f9bd50df991318031a17a website/src/components/VendorChip.vue
1569
- 3f01a6cc74a08ed74db05aa37deae44b72814016e4f3ac3da957497c8ecee84d website/src/data/Token_price.json
1570
- 182b13ff6955f4390e252c22aed1e2358d23027c717f0ae1f01b4eb0af87ccdb website/src/data/bundle.json
1571
- 815604fa13862892c962fea679245db28ec3007d5db942db69a0f765973ebe30 website/src/data/eval_datasets_summary/claude-code-claude-opus-5.json
1572
- a0cccb748e4db8445bb5fddba7c06c9f3f8a552283be9abacf9a120cdc09847d website/src/data/eval_datasets_summary/claude-code-deepseek-v4-flash-bf16.json
1573
- 447f037726f95ed89a0dc84e44673f85f871a9210f6e50aec5d721fd8f89a35a website/src/data/eval_datasets_summary/claude-code-glm-5-3.json
1574
- 7e3b88c30909c63492bb87d351de7ba68a72abc182d986208d8df2222a20c7f2 website/src/data/eval_datasets_summary/claude-code-gpt-5-6-terra.json
1575
- 778e8f59217c31b4bc605d657beb7ee601515046adfa814ddf5208e162c5636d website/src/data/eval_datasets_summary/claude-code-qwen3-5-122b.json
1576
- 8c153f5b1d35d2405db7a833ca25caae49722f80f24e4b855decfaa6d5ce4268 website/src/data/eval_datasets_summary/claude-code-qwen3-6-35b.json
1577
- c7d1c8cf95a47dc771c6035a9a1e5601d67d2668df17140d8a4c3bac1eb4d689 website/src/data/eval_datasets_summary/codex-deepseek-v4-flash-bf16.json
1578
- ba13134bd2733d750c12416af324c6b2a60972f672e827e1ea625ee301e30ab7 website/src/data/eval_datasets_summary/codex-glm-5-3.json
1579
- 80cd2e62a614a0dad56ea99907143cca58df4e0b19ec18bcd3e97fa207a6feb1 website/src/data/eval_datasets_summary/codex-gpt-5-6-terra.json
1580
- 69fd43c18a9bb002dff8443c10378a34ac70f3ba45d0d97a29eacc8c22e296d1 website/src/data/eval_datasets_summary/codex-qwen3-5-122b.json
1581
- a3c9aee04bbb2b243185fd4919840a2711bad0cb4728e8a43020f94c546166b9 website/src/data/eval_datasets_summary/codex-qwen3-6-35b.json
1582
- 7e527e4238d7291aa57e6384a53c3a117ddcc528257f476568ef6e8cf742ddbd website/src/data/eval_datasets_summary/opencode-claude-opus-5.json
1583
- 122d163d7d591eabc7226c57006eb68f32a5be3c0cd2759e13456ff1f42c6493 website/src/data/eval_datasets_summary/opencode-deepseek-v4-flash-bf16.json
1584
- 4c1d9d54f6b13fa08c16ea20a5de436f4473ab4dba81b4fb8292ea12635055b2 website/src/data/eval_datasets_summary/opencode-gpt-5-6-terra.json
1585
- 05d6b53e61e4780d7d9d2e7a9da70ab4210d4b68a059908cb7d3618b35b0bda7 website/src/data/eval_datasets_summary/opencode-qwen3-5-122b.json
1586
- 2ee4f21ded211559310a7f82675557b57abb8c07e22784d7d13acddcafcee3f0 website/src/data/eval_datasets_summary/opencode-qwen3-6-35b.json
1587
- 1188d9290d637b54d910580a99d3ce36ba6480434a02798f3be70d3df894ec11 website/src/data/evaluation_results_core.json
1588
- 1b7a7878552c7d9e7d6539b8a1f8dab0e9879945f267796e8a176b1561886d05 website/src/data/operator_bundle.json
1589
- a63167126db7a826dc1cdf2b16ca9823834d08b0134d1bcdcd4fffec494e15bf website/src/data/pricing.json
1590
- ef6c68984dae7ec679f0b304dac8628891354a746e8f691d88633f4d3430ebf9 website/src/data/shape_results.json
1591
- 99d87f6d5acd07af9dc6b6bdda205cc0f759437b6fc600b5fa9707f7628dac52 website/src/data/supplement_bundle.json
1592
- 7ea4ae8d90e97bcf63aee37ce903425f7458de0ac3d8ccc45949acef712c8d24 website/src/data/supplement_manifest.json
1593
- 3257dc996b91afffedb26be7703363b1cc9d0d1761d825c9a91bfad58eacf516 website/src/lib/datasets.js
1594
- b241aff8027dc3cc9ce23e642f6228979ae9aeedc1f8eb069a7e1092e93b7e46 website/src/lib/displayNames.js
1595
- 01026d5525cee2df885c8966d6438a7df3dcd8c7e21c674f7429b367115542e0 website/src/lib/displayScale.js
1596
- a11aaed241ac259c4fb45cab6f383b80db019c02fcb5d5bdf61b6fd1275a0c31 website/src/lib/format.js
1597
- 7b9c002af5b8e52e4e96496fa7c5fd097fc581b2322dbf36d4a98e7f1ae8fbbd website/src/lib/geometry.js
1598
- efa285dd1625b7808174fcd6a3911cda9520ba99f8e3e3c35089fa7552ebdd7a website/src/lib/markup.js
1599
- 8d31191ece56c154c061b55127f07499735d42901701bac152b5a24d0f9ffe86 website/src/lib/mathml.js
1600
- 98182cdeb95d4cb47a8d842c1e880766928c6e25d91424f55136d114f835ea42 website/src/lib/meta.js
1601
- 3e6964faff49653869802d60127274d64490a45688f07c9ab7c6293c9fecadd3 website/src/lib/metrics.js
1602
- 829f392d1f589e549283b76a64119b44f9920c29adf66a3d516660d995c74e9e website/src/lib/operatorProcess.js
1603
- 92cfcc7a4ea5eff45cbf3b67cf2200f043c336f0ca6b396862689cc9651d6893 website/src/lib/precision.js
1604
- 99b0eaed74699b9fe2b6186369afe35f3238a5e50cdaf089c0e16d7de20d4e24 website/src/lib/pricing.js
1605
- 48f0c1b4ecc5bd2b4104ff0ec86825a8ff6fc1e73d172d96ed7546d165881711 website/src/lib/reportMath.js
1606
- cda44dcc5eb5411e9e8d03f42a9c8131ea4a6fde9372bbfd8da48b56008c6884 website/src/lib/sections.js
1607
- b5b4d4ea394a54a1483f14a2182a20098a7b8a9261441dfb4aced1f223e0f53d website/src/lib/shapeResults.js
1608
- d151c9d876c684c0d81b23ac1ebc9f6a33533d4e36eaf6a28b9b83db59a88422 website/src/lib/storage.js
1609
- fb520e18ad0c1dc1104d0b399dc462cd2532b1914fa3aaf72a77d0f4e80dbc76 website/src/lib/submission.js
1610
- cd629bd24743257b26df8731a1c6d9ee2855737e1e625e4b11a9697ff718f1b7 website/src/lib/summary.js
1611
- 796b9100d03de7460fdfc9325f9a4a464944d4d76bcb8a1666d532c59706d279 website/src/lib/usePagination.js
1612
- 097099b270deb3005f50d89efa8d56e53670362d855b65f41e0c2a2b8a9e3211 website/src/lib/useSectionNav.js
1613
- cdceaa66130b07be49e97aaa60d2a0bba0363112b477a560197793ec9056c150 website/src/lib/vendors.js
1614
- 6bb20573dfb388234f92bb190b97cab16f1764ae87094d51c2da3f100f599a16 website/src/main.js
1615
- 4d102e4f206d5a6eeae84bf0086ff64ebdd684f19adf50781591dd847dad771f website/src/pages/DesignSystemPage.vue
1616
- fc6e16b0aa287b860674be710434573f14a5065f5ea1cf6d1fd01f69ca59fe9f website/src/pages/ImportsPage.vue
1617
- 9fd9583da6720cba8029c7364910cfa09d0253880eb1659d5096804c40c35975 website/src/pages/LeaderboardPage.vue
1618
- 7f9f3ea118603a000aba6dd276e1f59646d5ef751a83b4a58d19f7e7ece5324e website/src/pages/MethodologyPage.vue
1619
- 54adb56b647be8660ad519a96a3729edb94699fd918047cfe75de690a90e027d website/src/router.js
1620
- b7d9a2ce786e9d55c5ff6fd0ebd7ffcbd632a4fef38662e899b9da2858f81adf website/src/services/api.js
1621
- 81210869109a3e142c01f4ace83168dfcbc7fca9ed5f9e1667e5edcf3f917c4e website/src/services/datasets.js
1622
- 2284deb52c44a452c5684edfdacbe11a92107886db93c03379a372a2027bb595 website/src/store/benchmark.js
1623
- 3a4d4b9aa7ebc7da1c68dbb03ecd743f1597ab897b5ea7627734a4c6d17b2033 website/src/store/datasets.js
1624
- d19b2f4e871cc1c71edc735ab02319c5ebb7b107ba048e011ec39b8b534a2533 website/src/store/ui.js
1625
- b509ec3190138f3053d87601cfd6a5d2cf4995699408d0a90127156731cc84d7 website/src/store/uploads.js
1626
- 6c095d5cf18c6197c8d6307431a550a49863bb288768195edbee16d873c2c12b website/vite.config.js
 
1
+ 9591a5f8dd77637f08ac80c621fba075f3b84502975814b572ed713d55975d02 .gitattributes
 
2
  b40089560df62b980d4767263f8393bcdf38cf414044fd94bd0a18abf989ead7 .gitignore
3
  c755d20abf6031b09024f4b6b5594a727111075bd97ef9b844ca4405fba98936 LICENSE
4
  43070e2d4e532684de521b885f385d0841030efa2b1a20bafb76133a5e1379c1 LICENSE-APACHE-2.0
5
  37dc4945c5ebcedfdcd78a6c266e177526ebbe4b8e0a127e139a3bd8ab14cc8a LICENSES.md
6
+ 75412c56854b363255433877d9821229f48f17202ebb08aaef7287b3f06bcdd4 README.md
7
+ 6d1940dbbf98c99add25e20bd35c21f858c34d2e8e02260f533e6ea88a9434b0 README_PROJECT.md
8
  5aaba4fd7cc4e4399e9792716cee7d67e05f651969c1a3369065d158d14f745a bridge/bare_bench/__init__.py
9
  5e4d6bdce75c36a5cafb1c874bb14488a5c938fb4c2c889c044c8b8165873fa1 bridge/bare_bench/__main__.py
10
  8fd7e8cebce8d008988cd2f4a23470065ebc350c1de7a618964ec6002aa42778 bridge/bare_bench/runner.py
 
1354
  1991f6e7e17f4acb990f27f7eda5f7a710c08225579cdb88c22c18aa01ec2165 datasets/tasks/preserved-golden/transformer__gather_pa_kv_cache/task.json
1355
  bbf5fc282264b5cc2869c35ec1884a106c17ab29e3e815aba56c91060f3b0c05 datasets/tasks/preserved-golden/transformer__kv_rms_norm_rope_cache/README.md
1356
  2d486840ce20d43cf636dae71e88f77af8ce0563ce0ce3c2f69a5f3689058584 datasets/tasks/preserved-golden/transformer__kv_rms_norm_rope_cache/task.json
1357
+ 009a2553e149622eeda439b9c67f7a90b4cd6cd3bea8cf8a4a1d530807cae75b docs/AscendGYM_TechReport.pdf
1358
  5b3beff07299df1018d1f8e20de2f10761faeca147f9700ef0f538bf1b232f66 docs/BRIDGE_CONFIG.md
1359
  1057d30bc4d89026341b4d7329744cc719767902a720ef43a82bd61550b7fe21 docs/BRIDGE_TOOLS.md
1360
  fd33d8922bd5573d750ff7e51db9f7da41941fe33ceac0c48300b130f2768fc7 docs/BRIDGE_USAGE.md
 
1367
  2da03251246b591cbcd4360ef9d65b73cf2b68b4db47e9e00148c8b5c3bff612 docs/SELF_HOSTED_ENVIRONMENT.md
1368
  80eaa4ebe9373b62d5b56878ce85b5447142edfe9406589a7172e9d27c50ce21 docs/assets/leaderboard/leaderboard-performance-cost.png
1369
  434e519414fc72e73eee0c6c116699c08c9b8ec7f008de2ab5216b2f5cedd52e docs/assets/leaderboard/leaderboard-sts.png
1370
+ 20dc01f4d47bc75a7c4a15066c787e5ab07b31634044ebc8e351f39ec7e0ea98 docs/assets/logo/mark-dark.svg
1371
+ 9afbc997275d345728f7b7430b55fae9973ef92458758da02d50f09fa6f3f497 docs/assets/logo/mark-primary.svg
1372
  f3edfb722789c407fc8538c4195889391f798d84147f111a4fe960bbc7ca9230 environment/public-dependencies.json
1373
  171fc2add32c705faf032b55225108143e67bc603c79f87f7896a2e11855b3d0 environment/reference-stack.json
1374
  8b0d0b526e961ae3d74a0729d7c2d516f7cc52ee4fbabcc70a655cac3d14faab environment/requirements-tests.txt
 
1448
  14442bf8facae08ddc7cf2e9af28db7b53d5523e042f708c325c903e36197ffa examples/README.md
1449
  845a604a94796afe16b636144b33b1f9949a915a406b2bd87fd19aa0ba7fdeac examples/env.sh
1450
  ff5de6d2da9b39acaaf840464751f0ed10b260be1ebefa3fb5534ced1a17e684 examples/run_all_serial.sh
 
 
 
 
1451
  5e9348deec725c24e52017aaddb70e579d445d619ff8dc117df1893e6274977e metadata/batch1.jsonl
1452
  e208f0486ba91fe11c35adf88c378d6eb09cfb3de8277917fefd0f4460064002 metadata/preserved_golden.jsonl
1453
+ cf258d6136982e1318eab59137e4ad4592973cf0a02c818150d9c8fa1462e143 metadata/release.json
1454
  1e7f6a17ca2aad79e3ee13f55b3b9695a19a046b9d91aee7d056e6c136ea2444 requirements.txt
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
README.md CHANGED
@@ -34,7 +34,7 @@ operator baseline snapshots, frozen correctness tests, golden tensors, and the
34
  evaluation client code, preserving the upstream directory layout.
35
 
36
  The benchmark targets kernel optimization in real-world training and inference
37
- workloads. The complete project is included for local task preparation.
38
  Compilation, correctness testing, and performance profiling require access to
39
  an Ascend execution environment. For the complete workflow, see
40
  [the original project instructions](README_PROJECT.md).
@@ -44,7 +44,7 @@ an Ascend execution environment. For the complete workflow, see
44
  - Active Batch 1 tasks: **17**
45
  - Additional preserved-golden tasks: **3** (outside the official Batch 1 evaluation)
46
  - Frozen tensor files: **384** across both subsets
47
- - [Technical report](website/public/AscendGYM_TechReport.pdf)
48
  - [Leaderboard](https://slai-aitp.github.io/AscendGYM/)
49
 
50
  ## Benchmark tasks
@@ -102,14 +102,12 @@ datasets/
102
  bridge/ # Task preparation and evaluation entry points
103
  eval/ # Native and bare harness materials
104
  environment/ # Reference stack and pinned upstream dependencies
105
- docs/ # Execution environment and protocol documentation
106
  examples/
107
- website/ # Leaderboard sources and technical report
108
  LICENSE # Component license summary and full license texts
109
  LICENSE-APACHE-2.0 # Unmodified original project license text
110
  LICENSES.md # Component scope and third-party notice links
111
  CHECKSUMS.sha256 # Release file integrity list
112
- hf/ # Offline preparation and verification tools
113
  ```
114
 
115
  ## Load the task index
@@ -147,11 +145,11 @@ snapshot_download(
147
  )
148
  ```
149
 
150
- Verify the downloaded release **before generating local task packages**:
151
 
152
  ```bash
153
- python3 AscendGYM/hf/prepare_release.py --verify AscendGYM
154
  cd AscendGYM
 
155
  source datasets/activate.sh
156
  python3 bridge/op_manager.py scan
157
  python3 bridge/op_manager.py build --all
 
34
  evaluation client code, preserving the upstream directory layout.
35
 
36
  The benchmark targets kernel optimization in real-world training and inference
37
+ workloads. The benchmark package includes the files needed for local task preparation.
38
  Compilation, correctness testing, and performance profiling require access to
39
  an Ascend execution environment. For the complete workflow, see
40
  [the original project instructions](README_PROJECT.md).
 
44
  - Active Batch 1 tasks: **17**
45
  - Additional preserved-golden tasks: **3** (outside the official Batch 1 evaluation)
46
  - Frozen tensor files: **384** across both subsets
47
+ - [Technical report](docs/AscendGYM_TechReport.pdf)
48
  - [Leaderboard](https://slai-aitp.github.io/AscendGYM/)
49
 
50
  ## Benchmark tasks
 
102
  bridge/ # Task preparation and evaluation entry points
103
  eval/ # Native and bare harness materials
104
  environment/ # Reference stack and pinned upstream dependencies
105
+ docs/ # Technical report, guides, and documentation assets
106
  examples/
 
107
  LICENSE # Component license summary and full license texts
108
  LICENSE-APACHE-2.0 # Unmodified original project license text
109
  LICENSES.md # Component scope and third-party notice links
110
  CHECKSUMS.sha256 # Release file integrity list
 
111
  ```
112
 
113
  ## Load the task index
 
145
  )
146
  ```
147
 
148
+ Check the downloaded file hashes **before generating local task packages**, then prepare the tasks:
149
 
150
  ```bash
 
151
  cd AscendGYM
152
+ sha256sum --check CHECKSUMS.sha256
153
  source datasets/activate.sh
154
  python3 bridge/op_manager.py scan
155
  python3 bridge/op_manager.py build --all
README_PROJECT.md CHANGED
@@ -1,7 +1,7 @@
1
  <p align="center">
2
  <picture>
3
- <source media="(prefers-color-scheme: dark)" srcset="website/public/assets/logo/mark-dark.svg">
4
- <img src="website/public/assets/logo/mark-primary.svg" width="104" height="104" alt="AscendGYM logo">
5
  </picture>
6
  </p>
7
 
@@ -15,7 +15,7 @@
15
  <p align="center">
16
  <a href="#sandbox-setup">环境准备</a> ·
17
  <a href="#quick-start">快速开始</a> ·
18
- <a href="website/public/AscendGYM_TechReport.pdf">技术报告</a> ·
19
  <a href="https://slai-aitp.github.io/AscendGYM/">Leaderboard</a> ·
20
  <a href="#evaluation-guide">参与评测</a> ·
21
  <a href="#documentation">文档</a> ·
@@ -64,9 +64,8 @@ AscendGYM/
64
  ├── eval/
65
  │ ├── native_harness/ # 原生 coding agent 自主循环
66
  │ └── bare_harness/ # 裸模型与工具池
67
- ├── docs/ # 跨模块详细文档
68
  ├── examples/ # 启动示例
69
- ├── website/ # Leaderboard、指标说明与技术报告
70
  ├── requirements.txt
71
  ├── LICENSE
72
  └── README.md
@@ -391,7 +390,7 @@ export CODEX_HOME=/path/to/codex-template
391
  | [OPERATOR_TEST_GUIDE](docs/OPERATOR_TEST_GUIDE.md) | 冻结测试脚本接口 |
392
  | [DATA_LAYOUT](docs/DATA_LAYOUT.md) | 自定义数据根布局 |
393
  | [examples](examples/README.md) | 多种 agent 的启动示例 |
394
- | [Leaderboard 开发与部署](website/README.md) | Leaderboard 本地预览、静态构建与 GitHub Pages 部署 |
395
 
396
  <a id="license"></a>
397
  ## 许可
 
1
  <p align="center">
2
  <picture>
3
+ <source media="(prefers-color-scheme: dark)" srcset="docs/assets/logo/mark-dark.svg">
4
+ <img src="docs/assets/logo/mark-primary.svg" width="104" height="104" alt="AscendGYM logo">
5
  </picture>
6
  </p>
7
 
 
15
  <p align="center">
16
  <a href="#sandbox-setup">环境准备</a> ·
17
  <a href="#quick-start">快速开始</a> ·
18
+ <a href="docs/AscendGYM_TechReport.pdf">技术报告</a> ·
19
  <a href="https://slai-aitp.github.io/AscendGYM/">Leaderboard</a> ·
20
  <a href="#evaluation-guide">参与评测</a> ·
21
  <a href="#documentation">文档</a> ·
 
64
  ├── eval/
65
  │ ├── native_harness/ # 原生 coding agent 自主循环
66
  │ └── bare_harness/ # 裸模型与工具池
67
+ ├── docs/ # 技术报告、跨模块文档和展示图片
68
  ├── examples/ # 启动示例
 
69
  ├── requirements.txt
70
  ├── LICENSE
71
  └── README.md
 
390
  | [OPERATOR_TEST_GUIDE](docs/OPERATOR_TEST_GUIDE.md) | 冻结测试脚本接口 |
391
  | [DATA_LAYOUT](docs/DATA_LAYOUT.md) | 自定义数据根布局 |
392
  | [examples](examples/README.md) | 多种 agent 的启动示例 |
393
+ | [Leaderboard 开发与部署](https://github.com/SLAI-AITP/AscendGYM/blob/main/website/README.md) | GitHub 仓库中的主页源码与部署说明 |
394
 
395
  <a id="license"></a>
396
  ## 许可
{website/public → docs}/AscendGYM_TechReport.pdf RENAMED
File without changes
{website/public → docs}/assets/logo/mark-dark.svg RENAMED
File without changes
{website/public → docs}/assets/logo/mark-primary.svg RENAMED
File without changes
hf/DATASET_CARD.md DELETED
@@ -1,216 +0,0 @@
1
- ---
2
- pretty_name: AscendGYM
3
- language:
4
- - en
5
- - zh
6
- license: other
7
- license_name: ascendgym-component-licenses
8
- license_link: LICENSE
9
- size_categories:
10
- - n<1K
11
- tags:
12
- - benchmark
13
- - code
14
- - kernel-optimization
15
- - ascend
16
- - npu
17
- configs:
18
- - config_name: batch1
19
- default: true
20
- data_files:
21
- - split: test
22
- path: metadata/batch1.jsonl
23
- - config_name: preserved_golden
24
- data_files:
25
- - split: test
26
- path: metadata/preserved_golden.jsonl
27
- ---
28
-
29
- # AscendGYM — Kernel Optimization Benchmark on Ascend
30
-
31
- AscendGYM evaluates the ability of models and coding-agent harnesses to optimize
32
- operators for Ascend hardware. This release provides task definitions,
33
- operator baseline snapshots, frozen correctness tests, golden tensors, and the
34
- evaluation client code, preserving the upstream directory layout.
35
-
36
- The benchmark targets kernel optimization in real-world training and inference
37
- workloads. The complete project is included for local task preparation.
38
- Compilation, correctness testing, and performance profiling require access to
39
- an Ascend execution environment. For the complete workflow, see
40
- [the original project instructions](README_PROJECT.md).
41
-
42
- - Upstream project: [SLAI-AITP/AscendGYM](https://github.com/SLAI-AITP/AscendGYM)
43
- - Source revision: [`@@SOURCE_COMMIT@@`](https://github.com/SLAI-AITP/AscendGYM/tree/@@SOURCE_COMMIT@@)
44
- - Active Batch 1 tasks: **@@ACTIVE_TASKS@@**
45
- - Additional preserved-golden tasks: **@@PRESERVED_GOLDEN_TASKS@@** (outside the official Batch 1 evaluation)
46
- - Frozen tensor files: **@@TENSOR_FILES@@** across both subsets
47
- - [Technical report](website/public/AscendGYM_TechReport.pdf)
48
- - [Leaderboard](https://slai-aitp.github.io/AscendGYM/)
49
-
50
- ## Benchmark tasks
51
-
52
- ### Official evaluation: Batch 1
53
-
54
- The default `batch1` subset contains **@@ACTIVE_TASKS@@ evaluation tasks**. They are
55
- grouped below by their source component; each operator links to its task guide.
56
- All tasks include frozen correctness tests.
57
-
58
- @@OFFICIAL_TASK_TABLE@@
59
-
60
- <details>
61
- <summary>Additional reference materials: @@PRESERVED_GOLDEN_TASKS@@ tasks outside the official benchmark</summary>
62
-
63
- The `preserved_golden` subset retains the following ops-transformer sources and
64
- golden references. These tasks are separate from the official Batch 1 evaluation.
65
-
66
- @@REFERENCE_TASK_LIST@@
67
-
68
- </details>
69
-
70
- ### Task indexes and files
71
-
72
- The Dataset Viewer and `load_dataset()` expose **task metadata**, one row per task.
73
- They do not load tensors, execute tests, or run the benchmark. Both subsets use
74
- the `test` split; no training split is provided. The underlying indexes are
75
- [Batch 1](metadata/batch1.jsonl) and [reference materials](metadata/preserved_golden.jsonl).
76
-
77
- Each index row includes the task ID, operator, batch, active status, backend,
78
- paths to the task specification and tests, case manifests, editable source
79
- prefixes, tensor count, baseline component license, and source revision.
80
- Paths are relative to the full downloaded repository root. License fields
81
- describe the baseline component; file-level notices continue to apply.
82
-
83
- The `golden_tensor_files` field counts packaged `.pt` files, not test cases.
84
- Tasks without packaged `.pt` files still include frozen tests.
85
-
86
- ```text
87
- README.md # This Dataset Card
88
- README_PROJECT.md # Original upstream project instructions
89
- metadata/ # Viewer indexes and release provenance
90
- datasets/
91
- registry.json
92
- tasks/ # Task specifications and task documentation
93
- ops_src/ # Operator baselines, tests and frozen tensors
94
- LICENSES/ # Third-party licenses and notices
95
- SOURCE_MANIFEST.json # Upstream file sizes and SHA-256 hashes
96
- bridge/ # Task preparation and evaluation entry points
97
- eval/ # Native and bare harness materials
98
- environment/ # Reference stack and pinned upstream dependencies
99
- docs/ # Execution environment and protocol documentation
100
- examples/
101
- website/ # Leaderboard sources and technical report
102
- LICENSE # Component license summary and full license texts
103
- LICENSE-APACHE-2.0 # Unmodified original project license text
104
- LICENSES.md # Component scope and third-party notice links
105
- CHECKSUMS.sha256 # Release file integrity list
106
- hf/ # Offline preparation and verification tools
107
- ```
108
-
109
- ## Load the task index
110
-
111
- Set `HF_REPO_ID` to the actual Dataset repository ID hosting this release:
112
-
113
- ```bash
114
- export HF_REPO_ID='SLAI-AITP/AscendGYM'
115
- python -m pip install datasets huggingface_hub
116
- ```
117
-
118
- ```python
119
- import os
120
- from datasets import load_dataset
121
-
122
- tasks = load_dataset(os.environ["HF_REPO_ID"], "batch1", split="test")
123
- preserved = load_dataset(os.environ["HF_REPO_ID"], "preserved_golden", split="test")
124
- print(tasks[0]["task_id"], tasks[0]["source_dir"])
125
- ```
126
-
127
- For reproducible use, pass `revision=` with the Hugging Face commit of this
128
- release. The GitHub source commit above is provenance and is not a Hugging Face
129
- revision identifier.
130
-
131
- ## Download the complete benchmark
132
-
133
- ```python
134
- import os
135
- from huggingface_hub import snapshot_download
136
-
137
- snapshot_download(
138
- repo_id=os.environ["HF_REPO_ID"],
139
- repo_type="dataset",
140
- local_dir="./AscendGYM",
141
- )
142
- ```
143
-
144
- Verify the downloaded release **before generating local task packages**:
145
-
146
- ```bash
147
- python3 AscendGYM/hf/prepare_release.py --verify AscendGYM
148
- cd AscendGYM
149
- source datasets/activate.sh
150
- python3 bridge/op_manager.py scan
151
- python3 bridge/op_manager.py build --all
152
- ```
153
-
154
- The scan and build commands prepare local files; they do not compile kernels,
155
- contact a model, or establish NPU correctness or performance. Evaluation client
156
- dependencies and the complete run procedure are in [README_PROJECT.md](README_PROJECT.md).
157
-
158
- ## Execution requirements
159
-
160
- The client requires Python 3.10+ and can run separately from the execution
161
- host. Actual compilation and profiling require a compatible Ascend NPU host,
162
- driver, CANN stack, operator dependencies, and a reachable execution service.
163
- Uploading or loading this dataset does not provision those resources.
164
-
165
- See the [container guide](docs/SANDBOX_IMAGE.md),
166
- [self-hosted environment guide](docs/SELF_HOSTED_ENVIRONMENT.md),
167
- [execution protocol](docs/EXECUTION_BACKEND.md),
168
- [reference stack](environment/reference-stack.json), and
169
- [pinned build dependencies](environment/public-dependencies.json).
170
- The reference stack is a record of selected checks, not a guarantee that every
171
- operator works on every Ascend installation.
172
-
173
- ## Provenance, integrity and limitations
174
-
175
- `metadata/release.json` records the GitHub source commit and release overlays.
176
- The original `datasets/SOURCE_MANIFEST.json` is preserved unchanged. Export
177
- verification checked **@@SOURCE_MANIFEST_FILES_VERIFIED@@** source files against
178
- that manifest. `CHECKSUMS.sha256` additionally covers the complete release
179
- payload, including the task indexes and this card.
180
-
181
- Golden tensors and frozen tests are correctness references. They are not model
182
- weights, model training data, or historical benchmark scores. Keep them unchanged
183
- when optimizing candidate operator code; allowed edits are listed in each task.
184
- Results depend on the hardware, software stack, model endpoint, harness and
185
- evaluation budget. Local release validation does not claim new NPU evaluation
186
- results. No train/test independence beyond the supplied benchmark design is
187
- claimed.
188
-
189
- ## Licenses
190
-
191
- Project-authored code, task definitions, tests, configurations, and documentation
192
- without another license declaration are licensed under
193
- [Apache License 2.0](LICENSE-APACHE-2.0), as stated by the upstream project.
194
- This file preserves the original project license text without modification.
195
-
196
- Included third-party materials retain their own licenses and copyright notices:
197
-
198
- - **ops-nn and ops-transformer:** the CANN Open Software License Agreement
199
- Version 2.0, including its stated Huawei AI Processor purpose restrictions,
200
- and the applicable third-party notices.
201
- - **vLLM-Ascend:** Apache License 2.0 and the existing third-party notices.
202
- - **Frozen inputs and golden tensors:** the scope described in the original
203
- [dataset provenance and licensing statement](datasets/README.md). Third-party
204
- or derived materials remain subject to their original terms.
205
-
206
- The project Apache license does not replace these component-specific terms.
207
- See [the component license overview](LICENSES.md) and
208
- [the full license texts](LICENSE) for details.
209
-
210
- ## Citation and updates
211
-
212
- When referring to this snapshot, cite the
213
- [upstream source revision](https://github.com/SLAI-AITP/AscendGYM/tree/@@SOURCE_COMMIT@@)
214
- and record the Hugging Face revision used for download. Refer to the upstream
215
- technical report for its author-provided citation details.
216
- Report benchmark issues through the [upstream issue tracker](https://github.com/SLAI-AITP/AscendGYM/issues).
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
hf/LICENSES.md DELETED
@@ -1,30 +0,0 @@
1
- # Component licenses for the AscendGYM HF release
2
-
3
- This document describes the existing license scopes in the upstream release.
4
- It does not grant new rights or replace any component's license, copyright
5
- notice, or third-party notice. The Hugging Face card uses `license: other`
6
- to avoid describing this whole distribution as Apache-2.0.
7
-
8
- | Material | Applicable license / declaration |
9
- |---|---|
10
- | Project-authored framework, configurations, tests, task definitions and documentation without another license declaration | [Apache-2.0](LICENSE-APACHE-2.0), within the scope stated in [README_PROJECT.md](README_PROJECT.md#许可) |
11
- | `datasets/ops_src/ops_nn/` baselines and relevant upstream materials | [CANN Open Software License Agreement Version 2.0](datasets/LICENSES/ops-nn/LICENSE), with [third-party notices](datasets/LICENSES/ops-nn/Third_Party_Open_Source_Software_Notice) |
12
- | `datasets/ops_src/ops_transformer/` baselines and relevant upstream materials | [CANN Open Software License Agreement Version 2.0](datasets/LICENSES/ops-transformer/LICENSE), with [third-party notices](datasets/LICENSES/ops-transformer/Third_Party_Open_Source_Software_Notice) |
13
- | `datasets/ops_src/vllm_ascend/` | [vLLM-Ascend Apache-2.0 license](datasets/LICENSES/vllm-ascend/LICENSE), with existing file-level notices |
14
- | Frozen inputs and golden tensors | Follow the [upstream provenance and scope statement](datasets/README.md#来源与许可); no blanket relicensing of third-party or derived material |
15
- | Images, documents and other third-party materials | Existing component and file-level declarations continue to apply |
16
-
17
- The CANN agreement's Sections 2.1 and 3.1 limit the purposes for which the covered
18
- software may be used, modified and distributed to the stated Huawei AI Processor
19
- and/or Software scope. Section 3.3 requires retaining notices and providing the
20
- agreement when distributing covered software or derivatives. Read the full
21
- component texts for their terms.
22
-
23
- The original project Apache-2.0 license is preserved verbatim as
24
- `LICENSE-APACHE-2.0`. The release root `LICENSE` contains this summary followed by
25
- the full Apache-2.0 and CANN component texts. Original licenses and third-party
26
- notices under `datasets/LICENSES/` are preserved unchanged.
27
-
28
- File hashes establish which bytes were packaged; they do not prove copyright
29
- ownership or additional authorization. This packaging adds no rights beyond
30
- those granted by the existing licenses.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
hf/README.md DELETED
@@ -1,80 +0,0 @@
1
- # Hugging Face 发布与校验
2
-
3
- 发布类型为 **Dataset**,采用保留完整工程目录的评测包。GitHub 继续作为开发上游;
4
- HF 提供固定版本、任务索引和完整文件下载。
5
-
6
- ## 1. 导出发布目录
7
-
8
- 在 GitHub 源码工作区根目录执行(Python 3.10+,仅使用标准库)。若源码工作区没有
9
- `hf/`,先复制本发布包的 `hf/` 目录。已下载的 HF 发布包用于下文的 `--verify` 校验,
10
- 不要将其作为再次导出的源码工作区。
11
-
12
- ```bash
13
- python3 hf/prepare_release.py
14
- ```
15
-
16
- 默认输出到工作区同级的 `<工作区名称>-release/`。也可以指定一个尚不存在的目录:
17
-
18
- ```bash
19
- python3 hf/prepare_release.py --output ../AscendGYM-hf-release-v2
20
- ```
21
-
22
- 脚本仅复制 Git 跟踪的文件和明确列出的 `hf/` 发布工具;要求已跟踪文件与 HEAD
23
- 一致,防止版本记录与导出内容不一致。不会上传 `.git/`、未跟踪的运行产物、虚拟环境
24
- 或本地凭据。不会创建 HF 仓库、登录、上传或调用 NPU/模型。
25
-
26
- 发布目录中的变化:
27
-
28
- - 根 `README.md` 替换为 Dataset Card;项目说明保存在 `README_PROJECT.md`,项目文档中的相关链接同步调整。
29
- - 原根 `LICENSE` 保存在 `LICENSE-APACHE-2.0`;新根 `LICENSE` 汇总各组件适用范围和
30
- 完整 Apache-2.0、CANN 协议文本,原第三方声明保留原路径。
31
- - `.gitattributes` 增加 `.pt` 的二进制存储规则;张量字节不变。
32
- - `metadata/batch1.jsonl`:17 个正式任务,作为默认 `test` 子集。
33
- - `metadata/preserved_golden.jsonl`:3 个额外保留任务,单独作为 `test` 子集。
34
- - `metadata/release.json`:来源提交、打包变化和校验统计。
35
- - `CHECKSUMS.sha256`:除自身外所有发布文件的 SHA-256。
36
-
37
- 任务数量根据 registry 自动生成;上面的数量对应本次源快照。Viewer 只展示任务元数据,
38
- 源码、测试和张量需要下载完整评测包后使用。
39
-
40
- ## 2. 离线校验
41
-
42
- ```bash
43
- python3 hf/prepare_release.py --verify ../AscendGYM-hf-release
44
- ```
45
-
46
- 校验发布文件哈希、`SOURCE_MANIFEST.json` 覆盖与哈希、任务引用路径、张量数量、正式
47
- 任务与 bridge catalog 的对应关系,以及两份 JSONL 索引。校验命令不依赖 Git 或网络,
48
- 也已随发布包提供。HF 下载或上传产生的 `.cache/huggingface/` 不属于发布内容。
49
-
50
- 请在生成 `datasets/packages/` 等运行文件前校验。该校验确认打包完整性,不替代昇腾
51
- 环境中的编译、正确性测试和 profiling;许可证范围说明也不替代材料来源授权。
52
-
53
- ## 3. 上传发布版本
54
-
55
- 目标为 [SLAI-AITP/AscendGYM Dataset](https://huggingface.co/datasets/SLAI-AITP/AscendGYM)。
56
- 确认发布目录通过上述校验后,使用具有目标仓库写权限的账户上传:
57
-
58
- ```bash
59
- python -m pip install 'huggingface_hub>=0.36'
60
- hf auth login
61
- export HF_REPO_ID='SLAI-AITP/AscendGYM'
62
- hf upload "$HF_REPO_ID" ../AscendGYM-hf-release . \
63
- --repo-type dataset \
64
- --exclude '.cache/**' \
65
- --commit-message 'Publish AscendGYM benchmark snapshot'
66
- ```
67
-
68
- 通过 `hf auth login` 交互登录,不要将 token 写进源码、文档或上传目录。上传不会改变
69
- 已有仓库的可见性;公开范围由仓库设置管理。
70
- 通过 HTTP 上传完整文件,不需要把 GitHub 的历史提交镜像推送到 HF,也不要求安装
71
- Git LFS 命令。
72
-
73
- 上传完成后检查 Dataset Card、两个 Viewer 子集、完整快照下载及哈希校验,并记录 HF
74
- 提交号。GitHub 提交号是来源标识,不能直接当作 HF 的 `revision`。
75
-
76
- 参考:[Dataset Card](https://huggingface.co/docs/hub/datasets-cards)、
77
- [Viewer 配置](https://huggingface.co/docs/hub/datasets-manual-configuration)、
78
- [组件许可元数据](https://huggingface.co/docs/hub/repositories-licenses)、
79
- [HTTP 上传](https://huggingface.co/docs/huggingface_hub/guides/upload)、
80
- [HF CLI](https://huggingface.co/docs/huggingface_hub/guides/cli)。
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
hf/prepare_release.py DELETED
@@ -1,302 +0,0 @@
1
- #!/usr/bin/env python3
2
- """Prepare or verify an offline Hugging Face Dataset release (Python 3.10+)."""
3
-
4
- import argparse
5
- import hashlib
6
- import json
7
- import os
8
- from pathlib import Path
9
- import re
10
- import shutil
11
- import subprocess
12
- import tempfile
13
-
14
-
15
- SOURCE_URL = "https://github.com/SLAI-AITP/AscendGYM"
16
- GROUPS = {"tasks": "batch1", "preserved_golden_tasks": "preserved_golden"}
17
- LICENSE_FILES = {
18
- "nn": "datasets/LICENSES/ops-nn/LICENSE",
19
- "transformer": "datasets/LICENSES/ops-transformer/LICENSE",
20
- "vllm_ascend": "datasets/LICENSES/vllm-ascend/LICENSE",
21
- }
22
- HELPERS = ("README.md", "DATASET_CARD.md", "LICENSES.md", "prepare_release.py")
23
-
24
-
25
- def read_json(path):
26
- return json.loads(path.read_text(encoding="utf-8"))
27
-
28
-
29
- def write_json(path, value):
30
- path.parent.mkdir(parents=True, exist_ok=True)
31
- path.write_text(json.dumps(value, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
32
-
33
-
34
- def sha256(path):
35
- digest = hashlib.sha256()
36
- with path.open("rb") as stream:
37
- for chunk in iter(lambda: stream.read(1024 * 1024), b""):
38
- digest.update(chunk)
39
- return digest.hexdigest()
40
-
41
-
42
- def inside(root, relative):
43
- path = root / relative
44
- if Path(relative).is_absolute() or not path.resolve().is_relative_to(root.resolve()):
45
- raise ValueError(f"Path leaves the release directory: {relative}")
46
- if path.is_symlink():
47
- raise ValueError(f"Symlinks are not release files: {relative}")
48
- return path
49
-
50
-
51
- def payload_files(root):
52
- return sorted(
53
- p for p in root.rglob("*")
54
- if p.is_file()
55
- and p.relative_to(root).as_posix() != "CHECKSUMS.sha256"
56
- and not p.relative_to(root).as_posix().startswith(".cache/huggingface/")
57
- )
58
-
59
-
60
- def rewrite_project_readme_links(root):
61
- """Keep project instructions reachable after the Dataset Card replaces README."""
62
- changed = []
63
- for path in sorted(root.rglob("*.md")):
64
- relative = path.relative_to(root).as_posix()
65
- if relative.startswith(("datasets/ops_src/", "eval/bare_harness/Backend/", "hf/")):
66
- continue
67
- old = Path(os.path.relpath(root / "README.md", path.parent)).as_posix()
68
- new = Path(os.path.relpath(root / "README_PROJECT.md", path.parent)).as_posix()
69
- pattern = re.compile(r'(\]\(|\bhref=[\"\x27])' + re.escape(old) + r'(?=[#)\"\x27])')
70
- original = path.read_text(encoding="utf-8")
71
- output, fence = [], None
72
- for line in original.splitlines(keepends=True):
73
- marker = re.match(r"^\s*(`{3,}|~{3,})", line)
74
- if marker:
75
- delimiter = marker.group(1)
76
- if fence is None:
77
- fence = delimiter
78
- elif delimiter[0] == fence[0] and len(delimiter) >= len(fence):
79
- fence = None
80
- output.append(line)
81
- continue
82
- output.append(line if fence else pattern.sub(lambda match: match.group(1) + new, line))
83
- updated = "".join(output)
84
- if updated != original:
85
- path.write_text(updated, encoding="utf-8")
86
- changed.append(relative)
87
- return changed
88
-
89
-
90
- def check_snapshot(root, commit):
91
- """Check frozen files and derive index rows without importing torch or tests."""
92
- manifest = read_json(root / "datasets/SOURCE_MANIFEST.json")
93
- seen = set()
94
- for record in manifest["files"]:
95
- relative = "datasets/" + record["path"]
96
- if relative in seen:
97
- raise ValueError(f"Duplicate source manifest entry: {relative}")
98
- seen.add(relative)
99
- path = inside(root, relative)
100
- if not path.is_file() or path.stat().st_size != record["bytes"]:
101
- raise ValueError(f"Missing file or size mismatch: {relative}")
102
- if sha256(path) != record["sha256"]:
103
- raise ValueError(f"Source manifest SHA-256 mismatch: {relative}")
104
- actual = {p.relative_to(root).as_posix() for p in (root / "datasets/ops_src").rglob("*") if p.is_file()}
105
- if actual != seen:
106
- raise ValueError(f"Source manifest coverage mismatch: {sorted(actual ^ seen)[:8]}")
107
-
108
- registry = read_json(root / "datasets/registry.json")
109
- catalog = read_json(inside(root, "datasets/" + registry["active_catalog"]))
110
- rows = {name: [] for name in GROUPS.values()}
111
- ids = set()
112
- for group, name in GROUPS.items():
113
- for entry in registry[group]:
114
- spec_path = "datasets/" + entry["spec"]
115
- spec = read_json(inside(root, spec_path))
116
- task_id = spec["task_id"]
117
- if task_id in ids or task_id != entry["id"]:
118
- raise ValueError(f"Duplicate or inconsistent task ID: {task_id}")
119
- ids.add(task_id)
120
- if spec["active"] != (group == "tasks"):
121
- raise ValueError(f"Task active status mismatch: {task_id}")
122
- source_dir = "datasets/" + spec["source_dir"]
123
- source = inside(root, source_dir)
124
- test_script = source_dir + "/" + spec["test_script"]
125
- case_manifests = [source_dir + "/" + p for p in spec["case_manifests"]]
126
- readme = str(Path(spec_path).with_name("README.md"))
127
- license_file = LICENSE_FILES[spec["ops_repo"]]
128
- for relative in [test_script, readme, license_file, *case_manifests]:
129
- if not inside(root, relative).is_file():
130
- raise ValueError(f"Task {task_id} references a missing file: {relative}")
131
- tensor_count = sum(1 for p in source.rglob("*.pt") if p.is_file())
132
- if tensor_count != spec["golden_tensor_files"]:
133
- raise ValueError(f"Task tensor count mismatch: {task_id}")
134
- rows[name].append({
135
- "task_id": task_id,
136
- "operator": spec["operator"],
137
- "batch": spec["batch"],
138
- "active": spec["active"],
139
- "backend": spec["ops_repo"],
140
- "task_spec": spec_path,
141
- "readme": readme,
142
- "source_dir": source_dir,
143
- "test_script": test_script,
144
- "case_manifests": case_manifests,
145
- "golden_tensor_files": tensor_count,
146
- "editable_prefixes": spec["editable_prefixes"],
147
- "component_license": "Apache-2.0" if spec["ops_repo"] == "vllm_ascend" else "CANN Open Software License Agreement Version 2.0",
148
- "component_license_file": license_file,
149
- "source_commit": commit,
150
- })
151
- if {r["operator"] for r in rows["batch1"]} != {op["name"] for op in catalog["ops"]}:
152
- raise ValueError("The active task index does not match the bridge catalog")
153
- return rows, {
154
- "source_manifest_files_verified": len(seen),
155
- "active_tasks": len(rows["batch1"]),
156
- "preserved_golden_tasks": len(rows["preserved_golden"]),
157
- "tensor_files": sum(r["golden_tensor_files"] for group in rows.values() for r in group),
158
- }
159
-
160
-
161
- def verify(root):
162
- root = root.resolve()
163
- checksums = {}
164
- for line in (root / "CHECKSUMS.sha256").read_text(encoding="utf-8").splitlines():
165
- digest, relative = line.split(" ", 1)
166
- if relative in checksums or len(digest) != 64:
167
- raise ValueError(f"Invalid checksum entry: {relative}")
168
- checksums[relative] = digest
169
- path = inside(root, relative)
170
- if not path.is_file() or sha256(path) != digest:
171
- raise ValueError(f"Release checksum mismatch: {relative}")
172
- actual = {p.relative_to(root).as_posix() for p in payload_files(root)}
173
- if actual != set(checksums):
174
- raise ValueError(f"Release file list mismatch: {sorted(actual ^ set(checksums))[:8]}")
175
- release = read_json(root / "metadata/release.json")
176
- rows, counts = check_snapshot(root, release["source_commit"])
177
- for name, expected in rows.items():
178
- path = root / f"metadata/{name}.jsonl"
179
- found = [json.loads(line) for line in path.read_text(encoding="utf-8").splitlines()]
180
- if found != expected:
181
- raise ValueError(f"Task index differs from task definitions: {name}")
182
- if counts != release["validation"]:
183
- raise ValueError("Release counts differ from the validated snapshot")
184
- if len(actual) != release["payload_files"]:
185
- raise ValueError("Release payload file count mismatch")
186
- print(json.dumps({"release": str(root), "checksummed_files": len(actual), **counts}, indent=2))
187
-
188
-
189
- def render_task_overview(rows):
190
- """Keep official tasks grouped by source and reference-only tasks separate."""
191
- labels = {"nn": "ops-nn", "transformer": "ops-transformer", "vllm_ascend": "vLLM-Ascend"}
192
- groups = {}
193
- for row in rows["batch1"]:
194
- groups.setdefault(row["backend"], []).append(row)
195
- table = ["| Source component | Operators |", "|---|---|"]
196
- for backend, group in groups.items():
197
- links = ", ".join(f"[{row['operator']}]({row['readme']})" for row in group)
198
- table.append(f"| **{labels.get(backend, backend)}** ({len(group)} tasks) | {links} |")
199
- references = "\n".join(f"- [{row['operator']}]({row['readme']})" for row in rows["preserved_golden"])
200
- return {"OFFICIAL_TASK_TABLE": "\n".join(table), "REFERENCE_TASK_LIST": references}
201
-
202
-
203
- def prepare(source, output):
204
- source, output = source.resolve(), output.resolve()
205
- if output == source or output.is_relative_to(source) or source.is_relative_to(output):
206
- raise ValueError("The output must be separate from the source repository")
207
- if output.exists():
208
- raise ValueError(f"Output already exists; choose a new directory: {output}")
209
-
210
- def git(*args):
211
- return subprocess.check_output(["git", "-C", str(source), *args])
212
-
213
- dirty = git("diff", "--name-only", "HEAD").decode().strip()
214
- if dirty:
215
- raise ValueError("Commit or restore tracked changes before exporting:\n" + dirty)
216
- commit = git("rev-parse", "HEAD").decode().strip()
217
- paths = [p.decode() for p in git("ls-files", "-z").split(b"\0") if p]
218
- paths = [p for p in paths if not p.startswith("hf/")]
219
- rows, counts = check_snapshot(source, commit)
220
-
221
- output.parent.mkdir(parents=True, exist_ok=True)
222
- with tempfile.TemporaryDirectory(prefix=output.name + "-", dir=output.parent) as temp:
223
- staging = Path(temp) / "payload"
224
- staging.mkdir()
225
- for relative in paths:
226
- original = inside(source, relative)
227
- target = inside(staging, relative)
228
- target.parent.mkdir(parents=True, exist_ok=True)
229
- shutil.copy2(original, target)
230
- for name in HELPERS:
231
- target = staging / "hf" / name
232
- target.parent.mkdir(parents=True, exist_ok=True)
233
- shutil.copy2(source / "hf" / name, target)
234
-
235
- (staging / "README.md").rename(staging / "README_PROJECT.md")
236
- rewrite_project_readme_links(staging)
237
- (staging / "LICENSE").rename(staging / "LICENSE-APACHE-2.0")
238
- summary = (source / "hf/LICENSES.md").read_text(encoding="utf-8")
239
- (staging / "LICENSES.md").write_text(summary, encoding="utf-8")
240
- licenses = [
241
- ("Apache-2.0 (project-authored material; see component scopes above)", "LICENSE-APACHE-2.0"),
242
- ("CANN license (ops-nn)", LICENSE_FILES["nn"]),
243
- ("CANN license (ops-transformer)", LICENSE_FILES["transformer"]),
244
- ]
245
- combined = summary + "\n\n## Full component license texts\n"
246
- for title, relative in licenses:
247
- combined += f"\n### {title}\n\n" + (staging / relative).read_text(encoding="utf-8") + "\n"
248
- (staging / "LICENSE").write_text(combined, encoding="utf-8")
249
-
250
- for name, group in rows.items():
251
- target = staging / f"metadata/{name}.jsonl"
252
- target.parent.mkdir(parents=True, exist_ok=True)
253
- target.write_text("".join(json.dumps(r, ensure_ascii=False) + "\n" for r in group), encoding="utf-8")
254
- card = (source / "hf/DATASET_CARD.md").read_text(encoding="utf-8")
255
- replacements = {"SOURCE_COMMIT": commit, **render_task_overview(rows), **{key.upper(): str(value) for key, value in counts.items()}}
256
- for key, value in replacements.items():
257
- card = card.replace("@@" + key + "@@", value)
258
- if "@@" in card:
259
- raise ValueError("Unresolved Dataset Card template variable")
260
- (staging / "README.md").write_text(card, encoding="utf-8")
261
- with (staging / ".gitattributes").open("a", encoding="utf-8") as stream:
262
- stream.write("\n# Hugging Face binary storage; preserve frozen tensor bytes.\n*.pt filter=lfs diff=lfs merge=lfs -text\n")
263
-
264
- release = {
265
- "schema_version": "ascendgym-hf-release-v1",
266
- "source_repository": SOURCE_URL,
267
- "source_commit": commit,
268
- "source_files": len(paths),
269
- "source_manifest_sha256": sha256(staging / "datasets/SOURCE_MANIFEST.json"),
270
- "validation": counts,
271
- "payload_files": len(payload_files(staging)) + 1,
272
- "release_overlays": ["README.md (original saved as README_PROJECT.md)", "Project README links retargeted to README_PROJECT.md", "LICENSE (original saved as LICENSE-APACHE-2.0)", "LICENSES.md", ".gitattributes (added binary storage rule)", "metadata/", "hf/"],
273
- "validation_scope": "File integrity, task references and index consistency; no NPU execution or model evaluation",
274
- }
275
- write_json(staging / "metadata/release.json", release)
276
- (staging / "CHECKSUMS.sha256").write_text(
277
- "".join(f"{sha256(p)} {p.relative_to(staging).as_posix()}\n" for p in payload_files(staging)), encoding="utf-8"
278
- )
279
- verify(staging)
280
- staging.rename(output)
281
- print(f"Prepared: {output}")
282
-
283
-
284
- def main():
285
- parser = argparse.ArgumentParser(description=__doc__)
286
- parser.add_argument("--output", type=Path, help="New directory for the release (never overwritten)")
287
- parser.add_argument("--verify", type=Path, help="Verify an existing release, without Git or network access")
288
- args = parser.parse_args()
289
- if args.verify and args.output:
290
- parser.error("Use either --verify or --output")
291
- try:
292
- if args.verify:
293
- verify(args.verify)
294
- else:
295
- source = Path(__file__).resolve().parents[1]
296
- prepare(source, args.output or source.with_name(source.name + "-release"))
297
- except (ValueError, OSError, KeyError, subprocess.CalledProcessError) as error:
298
- parser.exit(1, f"Release check failed: {error}\n")
299
-
300
-
301
- if __name__ == "__main__":
302
- main()
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
metadata/release.json CHANGED
@@ -10,16 +10,16 @@
10
  "preserved_golden_tasks": 3,
11
  "tensor_files": 384
12
  },
13
- "payload_files": 1626,
14
  "release_overlays": [
15
  "README.md (original saved as README_PROJECT.md)",
16
  "LICENSE (original saved as LICENSE-APACHE-2.0)",
17
  "LICENSES.md",
18
  ".gitattributes (added binary storage rule)",
19
  "metadata/",
20
- "hf/",
21
  ".gitattributes (Hugging Face generated additional binary storage rules)",
22
- "Documentation and Leaderboard resource-link corrections; benchmark inputs unchanged"
 
23
  ],
24
  "validation_scope": "File integrity, task references and index consistency; no NPU execution or model evaluation",
25
  "huggingface_repository": "SLAI-AITP/AscendGYM",
@@ -46,5 +46,19 @@
46
  "card_presentation_revision": {
47
  "base_huggingface_commit": "961d5d44af6f50712dc0adb4d60ce15a5ca01946",
48
  "scope": "README task overview and export template only; Dataset Viewer configuration and task index data unchanged."
 
 
 
 
 
 
 
 
 
 
 
 
 
 
49
  }
50
  }
 
10
  "preserved_golden_tasks": 3,
11
  "tensor_files": 384
12
  },
13
+ "payload_files": 1454,
14
  "release_overlays": [
15
  "README.md (original saved as README_PROJECT.md)",
16
  "LICENSE (original saved as LICENSE-APACHE-2.0)",
17
  "LICENSES.md",
18
  ".gitattributes (added binary storage rule)",
19
  "metadata/",
 
20
  ".gitattributes (Hugging Face generated additional binary storage rules)",
21
+ "Documentation and Leaderboard resource-link corrections; benchmark inputs unchanged",
22
+ "HF package excludes website sources, publishing helpers, and the GitHub Pages workflow; report and README logos retained under docs/"
23
  ],
24
  "validation_scope": "File integrity, task references and index consistency; no NPU execution or model evaluation",
25
  "huggingface_repository": "SLAI-AITP/AscendGYM",
 
46
  "card_presentation_revision": {
47
  "base_huggingface_commit": "961d5d44af6f50712dc0adb4d60ce15a5ca01946",
48
  "scope": "README task overview and export template only; Dataset Viewer configuration and task index data unchanged."
49
+ },
50
+ "package_layout_revision": {
51
+ "base_huggingface_commit": "a76318cfb67b61a49baf05b34527246f5d89d028",
52
+ "removed_directories": [
53
+ "hf/",
54
+ "website/"
55
+ ],
56
+ "removed_workflow": ".github/workflows/pages.yml",
57
+ "relocated_files": {
58
+ "website/public/AscendGYM_TechReport.pdf": "docs/AscendGYM_TechReport.pdf",
59
+ "website/public/assets/logo/mark-primary.svg": "docs/assets/logo/mark-primary.svg",
60
+ "website/public/assets/logo/mark-dark.svg": "docs/assets/logo/mark-dark.svg"
61
+ },
62
+ "scope": "HF package layout and documentation only; task indexes, frozen inputs, evaluation code, and GitHub deployment unchanged."
63
  }
64
  }
website/.gitattributes DELETED
@@ -1,5 +0,0 @@
1
- # Some upstream source files use CRLF. Accept that line ending in whitespace checks.
2
- * whitespace=blank-at-eol,blank-at-eof,space-before-tab,cr-at-eol
3
-
4
- # The shape bundle records a SHA-256 of these CSV bytes; do not normalize them.
5
- public/templates/*.csv -text
 
 
 
 
 
 
website/.gitignore DELETED
@@ -1,15 +0,0 @@
1
- node_modules/
2
- dist/
3
- .npm-cache/
4
- .local-deploy/
5
- tmp/
6
- screenshots/
7
- __pycache__/
8
- *.py[cod]
9
- .DS_Store
10
- ._*
11
- __MACOSX/
12
- .env
13
- .env.*
14
- !.env.example
15
- !.env.template
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
website/README.md DELETED
@@ -1,81 +0,0 @@
1
- # AscendGYM 项目主页
2
-
3
- 项目主页与评测榜单,使用 Vue 3、Vite 和随包测量数据生成静态网站。评测框架的安装、模型配置和使用方法见[根 README](../README_PROJECT.md)。
4
-
5
- ## 本地预览
6
-
7
- 环境:Node.js 22、npm、Python 3.12(构建前的 shape 数据校验使用 Python 标准库)。
8
-
9
- ```bash
10
- cd website
11
- npm ci
12
- npm run dev -- --host 127.0.0.1
13
- ```
14
-
15
- 开发地址为 `http://127.0.0.1:5173`。当前默认数据来自 `src/data/supplement_bundle.json` 与 `src/data/shape_results.json`;静态展示不需要模型密钥、AscendForge 服务或 NPU。
16
-
17
- ## 构建
18
-
19
- ```bash
20
- npm run build
21
- npm run preview -- --host 127.0.0.1
22
- ```
23
-
24
- 构建产物为 `dist/`,预览端口为 4174。构建前会从随包的 `public/templates/shape-results.csv` 校验并生成 shape 数据。默认相对资源路径及 hash 路由支持静态托管。
25
-
26
- 针对本项目 GitHub Pages 子路径的本地验证命令:
27
-
28
- ```bash
29
- npm run build -- --base /AscendGYM/
30
- npm run preview -- --host 127.0.0.1 --base /AscendGYM/
31
- ```
32
-
33
- 随后访问 `http://127.0.0.1:4174/AscendGYM/`。正式部署由工作流读取 Pages 的实际 base path,不在组件中写死站点根路径。
34
-
35
- ## 部署到 GitHub Pages
36
-
37
- 仓库已经提供 [pages.yml](../.github/workflows/pages.yml)。当前仅允许手动触发,普通代码推送不会自动发布网站。
38
-
39
- 1. 将主页与工作流提交并合入 `main`。
40
- 2. 在仓库 **Settings → Pages → Build and deployment → Source** 选择 **GitHub Actions**。设置页面需要相应的仓库管理权限。
41
- 3. 在 **Actions → Deploy project website → Run workflow** 选择 `main`,运行构建和部署。
42
- 4. 在成功的运行记录中打开部署地址,核对首页、Methodology、Imports、logo、技术报告和仓库按钮。
43
-
44
- Leaderboard 地址:[AscendGYM Leaderboard](https://slai-aitp.github.io/AscendGYM/)。更新网站后,重新运行上述部署工作流发布修改。
45
-
46
- 工作流只上传 `website/dist/`,使用 GitHub 自动提供的授权完成 Pages 部署,不需要把个人 token 写入仓库。参见 [GitHub Pages 官方配置说明](https://docs.github.com/en/pages/getting-started-with-github-pages/using-custom-workflows-with-github-pages) 和 [Vite 部署说明](https://vite.dev/guide/static-deploy.html#github-pages)。
47
-
48
- ## 文件和数据
49
-
50
- | 路径 | 内容 |
51
- |---|---|
52
- | `src/components/`、`src/pages/` | 页面和图表组件 |
53
- | `src/data/` | 当前展示所用的随包数据与历史辅助数据 |
54
- | `public/assets/logo/` | 项目 logo,根 README 直接引用同一份资源 |
55
- | `public/assets/` | 图标与来源说明 |
56
- | `public/AscendGYM_TechReport.pdf` | 技术报告 |
57
- | `public/templates/` | shape 数据和导入模板 |
58
- | `scripts/` | 构建、数据转换和检查工具 |
59
- | `evaluate/` | 原主页项目携带的离线评估辅助脚本 |
60
-
61
- `npm ci` 和生产构建使用随包数据即可。`data:sync`、`test:supplement` 需要额外的原始测量归档 `../eval_plot/png_600dpi_v2/`;这些归档不属于本站的构建依赖,也没有随本次主页源码复制。原有 `npm test` 包含该归档核对,因此不能把缺少归档时的执行结果表述为完整验收通过。
62
-
63
- 可以单独运行不依赖外部归档的现有检查:
64
-
65
- ```bash
66
- npm run test:anchors
67
- npm run test:mapping
68
- npm run test:precision
69
- npm run test:pricing
70
- npm run test:pagination
71
- npm run test:shapes
72
- npm run test:submission
73
- ```
74
-
75
- 上述检查的范围与独立原始数据复核不同。更新测量值或重算指标时,还需核对相应原始记录及评测定义。
76
-
77
- ## 发布范围
78
-
79
- 源码、锁文件、随包数据、公开资源及来源声明纳入 Git;`node_modules/`、`dist/`、截图、临时日志、macOS 元数据和本地部署缓存不纳入。页面的 Hugging Face 入口指向 [SLAI-AITP/AscendGYM 数据集](https://huggingface.co/datasets/SLAI-AITP/AscendGYM)。
80
-
81
- 许可范围见[主仓许可说明](../README_PROJECT.md#license),第三方图标及其他材料保留各自来源与声明。
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
website/docs/api-cost-estimates-2026-10-26.csv DELETED
@@ -1,22 +0,0 @@
1
- entry_id,configuration,cohort,tokens,sts,estimate_usd_10pct_output_no_cache,scenario_usd_5pct_output_no_cache,scenario_usd_20pct_output_no_cache,scenario_usd_10pct_output_80pct_input_cache,rate_basis,official_source,checked
2
- "claude-code+claude-opus-5","Claude Code × Claude-Opus-5","harness","1389784009","74.03717915621269","9728.488063","8338.704054","12508.056080999999","5225.58787384","exact_model","https://platform.claude.com/docs/en/about-claude/pricing","2026-10-26"
3
- "claude-code+gpt-5.6-terra","Claude Code × GPT-5.6 Terra","harness","490526582","7.027841914037475","1471.5797460000001","1226.3164550000001","1962.106328","835.8572957280002","exact_model","https://developers.openai.com/api/docs/models/gpt-5.6-terra","2026-10-26"
4
- "claude-code+glm-5.3","Claude Code × GLM-5.3","harness","582735528","44.03181421251067","990.6503976000001","903.2400683999999","1165.471056","512.3410762176001","exact_model","https://docs.z.ai/guides/overview/pricing","2026-10-26"
5
- "claude-code+Qwen3.5-122B","Claude Code × Qwen3.5-122B-A10B","harness","232751513","6.930661924692416","45.4330953376","36.0997596663","64.0997666802","","exact_model","https://www.alibabacloud.com/help/en/model-studio/qwen3-5-122b-a10b","2026-10-26"
6
- "claude-code+Qwen3.6-35B","Claude Code × Qwen3.6-35B-A3B","harness","324953108","4.921939382511611","120.7850702436","100.68672051379998","160.9817697032","","exact_model","https://www.alibabacloud.com/help/en/model-studio/qwen3-6-35b-a3b","2026-10-26"
7
- "claude-code+DeepSeek-V4-Flash-BF16","Claude Code × DeepSeek-V4-Flash","harness","820864004.0000001","10.314318956397646","320.1369615600001","283.19808138","394.01472192000006","146.37646919328","successor_proxy","https://api-docs.deepseek.com/quick_start/pricing","2026-10-26"
8
- "codex+gpt-5.6-terra","Codex × GPT-5.6 Terra","harness","2387282525","58.36028333429681","7161.847575","5968.2063125","9549.1301","4067.9294226000006","exact_model","https://developers.openai.com/api/docs/models/gpt-5.6-terra","2026-10-26"
9
- "codex+glm-5.3","Codex × GLM-5.3","harness","1658804780","5.415272052537867","2819.968126","2571.1474089999997","3317.60956","1458.421162576","exact_model","https://docs.z.ai/guides/overview/pricing","2026-10-26"
10
- "codex+Qwen3.5-122B","Codex × Qwen3.5-122B-A10B","harness","926983058.0000001","3.6385936059254735","180.94709292160005","143.77507229580004","255.29113417320008","","exact_model","https://www.alibabacloud.com/help/en/model-studio/qwen3-5-122b-a10b","2026-10-26"
11
- "codex+Qwen3.6-35B","Codex × Qwen3.6-35B-A3B","harness","447902087","6.470042353818354","166.4852057379","138.78246165694998","221.89069389980003","","exact_model","https://www.alibabacloud.com/help/en/model-studio/qwen3-6-35b-a3b","2026-10-26"
12
- "codex+DeepSeek-V4-Flash-BF16","Codex × DeepSeek-V4-Flash","harness","1060282342","11.517911130385038","413.51011338","365.79740798999995","508.93552415999994","189.06954722543998","successor_proxy","https://api-docs.deepseek.com/quick_start/pricing","2026-10-26"
13
- "opencode+claude-opus-5","OpenCode × Claude-Opus-5","harness","247158861","47.06815928221053","1730.1120270000001","1482.953166","2224.429749","929.31731736","exact_model","https://platform.claude.com/docs/en/about-claude/pricing","2026-10-26"
14
- "opencode+gpt-5.6-terra","OpenCode × GPT-5.6 Terra","harness","51899132","6.624852911018037","155.697396","129.74783","207.596528","88.43612092800001","exact_model","https://developers.openai.com/api/docs/models/gpt-5.6-terra","2026-10-26"
15
- "opencode+glm-5.3","OpenCode × GLM-5.3","harness","598640013","31.601038826152745","1017.6880221","927.8920201499998","1197.280026","526.3242994295999","exact_model","https://docs.z.ai/guides/overview/pricing","2026-10-26"
16
- "opencode+Qwen3.5-122B","OpenCode × Qwen3.5-122B-A10B","harness","439008108","7.9019840608363605","85.6943826816","68.0901575508","120.90283294320001","","exact_model","https://www.alibabacloud.com/help/en/model-studio/qwen3-5-122b-a10b","2026-10-26"
17
- "opencode+Qwen3.6-35B","OpenCode × Qwen3.6-35B-A3B","harness","406413662","7.161371031002409","151.0639581654","125.92727317069998","201.33732815480002","","exact_model","https://www.alibabacloud.com/help/en/model-studio/qwen3-6-35b-a3b","2026-10-26"
18
- "opencode+DeepSeek-V4-Flash-BF16","OpenCode × DeepSeek-V4-Flash","harness","219336522","9.942030712770716","85.54124358","75.67110009","105.28153056","39.11208860303999","successor_proxy","https://api-docs.deepseek.com/quick_start/pricing","2026-10-26"
19
- "kl-luna","Default Cold × GPT-5.6 Luna","default","6488778357","22.58287459833043","1946.6335071000003","1622.19458925","2595.5113428000004","1105.6878320328","exact_model","https://developers.openai.com/api/docs/models/gpt-5.6-luna","2026-10-26"
20
- "kl-warm","Default Warm × GPT-5.6 Luna","default","970820498","34.695288712084285","291.24614940000004","242.7051245","388.32819920000003","165.4278128592","exact_model","https://developers.openai.com/api/docs/models/gpt-5.6-luna","2026-10-26"
21
- "kl-nl","Default No Lessons × GPT-5.6 Luna","default","1881228673","10.889903369737157","564.3686019000002","470.30716825","752.4914692000001","320.5613658792","exact_model","https://developers.openai.com/api/docs/models/gpt-5.6-luna","2026-10-26"
22
- "kl-cl","Default Code Only × GPT-5.6 Luna","default","1469735068","13.96391490898753","440.92052040000004","367.433767","587.8940272","250.4428555872","exact_model","https://developers.openai.com/api/docs/models/gpt-5.6-luna","2026-10-26"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
website/docs/perf-cost-analysis-2026-10-26.md DELETED
@@ -1,69 +0,0 @@
1
- # Perf × Cost 分析与 API 等价成本估算
2
-
3
- 核验日期:2026-10-26。价格单位为 USD / 100 万 token。
4
-
5
- ## A/B 是什么关系
6
-
7
- A 的纵轴为 STS,横轴为累计 token 或估算 API 成本;越靠左上越好,Pareto 前沿表示无法同时提高 STS 并降低成本的配置。STS 自身已包含 token 延迟惩罚,因此它不是单独的最终 GM 加速比。
8
-
9
- B 保留 STS 纵轴,将横轴替换为 PAOS(无量纲的场景相关 profiling 收益指标)。它是独立诊断视角,不是另一种 API 成本,也不是 A 的下一阶段。旧指标设计采用 PAOS = sqrt(PKAI × PTEI),结合 recipe 加权算子收益与估算 step 级收益;PAOS=1 是基准参考。技术报告指出,默认单算子 recipe 下这些量会退化为内核加速比,不能据此声称真实工作负载吞吐提升。当前 21 个配置的 PAOS 均未导出,所以 B 没有可用散点。
10
-
11
- ## 官方单价
12
-
13
- | 模型 | 输入 | 缓存输入 | 输出 | 来源 |
14
- | --- | ---: | ---: | ---: | --- |
15
- | Claude-Opus-5 | 5 | 0.5 | 25 | [官方页面](https://platform.claude.com/docs/en/about-claude/pricing) |
16
- | GPT-5.6 Terra | 2 | 0.2 | 12 | [官方页面](https://developers.openai.com/api/docs/models/gpt-5.6-terra) |
17
- | GLM-5.3 | 1.4 | 0.26 | 4.4 | [官方页面](https://docs.z.ai/guides/overview/pricing) |
18
- | Qwen3.5-122B-A10B | 0.115 | 不支持/未公布 | 0.917 | [官方页面](https://www.alibabacloud.com/help/en/model-studio/qwen3-5-122b-a10b) |
19
- | Qwen3.6-35B-A3B | 0.248 | 不支持/未公布 | 1.485 | [官方页面](https://www.alibabacloud.com/help/en/model-studio/qwen3-6-35b-a3b) |
20
- | DeepSeek-V4-Flash | 0.3 | 0.006 | 1.2 | [官方页面](https://api-docs.deepseek.com/quick_start/pricing) |
21
- | GPT-5.6 Luna | 0.2 | 0.02 | 1.2 | [官方页面](https://developers.openai.com/api/docs/models/gpt-5.6-luna) |
22
-
23
- - **Claude-Opus-5**:Standard API; 5-minute cache writes. 1-hour writes cost $10/1M and are not used in the scenario.
24
- - **GPT-5.6 Terra**:Standard API, input <=272K per request; longer prompts are billed at 2x input and 1.5x output.
25
- - **GLM-5.3**:Standard USD API pricing; cached-input storage is temporarily free.
26
- - **Qwen3.5-122B-A10B**:China (Beijing), input <=128K; 128K–256K input/output rates are $0.287/$2.294. Context caching is unsupported.
27
- - **Qwen3.6-35B-A3B**:China (Beijing). Context caching is unsupported. Singapore input/output rates are $0.375/$2.25.
28
- - **DeepSeek-V4-Flash**:Peak rates. Original V4-Flash has retired; the legacy API name now serves V4.1-Flash. This is a successor-rate proxy, not the historical BF16 model bill. Off-peak rates are half.
29
- - **GPT-5.6 Luna**:Standard API, input <=272K per request; longer prompts are billed at 2x input and 1.5x output.
30
-
31
- ## 估算方法
32
-
33
- C = T / 1e6 × (0.9 × P_input + 0.1 × P_output)。T 使用当前导出的累计 token 总量,含失败尝试,不截断为 STS 的 50M 预算。所有配置统一假设 90% 输入、10% 输出、无缓存;这不是从日志恢复的实际用量比例。
34
-
35
- CSV 另外提供 5% / 20% 输出占比和 80% 输入缓存命中场景;它们是敏感性情景,不是成本上下界或置信区间。Qwen 官方不支持缓存,因此缓存场景留空。DeepSeek 统一采用当前 Flash 后继版本的 peak 单价;off-peak 为一半,且原 V4-Flash/BF16 的价格不可据此确认为相同。
36
-
37
- Claude 5 分钟缓存写入 $6.25/1M,1 小时 $10/1M;OpenAI 缓存写入按输入单价 1.25 倍。主情景未假设任何缓存读写。Qwen 使用北京区域,Qwen3.5 输入≤128K;OpenAI 输入≤272K。长上下文附加费、工具、税、订阅、加速卡与环境费用均未计入。
38
-
39
- 当前 UI bundle 把计费分类留空。上游 normalized_runs 虽有分项字段,但含估算和不同计数口径,例如部分 Claude 记录的 cached_input 大于 input;在完成分类口径核对前,不将它们简单相加或称为真实账单。T 本身保留补测源数据中的估算属性。没有把预估成本写回原始实验指标或改变 STS 排名。
40
-
41
- ## 配置估算(费用显示两位小数,CSV 保留计算精度)
42
-
43
- | 配置 | Cohort | 累计 token(M) | API 等价估算(USD) |
44
- | --- | --- | ---: | ---: |
45
- | Claude Code × Claude-Opus-5 | harness | 1,389.784 | 9,728.49 |
46
- | Claude Code × GPT-5.6 Terra | harness | 490.527 | 1,471.58 |
47
- | Claude Code × GLM-5.3 | harness | 582.736 | 990.65 |
48
- | Claude Code × Qwen3.5-122B-A10B | harness | 232.752 | 45.43 |
49
- | Claude Code × Qwen3.6-35B-A3B | harness | 324.953 | 120.79 |
50
- | Claude Code × DeepSeek-V4-Flash | harness | 820.864 | 320.14 |
51
- | Codex × GPT-5.6 Terra | harness | 2,387.283 | 7,161.85 |
52
- | Codex × GLM-5.3 | harness | 1,658.805 | 2,819.97 |
53
- | Codex × Qwen3.5-122B-A10B | harness | 926.983 | 180.95 |
54
- | Codex × Qwen3.6-35B-A3B | harness | 447.902 | 166.49 |
55
- | Codex × DeepSeek-V4-Flash | harness | 1,060.282 | 413.51 |
56
- | OpenCode × Claude-Opus-5 | harness | 247.159 | 1,730.11 |
57
- | OpenCode × GPT-5.6 Terra | harness | 51.899 | 155.70 |
58
- | OpenCode × GLM-5.3 | harness | 598.64 | 1,017.69 |
59
- | OpenCode × Qwen3.5-122B-A10B | harness | 439.008 | 85.69 |
60
- | OpenCode × Qwen3.6-35B-A3B | harness | 406.414 | 151.06 |
61
- | OpenCode × DeepSeek-V4-Flash | harness | 219.337 | 85.54 |
62
- | Default Cold × GPT-5.6 Luna | default | 6,488.778 | 1,946.63 |
63
- | Default Warm × GPT-5.6 Luna | default | 970.82 | 291.25 |
64
- | Default No Lessons × GPT-5.6 Luna | default | 1,881.229 | 564.37 |
65
- | Default Code Only × GPT-5.6 Luna | default | 1,469.735 | 440.92 |
66
-
67
- 只可比较相同 cohort 和覆盖范围。Default 模式有重复运行,不应将其累计费用与单次 Harness 数据当成严格受控对照。
68
-
69
- 数据:src/data/supplement_bundle.json;价格:src/data/Token_price.json;公式:src/lib/pricing.js。运行 node scripts/export-api-cost-estimates.mjs 可重建 CSV 和本文。
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
website/docs/shape-results-guide.zh.md DELETED
@@ -1,45 +0,0 @@
1
- # Shape 数据填写模板
2
-
3
- 这次只填写一个 CSV,不需要之前的多层 JSON、协议块或候选信息。
4
-
5
- ## 使用方法
6
-
7
- 1. 打开 `shape-results.csv`。已预填当前 21 个配置、353 个“配置×算子”的 entry_id、operator 和 run_id,与当前图表对应。
8
- 2. **每行填写一个 shape / case。** 同一算子有多少个 case,就复制该行多少次,保留前三列不变,逐行填写 case_id、shape、dtype。
9
- 3. 有逐 shape 性能数据时,再填两列耗时和 correctness;若暂时只采集 shape,耗时留空,correctness 保持 not_measured。
10
- 4. 只采集部分算子也可以:提交已填写的行即可,其他占位行可以删除。保持 CSV UTF-8 编码,不要合并单元格或改动表头。
11
-
12
- 也提供只有表头的 `shape-results.blank.csv`。`shape-results.example.csv` 只展示格式,里面所有尺寸、耗时与标识均为演示,不能作为真实结果提交。
13
-
14
- ## 九列分别填什么
15
-
16
- | 列名 | 填写方式 |
17
- | --- | --- |
18
- | entry_id | 已预填,保持不变;它标识 Harness×Model 配置,不能换成页面展示名。 |
19
- | operator | 已预填,保持不变。 |
20
- | run_id | 已预填,保持不变;对应当前图表使用的那个 run。若实际采集其他 run,请填写真实 run_id 并说明。 |
21
- | case_id | 环境中的 case ID;同一个配置、算子、run 内不能重复。若环境没有 ID,可依 manifest 顺序固定编号 case_001、case_002 等。 |
22
- | shape | 具体输入尺寸。单输入例如 x=[8,4096];多输入例如 x=[8,4096]; weight=[4096]。所有输入都列出;标量使用 []。case 若由额外参数区分,可追加 axis=-1; epsilon=1e-5 等。 |
23
- | dtype | 单输入例如 float16;多输入例如 x=bfloat16; weight=float16。使用环境实际 dtype。 |
24
- | t_base_us | 可选,当前 case 的基线耗时,纯数字,单位 µs。 |
25
- | t_best_us | 可选,当前 case 在选定候选内核上的耗时,纯数字,单位 µs。 |
26
- | correctness | pass / fail / error / not_measured。没有测正确性时保留 not_measured,不能填 pass。 |
27
-
28
- 例:
29
-
30
- ```text
31
- shape: x=[8,4096]; weight=[4096]
32
- dtype: x=float16; weight=float16
33
- ```
34
-
35
- CSV 中尺寸包含逗号,使用 Excel / Numbers 正常保存 CSV 即可自动加引号。若用脚本导出,请用 Python csv.DictWriter,不要手动用逗号拼接字符串。
36
-
37
- ## 注意
38
-
39
- - 耗时未知就留空,不填 0,不写“100us”等带单位文本;ms 转 µs 乘 1000。
40
- - 同一个 run 的不同 shape 应使用同一个选定候选内核,不要逐 shape 换成不同候选的最好值。
41
- - 失败、报错和未测 case 也保留对应行,以便展示完整 shape 分布。
42
- - 无需手填加速比:仅在 correctness=pass 且两项耗时有效时,页面计算 t_base_us / t_best_us。
43
- - 每个算子的 case 数来自环境的完整 case 清单,不是全站统计中的 951。
44
-
45
- 完成后把填写好的 CSV 发回即可接入现在的行内展开列表。当前 Imports 页面尚不直接支持此 CSV;无需将其转换为旧版 summary JSON。
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
website/docs/shape-results-kit.zip DELETED
@@ -1,3 +0,0 @@
1
- version https://git-lfs.github.com/spec/v1
2
- oid sha256:796fd0e5349d447644c2dac9a4a0db7e72822dab60d12a7f9ea9967a65da1883
3
- size 5227
 
 
 
 
website/evaluate/OPERATOR_EVAL_SPEC.md DELETED
@@ -1,467 +0,0 @@
1
- # Operator 算子评测体系设计
2
-
3
- > 本文是AscendGYM的评测方案,该评测集主要回答:"在最终峰值性能和token效率上,Agent 能在多大程度上改进主流工作负载下的AscendC的内核效率?"
4
-
5
- ---
6
-
7
- ## 1. 设计目标
8
-
9
- 这套评测体系需要回答2个工程问题:
10
-
11
- | 子问题 | 体系 | 回答 |
12
- |---|---|---|
13
- | Agent在多少Token成本下,能拿到多少**算子性能收益**?是否可靠、是否通用、需要多长时间? | **STS**(Speedup–Token Surface) | Agent的**效率与能力** |
14
- | 生成的算子能否带**模型token吞吐**提升?是否逼近**硬件理论性能上限**? | **PAOS**(Profile Agent Optimization Score) | 优化的**价值与质量** |
15
-
16
-
17
- 因此,评测系统需要输入以下数据:
18
-
19
- - 算子生成准确率
20
- - 算子生成任务的Token消耗数量
21
- - 算子生成任务总时长
22
- - 算子利用率(被调用的次数)
23
- - 算子性能收益
24
-
25
- ---
26
-
27
- ## 2. 数据源:必须来自SandBox的实际评测数据(本project归档至operator文档)
28
-
29
- 所有评测逻辑都必须从operator目录中读取原始产物,不能“靠估算”填充缺失数据。
30
-
31
- ### 2.1 必需输入
32
-
33
- | 文件 | 用途 | 关键字段 |
34
- | --- | --- | --- |
35
- | operator/<run_id>/result.json | 任务运行总览 | run_id, baseline_duration_ms, best_duration_ms, speedup, token_usage, llm_rounds |
36
- | operator/<run_id>/candidates.jsonl | 候选实现明细 | status, duration_ms, correctness_passed, cumulative_tokens, solution_elapsed_sec |
37
- | operator/<run_id>/metrics.json | 任务级聚合摘要 | STS_token, STS_time, official_ranking_eligible, run_validity |
38
- | operator/<run_id>/input_manifest.json | 输入快照和任务来源 | op_name, source_hash, files |
39
- | operator/<run_id>/profiling/E000x/response.json | 功能验证结果 | status, passed/failed verification |
40
-
41
- ### 2.2 可选输入
42
-
43
- | 文件 | 用途 |
44
- | --- | --- |
45
- | operator/<run_id>/recipe.json | PAOS 计算的工作负载描述 |
46
- | operator/recipes/*.json | 统一 recipe 配置 |
47
- | operator/<run_id>/profiling/*/report.md | profiling 诊断、硬件分析 |
48
-
49
- ### 2.3 数据来源约束
50
-
51
- - result.json 负责任务状态和总览信息;
52
- - candidates.jsonl 负责 candidate 层的细粒度性能数据;
53
- - recipe.json 负责模型级工作负载价值,不可缺失时隐式填 1;
54
- - `G` 是运行时场景量,必须由 benchmark 方指定:decode 场景 `G=batch_size`,prefill 场景 `G=batch_size×chunk_len`。缺少场景或场景参数时输出 `missing_runtime_G`,不静默补默认值。
55
-
56
- ---
57
-
58
- ## 3. 核心评测维度
59
-
60
- 评测体系分为 4 层:
61
-
62
- ### 3.1 正确性维度
63
-
64
- - TaskPassRate:至少一个候选通过功能测试的任务占比
65
- - SolutionPassRate:通过功能测试的候选解占候选总数比例
66
- - CorrectnessGate:任务级硬门槛,只有存在通过解才允许进入性能榜
67
-
68
- ### 3.2 过程效率维度(STS)
69
-
70
- - Speedup:已验证最佳实现相对 baseline 的性能放大倍数
71
- - VLR:Validated Latency Reduction,延迟降低率
72
- - AULS_token:在 token 预算下的 anytime 面积
73
- - AULS_time:在 wall-clock 预算下的 anytime 面积
74
- - TokensTo1% / TimeTo1%、TokensTo5% / TimeTo5%、TokensTo10% / TimeTo10%:首次达到 1%、5%、10% 收益所需 token 或时间
75
- - STS_token / STS_time:综合主分数
76
- - EBGS:跨任务几何平均加速比
77
- - EIR_1% / EIR_5% / EIR_10%:分别至少达到 1%、5%、10% 收益的任务比例
78
-
79
- ### 3.3 工作负载价值维度(PAOS)
80
-
81
- - PTEI:模型 step 吞吐增益
82
- - PKAI:token-time 加权有效算子加速
83
- - PAOS_base:工作负载价值基础综合分
84
-
85
- PAOS 在正确性门槛通过时启用。recipe 优先提供真实 workload 参数,缺失参数由默认值或 CLI 手工参数补齐,并在输出中显式标注来源。
86
-
87
- ### 3.4 可靠性维度
88
-
89
- - 多 run 达标率
90
- - 中位数/分布/置信区间
91
- - 单 run 只能作为点估计,不得声明稳定性结论
92
-
93
- ---
94
-
95
- ## 4. 指标清单与数据来源、公式
96
-
97
- 下面给出“每一个指标的数据源 + 计算公式”列表,方便开发者直接对照代码和原始文件。
98
-
99
- | 指标 | 维度 | 数据来源 | 计算公式 / 语义 |
100
- | --- | --- | --- | --- |
101
- | TaskPassRate | 正确性 | candidates.jsonl + result.json | 任务中是否至少存在一个 status=passed 或 correctness_passed=true。任务级占比。 |
102
- | SolutionPassRate | 正确性 | candidates.jsonl | 通过的候选数 / 完成的候选总数。 |
103
- | CorrectnessGate | 正确性 | candidates.jsonl | Gate_i = 1 if ∃ passed candidate else 0。未通过则不能进入性能榜。 |
104
- | Speedup | 性能 | result.json + candidates.jsonl | S_i = L_0 / L_best,其中 L_0 为 baseline_duration_ms,L_best 为已通过最优 candidate 的 duration_ms。 |
105
- | VLR | 性能 | result.json + candidates.jsonl | VLR_i = 1 - 1 / S_i。表示延迟降低比例。 |
106
- | AULS_token | STS | candidates.jsonl | 在 token 预算轴上,对 best-so-far 曲线做 log 空间归一化面积积分。 |
107
- | AULS_time | STS | candidates.jsonl | 在 wall-clock 轴上,对 best-so-far 曲线做 log 空间归一化面积积分。 |
108
- | TokensTo1% / TokensTo5% / TokensTo10% | STS | candidates.jsonl | 最早满足 VLR ≥ 1% / 5% / 10% 的 cumulative_tokens。若未达成,则记为 null/NA。 |
109
- | TimeTo1% / TimeTo5% / TimeTo10% | STS | candidates.jsonl | 最早满足 VLR ≥ 1% / 5% / 10% 的 solution_elapsed_sec。若未达成,则记为 null/NA。 |
110
- | EBGS | STS 汇总 | runs[].speedup_final | $\mathrm{EBGS}=\exp\left(\frac{1}{N}\sum_i\ln S_i\right)$,用于跨任务几何平均收益。 |
111
- | EIR_1% / EIR_5% / EIR_10% | STS 汇总 | runs[].sts.VLR | $\mathrm{EIR}_{\Delta}=\frac{1}{N}\sum_i\mathbf{1}[\mathrm{VLR}_i\ge\Delta]$,其中 $\Delta\in\{0.01,0.05,0.10\}$。 |
112
- | STS_token | STS | candidates.jsonl | $\mathrm{STS}_{\mathrm{token}}=100(0.5\,\mathrm{EndGain}+0.5\,\mathrm{AULS}_{\mathrm{token}})$。 |
113
- | STS_time | STS | candidates.jsonl | $\mathrm{STS}_{\mathrm{time}}=100(0.5\,\mathrm{EndGain}+0.5\,\mathrm{AULS}_{\mathrm{time}})$。 |
114
- | PTEI | PAOS | recipe.json | PTEI = T_step^0 / T_step^1,表示 step 吞吐增益。 |
115
- | PKAI | PAOS | recipe.json + candidate speedup | $\mathrm{PKAI}=\exp\left(\sum_i w_i\ln S_i\right)$,对权重较大的算子进行加权。 |
116
- | PAOS_base | PAOS | recipe.json + PTEI + PKAI | $\mathrm{PAOS}_{\mathrm{base}}=\sqrt{\mathrm{PKAI}\times\mathrm{PTEI}}$。 |
117
- | Reliability@Target | 可靠性 | 多 run 的结果 | 在 K 次独立 run 中,达到给定目标的比例。当前 operator 多为单 run,故仅可作为 partial。 |
118
-
119
- ### 4.1 正确性门槛公式
120
-
121
- 任务级硬门槛:
122
-
123
- $$
124
- \mathrm{Gate}_i =
125
- \begin{cases}
126
- 1, & \text{if any candidate passes},\\
127
- 0, & \text{otherwise}.
128
- \end{cases}
129
- $$
130
-
131
- 工程实现规则:
132
-
133
- - 只要某个 candidate 的 status == "passed" 或 correctness_passed == true,就允许进入性能统计。
134
- - 若所有 candidate 都失败,则任务状态为 failed 或 blocked,不能计入 speedup。
135
-
136
- ### 4.2 最终加速与 VLR 公式
137
-
138
- 其中,$L_0$ 为 `baseline_duration_ms`,$L_{\mathrm{best}}$ 为已通过候选中的最小 `duration_ms`。
139
-
140
- Speedup:
141
-
142
- $$
143
- S_i = \frac{L_0}{L_{\mathrm{best}}}
144
- $$
145
-
146
- VLR:
147
-
148
- $$
149
- \mathrm{VLR}_i = 1 - \frac{1}{S_i}
150
- = 1 - \frac{L_{\mathrm{best}}}{L_0}
151
- $$
152
-
153
- 这两个指标都来自结果层,而不是从代码中推断。
154
-
155
- ### 4.3 AULS 公式
156
-
157
- AULS 的核心思想:收益越早出现、且后续有持续保留,分值越高。
158
-
159
- 定义 best-so-far 曲线:
160
-
161
- $$
162
- L_{\mathrm{best}}(b)
163
- = \min\left\{L_j \mid j \text{ passed},\; \mathrm{axis}_j \le b\right\}
164
- $$
165
-
166
- 再定义:
167
-
168
- $$
169
- S(b) = \frac{L_0}{L_{\mathrm{best}}(b)}
170
- $$
171
-
172
- 归一化收益函数:
173
-
174
- $$
175
- q(S) = \frac{\log_2(S)}{1 + \log_2(S)}
176
- $$
177
-
178
- 最后采用 trapezoid-like area integration:
179
-
180
- $$
181
- \mathrm{AULS}
182
- = \sum_{k=1}^{K}
183
- (x_k - x_{k-1})
184
- \frac{q(S_{k-1}) + q(S_k)}{2}
185
- $$
186
-
187
- 这里的 x_k 是对数预算空间的归一化坐标。
188
-
189
- ### 4.4 STS 公式
190
-
191
- STS 结合“最终收益”和“过程效率”两部分:
192
-
193
- $$
194
- \mathrm{EndGain} = q\left(S_{B_{\max}}\right)
195
- $$
196
-
197
- $$
198
- \mathrm{STS}
199
- = 100\left(0.5\,\mathrm{EndGain} + 0.5\,\mathrm{AULS}\right)
200
- $$
201
-
202
- 分别在 token 和 time 轴上计算:
203
-
204
- - STS_token:token 预算轴
205
- - STS_time:wall-clock 轴
206
-
207
- ### 4.5 PAOS 公式
208
-
209
- 脚本优先使用 recipe 中的真实场景字段;`G` 按 decode/prefill 规则计算,或由 benchmark 通过 `--paos-g` 显式指定。`T_other_ms` 和 `calls_per_step` 仍支持默认值与命令行覆盖,但 `G` 不使用静默默认值。
210
-
211
- $$
212
- T_{\mathrm{step}} = T_{\mathrm{other}} + \sum_i n_i t_i
213
- $$
214
-
215
- $$
216
- \mathrm{PTEI} = \frac{T_{\mathrm{step}}^0}{T_{\mathrm{step}}^1}
217
- $$
218
-
219
- $$
220
- \mathrm{PKAI} = \exp\left(\sum_i w_i\ln S_i\right)
221
- $$
222
-
223
- $$
224
- \mathrm{PAOS}_{\mathrm{base}}
225
- = \sqrt{\mathrm{PKAI}\times\mathrm{PTEI}}
226
- $$
227
-
228
- 输出必须包含 `input_source`,区分 `recipe`、`cli` 和 `default`。默认值不是外部估算,但使用默认值时应在结果中标注,不能将其解读为完整 workload recipe。
229
-
230
- #### PAOS 输入变量规范
231
-
232
- | 变量 | 含义 | 单位/合法范围 | 输入要求 |
233
- | --- | --- | --- | --- |
234
- | `G` | 当前运行时场景每个 workload step 对应的总 token 数,用于计算 `eTPS=G/T_step` | token/step;`>0` | decode:`G=batch_size`;prefill:`G=batch_size×chunk_len`;也可由 `--paos-g` 显式指定 |
235
- | `runtime_scenario` | benchmark 当前运行时场景 | `decode` 或 `prefill` | 必须由 benchmark 方指定,可由 recipe 或 `--paos-scenario` 提供 |
236
- | `batch_size` | 每个 workload step 同时处理的样本数 | sample;`>0`,通常为整数 | decode/prefill 均需与实际场景一致,可由 `--paos-batch-size` 提供 |
237
- | `chunk_len` | prefill 当前 chunk 每个样本处理的 token 数 | token/sample;prefill 时 `>0` | 仅 prefill 使用,可由 recipe 或 `--paos-chunk-len` 提供 |
238
- | `calls_per_step` | 每 step 调用被优化算子的次数 `n_i` | 次/step;`>=0`,通常为整数 | 必须是 workload 调用次数,不是候选数、profiling 次数或 LLM 调用次数;缺失时默认 `1`,可由 `--paos-calls-per-step` 覆盖 |
239
- | `T_other_ms` | 不属于当前优化算子的其余 step 耗时 | ms/step;`>=0` | 不得包含 `n_i × t_i`;缺失时默认 `0`,可由 `--paos-t-other-ms` 覆盖 |
240
- | `core_count` | profiling active time 归一化所使用的 AI Core 数量 | 个;正整数 | 只能来自实际 profiling/硬件元数据;不要用 Block Num、Mix Block Num 或猜测的设备总核数代替;当前不参与 PAOS 简洁版主公式 |
241
-
242
- 输入优先级为:显式 `G`/真实 recipe 场景字段 > CLI 手工参数 > 默认值。`runtime_scenario`、`batch_size`、`chunk_len`、`calls_per_step`、`T_other_ms` 必须使用同一 workload 的 shape、batch、数据类型和 step 定义;没有有效场景量时不计算 PAOS。脚本输出的 `input_source` 是审计依据。
243
-
244
- ---
245
-
246
- ## 5. 自动化脚本要求
247
-
248
- ### 5.1 评测脚本入口
249
-
250
- 建议入口脚本:
251
-
252
- ```bash
253
- python evaluate/evaluate_operator_runs.py --root operator --paos-scenario decode --paos-batch-size 1
254
- ```
255
-
256
- ### 5.2 脚本职责
257
-
258
- 脚本必须完成以下 7 项:
259
-
260
- 1. 扫描 operator 目录中的所有 run 目录;
261
- 2. 识别 result.json 和 candidates.jsonl;
262
- 3. 解析正确性状态;
263
- 4. 计算 speedup、VLR、AULS、STS;
264
- 5. 汇总 suite 级指标;
265
- 6. 读取 recipe 或默认/手工输入,计算 PAOS 简洁版并输出各字段及输入来源;
266
- 7. 按 (agent × 驱动模型) 拆分结果,并根据身份自动推导文件名写入数据集目录
267
- (`slug(agent)-slug(model).json`),使产物可直接被看板读取,无需手工改名。
268
-
269
- ### 5.3 必须支持的输出结构
270
-
271
- ```json
272
- {
273
- "root": "operator",
274
- "n_runs": 3,
275
- "accuracy": {
276
- "TaskPassRate": 0.667,
277
- "GateRate": 0.333
278
- },
279
- "sts": {
280
- "EBGS": 1.006,
281
- "EIR_1%": 1.0,
282
- "EIR_5%": 0.0,
283
- "EIR_10%": 0.0,
284
- "mean_STS_token": 0.67,
285
- "mean_STS_time": 0.62,
286
- "median_TokensTo1%": 12000,
287
- "median_TimeTo1%": 38.5,
288
- "median_TokensTo5%": null,
289
- "median_TimeTo5%": null,
290
- "median_TokensTo10%": null,
291
- "median_TimeTo10%": null
292
- },
293
- "paos": {
294
- "status": "computed",
295
- "runtime_scenario": "decode",
296
- "G": 1,
297
- "batch_size": 1,
298
- "T_other_ms": 0,
299
- "calls_per_step": [1],
300
- "T_step_base_ms": 10.0,
301
- "T_step_agent_ms": 9.87,
302
- "eTPS_base": 0.1,
303
- "eTPS_agent": 0.1013,
304
- "S_i_lat": 1.013,
305
- "PTEI": 1.013,
306
- "PKAI": 1.013,
307
- "PAOS_base": 1.013,
308
- "input_source": {"recipe": "default", "G": "default", "T_other_ms": "default", "calls_per_step": "default"}
309
- },
310
- "runs": [
311
- {
312
- "task_id": "fused_gdn_gating-...",
313
- "status": "passed",
314
- "accuracy": {"CorrectnessGate": 1},
315
- "sts": {
316
- "Speedup": 1.013,
317
- "VLR": 0.013,
318
- "AULS_token": 0.0083,
319
- "AULS_time": 0.0064,
320
- "STS_token": 1.34,
321
- "STS_time": 1.25,
322
- "TokensTo1%": 12000,
323
- "TimeTo1%": 38.5,
324
- "TokensTo5%": null,
325
- "TimeTo5%": null,
326
- "TokensTo10%": null,
327
- "TimeTo10%": null
328
- },
329
- "paos": {
330
- "status": "computed",
331
- "runtime_scenario": "decode",
332
- "G": 1,
333
- "batch_size": 1,
334
- "T_other_ms": 0,
335
- "calls_per_step": [1],
336
- "T_step_base_ms": 10.0,
337
- "T_step_agent_ms": 9.87,
338
- "eTPS_base": 0.1,
339
- "eTPS_agent": 0.1013,
340
- "S_i_lat": 1.013,
341
- "PTEI": 1.013,
342
- "PKAI": 1.013,
343
- "PAOS_base": 1.013
344
- }
345
- }
346
- ]
347
- }
348
- ```
349
-
350
- ---
351
-
352
- ## 6. 运行判定规则与输出结构
353
-
354
- 脚本输出固定分为 suite 级 `accuracy`、`sts`、`paos` 和逐任务 `runs[]`,指标不平铺。run 级 `sts` 包含 `Speedup`、`VLR`、`AULS_token`、`AULS_time`、`STS_token`、`STS_time`,以及三档 `TokensTo1%/5%/10%`、`TimeTo1%/5%/10%`。suite 级 `sts` 额外包含 `EBGS`、三档 `EIR_1%/5%/10%` 和三档 `median_TokensTo`、`median_TimeTo`。
355
-
356
- 缺失或未达标统一输出 `null`,PAOS 通过 `status=unavailable` 和 `reason` 说明原因。汇总时仅使用对应指标非 `null` 的 run;失败任务不从 `runs[]` 删除。
357
-
358
- 推荐命令:
359
-
360
- ```bash
361
- # 默认:按 agent × 驱动模型自动拆分并命名,写入 src/data/eval_datasets_summary/
362
- python evaluate/evaluate_operator_runs.py --root operator --paos-scenario decode --paos-batch-size 1
363
-
364
- # 合并为单个 summary(可视化脚本的输入)
365
- python evaluate/evaluate_operator_runs.py --root operator --no-split --output operator/eval_summary.json
366
- python evaluate/evaluate_operator_runs.py --root operator --no-split --targets 0.01 0.05 0.10
367
- python evaluate/visualize_operator_metrics.py --input operator/eval_summary.json --output-dir operator/export
368
- ```
369
-
370
- 拆分文件名由 run 身份推导(`slug(agent)-slug(model).json`),不需要显式指定输出文件名;
371
- `--dry-run` 可先查看将要写入的文件清单,`--out-dir` 可改写入位置。
372
- 身份解析不出的 run 不会被丢弃,文件名退化为 `{agent}.json` 或 `unidentified.json`。
373
-
374
- PAOS 手工输入参数:
375
-
376
- ```bash
377
- python evaluate/evaluate_operator_runs.py --root operator --paos-scenario decode --paos-batch-size 1 --paos-t-other-ms 0 --paos-calls-per-step 1
378
- ```
379
-
380
- ### 6.1 任务判定
381
-
382
- - status == "passed" 或 correctness_passed == true:有效候选;
383
- - 若无有效通过候选:任务进入 failed 或 blocked 状态;
384
- - 不通过的 candidate 不能用作 speedup 计算;
385
- - 结果要保留失败记录,避免“看起来很成功”的误判。
386
-
387
- ### 6.2 单 run 与多 run
388
-
389
- - 单 run:只能计算点估计,不能宣称稳定性;
390
- - 多 run(>=3):可启动 Reliability 计算;
391
- - 当前 operator 样例主要为单 run,因此要明确标注为 partial,不可误报为可靠结论。
392
-
393
- ### 6.3 排名规则
394
-
395
- 推荐的排序逻辑:
396
-
397
- 1. CorrectnessGate
398
- 2. STS 主分数
399
- 3. 可靠性下界
400
- 4. 成本(token/time)
401
-
402
- 但排序逻辑不应写死进脚本,脚本主职责是产出原始指标,不做“高明”但不透明的排序决策。
403
-
404
- ---
405
-
406
- ## 7. 工程落地顺序
407
-
408
- ### P0:基础可用
409
-
410
- - 读取 result.json 和 candidates.jsonl;
411
- - 计算 correctness gate、speedup、VLR、AULS、STS;
412
- - 输出 run 级和 suite 级 JSON。
413
-
414
- ### P1:稳定性增强
415
-
416
- - 多 run 聚合;
417
- - 统计达标 run 数量;
418
- - 计算 reliability。
419
-
420
- ### P2:PAOS 扩展
421
-
422
- - 引入 recipe.json;
423
- - 提供 G、n_i、T_other;
424
- - 正确性门槛通过即可输出 PAOS 简洁版;使用默认参数时必须结合 `input_source` 解读,不能视作完整 workload 结论。
425
-
426
- ### P3:profiling 深化
427
-
428
- - 已从 `profiling/<evaluation_id>/profile_output/**/op_summary_*.csv` 读取 active time、Vector/Cube/MTE2 比例;多样本使用中位数;
429
- - 输出 `profiling.S_i_aicore`、`profiling.U_i`、`profiling.U_step` 及 baseline/agent active time;
430
- - `core_count` 若文件中没有可靠字段则保持缺失;运行时场景和 `G` 必须由 recipe 或 CLI 指定,不使用默认值;`n_i`、`T_other_ms` 仍可使用 recipe、CLI 或默认值并记录来源;
431
- - 在可用时继续扩展 roofline PPUI。
432
-
433
- ---
434
-
435
- ## 8. CSV 导出
436
-
437
- 评估结果通过 `visualize_operator_metrics.py` 导出为单一 CSV 文件,不再生成图表、图片或额外的可视化汇总 JSON:
438
-
439
- ```bash
440
- python evaluate/visualize_operator_metrics.py \
441
- --input operator/eval_summary.json \
442
- --output-dir operator/export
443
- ```
444
-
445
- 输出文件为 `operator/export/operator_metrics.csv`,包含任务状态、CorrectnessGate、Speedup、VLR、AULS、STS 以及 1%、5%、10% 三档 TokensTo/TimeTo 指标。缺失值保留为空,不用 0 替代。
446
-
447
- ---
448
-
449
- ## 9. 最终结论
450
-
451
- 这套评测体系的核心设计原则是:
452
-
453
- - 正确性永远先于性能;
454
- - 用真实 operator 数据来计算,不做隐式补值;
455
- - 既看单任务收益,也看过程效率;
456
- - 对 PAOS 采用“有 recipe 才计算”的严格门槛;
457
- - 缺失数据时明确降级,而不是输出假装完整的结果。
458
-
459
- 这使得它既保留了 metrics_master_optimized.md 的科学性,又具备工程落地性,能直接支撑开发者评审、性能对比和后续 benchmark 迭代
460
-
461
- Token消耗
462
- 时间戳
463
- 性能提升
464
- STS考察Token效率(100W Token得到多少性能提升,核算ROI)
465
- PAOS 考察 Coding 与Agent的组合,计算算子性能优化的上限与性能优化空间
466
-
467
- 优化空间:基于硬件计算算子理论计算性能上限
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
website/evaluate/Operator_eval.md DELETED
@@ -1,567 +0,0 @@
1
- # 算子生成能力评测方案
2
-
3
- 本文只定义基于算子真实轨迹下,可计算的评测指标,输出维度分别为:**准确率、STS、PAOS**。缺失字段统一输出 `null`,不得用 0 或估算值替代。
4
-
5
- ## 1. 最小数据采集
6
-
7
- ### 1.1 轨迹字段
8
-
9
- | 采集项 | 来源 | 用途 | 中文译名 |
10
- | --- | --- | --- | --- |
11
- | `run_id`、`success`、`operational_status`、`termination_reason` | `result.json` | 任务标识和运行状态 | 运行标识、成功状态、运行状态、终止原因 |
12
- | `baseline_duration_ms` | `result.json`;回退到 `candidates.jsonl` 中 `evaluation_kind=baseline` 的 `duration_ms` | 性能基线 `L0` | 基线耗时(毫秒) |
13
- | `status`、`correctness_passed`、`evaluation_kind` | `candidates.jsonl` | 正确性门槛 | 候选状态、正确性通过、评测类型 |
14
- | `duration_ms` | `candidates.jsonl` | 候选 kernel latency | 候选算子耗时(毫秒) |
15
- | `cumulative_tokens` | `candidates.jsonl` 或 `solutions.jsonl` | STS token 横轴 | 累计 token 数 |
16
- | `solution_elapsed_sec` | `candidates.jsonl` 或 `solutions.jsonl` | STS 时间横轴 | 解生成累计耗时(秒) |
17
- | `token_usage.total_tokens`、`llm_rounds`、`tool_calls` | `result.json` | 成本统计 | 总 token 数、LLM 轮数、工具调用数 |
18
- | `profiling.active_time_us`、`aicore_time_us`、`aiv_time_us`、`core_count`、`vec_ratio`、`cube_ratio`、`mte2_ratio` | 候选 `profiling` | PAOS 资源指标;缺失则为 `null` | 活跃时间、AI Core 时间、AI Vector 时间、核数及硬件利用率比例 |
19
-
20
- ### 1.2 PAOS 最小 recipe
21
-
22
- PAOS 使用简洁版,只需要一个 workload recipe。默认规则如下:
23
-
24
- ```json
25
- {
26
- "runtime_scenario": "decode",
27
- "batch_size": 1,
28
- "chunk_len": null,
29
- "T_other_ms": 0,
30
- "ops": [
31
- {
32
- "task_id": "<operator-run-name>",
33
- "calls_per_step": "<n_i>"
34
- }
35
- ]
36
- }
37
- ```
38
-
39
- - `G` 是运行时场景量,不是固定的模型常量:decode 场景 `G = batch_size`,prefill 场景 `G = batch_size × chunk_len`。
40
- - `runtime_scenario`、`batch_size` 和 prefill 的 `chunk_len` 必须由 benchmark 方按实际评测场景指定;不能用 Agent token 或候选累计 token 代替。
41
- - `calls_per_step=n_i` 优先来自真实 workload recipe;缺少时默认使用 `1`,并在 PAOS 输出的 `input_source` 中标明,可由 CLI 覆盖。
42
- - `T_other_ms` 表示未优化部分的 step 时间;缺失时使用 `0`,并标记为默认假设,不能把该结果解读为完整 workload 结论。
43
-
44
- ### 1.3 PAOS 输入变量规范
45
-
46
- | 变量 | 含义 | 单位/类型 | 规范输入 | 推荐来源与默认值 |
47
- | --- | --- | --- | --- | --- |
48
- | `G` | 运行时场景下每 Step 实际处理的 token 数,用于把 step 时间换算为 token 吞吐 | token/step;正数 | decode:`G = batch_size`;prefill:`G = batch_size × chunk_len` | 优先 benchmark 显式 `G` 或由场景字段计算;可用 `--paos-g` |
49
- | `runtime_scenario` | 当前 benchmark 的运行时场景 | `decode` 或 `prefill` | 必须由 benchmark 方指定,不能从 Agent token usage 推断 | recipe 的 `runtime_scenario`/`scenario` 或 `--paos-scenario` |
50
- | `batch_size` | 一个 workload step 同时处理的样本数 | sample;正数,通常为整数 | 应填写当前场景实际 batch,不要填写候选数或 profiling 样本数 | recipe 或 `--paos-batch-size` |
51
- | `chunk_len` | prefill 场景每个样本在当前 chunk 中处理的 token 数 | token/sample;prefill 时为正数 | decode 场景不需要;prefill 必须指定 | recipe 或 `--paos-chunk-len` |
52
- | `calls_per_step` | 当前算子在一个 workload step 中被调用的次数,即公式中的 `n_i` | 次/step;非负数,通常为整数 | 应填写真实调用次数;同一算子多次调用应填写总次数,不要填写候选数或 profiling 样本数 | 优先 `recipe.ops[].calls_per_step`/`n_i`,可用 `--paos-calls-per-step` 覆盖,默认 `1` |
53
- | `T_other_ms` | step 中除当前优化算子外的未优化耗时,包括其他算子和已知固定开销 | ms/step;非负数 | 必须与 `T_step`、`duration_ms` 使用相同的 step 和测量口径;未知时填 `0` 但必须保留默认来源标记 | 优先 `recipe.T_other_ms`,可用 `--paos-t-other-ms` 覆盖,默认 `0` |
54
- | `core_count` | profiling 时参与执行或用于 active time 归一化的 AI Core 数量 | 个;正整数 | 应填写 profiling 实际设备和采集口径对应的核数;不能填写 Block Num、Mix Block Num 或设备总核数的猜测值 | 优先 profiling/硬件元数据;当前简洁版主公式不使用,缺失保持 `null` |
55
-
56
- 输入注意事项:
57
-
58
- - `runtime_scenario`、`batch_size`、`chunk_len`、`calls_per_step`、`T_other_ms` 必须对应同一个 workload 场景,不能跨 batch、shape 或数据类型拼接。
59
- - benchmark 方必须明确提供场景;没有场景或有效 `G` 时,脚本输出 `missing_runtime_G`,不使用静默默认值。
60
- - `T_other_ms` 不能重复��含被优化算子的 `n_i × t_i`,否则会重复计算 step 时间。
61
- - `core_count` 不改变 `PAOS_base`、`PTEI` 或 `PKAI`;它只服务于 active time、`U_i` 等 profiling 资源指标的归一化。
62
- - 手工输入优先级高于 recipe;脚本会在 `paos.input_source` 中记录 `recipe`、`cli` 或 `default`。不确定时宁可使用默认值并标注,不要填入未经测量的估算值。
63
- - `t_i^0` 和 `t_i^1` 分别取 baseline 与最佳通过候选的 `duration_ms`,不从代码推断。
64
-
65
- ## 2. 记录筛选规则
66
-
67
- ### 2.1 有效 baseline
68
-
69
- baseline 必须同时满足:
70
-
71
- ```text
72
- evaluation_kind == "baseline"
73
- status == "passed"
74
- correctness_passed == true
75
- duration_ms > 0
76
- ```
77
-
78
- 优先使用 `result.baseline_duration_ms`,但只有它对应的 baseline 正确性有效时才打开性能门槛。没有有效 baseline 时,STS 和 PAOS 均为 `unavailable`。
79
-
80
- ### 2.2 有效候选
81
-
82
- 候选 `j` 必须满足:
83
-
84
- ```text
85
- evaluation_kind == "candidate"
86
- status == "passed"
87
- correctness_passed == true
88
- duration_ms > 0
89
- ```
90
-
91
- 失败候选计入候选总数和准确率分母,但不进入性能曲线。通过候选中最小的 `duration_ms` 为 `L_best`。
92
-
93
- ## 3. 准确率指标
94
-
95
- ### 3.1 Run 级指标
96
-
97
- | 指标 | 公式 | 中文译名 |
98
- | --- | --- | --- |
99
- | `TaskPass` | 存在有效 baseline 或有效候选时为 `1`,否则为 `0` | 任务通过 |
100
- | `CandidatePassRate` | `N_passed / N_completed`;分母为 0 时为 `null` | 候选通过率 |
101
- | `CorrectnessGate` | 有效 baseline 且 `N_passed > 0` 时为 `1`,否则为 `0` | 正确性门槛 |
102
- | `RunCompleted` | `operational_status == "completed"` 时为 `1`,否则为 `0` | 运行完成 |
103
- | `FailureReason` | 原样保留 `termination_reason` | 失败原因 |
104
-
105
- 只有 `CorrectnessGate=1` 的 run 才进入 STS 和 PAOS 性能汇总。
106
-
107
- ### 3.2 Suite 级指标
108
-
109
- 令 `N` 为 run 总数:
110
-
111
- $$
112
- TaskPassRate = \frac{\sum_i TaskPass_i}{N},\qquad
113
- GateRate = \frac{\sum_i CorrectnessGate_i}{N}
114
- $$
115
-
116
- ## 4. STS 指标(metrics_master.md 2.1)
117
-
118
- 令 `L0` 为 baseline latency,`L_best(B)` 为 token 或时间预算 `B` 内所有有效候选的 best-so-far latency:
119
-
120
- $$
121
- S_i(B)=\frac{L_0}{L_{\mathrm{best}}(B)},\qquad
122
- \mathrm{VLR}_i(B)=1-\frac{1}{S_i(B)}
123
- =1-\frac{L_{\mathrm{best}}(B)}{L_0}
124
- $$
125
-
126
- ### 4.1 最小 STS 输出
127
-
128
- | 指标 | 计算公式 | 数据来源 | 中文译名 |
129
- | --- | --- | --- | --- |
130
- | `Speedup` | $S_i=\frac{L_0}{L_{\mathrm{best}}}$ | baseline/candidate `duration_ms` | 加速比 |
131
- | `VLR` | $\mathrm{VLR}_i=1-\frac{1}{S_i}=1-\frac{L_{\mathrm{best}}}{L_0}$ | `Speedup` | 延迟降低率 |
132
- | `TokensTo1%` / `TokensTo5%` / `TokensTo10%` | $\min\{T_j\mid \mathrm{VLR}_j\ge\Delta\}$,其中 $\Delta\in\{0.01,0.05,0.10\}$ | `cumulative_tokens`、`duration_ms` | 达到目标收益所需 token 数 |
133
- | `TimeTo1%` / `TimeTo5%` / `TimeTo10%` | $\min\{W_j\mid \mathrm{VLR}_j\ge\Delta\}$,其中 $\Delta\in\{0.01,0.05,0.10\}$ | `solution_elapsed_sec`、`duration_ms` | 达到目标收益所需时间 |
134
- | `AULS_token` | 对数 token 轴上的 best-so-far 面积 | `cumulative_tokens` | token 轴加速曲线面积 |
135
- | `AULS_time` | 对数时间轴上的 best-so-far 面积 | `solution_elapsed_sec` | 时间轴加速曲线面积 |
136
- | `STS_token` | $100\left(0.5\log_2 S(B_{\max})+0.5\,\mathrm{AULS}_{\mathrm{token}}\right)$ | token 曲线 | token 效率综合分 |
137
- | `STS_time` | $100\left(0.5\log_2 S(B_{\max})+0.5\,\mathrm{AULS}_{\mathrm{time}}\right)$ | 时间曲线 | 时间效率综合分 |
138
- | `S_equiv` | $2^{\mathrm{STS}_{\mathrm{token}}/100}$ | `STS_token` | 等效加速比 |
139
-
140
- 默认 `Δ ∈ {0.01, 0.05, 0.10}`,对应输出字段 `TokensTo1%`、`TokensTo5%`、`TokensTo10%` 及三档 `TimeTo`。未达标时字段为 `null`。可通过 `--targets` 覆盖默认阈值,阈值以小数传入。
141
-
142
- ### 4.2 AULS 公式
143
-
144
- 按横轴递增排列有效点,加入 baseline 起点,令:
145
-
146
- $$
147
- g_k=\log_2\left(S_i(B_k)\right),\qquad
148
- \gamma_k=\frac{\ln(B_k/B_{k-1})}{\ln(B_K/B_0)}
149
- $$
150
-
151
- $$
152
- AULS_i=\sum_{k=1}^{K}\gamma_k\frac{g_{k-1}+g_k}{2}
153
- $$
154
-
155
- 其中横轴分别取 `cumulative_tokens` 和 `solution_elapsed_sec`。横轴缺失、非正数或有效点不足两个时,AULS 为 `null`。
156
-
157
- ### 4.3 Suite 级 STS 汇总
158
-
159
- 只对通过 `CorrectnessGate` 且对应指标非空的 run 汇总:
160
-
161
- $$
162
- \mathrm{EBGS}=\exp\left(\frac{1}{M}\sum_{i=1}^{M}\ln\left(S_i(B_{\max})\right)\right)
163
- $$
164
-
165
- $$
166
- \mathrm{EIR}_{\Delta}=\frac{1}{M}\sum_{i=1}^{M}\mathbf{1}\left[\mathrm{VLR}_i(B_{\max})\ge\Delta\right]
167
- $$
168
-
169
- | 指标 | 计算公式 | 中文译名 |
170
- | --- | --- | --- |
171
- | `EBGS` | `exp(mean(log(S_i(Bmax))))` | 专家基线几何平均加速比 |
172
- | `EIR_Δ` | `mean(1[VLR_i(Bmax) >= Δ])` | 达到目标收益的任务比例 |
173
- | `mean(STS_token)` | 通过门槛 run 的 `STS_token` 算术平均 | 平均 token 效率综合分 |
174
- | `mean(STS_time)` | 通过门槛 run 的 `STS_time` 算术平均 | 平��时间效率综合分 |
175
- | `median_TokensTo1%` / `median_TokensTo5%` / `median_TokensTo10%` | 对应达标 run 的 `TokensTo` 中位数 | 各档达标 token 数中位数 |
176
- | `median_TimeTo1%` / `median_TimeTo5%` / `median_TimeTo10%` | 对应达标 run 的 `TimeTo` 中位数 | 各档达标时间中位数 |
177
-
178
- ## 5. PAOS 简洁版指标(metrics_master.md 2.2)
179
-
180
- PAOS 在 baseline 和最佳候选均通过、且 benchmark 提供有效
181
-
182
-
183
-
184
-
185
-
186
- 运行时场景量 `G` 时计算。脚本优先读取 recipe;`T_other_ms` 和 `calls_per_step` 仍可使用默认值,但 `G` 必须按 decode/prefill 场景由 benchmark 指定或计算,并在 `input_source` 中标明来源。
187
-
188
- ### 5.1 单算子 step 合成
189
-
190
- 对 recipe 中每个算子 `i`:
191
-
192
- $$
193
- t_i^0=\text{baseline duration},\qquad
194
- t_i^1=\text{best passed duration},\qquad
195
- S_i^{lat}=\frac{t_i^0}{t_i^1}
196
- $$
197
-
198
- 令 `n_i=calls_per_step`,`T_other` 为 recipe 中的未优化 step 时间:
199
-
200
- $$
201
- T_{step}^0=T_{other}+\sum_i n_i t_i^0,
202
- \qquad
203
- T_{step}^1=T_{other}+\sum_i n_i t_i^1
204
- $$
205
-
206
- $$
207
- PTEI=eTPSGain=\frac{T_{step}^0}{T_{step}^1},
208
- \qquad eTPS^r=\frac{G}{T_{step}^r}
209
- $$
210
-
211
- `eTPS` 的单位为 token/ms;如需 token/s,乘以 `1000`。`G` 是当前 decode/prefill 运行时场景量。
212
-
213
- ### 5.2 PKAI 与 PAOS
214
-
215
- 以 baseline token-time 为权重:
216
-
217
- $$
218
- w_i=\frac{n_i t_i^0}{T_{step}^0},\qquad
219
- PKAI=\exp\left(\sum_i w_i\ln(S_i^{lat})\right)
220
- $$
221
-
222
- PAOS 简洁版:
223
-
224
- $$
225
- PAOS_{base}=\sqrt{PKAI\times PTEI}
226
- $$
227
-
228
- | 指标 | 计算公式 | 中文译名 |
229
- | --- | --- | --- |
230
- | `S_i^lat` | `t_i^0 / t_i^1` | 算子延迟加速比 |
231
- | `T_step^0` | `T_other + Σ(n_i × t_i^0)` | 优化前 step 总耗时 |
232
- | `T_step^1` | `T_other + Σ(n_i × t_i^1)` | 优化后 step 总耗时 |
233
- | `eTPS^0` | `G / T_step^0` | 优化前估算 token 吞吐 |
234
- | `eTPS^1` | `G / T_step^1` | 优化后估算 token 吞吐 |
235
- | `PTEI` | `T_step^0 / T_step^1` | token 效率提升指数 |
236
- | `PKAI` | `exp(Σ(w_i × ln(S_i^lat)))` | 加权算子加速指数 |
237
- | `PAOS_base` | `sqrt(PKAI × PTEI)` | PAOS 简洁版综合分 |
238
-
239
- 对于单算子 recipe 且 `T_other_ms=0`,有:
240
-
241
- $$
242
- PKAI=S_i^{lat},\qquad PTEI=S_i^{lat},\qquad PAOS_{base}=S_i^{lat}
243
- $$
244
-
245
- ### 5.3 PAOS 资源指标
246
-
247
- 脚本会按候选的 `evaluation_id` 自动读取 `profiling/<evaluation_id>/profile_output/**/op_summary_*.csv`。当前文件可提供 `Task Duration(us)`、`aicore_time(us)`、`aiv_time(us)`、`aiv_vec_ratio`、`aic_mte2_ratio` 和 `cube_utilization(%)`。多条 profiling 样本使用中位数聚合。若字段或 CSV 不存在,则对应值为 `null`。
248
-
249
- 可读取并输出的资源指标:
250
-
251
- | 指标 | 公式 | 中文译名 |
252
- | --- | --- | --- |
253
- | `profiling.baseline.ActiveTimeUs` / `profiling.agent.ActiveTimeUs` | CSV 的 `aicore_time(us)`,无值时回退 `aiv_time(us)` | baseline/优化后 AI 活跃时间(微秒) |
254
- | `profiling.S_i_aicore` | `baseline.ActiveTimeUs / agent.ActiveTimeUs` | AI Core 活跃时间加速比 |
255
- | `profiling.U_i` | `agent.ActiveTimeUs / (CSV Task Duration(us) × 1000)` | 算子 AI 活跃利用率 |
256
- | `profiling.U_step` | `Σ(n_i × agent.ActiveTimeUs_i) / T_step_agent` | step 级 AI 活跃利用率 |
257
- | `profiling.VecRatio/CubeRatio/MTE2Ratio` | 读取 `aiv_vec_ratio`、`cube_utilization(%)`、`aic_mte2_ratio`/`aiv_mte2_ratio` | Vector/Cube/MTE2 利用率比例 |
258
-
259
- 当前轨迹的 `candidates.jsonl` 中硬件字段为 `null`,但 profiling CSV 可以提供上述资源指标;`core_count`、运行时场景和 `G` 当前仍无法从文件可靠获得,必须由 benchmark recipe/CLI 指定;`n_i`、`T_other_ms` 仍可按 recipe/CLI/默认值策略处理。
260
-
261
- ### 5.4 PAOS 缺失规则
262
-
263
- | 情况 | PAOS 输出 | 中文译名 |
264
- | --- | --- | --- |
265
- | 缺场景或 `G` | `status=unavailable`,原因 `missing_runtime_G` | benchmark 未提供运行时场景量 |
266
- | 缺 `calls_per_step` | 使用 `calls_per_step=1`,并标注来源 | 缺少每 step 调用次数 |
267
- | 缺有效 baseline 或通过候选 | `status=unavailable`,原因 `correctness_gate_failed` | 未通过正确性门槛 |
268
- | 缺 `T_other_ms` | 使用 `T_other_ms=0`,并标注来源 | 缺少未优化 step 耗时,结果为默认假设 |
269
- | baseline 和候选完整 | 输出 `T_step^0/T_step^1`、`eTPS^0/eTPS^1`、`S_i^lat`、`PTEI`、`PKAI`、`PAOS_base` | 可计算 PAOS 简洁版 |
270
-
271
- ## 6. 推荐输出结构
272
-
273
- 脚本输出分为 suite 级 `accuracy`、`cost`、`sts`、`paos`、`catalog` 和 run 级 `runs[]`。所有不可计算项使用 `null` 或 `status=unavailable`,不使用 0 代替。每个 run 和 suite 都按大类组织,不把指标平铺混在一起。
274
-
275
- 输出同时面向两类消费者:
276
-
277
- - **数值消费者**(看板、报表):直接读取 `accuracy` / `cost` / `sts` / `paos` / `durations`,不需要自己再算聚合。
278
- - **展示消费者**(前端 UI):读取 `catalog` 拿到指标的中文名、单位、方向与友好定义,避免前端自行维护一套可能漂���的文案。
279
-
280
- ### 6.1 cost 的计费明细
281
-
282
- `cost` 同时给出「消耗」与「可计费」两类信息,后者是 API 价格换算的输入:
283
-
284
- | 字段 | 含义 |
285
- | --- | --- |
286
- | `InputTokens` | **未命中缓存**的输入 token(`uncached_input_tokens`,缺失回退 `input_tokens`) |
287
- | `CachedInputTokens` | 命中缓存的输入 token(`cache_read_tokens`) |
288
- | `CacheWriteTokens` | 写入缓存的 token(`cache_write_tokens`) |
289
- | `OutputTokens` | 输出 token(`output_tokens`) |
290
- | `UsageAvailable` | 原始 `token_usage.available`,用于区分“真的是 0”与“未采集” |
291
-
292
- suite 级对应 `TotalInputTokens` / `TotalCachedInputTokens` / `TotalCacheWriteTokens` / `TotalOutputTokens`。
293
-
294
- 计费在下游完成,脚本**不预设任何价格**:
295
-
296
- ```text
297
- cost = InputTokens/1e6 · input_price
298
- + CachedInputTokens/1e6 · cached_input_price
299
- + CacheWriteTokens/1e6 · cache_write_price
300
- + OutputTokens/1e6 · output_price
301
- ```
302
-
303
- 价格来源为使用者导入的价格表(前端 `src/lib/pricing.js` 或构建期 `scripts/pricing.json`)。
304
- 未配置价格的模型,`api_cost_usd` 保持 `null` 并给出 `api_cost_usd_reason`,不得用 0 或估算值替代。
305
-
306
- ### 6.2 catalog 的展示契约
307
-
308
- `catalog` 携带展示层所需的全部元数据,前端据此渲染标签与释义:
309
-
310
- ```json
311
- {
312
- "catalog": {
313
- "version": "1.0",
314
- "categories": [{ "key": "sts", "label": "STS · token 效率轴", "subtitle": "Speedup–Token Surface" }],
315
- "metrics": {
316
- "mean_STS_token": { "label": "STS Token", "short": "STS", "unit": "", "direction": "up", "category": "sts",
317
- "definition": "主榜分:100·[0.5·log₂S(B_max) + 0.5·AULS_token]。" }
318
- }
319
- }
320
- }
321
- ```
322
-
323
- `direction` 取 `up` / `down` / `context`;`context` 表示仅作诊断、不参与排序。
324
-
325
- JSON 结构如下:
326
-
327
- ```json
328
- {
329
- "root": "operator",
330
- "n_runs": 3,
331
- "accuracy": {
332
- "TaskPassRate": 0.667,
333
- "GateRate": 0.333,
334
- "RunCompletedRate": 0.333
335
- },
336
- "cost": {
337
- "TotalTokens": 3716205.0,
338
- "median_TotalTokens": 942649.0,
339
- "TotalWallClockSec": 6876.33,
340
- "median_WallClockSec": 1659.63,
341
- "mean_LLMRounds": 19.0,
342
- "median_LLMRounds": 12.0,
343
- "TotalToolCalls": 55.0,
344
- "median_ToolCalls": 12.0,
345
- "TotalInputTokens": 3160765.0,
346
- "TotalCachedInputTokens": 0.0,
347
- "TotalCacheWriteTokens": 0.0,
348
- "TotalOutputTokens": 79129.0
349
- },
350
- "sts": {
351
- "EBGS": 1.013,
352
- "mean_Speedup": 1.013,
353
- "median_Speedup": 1.013,
354
- "max_Speedup": 1.013,
355
- "mean_VLR": 0.013,
356
- "mean_AULS_token": 0.0114,
357
- "mean_AULS_time": 0.009,
358
- "mean_STS_token": 1.514,
359
- "mean_STS_time": 1.396,
360
- "EIR_1%": 1.0,
361
- "EIR_5%": 0.0,
362
- "EIR_10%": 0.0,
363
- "median_TokensTo1%": 1021280.0,
364
- "median_TimeTo1%": 2451.44,
365
- "median_TokensTo5%": null,
366
- "median_TimeTo5%": null,
367
- "median_TokensTo10%": null,
368
- "median_TimeTo10%": null
369
- },
370
- "paos": {
371
- "status": "computed",
372
- "computed_runs": 1,
373
- "runtime_scenario": "decode",
374
- "G": 1,
375
- "batch_size": 1,
376
- "chunk_len": null,
377
- "T_other_ms": 0,
378
- "calls_per_step": [1],
379
- "input_source": {
380
- "recipe": "default",
381
- "G": "computed",
382
- "runtime_scenario": "cli",
383
- "batch_size": "cli",
384
- "chunk_len": "missing",
385
- "T_other_ms": "default",
386
- "calls_per_step": "default"
387
- },
388
- "mean_PAOS_base": 1.0132,
389
- "mean_PTEI": 1.0132,
390
- "mean_PKAI": 1.0132,
391
- "mean_eTPS_base": 120.18,
392
- "mean_eTPS_agent": 121.76,
393
- "mean_S_i_aicore": 0.9312,
394
- "mean_U_i": 0.00058,
395
- "mean_U_step": 0.4727,
396
- "mean_VecRatio": 0.118,
397
- "mean_CubeRatio": 0.0,
398
- "mean_MTE2Ratio": 0.0
399
- },
400
- "catalog": {
401
- "version": "1.0",
402
- "categories": [
403
- { "key": "accuracy", "label": "正确性", "subtitle": "Correctness" },
404
- { "key": "cost", "label": "成本", "subtitle": "Cost & effort" },
405
- { "key": "sts", "label": "STS · token 效率轴", "subtitle": "Speedup–Token Surface" },
406
- { "key": "paos", "label": "PAOS · 价值轴", "subtitle": "Profile Agent Optimization Score" },
407
- { "key": "durations", "label": "延迟", "subtitle": "Latency" }
408
- ],
409
- "metrics": {
410
- "PAOS_base": {
411
- "label": "PAOS",
412
- "short": "PAOS",
413
- "unit": "×",
414
- "direction": "up",
415
- "category": "paos",
416
- "definition": "价值轴主分 √(PKAI × PTEI):既奖励算得快,也奖励优化在关键路径上。"
417
- }
418
- }
419
- },
420
- "runs": [
421
- {
422
- "task_id": "fused_gdn_gating-agent-20260918-185348",
423
- "status": "passed",
424
- "accuracy": {
425
- "TaskPass": 1,
426
- "CandidatePassRate": 1.0,
427
- "CorrectnessGate": 1,
428
- "RunCompleted": 1,
429
- "FailureReason": "agent_finished",
430
- "CandidatesTotal": 12,
431
- "CandidatesPassed": 12,
432
- "BaselineEvidence": "candidate_trace"
433
- },
434
- "cost": {
435
- "TotalTokens": 2297245.0,
436
- "TotalWallClockSec": 4243.54,
437
- "SearchWallClockSec": 4013.97,
438
- "LLMRounds": 33.0,
439
- "ToolCalls": 32.0
440
- },
441
- "durations": {
442
- "BaselineMs": 0.00832108,
443
- "BestMs": 0.00821284,
444
- "Speedup": 1.0132,
445
- "BaselineEvaluationId": "E0000",
446
- "BestEvaluationId": "E0007",
447
- "BaselineEvidence": "candidate_trace"
448
- },
449
- "sts": {
450
- "status": "computed",
451
- "Speedup": 1.0132,
452
- "VLR": 0.013,
453
- "AULS_token": 0.0114,
454
- "AULS_time": 0.009,
455
- "STS_token": 1.514,
456
- "STS_time": 1.396,
457
- "S_equiv": 1.0106,
458
- "TokensTo1%": 1021280.0,
459
- "TimeTo1%": 2451.44,
460
- "TokensTo5%": null,
461
- "TimeTo5%": null,
462
- "TokensTo10%": null,
463
- "TimeTo10%": null,
464
- "VerificationPassRate": 1.0,
465
- "VerificationMedianMs": 0.0084,
466
- "VerificationCV": 0.0122,
467
- "SpeedupConfirmed": 0.9851,
468
- "VLRConfirmed": -0.0151
469
- },
470
- "paos": {
471
- "status": "computed",
472
- "G": 1.0,
473
- "runtime_scenario": "decode",
474
- "batch_size": 1.0,
475
- "chunk_len": null,
476
- "T_other_ms": 0.0,
477
- "calls_per_step": [1.0],
478
- "T_step_base_ms": 0.0083,
479
- "T_step_agent_ms": 0.0082,
480
- "eTPS_base": 120.18,
481
- "eTPS_agent": 121.76,
482
- "S_i_lat": 1.0132,
483
- "PTEI": 1.0132,
484
- "PKAI": 1.0132,
485
- "PAOS_base": 1.0132,
486
- "profiling": {
487
- "S_i_aicore": 0.9312,
488
- "U_i": 0.00058,
489
- "U_step": 0.4727,
490
- "VecRatio": 0.118,
491
- "CubeRatio": 0.0,
492
- "MTE2Ratio": 0.0
493
- }
494
- },
495
- "anytime": {
496
- "BaselineMs": 0.00832108,
497
- "token": [{ "Tokens": 1021280.0, "BestMs": 0.00821284, "Speedup": 1.0132 }],
498
- "time": [{ "Seconds": 2451.44, "BestMs": 0.00821284, "Speedup": 1.0132 }]
499
- }
500
- }
501
- ]
502
- }
503
- ```
504
-
505
- 下面逐块说明新增字段的用途:
506
-
507
- | 块 | 字段 | 用途 |
508
- | --- | --- | --- |
509
- | suite `cost` | `TotalTokens`、`TotalWallClockSec`、`mean/median_LLMRounds`、`TotalToolCalls`、`median_ToolCalls` | 成本维度:token 量、wall-clock、轮数、工具调用,供 UI 直接展示而无需自行求和 |
510
- | suite `sts` | `mean/median/max_Speedup` | 加速比聚合,供 UI 展示 `S_equiv` 与延迟摘要 |
511
- | suite `paos` | `runtime_scenario`、`G`、`batch_size`、`chunk_len`、`T_other_ms`、`calls_per_step`、`input_source` | 场景可信度:标注 PAOS 是在哪个运行时场景下算出来的,以及哪些参数来自 recipe / CLI / 默认假设 |
512
- | run `durations` | `BaselineMs`、`BestMs`、`Speedup`、`BaselineEvaluationId`、`BestEvaluationId` | 原始延迟与评测证据 ID,让延迟类展示不再需要反推 |
513
- | run `anytime` | `token[]`、`time[]`(含 `Tokens`/`Seconds`/`BestMs`/`Speedup`) | 真实的 best-so-far 轨迹点,供 anytime 曲线直接绘制,不需要插值 |
514
- | `catalog` | `categories`、`metrics{label,short,unit,direction,category,definition}` | 指标展示契约:中文名、单位、排序方向与友好定义,前后端共用一份文案 |
515
-
516
- suite `paos` 中的场景字段取自第一条 PAOS 可计算的 run;当所有 run 都不可计算时这些字段为 `null`。`catalog` 与数值无关,仅描述指标语义,可被前端直接消费。
517
-
518
- `sts` 包含三档 `EIR`、三档 `median_TokensTo` 和 `median_TimeTo`,档位由 `--targets` 决定;改变阈值后 `EIR_*` / `TokensTo*` / `TimeTo*` / `catalog` 之外的键名会同步变化。所有汇总分母只统计对应指标非 `null` 的 run;单 run 只能解释为点估计,不能据此声明可靠性。
519
-
520
- ### 6.1 运行命令
521
-
522
- 在项目根目录执行:
523
-
524
- ```bash
525
- python evaluate/evaluate_operator_runs.py --root operator --output operator/eval_summary.json
526
- ```
527
-
528
- **默认行为(按身份拆分)**:不给 `--no-split` 时,脚本会把 run 按 **agent × 驱动模型**拆成多个文件,
529
- 文件名由 run 里读到的身份自动推导(`slug(agent)-slug(model).json`,如 `opencode-qwen3-5-122b.json`),
530
- 写入看板数据集目录 `src/data/eval_datasets_summary/`:
531
-
532
- ```bash
533
- # 采集 + 拆分 + 自动命名 + 落到看板目录
534
- python evaluate/evaluate_operator_runs.py --root operator --paos-scenario decode --paos-batch-size 1
535
-
536
- # 只列出将写入的文件名
537
- python evaluate/evaluate_operator_runs.py --root operator --dry-run
538
-
539
- # 写到别处
540
- python evaluate/evaluate_operator_runs.py --root operator --out-dir operator/export
541
- ```
542
-
543
- 拆分**不丢 run**:身份解析不出的 run 仍会写入(文件名退化为 `{agent}.json` 或 `unidentified.json`)。
544
- 要回到「单个合并 summary」(`visualize_operator_metrics.py` 的输入),用 `--no-split`;
545
- 省略 `--output` 时直接打印 JSON;使用自定义目标时传入小数阈��:
546
-
547
- ```bash
548
- python evaluate/evaluate_operator_runs.py --root operator --targets 0.01 0.05 0.10
549
- ```
550
-
551
- 手工指定 decode 场景 PAOS 输入:
552
-
553
- ```bash
554
- python evaluate/evaluate_operator_runs.py --root operator --paos-scenario decode --paos-batch-size 1 --paos-t-other-ms 0 --paos-calls-per-step 1
555
- ```
556
-
557
- 手工指定 prefill 场景:
558
-
559
- ```bash
560
- python evaluate/evaluate_operator_runs.py --root operator --paos-scenario prefill --paos-batch-size 1 --paos-chunk-len 512
561
- ```
562
-
563
- 导出与 `evaluate_operator_runs.py` run 级输出维度对齐的 CSV。CSV 按一个 run 一行,覆盖 `accuracy`、`cost`、完整 `sts`、`paos` 和 profiling 字段;列表或对象字段以 JSON 字符串保存,缺失值保持为空:
564
-
565
- ```bash
566
- python evaluate/visualize_operator_metrics.py --input operator/eval_summary.json --output-dir operator/export
567
- ```
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
website/evaluate/evaluate_operator_runs.py DELETED
The diff for this file is too large to render. See raw diff
 
website/evaluate/visualize_operator_metrics.py DELETED
@@ -1,136 +0,0 @@
1
- #!/usr/bin/env python3
2
- """Export operator evaluation summary into a CSV file.
3
-
4
- Usage:
5
- python evaluate/visualize_operator_metrics.py --input operator/eval_summary.json --output-dir operator/export
6
- """
7
-
8
- from __future__ import annotations
9
-
10
- import argparse
11
- import csv
12
- import json
13
- from pathlib import Path
14
- from typing import Any, Dict, List
15
-
16
- def load_json(path: Path) -> Dict[str, Any]:
17
- with path.open("r", encoding="utf-8") as fh:
18
- return json.load(fh)
19
-
20
-
21
- def compute_summary_rows(summary: Dict[str, Any]) -> List[Dict[str, Any]]:
22
- rows: List[Dict[str, Any]] = []
23
- for run in summary.get("runs", []):
24
- accuracy = run.get("accuracy", {})
25
- cost = run.get("cost", {})
26
- sts = run.get("sts", {})
27
- paos = run.get("paos", {})
28
- profiling = paos.get("profiling", {})
29
- baseline_profile = profiling.get("baseline", {})
30
- agent_profile = profiling.get("agent", {})
31
- rows.append({
32
- "task_id": run.get("task_id"),
33
- "status": run.get("status", "unknown"),
34
- "TaskPass": accuracy.get("TaskPass"),
35
- "CandidatePassRate": accuracy.get("CandidatePassRate"),
36
- "CorrectnessGate": accuracy.get("CorrectnessGate"),
37
- "RunCompleted": accuracy.get("RunCompleted"),
38
- "FailureReason": accuracy.get("FailureReason"),
39
- "CandidatesTotal": accuracy.get("CandidatesTotal"),
40
- "CandidatesPassed": accuracy.get("CandidatesPassed"),
41
- "BaselineEvidence": accuracy.get("BaselineEvidence"),
42
- "TotalTokens": cost.get("TotalTokens"),
43
- "TotalWallClockSec": cost.get("TotalWallClockSec"),
44
- "SearchWallClockSec": cost.get("SearchWallClockSec"),
45
- "LLMRounds": cost.get("LLMRounds"),
46
- "ToolCalls": cost.get("ToolCalls"),
47
- "STS_status": sts.get("status"),
48
- "STS_reason": sts.get("reason"),
49
- "Speedup": sts.get("Speedup"),
50
- "VLR": sts.get("VLR"),
51
- "AULS_token": sts.get("AULS_token"),
52
- "AULS_time": sts.get("AULS_time"),
53
- "SpeedupSearch": sts.get("SpeedupSearch"),
54
- "VLRSearch": sts.get("VLRSearch"),
55
- "STS_token": sts.get("STS_token"),
56
- "STS_time": sts.get("STS_time"),
57
- "S_equiv": sts.get("S_equiv"),
58
- "VerificationPassRate": sts.get("VerificationPassRate"),
59
- "VerificationMedianMs": sts.get("VerificationMedianMs"),
60
- "VerificationCV": sts.get("VerificationCV"),
61
- "SpeedupConfirmed": sts.get("SpeedupConfirmed"),
62
- "VLRConfirmed": sts.get("VLRConfirmed"),
63
- "TokensTo1%": sts.get("TokensTo1%"),
64
- "TimeTo1%": sts.get("TimeTo1%"),
65
- "TokensTo5%": sts.get("TokensTo5%"),
66
- "TimeTo5%": sts.get("TimeTo5%"),
67
- "TokensTo10%": sts.get("TokensTo10%"),
68
- "TimeTo10%": sts.get("TimeTo10%"),
69
- "PAOS_status": paos.get("status"),
70
- "PAOS_reason": paos.get("reason"),
71
- "G": paos.get("G"),
72
- "runtime_scenario": paos.get("runtime_scenario"),
73
- "batch_size": paos.get("batch_size"),
74
- "chunk_len": paos.get("chunk_len"),
75
- "T_other_ms": paos.get("T_other_ms"),
76
- "calls_per_step": paos.get("calls_per_step"),
77
- "T_step_base_ms": paos.get("T_step_base_ms"),
78
- "T_step_agent_ms": paos.get("T_step_agent_ms"),
79
- "eTPS_base": paos.get("eTPS_base"),
80
- "eTPS_agent": paos.get("eTPS_agent"),
81
- "S_i_lat": paos.get("S_i_lat"),
82
- "PTEI": paos.get("PTEI"),
83
- "PKAI": paos.get("PKAI"),
84
- "PAOS_base": paos.get("PAOS_base"),
85
- "PAOS_recipe": paos.get("recipe"),
86
- "PAOS_input_source": paos.get("input_source"),
87
- "profiling_baseline_status": baseline_profile.get("status"),
88
- "profiling_baseline_active_time_us": baseline_profile.get("ActiveTimeUs"),
89
- "profiling_baseline_source": baseline_profile.get("source"),
90
- "profiling_agent_status": agent_profile.get("status"),
91
- "profiling_agent_active_time_us": agent_profile.get("ActiveTimeUs"),
92
- "profiling_agent_source": agent_profile.get("source"),
93
- "S_i_aicore": profiling.get("S_i_aicore"),
94
- "U_i": profiling.get("U_i"),
95
- "U_step": profiling.get("U_step"),
96
- "VecRatio": profiling.get("VecRatio"),
97
- "CubeRatio": profiling.get("CubeRatio"),
98
- "MTE2Ratio": profiling.get("MTE2Ratio"),
99
- })
100
- return rows
101
-
102
-
103
- def write_csv(rows: List[Dict[str, Any]], output_path: Path):
104
- fieldnames = list(rows[0].keys()) if rows else []
105
-
106
- def csv_value(value: Any) -> Any:
107
- if isinstance(value, (dict, list)):
108
- return json.dumps(value, ensure_ascii=False, separators=(",", ":"))
109
- return value
110
-
111
- with output_path.open("w", newline="", encoding="utf-8") as fh:
112
- writer = csv.DictWriter(fh, fieldnames=fieldnames)
113
- writer.writeheader()
114
- for row in rows:
115
- writer.writerow({k: csv_value(row.get(k)) for k in fieldnames})
116
-
117
-
118
- def main() -> None:
119
- parser = argparse.ArgumentParser(description="Export operator evaluation summary JSON into CSV.")
120
- parser.add_argument("--input", type=Path, required=True, help="Path to eval_summary.json")
121
- parser.add_argument("--output-dir", type=Path, default=None, help="Folder for the CSV output")
122
- args = parser.parse_args()
123
-
124
- summary = load_json(args.input)
125
- rows = compute_summary_rows(summary)
126
- out_dir = args.output_dir or args.input.parent / "export"
127
- out_dir.mkdir(parents=True, exist_ok=True)
128
-
129
- csv_path = out_dir / "operator_metrics.csv"
130
- write_csv(rows, csv_path)
131
-
132
- print(f"CSV exported: {csv_path}")
133
-
134
-
135
- if __name__ == "__main__":
136
- main()
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
website/index.html DELETED
@@ -1,15 +0,0 @@
1
- <!doctype html>
2
- <html lang="en" data-theme="light">
3
- <head>
4
- <meta charset="utf-8">
5
- <meta name="viewport" content="width=device-width,initial-scale=1">
6
- <meta name="color-scheme" content="light dark">
7
- <title>AscendGYM — Kernel Optimization Benchmark</title>
8
- <meta name="description" content="AscendGYM — AscendC 算子优化 Agent 评测集:STS(token 效率)× PAOS(profiling 价值)双轴单页 dashboard + entry 详情抽屉。">
9
- <link rel="icon" href="%BASE_URL%assets/logo/favicon.svg">
10
- </head>
11
- <body data-preview-state="default">
12
- <div id="app"></div>
13
- <script type="module" src="/src/main.js"></script>
14
- </body>
15
- </html>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
website/metrics_master.md DELETED
@@ -1,389 +0,0 @@
1
- # Ascend 内核优化 Benchmark 指标总纲
2
-
3
- > 本文档是评测集指标体系的**总纲**:记录两套指标体系(STS / PAOS)的选择原因、选定指标、计算公式与递推关系、所需数据及其就绪度。
4
- > 依据:`孟博gpt对话_中文.txt`(STS 体系)、`基于Profiling的Ascend内核优化Benchmark指标设计.md`(PAOS 体系)、`EvalSet/leaderboard/seven_dimensions.md`(七维度业务面)、agent 实际采集数据(result.json / solutions.jsonl / op_summary.csv)实证。
5
- > 配套:`EvalSet/leaderboard/eval_schema.md`(榜单列定义)、`docs/eval_design.md`(埋点设计)。
6
-
7
- ---
8
-
9
- ## 0. 一句话定位
10
-
11
- Benchmark 要回答:**"在最终峰值性能和 token 效率方面,Agent 能在多大程度上改进主流工作负载下的 Ascend C 内核?"**
12
-
13
- 这句话天然拆成两个子问题,分别由两套体系回答:
14
-
15
- | 子问题 | 体系 | 回答 |
16
- |---|---|---|
17
- | Agent 在 token 预算下能拿到多少**已验证加速**?多早、多广、多可靠? | **STS**(Speedup–Token Surface) | Agent 的**效率与能力** |
18
- | 加速是否转化为**模型 token 吞吐**提升?是否逼近**硬件峰值**? | **PAOS**(Profile Agent Optimization Score) | 优化的**价值与质量** |
19
-
20
- **七维度**(正确性/优化率/token/轮数/时间/利用率/可靠性)是这两套体系面向读者的**业务投影**,不是独立计算层。
21
-
22
- ---
23
-
24
- ## 1. 为什么是这两套体系(选择原因)
25
-
26
- ### 1.1 为什么需要 STS(token 效率轴)
27
-
28
- 内核优化是一个**有预算约束的 anytime 问题**:给定初始正确的 Ascend C 实现,Agent 消耗 t 个 token 能拿到多少已验证提升。仅报"最终加速比"或"每 token 加速比"都不够:
29
-
30
- - **最终加速比**忽略了 Agent 多早找到优化——同样 1.2× 加速,1 万 token 找到 vs 400 万 token 找到,价值天差地别。
31
- - **每 token 加速比**会让"快速肤浅优化"和"慢速深度优化"不可区分。
32
- - 大加速(24h 后 10×)会被有限阈值集合(1.05/1.2/1.5/2.0×)**饱和**截断。
33
-
34
- STS 的解法:**连续 `log₂ S`(无界、不饱和)+ 对数 token 轴下的 anytime 曲线面积(AULS)**,并按 `λ=0.5` 混合"峰值性能"与"尽早改进"。这是 Agent 能力的主排名轴。
35
-
36
- ### 1.2 为什么需要 PAOS(profiling 价值轴)
37
-
38
- STS 只看"内核变快了多少倍",但有两个盲区:
39
-
40
- - **加速可能不在关键路径上**:一个占 step 时间 0.1% 的内核拿到 5× 加速,远不如占 20% 的内核拿到 5% 加速。STS 的单任务 `S_i` 不含 workload 权重。
41
- - **加速可能以资源为代价**:变快但吃满内存、或离硬件峰值仍很远,价值打折。
42
-
43
- PAOS 的解法:用 **model cluster profile recipe**(算子调用图谱 + tokens/step)把单算子 profiling 合成**模型级 token 吞吐**(eTPS),并分解为 `eTPSGain = 利用率增益 × AI Core token cost 增益`,再合层 `√(PKAI·PTEI)`。这把"内核加速"翻译成"模型真实收益 + 硬件逼近度"。
44
-
45
- ### 1.3 两套的耦合点
46
-
47
- - **共享原子量**:`t_i`(kernel_time)、`status`(正确性)、`S_i = t_base/t_agent`(kernel 加速比)。
48
- - **分叉点**:STS 把 `S_i` 喂进 anytime 曲线(横轴 = token);PAOS 把 `t_i` 喂进 model step 合成(横轴 = AI Core time / utilization)。
49
- - **正交**:一个 Agent 可以 STS 高(很快找到加速)但 PAOS 低(优化的是非关键算子);反之亦然。两者**都保留、联合发布**。
50
-
51
- ---
52
-
53
- ## 2. 选定指标清单
54
-
55
- ### 2.1 STS 体系(token 效率轴)
56
-
57
- | 指标 | 含义 | 角色 |
58
- |---|---|---|
59
- | `S_i(B)` | 已验证加速比 `L⁰/L^{best}(B)` | 原子量 |
60
- | `VLR` | 延迟降低率 `1−1/S_i` | 可读量 |
61
- | `EBGS@B` | 专家基线几何平均加速比 `exp(Σw·ln S_i)` | 跨任务峰值 |
62
- | `EIR_Δ@B` | 改进广度(Δ∈{1,3,5,10}%)`Σw·1[VLR≥Δ]` | 广度 |
63
- | `AULS_i` | 对数 token 轴下加速比曲线面积 | anytime 效率 |
64
- | `STS_token` | 主榜分 `100·Σw·[0.5·log₂S(Bmax)+0.5·AULS]` | **主排名** |
65
- | `STS_time` | 同上,横轴换 wall-clock | 基础设施效率诊断 |
66
- | `Tokens-to-Δ` | 首达 Δ% 的累计 token | token 效率 |
67
- | `S_equiv` | `2^(STS/100)` | 可解释等效加速比 |
68
-
69
- ### 2.2 PAOS 体系(profiling 价值轴)
70
-
71
- | 指标 | 含义 | 角色 |
72
- |---|---|---|
73
- | `S_i^lat` | kernel latency speedup(= STS 的 `S_i`) | 原子量 |
74
- | `S_i^aicore` | AI Core active time speedup `a_base/a_agent` | 资源加速 |
75
- | `U_i` / `U_step` | AI Core 利用率 `a_i/t_i` / step 级 | 利用率 |
76
- | `eTPS` / `eTPSGain` | 估算 tokens/s 及其增益 `T_step^base/T_step^agent` | **token 吞吐** |
77
- | `AICoreTokenCostGain` | 每 token AI Core cost 降低 `A_token^base/A_token^agent` | 分解量 |
78
- | `PKAI` | token-time 加权 kernel speedup `exp(Σw̃·ln S_i^lat,eff)` | kernel 加速能力 |
79
- | `PTEI` | = `eTPSGain` | token 效率 |
80
- | `PPUI` | peak utilization index `1+RHC` | 峰值逼近 |
81
- | `PAOS` | 综��分 `√(PKAI·PTEI)`(简洁版)/ `PKAI^0.4·PTEI^0.4·PPUI^0.2`(完整版) | **主排名** |
82
-
83
- ### 2.3 七维度 → 体系映射
84
-
85
- | 七维度 | 来自 STS | 来自 PAOS |
86
- |---|---|---|
87
- | ① 正确性 | FuncPassRate(硬门槛,单独发布) | — |
88
- | ② 优化率 | `S_i` / EBGS / VLR / STS | `S_i^lat` / PKAI |
89
- | ③ token 量 | cumulative_tokens / AULS / Tokens-to-Δ | — |
90
- | ④ 交互轮数 | llm_rounds / wasted_round_rate | — |
91
- | ⑤ 优化时间 | wall_clock / STS_time | — |
92
- | ⑥ 利用率 | — | `U_i` / Vec·Cube Ratio / PPUI |
93
- | ⑦ 可靠性 | Reliability@Δ(K≥3) | — |
94
-
95
- ---
96
-
97
- ## 3. 计算公式与递推关系
98
-
99
- ### 3.1 原子层(agent 采集 + profiling)
100
-
101
- ```
102
- 单解 j(一条 solution):
103
- status_{i,j} ∈ {passed, compile_failed, test_failed, test_mismatch, ...}
104
- t_{i,j} = duration_ms # kernel_time(passed 才有效)
105
- T_{i,j} = cumulative_tokens # 生成该解累计 token
106
- W_{i,j} = solution_elapsed_sec # 生成该解累计 wall-clock
107
- a_{i,j} = ai_core_time 或 aiv_time(见 §5) # AI Core active time
108
-
109
- 单任务 i:
110
- L_i^0 = baseline_duration_ms
111
- L_i^{best}(t) = min(L_i^0, min{ t_{i,j} : T_{i,j}≤t ∧ status=passed }) # 不过 correctness 回退 baseline
112
- S_i(t) = L_i^0 / L_i^{best}(t) # 已验证加速比(≥1)
113
- ```
114
-
115
- ### 3.2 STS 递推(token 效率轴)
116
-
117
- ```
118
- VLR_i(t) = 1 − 1/S_i(t) # 延迟降低率
119
- EBGS@B = exp( Σ_i w_i · ln S_i(B) ) # 跨任务几何均值(排除 abnormal/saturated)
120
-
121
- AULS_i = Σ_k γ_k · ( log₂S_i(B_{k-1}) + log₂S_i(B_k) ) / 2 # 对数 token 检查点梯形
122
- γ_k = ln(B_k/B_{k-1}) / ln(B_max/B_min)
123
- (检查点 B_0<...<B_K=B_max 对数间隔)
124
-
125
- STS_token = 100 · Σ_i w_i · [ 0.5·log₂S_i(B_max) + 0.5·AULS_i ] # λ=0.5
126
- STS_time = 同上,T_{i,j} 换 W_{i,j}
127
- S_equiv = 2^(STS/100) # 等效加速比
128
-
129
- EIR_Δ@B = Σ_i w_i · 1[ VLR_i^LCB(B) ≥ Δ ] # 改进广度(LCB=置信下界,单 run 用点估计)
130
- Tokens-to-Δ = min{ t : VLR_i(t) ≥ Δ } # 首达 token
131
- ```
132
-
133
- **递推链**:`status+t_{i,j}+T_{i,j}` → `L_i^{best}(t)` → `S_i(t)` → `{EBGS, VLR, AULS, STS, EIR, Tokens-to-Δ}`。
134
-
135
- ### 3.3 PAOS 递推(profiling 价值轴)
136
-
137
- 需先有 **recipe**:每模型簇每场景的 `{G(tokens/step), ops:[{kernel, shape_id, calls_per_step:n_i}]}`。
138
-
139
- ```
140
- # kernel 级
141
- S_i^lat = t_i^base / t_i^agent
142
- S_i^aicore = a_i^base / a_i^agent
143
- U_i = a_i^norm / t_i (a_i^norm = a_i^sum / core_count,若 ai_core_time 已是单核则不除)
144
- ΔU_i = U_i^agent − U_i^base
145
-
146
- # token 映射(kernel → 每 token 贡献)
147
- τ_i = n_i · t_i / G (token time per kernel)
148
- Δτ_i = n_i/G · (t_i^base − t_i^agent)
149
- α_i = n_i · a_i / G (AI Core token cost per kernel)
150
-
151
- # step 级(合成模型 token step)
152
- T_step = Σ_i n_i·t_i + T_other (未优化算子 speedup=1;无 overhead 数据时 T_other=0,标注 profile-only)
153
- T_token = T_step / G
154
- eTPS = G / T_step = 1/T_token
155
- eTPSGain = T_step^base / T_step^agent
156
- A_token = Σ_i n_i·a_i / G
157
- AICoreTokenCostGain = A_token^base / A_token^agent
158
- U_step = A_token / T_token
159
- eTPSGain = (U_step^agent/U_step^base) · (A_token^base/A_token^agent) # 分解恒等式
160
-
161
- # 加权聚合
162
- w_i = n_i·t_i^base / T_step^base (token-time 权重)
163
- w̃_i = w_i / Σ_j w_j (评测 kernel 内归一化)
164
- S_i^lat,eff = S_i^lat if pass else 1.0
165
- PKAI = exp( Σ_i w̃_i · ln S_i^lat,eff )
166
- PTEI = eTPSGain
167
-
168
- # 峰值(需 utilization 或 roofline)
169
- RHC_U = clip( (U_step^agent − U_step^base) / (1 − U_step^base), −1, 1 )
170
- PPUI = 1 + RHC_U (或 roofline 版 1+RHC_roofline)
171
-
172
- # 综合
173
- PAOS = √(PKAI · PTEI) # 简洁版(推荐,不需 PPUI)
174
- PAOS_full = PKAI^0.4 · PTEI^0.4 · PPUI^0.2 # 完整版
175
- PAOS_final = PAOS · FuncPassRate^λ (λ=2,非正式榜 penalty)
176
- PAOS_overall = exp( Σ_c α_c · ln PAOS_c ) # 跨模型簇几何均值
177
- ```
178
-
179
- **递推链**:`t_i+a_i(可选)+n_i+G+status` → `{S_i^lat, U_i, eTPS}` → `{PKAI, PTEI, PPUI}` → `PAOS`。
180
-
181
- ---
182
-
183
- ## 4. 所需数据清单
184
-
185
- | 数据 | 用途 | 来源 |
186
- |---|---|---|
187
- | `t_i` kernel_time | STS+PAOS 原子量 | `eval_result.duration_ms` / op_summary `Task Duration(us)` |
188
- | `status` 正确性 | 硬门槛 | `eval_result.status` |
189
- | `T_{i,j}` cumulative_tokens | STS anytime 横轴 | `solutions.jsonl.cumulative_tokens` |
190
- | `W_{i,j}` wall-clock | STS_time 横轴 | `solution_elapsed_sec` |
191
- | `a_i` AI Core active time | PAOS 利用率/分解 | op_summary `aicore_time(us)`(cube)/ `aiv_time(us)`(vector)|
192
- | `Vec/Cube/MTE2 Ratio` | 利用率诊断 | op_summary `aiv_vec_ratio` / `aic_mac_ratio` / `aic_mte2_ratio` |
193
- | `core_count` | `a_i` 归一化 | op_summary `Block Num` |
194
- | `n_i` calls/step | PAOS step 合成 | **recipe**(需构造)|
195
- | `G` tokens/step | PAOS step 合成 | **recipe**(需人工指定场景)|
196
- | `FLOPs` / memory bytes | roofline(PPUI 完整版) | msprof `ArithmeticUtilization.csv` / `Memory.csv`(**当前未采集**)|
197
-
198
- ---
199
-
200
- ## 5. 数据就绪度(实证:当前 agent 数据能算什么)
201
-
202
- > 实证来源:Swiglu run 的 `result.json` / `solutions.jsonl` / `op_summary_*.csv`(43 列真实表头)/ `models_configs/`。
203
-
204
- ### 5.1 STS 体系:**全部就绪** ✅
205
-
206
- `status` + `duration_ms` + `cumulative_tokens` + `solution_elapsed_sec` 全部已在 `solutions.jsonl`。STS_token / AULS / EBGS / EIR / Tokens-to-Δ 可立即计算。
207
-
208
- ### 5.2 PAOS 体系:**简洁版可算,完整版受限**
209
-
210
- | PAOS 指标 | 就绪度 | 说明 |
211
- |---|---|---|
212
- | `S_i^lat` / PKAI | 🟡 缺 recipe 的 `n_i` | `t_i` 有;token-time 权重 `w_i=n_i·t_i^base/T_step` 需 `n_i` |
213
- | eTPS / PTEI | 🟡 缺 recipe 的 `n_i`+`G` | step 合成必需 |
214
- | `S_i^aicore` / `U_i` / `U_step` / AICoreTokenCostGain | 🟡 `a_i` 部分有 | 见 §5.3 |
215
- | PPUI(utilization 版) | 🟡 依赖 `a_i` | 同上 |
216
- | PPUI(roofline 版)/ RHC_roofline | 🔴 缺 FLOPs+bytes | msprof 细分表未采集 |
217
- | **PAOS 简洁版 `√(PKAI·PTEI)`** | 🟡 补 recipe 即可算 | 不需 PPUI |
218
- | PAOS 完整版(含 PPUI/roofline) | 🔴 需重新 profiling | 见 §5.4 |
219
-
220
- ### 5.3 `a_i`(AI Core active time)的实证情况
221
-
222
- op_summary 真实表头(43 列)含:`aicore_time(us)`(AI Cube Core)、`aiv_time(us)`(AI Vector Core)、`aiv_vec_time(us)`、`aic_mac_time(us)` 等。
223
-
224
- - **Swiglu 是纯 vector 算子**:`aicore_time=0.0`(64/64 行全 0),主时间是 `aiv_time`。`select_metric_columns` 过滤掉全零列 → `result.json.hardware_metrics` 只剩 `{Task Duration, AI Vector Time, Vec Ratio}` 3 个 key。
225
- - **结论**:PAOS 的 `a_i` 不能统一用 `aicore_time`——**vector 算子必须用 `aiv_time`**,cube 算子用 `aicore_time`。这是 Ascend 双核(Cube+Vector)架构的固有要求。当前 `hardware_metrics` 已含 `AI Vector Time`,对 Swiglu 可算 `U_i = aiv_vec_time/t_i`(近似)和 Vec Ratio gain。
226
-
227
- ### 5.4 FLOPs / memory bytes:**当前完全缺失** 🔴
228
-
229
- - op_summary.csv(43 列)**无任何 bytes/FLOPs 列**(只有 time + ratio + cube_utilization)。
230
- - msprof 的 `Memory.csv`(HBM read/write bytes)、`ArithmeticUtilization.csv`(cube/vec FLOPs)**当前 profiling 流程未产出**——`find` 全树零命中。
231
- - `extract_msprof_fields.py` 能解析这 10 张细分表(字段映射准确),但(a)数据不存在、(b)它不在主流水线(`ascendc_code_profiling_service.py`)上,只被独立 service 调用。
232
- - **解锁路径**:profiling 命令追加 `--analysis_type` 等参数重跑,产出 Memory.csv / ArithmeticUtilization.csv,再激活 extract_msprof_fields 接入主流水线。在此之前,roofline 类指标(RHC_roofline、Token Roofline Efficiency)算不了。
233
-
234
- ---
235
-
236
- ## 6. 缺失的 recipe 数据:能否从 model_config / profiling 派生?(实证)
237
-
238
- recipe 需要两个量:`tokens_per_step (G)` 和 `calls_per_step (n_i)`。
239
-
240
- ### 6.1 `tokens_per_step (G)`:**不能自动派生,必须人工指定** ❌
241
-
242
- - `models_configs/*.json`(deepseek_v4_pro / kimi_k3 / qwen3.6_27b)通篇是**静态结构字段**(hidden_size / num_hidden_layers / moe_intermediate_size 等),**无 batch_size / chunk_len / seq_len / scenario**。
243
- - G 是**运行时场景量**:decode 场景 `G = batch_size`;prefill 场景 `G = batch_size × chunk_len`。须由 benchmark 方按场景指定。
244
- - golden case 的 token 维(SwiGlu 的 `1×{1,6,32,512,1024,2048}×6144`,第二维 = token 数)是**算子测试用例的枚举**,选哪个对应哪个场景是人工决策,不是派生。
245
-
246
- ### 6.2 `calls_per_step (n_i)`:**结构部分可半自动派生,运行时部分不可派生** 🟡
247
-
248
- 以 SwiGlu @ DeepSeek-V4-Pro 为例(transformers 源码 `modeling_deepseek_v4.py` 实证):
249
-
250
- - 单个 decoder layer 内 SwiGLU 调用点:① 路由专家 `_apply_gate`(L1023-1030,每个命中专家 1 次)+ ② 共享专家 `DeepseekV4MLP.forward`(L986-989,固定 1 次)。
251
- - **结构骨架(可派生)**:`n_SwiGLU/token/layer ≈ num_experts_per_tok(6) + 1 = 7`;整模型 `≈ 7 × num_hidden_layers(61) = 427`。`num_experts_per_tok`、`num_hidden_layers` 都在 config。
252
- - **运行时精确值(不可派生)**:精确次数依赖 router 决策(每 step 实际命中哪些专家、batch 内 token 分布到 384 专家的情况),需 profiling 或负载假设。
253
- - **可派生 vs 不可派生的分界**:层内固定次数算子(如 rmsnorm:`input_layernorm` + `post_attention_layernorm` + 最终 `norm` = `2×61+1=123`)能纯静态推出;**路由相关算子(SwiGLU/MoE)只能给上界/期望值**。
254
-
255
- ### 6.3 recipe 现状:**EvalSet 完全无此概念** ❌
256
-
257
- - 全项目 grep `calls_per_step|tokens_per_step|model_cluster|step_profile|recipe` → 零命中。
258
- - `eval_meta.json` 固定 4 字段 `{op_name, golden_root, default_model, default_shape}`(README 契约),无 recipe 扩展。
259
- - **recipe 是需新建的层**:建议新增 `EvalSet/recipes/<model_cluster>_<scenario>.json`,不破坏 eval_meta 契约。
260
-
261
- ### 6.4 派生可行性总表
262
-
263
- | 量 | 可派生? | 来源 |
264
- |---|---|---|
265
- | `G` (tokens/step) | ❌ 必须人工指定 | 场景(decode=batch / prefill=batch×chunk) |
266
- | `n_i` 结构骨架 | 🟡 半自动 | config `num_hidden_layers`/`num_experts_per_tok` + transformers 调用点统计 |
267
- | `n_i` 运行时精确 | ❌ 需 profiling/假设 | router 决策 + token 分布 |
268
- | shape 对齐(kernel↔模型) | ✅ 可派生 | config 维度(如 `6144=2×moe_intermediate_size`)+ golden case shape |
269
-
270
- ---
271
-
272
- ## 7. 落地优先级
273
-
274
- 1. **P0(数据全就绪,立即实现)**:STS 体系全套——`Eval/sts.py`,读 `solutions.jsonl` 算 STS_token / AULS / EBGS / EIR / Tokens-to-Δ。这是当前能完整跑通的主排名轴。
275
- 2. **P1(补 recipe 即可)**:PAOS 简洁版 `√(PKAI·PTEI)`——新建 `EvalSet/recipes/`,人工填 `G`、半自动派生 `n_i` 结构骨架;`a_i` 调 `GroupStats.active_time_us()`(坑2已修,自动按 cube/vector 选列)。
276
- 3. **P2(需重新 profiling)**:PAOS 完整版 PPUI/roofline——profiling 命令加 `--analysis_type` 产出 Memory.csv / ArithmeticUtilization.csv,激活 `extract_msprof_fields.py` 接入主流水线。
277
- 4. **P2(需多次 run)**:Reliability@Δ(K≥3 独立 run)。
278
-
279
- ---
280
-
281
- ## 8. 已知数据质量坑(实证发现)
282
-
283
- 1. **【✅ 已修 2026-08-06】`profile_op_summary.py` 列名不匹配**:候选列 csv 列名 `Block Dim`/`Mix Block Dim` 与真实表头 `Block Num`/`Mix Block Num` 不一致 → `core_count` 取不到。已把 CANDIDATE_METRIC_COLUMNS 的 csv 列名改为 `Block Num`/`Mix Block Num`(friendly label `Block Dim`/`Mix Block Dim` 保留不变,下游 key 不变)。实测:Swiglu `Block Num` 样本值=48 现已能进 groups。
284
- 2. **【✅ 已修 2026-08-06】vector 算子 `a_i` 用错列**:PAOS 若统一用 `aicore_time`,Swiglu 这类纯 vector 算子恒为 0(实测 64/64 行 `aicore_time=0.0`)。已在 `GroupStats` 新增 `active_time_label()`/`active_time_us()`:cube 优先 `AI Core Time (us)`,vector 回退 `AI Vector Time (us)`。实测 Swiglu 返回 `'AI Vector Time (us)'`、值 13.375us(非零)。**PAOS 计算层应调 `group.active_time_us()` 取 `a_i`,不要硬编码列名。**
285
- 3. **【⏳ 观望,暂不动】hw_metrics 白名单过滤**:主流水线 `_serialize_hw_metrics` 的 `HW_CORE_LABELS`(8 个 label,`ascendc_code_profiling_service.py:590`)只回传白名单字段,叠加 `select_metric_columns` 砍全零列。对纯 vector 算子(Swiglu)合理;但 cube/mixed 算子可能丢有用字段(如 Scalar Ratio)。**暂观望**:等真正有 cube 算子入库、且发现字段缺失影响 PAOS 计算时,再决定是否放宽白名单到 CANDIDATE_METRIC_COLUMNS 全集(18 个)。
286
-
287
- ---
288
-
289
- ## 9. 数据缺口建设待办(获取路径详记)
290
-
291
- > 三类数据缺口挡住 PAOS 的不同部分。STS 全套不受影响(数据全就绪)。PAOS 简洁版 `√(PKAI·PTEI)` + utilization 版 PPUI 只需缺口 B+C;完整版 roofline 才需缺口 A。
292
-
293
- ### 缺口 A:FLOPs + memory bytes(解锁 PAOS roofline 指标)
294
-
295
- - **缺什么**:`aic_cube_fops`/`aiv_vec_fops`(FLOPs)、`read_main_memory_datas`/`write_main_memory_datas`(HBM bytes)。用于 roofline 上限计算(`RHC_roofline`、Token Roofline Efficiency)。
296
- - **为何缺**:当前 profiling 只产 4 个 csv(op_summary/op_statistic/api_statistic/task_time),均无 bytes/FLOPs 列(op_summary 43 列已确认)。这些量在 msprof 的 `Memory.csv`/`ArithmeticUtilization.csv` 里,但当前 profiling 流程未产出(全树零命中)。根因:`ascendc_code_profiling_service.py:291` 的 profiling 命令调 sandbox `main.py`,只传 `--profile --profiling-runs`,未传控制细分分析的参数。
297
- - **获取路径(按推荐度)**:
298
- 1. **【推荐】sandbox profiling 加 `--analysis_type` 重跑**:改 `ascendc_code_profiling_service.py:291` 的 `profiling_cmd` 加 `--analysis_type ...`,并确认 sandbox `main.py`(`PROFILING_MAIN_PY_PATH=./main.py`)透传给底层 msprof。**前置不确定点:sandbox 的 main.py 是否接受 `--analysis_type`,需先核实 sandbox 代码**。
299
- 2. **激活 `extract_msprof_fields.py`**:它已能解析 10 张细分表(字段映射已验证准确),接入主流水线 `_extract_and_build_report`,产出后让 result.json 多一组 `msprof_detail_metrics`。
300
- 3. **手动 msprof 兜底**:脱离 agent 自动流程,编译运行后手动 `msprof --application ... --analysis_type=ai-core`,适合先验证指标可行性。
301
- - **影响范围**:只挡 PAOS roofline 版 PPUI。**PAOS 简洁版不依赖它**,可用 utilization 版 PPUI(`1+RHC_U`)顶上。
302
-
303
- ### 缺口 B:recipe 的 `tokens_per_step (G)`(解锁 PAOS step 合成)
304
-
305
- - **缺什么**:每个 model cluster × scenario 的 `G`(一个推理 step 处理/生成的 token 数)。
306
- - **为何缺**:`models_configs/*.json` 通篇静态结构(hidden_size/num_layers),无 batch_size/chunk_len/scenario。G 是运行时场景量。
307
- - **获取路径**:**人工指定,不可自动派生**。decode `G=batch_size`;prefill `G=batch_size×chunk_len`。落地新建 `EvalSet/recipes/<model_cluster>_<scenario>.json`,字段 `{model_cluster, scenario, tokens_per_step, ops:[...]}`。SwiGlu golden case 已枚举 token 维(1/6/32/512/1024/2048),选哪个 G 对应哪个 scenario。
308
-
309
- ### 缺口 C:recipe 的 `calls_per_step (n_i)`(解锁 PAOS step 合成)
310
-
311
- - **缺什么**:每算子在一个 step 里的调用次数 `n_i`。EvalSet 完全无 recipe 概念(全项目 grep 零命中)。
312
- - **获取路径(半自动派生)**:
313
- - **结构骨架可派生**:`num_hidden_layers`(config) × 算子层内调用次数(transformers 代码)。SwiGlu 已查清:单 token 单层 ≈ `num_experts_per_tok(6) + 1(共享专家) = 7`,整模型 `≈ 7 × num_hidden_layers(61) = 427`。
314
- - **运行时精确值不可派生**:路由相关算子(SwiGlu/MoE)精确次数依赖 router 决策,需 profiling 真实负载或给假设。层内固定算子(如 rmsnorm `2×61+1=123`)可纯静态推出。
315
- - 落地:recipe 给期望值/上界(`n_i = 7 × num_layers`)即可,PAOS 的 token-time 权重 `w_i = n_i·t_i^base/T_step` 用期望值足够。
316
-
317
- ### 缺口建设优先级
318
-
319
- | 缺口 | 工作量 | 解锁 | 前置 |
320
- |---|---|---|---|
321
- | B + C(recipe 文件) | ~1h(含派生 n_i) | PAOS 简洁版主轴 | 无 |
322
- | A(细分表) | 改 sandbox `gen_prof_code_application.py:25` 的 msprof 命令加 `--analysis-type`(已核实) | PAOS roofline | 改的是 sandbox 仓库(非本项目),影响面大,**已搁置** |
323
-
324
- ---
325
-
326
- *总纲依据实证:op_summary.csv 43 列真实表头、Swiglu hardware_metrics 实际 key、models_configs 结构字段、transformers V4 调用点行号、extract_msprof_fields.py 10 表映射。详见会话调查记录。*
327
-
328
- ---
329
-
330
- ## 10. `calls_per_step (n_i)` 推导方法论(缺口 C 补齐方案)
331
-
332
- > 目标:对每个算子,推出"一个推理 step 内它被调用的次数 `n_i`",喂给 PAOS 的 step 合成(`T_step = Σ n_i·t_i`)。
333
- > 不写统计脚本,只给可复用的方法论 + SwiGlu 实例验证。
334
-
335
- ### 10.1 口径定义(必须先钉死,否则 T_step 会错)
336
-
337
- PAOS 设计文档 §6.5 的 `τ_i = n_i·t_i/G` 定义了三者关系,由此钉死口径:
338
-
339
- - **`t_i`** = profiled **单次算子调用**的耗时(op_summary 里 Task Duration)。这次调用可能处理 1 个 token,也可能处理 4096 个 token——t_i 已含其处理的全部 token,**不再除以 token 数**。
340
- - **`n_i`** = 一个 step 内该算子被 **launch(调用)的次数**(profiler 视角:op_summary 里该 op_type 出现的行数)。
341
- - **`G`** = 一个 step 处理/生成的 token 数。
342
-
343
- **关键推论**:`n_i` 是"算子调用次数",**不是"每 token 算子次数 × token 数"**。这两者在共享专家这类"一次调用处理全部 token"的算子上最容易混淆——共享专家一层只 launch 1 次(不管多少 token),`n_i=1`,不是 `n_i=token数`。
344
-
345
- ### 10.2 推导四步法(适用任意算子)
346
-
347
- 1. **读 config 拿层结构**:`num_hidden_layers`(决定纵向放大倍数)、算子相关字段(如 MoE 的 `num_experts_per_tok`/`n_shared_experts`)。
348
- 2. **读 transformers 源码数"一层内的调用点"**:在 `modeling_<model>.py` 的 decoder layer forward 里,grep 该算子的数学语义(如 SwiGlu = `act(gate)*up` / `chunk(2,...)`),统计**一层 forward 内它执行几次**。区分两种调用模式:
349
- - **固定调用**(与 token 数无关):如 RMSNorm 每 layer 固定 2-3 次、共享专家每层 1 次。`n_i = 每层级数 × num_layers`。
350
- - **token/路由相关调用**(随负载变):如 MoE 路由专家,`n_i = 该层实际命中专家数`。
351
- 3. **判断是否全层都有该算子**:查 `mlp_layer_types`/`first_k_dense_replace` 等,确认 61 层是否都用同种 FFN(V4 全 MoE,无 dense 层)。若有 dense/MoE 混合,按层类型分段计。
352
- 4. **纵向聚合**:`n_i = 每层级数 × num_layers`(固定调用);路由调用按场景给期望值(见 10.3)。
353
-
354
- ### 10.3 decode vs prefill:路由算子 n_i 的场景差异(关键)
355
-
356
- 路由相关算子(MoE 路由专家)的 n_i **随场景变**,这是 n_i 推导里唯一的"非静态"部分:
357
-
358
- | 场景 | G | 路由专家每层命中数(= launch 次数) | 说明 |
359
- |---|---|---|---|
360
- | **decode b1** | 1 | = `num_experts_per_tok`(6) | 1 个 token 命中 6 个专家,6 次 launch |
361
- | **decode b32** | 32 | ≤ 384,期望 = min(384, 32×6 去重) | 32 token 各命中 6 专家,去重后被激活的专家数 |
362
- | **prefill b1 chunk4096** | 4096 | ≤ 384,期望接近 384(但受 top_k=6 约束) | 4096 token 散开,绝大多数专家被至少 1 token 选中 |
363
-
364
- - **共享专家**不受场景影响:每层永远 launch 1 次(一次吃全部 token)。
365
- - **精确命中数需真实负载 profiling**(取决于 router 决策 + token 分布)。**期望值对 PAOS 加权足够**——PAOS 用 n_i 算 token-time 权重 `w_i = n_i·t_i^base/T_step`,权重的小偏差不影响排名结论。
366
- - **取值建议**:recipe 里 decode 用确定值(`num_experts_per_tok + n_shared`);prefill 给"期望命中数"并标注口径,或直接用 decode 的 n_i 近似(保守下界)。
367
-
368
- ### 10.4 实例验证:SwiGlu @ DeepSeek-V4-PRO
369
-
370
- **第1步 config**:`num_hidden_layers=61`、`num_experts_per_tok=6`、`n_shared_experts=1`、`mlp_layer_types=("hash_moe","moe")`(全 61 层 MoE)。
371
-
372
- **第2步 数调用点**(`modeling_deepseek_v4.py` 实证):
373
- - `DeepseekV4Experts._apply_gate`(L1027-1030):`gate,up=chunk(2); act(gate)*up` —— 每个**命中**的路由专家执行 1 次。
374
- - `DeepseekV4MLP.forward`(L989):`act(gate)*up` —— 共享专家,一层固定 1 次(处理全部 token)。
375
- - 每 MoE block = 路由专家集合 + 1 共享专家(L1098 `routed + shared_experts(residual)`)。
376
-
377
- **第3步 全层确认**:61 层全是 MoE(前若干 hash_moe + 其余 moe,top_k 均 6),无 dense FFN。
378
-
379
- **第4步 聚合**:
380
- - **decode b1**:每层命中 6 路由 + 1 共享 = 7 → `n_i = 7 × 61 = 427`(确定值)。
381
- - **prefill b1 chunk4096**:每层命中专家数期望 ≈ 接近 384(受 top_k=6 约束的实际上界)+ 1 共享 → `n_i` 期望 ≈ `(命中专家数+1) × 61`,范围 `[7×61, 385×61] = [427, 23485]`,需负载假设取期望。
382
-
383
- **结论**:SwiGlu decode n_i=427 写进 recipe(确定值);prefill 标注口径取期望或用 decode 近似。完整 recipe 见 `EvalSet/recipes/DeepSeek-V4.json`。
384
-
385
- ### 10.5 方法论适用范围与边界
386
-
387
- - ✅ **适用**:所有 transformer 算子(RMSNorm/Attention/RoPE/SwiGlu/MoE 路由)。固定调用算子(RMSNorm/RoPE)能纯静态推出精确值;路由算子给期望值。
388
- - ⚠️ **caveat**:n_i 是"期望/上界",不是运行时精确值。对 PAOS 排名足够;若需精确模型级 eTPS,需真实 serving profiling(属缺口 A 同级的"重新 profiling"范畴)。
389
- - 🔄 **可复用**:换算子/换模型时,重走四步法(config 层数 → 源码数调用点 → 全层确认 → 聚合),不需写统计脚本——源码数调用点靠 grep + 精读单层 forward 即可。
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
website/package-lock.json DELETED
@@ -1,1388 +0,0 @@
1
- {
2
- "name": "ascendgym-ux-v2-vue",
3
- "version": "2.0.0",
4
- "lockfileVersion": 3,
5
- "requires": true,
6
- "packages": {
7
- "": {
8
- "name": "ascendgym-ux-v2-vue",
9
- "version": "2.0.0",
10
- "dependencies": {
11
- "katex": "0.19.1",
12
- "vue": "^3.5.13",
13
- "vue-router": "^4.5.0"
14
- },
15
- "devDependencies": {
16
- "@vitejs/plugin-vue": "^5.2.1",
17
- "vite": "^6.1.0"
18
- }
19
- },
20
- "node_modules/@babel/helper-string-parser": {
21
- "version": "7.29.7",
22
- "resolved": "https://registry.npmjs.org/@babel/helper-string-parser/-/helper-string-parser-7.29.7.tgz",
23
- "integrity": "sha512-Pb5ijPrZ89GDH8223L4UP8i6QApWxs04RbPQJTeWDV0/keR2E36MeKnyr6LYmUUvqRRI+Iv87SuF1W6ErINzYw==",
24
- "license": "MIT",
25
- "engines": {
26
- "node": ">=6.9.0"
27
- }
28
- },
29
- "node_modules/@babel/helper-validator-identifier": {
30
- "version": "7.29.7",
31
- "resolved": "https://registry.npmjs.org/@babel/helper-validator-identifier/-/helper-validator-identifier-7.29.7.tgz",
32
- "integrity": "sha512-qehxGkRj55h/ff8EMaJ+cYhyaKlHIxqYDn682wQD7RNp9UujOQsHog2uS0r2vzr4pW+sXf90NeeayjcNaX3fFg==",
33
- "license": "MIT",
34
- "engines": {
35
- "node": ">=6.9.0"
36
- }
37
- },
38
- "node_modules/@babel/parser": {
39
- "version": "7.29.9",
40
- "resolved": "https://registry.npmjs.org/@babel/parser/-/parser-7.29.9.tgz",
41
- "integrity": "sha512-CjXrNHTnvqBVqHgdBysY3vk2T8tpJHb5/RMeHJBTyVa9xgugCB0CJTx/3oO8RV2QRQP391RWpB7D6hLjm8V9uA==",
42
- "license": "MIT",
43
- "dependencies": {
44
- "@babel/types": "^7.29.8"
45
- },
46
- "bin": {
47
- "parser": "bin/babel-parser.js"
48
- },
49
- "engines": {
50
- "node": ">=6.0.0"
51
- }
52
- },
53
- "node_modules/@babel/types": {
54
- "version": "7.29.8",
55
- "resolved": "https://registry.npmjs.org/@babel/types/-/types-7.29.8.tgz",
56
- "integrity": "sha512-Vj1jF3cPfxg7OAfoI7QnVKLoILlm2JF9pnVHrX8qx7AHMiYWT+NDAA7jChlNgRS4WTLc/fD1lXLmPixluj+3Gg==",
57
- "license": "MIT",
58
- "dependencies": {
59
- "@babel/helper-string-parser": "^7.29.7",
60
- "@babel/helper-validator-identifier": "^7.29.7"
61
- },
62
- "engines": {
63
- "node": ">=6.9.0"
64
- }
65
- },
66
- "node_modules/@esbuild/aix-ppc64": {
67
- "version": "0.25.12",
68
- "resolved": "https://registry.npmjs.org/@esbuild/aix-ppc64/-/aix-ppc64-0.25.12.tgz",
69
- "integrity": "sha512-Hhmwd6CInZ3dwpuGTF8fJG6yoWmsToE+vYgD4nytZVxcu1ulHpUQRAB1UJ8+N1Am3Mz4+xOByoQoSZf4D+CpkA==",
70
- "cpu": [
71
- "ppc64"
72
- ],
73
- "dev": true,
74
- "license": "MIT",
75
- "optional": true,
76
- "os": [
77
- "aix"
78
- ],
79
- "engines": {
80
- "node": ">=18"
81
- }
82
- },
83
- "node_modules/@esbuild/android-arm": {
84
- "version": "0.25.12",
85
- "resolved": "https://registry.npmjs.org/@esbuild/android-arm/-/android-arm-0.25.12.tgz",
86
- "integrity": "sha512-VJ+sKvNA/GE7Ccacc9Cha7bpS8nyzVv0jdVgwNDaR4gDMC/2TTRc33Ip8qrNYUcpkOHUT5OZ0bUcNNVZQ9RLlg==",
87
- "cpu": [
88
- "arm"
89
- ],
90
- "dev": true,
91
- "license": "MIT",
92
- "optional": true,
93
- "os": [
94
- "android"
95
- ],
96
- "engines": {
97
- "node": ">=18"
98
- }
99
- },
100
- "node_modules/@esbuild/android-arm64": {
101
- "version": "0.25.12",
102
- "resolved": "https://registry.npmjs.org/@esbuild/android-arm64/-/android-arm64-0.25.12.tgz",
103
- "integrity": "sha512-6AAmLG7zwD1Z159jCKPvAxZd4y/VTO0VkprYy+3N2FtJ8+BQWFXU+OxARIwA46c5tdD9SsKGZ/1ocqBS/gAKHg==",
104
- "cpu": [
105
- "arm64"
106
- ],
107
- "dev": true,
108
- "license": "MIT",
109
- "optional": true,
110
- "os": [
111
- "android"
112
- ],
113
- "engines": {
114
- "node": ">=18"
115
- }
116
- },
117
- "node_modules/@esbuild/android-x64": {
118
- "version": "0.25.12",
119
- "resolved": "https://registry.npmjs.org/@esbuild/android-x64/-/android-x64-0.25.12.tgz",
120
- "integrity": "sha512-5jbb+2hhDHx5phYR2By8GTWEzn6I9UqR11Kwf22iKbNpYrsmRB18aX/9ivc5cabcUiAT/wM+YIZ6SG9QO6a8kg==",
121
- "cpu": [
122
- "x64"
123
- ],
124
- "dev": true,
125
- "license": "MIT",
126
- "optional": true,
127
- "os": [
128
- "android"
129
- ],
130
- "engines": {
131
- "node": ">=18"
132
- }
133
- },
134
- "node_modules/@esbuild/darwin-arm64": {
135
- "version": "0.25.12",
136
- "resolved": "https://registry.npmjs.org/@esbuild/darwin-arm64/-/darwin-arm64-0.25.12.tgz",
137
- "integrity": "sha512-N3zl+lxHCifgIlcMUP5016ESkeQjLj/959RxxNYIthIg+CQHInujFuXeWbWMgnTo4cp5XVHqFPmpyu9J65C1Yg==",
138
- "cpu": [
139
- "arm64"
140
- ],
141
- "dev": true,
142
- "license": "MIT",
143
- "optional": true,
144
- "os": [
145
- "darwin"
146
- ],
147
- "engines": {
148
- "node": ">=18"
149
- }
150
- },
151
- "node_modules/@esbuild/darwin-x64": {
152
- "version": "0.25.12",
153
- "resolved": "https://registry.npmjs.org/@esbuild/darwin-x64/-/darwin-x64-0.25.12.tgz",
154
- "integrity": "sha512-HQ9ka4Kx21qHXwtlTUVbKJOAnmG1ipXhdWTmNXiPzPfWKpXqASVcWdnf2bnL73wgjNrFXAa3yYvBSd9pzfEIpA==",
155
- "cpu": [
156
- "x64"
157
- ],
158
- "dev": true,
159
- "license": "MIT",
160
- "optional": true,
161
- "os": [
162
- "darwin"
163
- ],
164
- "engines": {
165
- "node": ">=18"
166
- }
167
- },
168
- "node_modules/@esbuild/freebsd-arm64": {
169
- "version": "0.25.12",
170
- "resolved": "https://registry.npmjs.org/@esbuild/freebsd-arm64/-/freebsd-arm64-0.25.12.tgz",
171
- "integrity": "sha512-gA0Bx759+7Jve03K1S0vkOu5Lg/85dou3EseOGUes8flVOGxbhDDh/iZaoek11Y8mtyKPGF3vP8XhnkDEAmzeg==",
172
- "cpu": [
173
- "arm64"
174
- ],
175
- "dev": true,
176
- "license": "MIT",
177
- "optional": true,
178
- "os": [
179
- "freebsd"
180
- ],
181
- "engines": {
182
- "node": ">=18"
183
- }
184
- },
185
- "node_modules/@esbuild/freebsd-x64": {
186
- "version": "0.25.12",
187
- "resolved": "https://registry.npmjs.org/@esbuild/freebsd-x64/-/freebsd-x64-0.25.12.tgz",
188
- "integrity": "sha512-TGbO26Yw2xsHzxtbVFGEXBFH0FRAP7gtcPE7P5yP7wGy7cXK2oO7RyOhL5NLiqTlBh47XhmIUXuGciXEqYFfBQ==",
189
- "cpu": [
190
- "x64"
191
- ],
192
- "dev": true,
193
- "license": "MIT",
194
- "optional": true,
195
- "os": [
196
- "freebsd"
197
- ],
198
- "engines": {
199
- "node": ">=18"
200
- }
201
- },
202
- "node_modules/@esbuild/linux-arm": {
203
- "version": "0.25.12",
204
- "resolved": "https://registry.npmjs.org/@esbuild/linux-arm/-/linux-arm-0.25.12.tgz",
205
- "integrity": "sha512-lPDGyC1JPDou8kGcywY0YILzWlhhnRjdof3UlcoqYmS9El818LLfJJc3PXXgZHrHCAKs/Z2SeZtDJr5MrkxtOw==",
206
- "cpu": [
207
- "arm"
208
- ],
209
- "dev": true,
210
- "license": "MIT",
211
- "optional": true,
212
- "os": [
213
- "linux"
214
- ],
215
- "engines": {
216
- "node": ">=18"
217
- }
218
- },
219
- "node_modules/@esbuild/linux-arm64": {
220
- "version": "0.25.12",
221
- "resolved": "https://registry.npmjs.org/@esbuild/linux-arm64/-/linux-arm64-0.25.12.tgz",
222
- "integrity": "sha512-8bwX7a8FghIgrupcxb4aUmYDLp8pX06rGh5HqDT7bB+8Rdells6mHvrFHHW2JAOPZUbnjUpKTLg6ECyzvas2AQ==",
223
- "cpu": [
224
- "arm64"
225
- ],
226
- "dev": true,
227
- "license": "MIT",
228
- "optional": true,
229
- "os": [
230
- "linux"
231
- ],
232
- "engines": {
233
- "node": ">=18"
234
- }
235
- },
236
- "node_modules/@esbuild/linux-ia32": {
237
- "version": "0.25.12",
238
- "resolved": "https://registry.npmjs.org/@esbuild/linux-ia32/-/linux-ia32-0.25.12.tgz",
239
- "integrity": "sha512-0y9KrdVnbMM2/vG8KfU0byhUN+EFCny9+8g202gYqSSVMonbsCfLjUO+rCci7pM0WBEtz+oK/PIwHkzxkyharA==",
240
- "cpu": [
241
- "ia32"
242
- ],
243
- "dev": true,
244
- "license": "MIT",
245
- "optional": true,
246
- "os": [
247
- "linux"
248
- ],
249
- "engines": {
250
- "node": ">=18"
251
- }
252
- },
253
- "node_modules/@esbuild/linux-loong64": {
254
- "version": "0.25.12",
255
- "resolved": "https://registry.npmjs.org/@esbuild/linux-loong64/-/linux-loong64-0.25.12.tgz",
256
- "integrity": "sha512-h///Lr5a9rib/v1GGqXVGzjL4TMvVTv+s1DPoxQdz7l/AYv6LDSxdIwzxkrPW438oUXiDtwM10o9PmwS/6Z0Ng==",
257
- "cpu": [
258
- "loong64"
259
- ],
260
- "dev": true,
261
- "license": "MIT",
262
- "optional": true,
263
- "os": [
264
- "linux"
265
- ],
266
- "engines": {
267
- "node": ">=18"
268
- }
269
- },
270
- "node_modules/@esbuild/linux-mips64el": {
271
- "version": "0.25.12",
272
- "resolved": "https://registry.npmjs.org/@esbuild/linux-mips64el/-/linux-mips64el-0.25.12.tgz",
273
- "integrity": "sha512-iyRrM1Pzy9GFMDLsXn1iHUm18nhKnNMWscjmp4+hpafcZjrr2WbT//d20xaGljXDBYHqRcl8HnxbX6uaA/eGVw==",
274
- "cpu": [
275
- "mips64el"
276
- ],
277
- "dev": true,
278
- "license": "MIT",
279
- "optional": true,
280
- "os": [
281
- "linux"
282
- ],
283
- "engines": {
284
- "node": ">=18"
285
- }
286
- },
287
- "node_modules/@esbuild/linux-ppc64": {
288
- "version": "0.25.12",
289
- "resolved": "https://registry.npmjs.org/@esbuild/linux-ppc64/-/linux-ppc64-0.25.12.tgz",
290
- "integrity": "sha512-9meM/lRXxMi5PSUqEXRCtVjEZBGwB7P/D4yT8UG/mwIdze2aV4Vo6U5gD3+RsoHXKkHCfSxZKzmDssVlRj1QQA==",
291
- "cpu": [
292
- "ppc64"
293
- ],
294
- "dev": true,
295
- "license": "MIT",
296
- "optional": true,
297
- "os": [
298
- "linux"
299
- ],
300
- "engines": {
301
- "node": ">=18"
302
- }
303
- },
304
- "node_modules/@esbuild/linux-riscv64": {
305
- "version": "0.25.12",
306
- "resolved": "https://registry.npmjs.org/@esbuild/linux-riscv64/-/linux-riscv64-0.25.12.tgz",
307
- "integrity": "sha512-Zr7KR4hgKUpWAwb1f3o5ygT04MzqVrGEGXGLnj15YQDJErYu/BGg+wmFlIDOdJp0PmB0lLvxFIOXZgFRrdjR0w==",
308
- "cpu": [
309
- "riscv64"
310
- ],
311
- "dev": true,
312
- "license": "MIT",
313
- "optional": true,
314
- "os": [
315
- "linux"
316
- ],
317
- "engines": {
318
- "node": ">=18"
319
- }
320
- },
321
- "node_modules/@esbuild/linux-s390x": {
322
- "version": "0.25.12",
323
- "resolved": "https://registry.npmjs.org/@esbuild/linux-s390x/-/linux-s390x-0.25.12.tgz",
324
- "integrity": "sha512-MsKncOcgTNvdtiISc/jZs/Zf8d0cl/t3gYWX8J9ubBnVOwlk65UIEEvgBORTiljloIWnBzLs4qhzPkJcitIzIg==",
325
- "cpu": [
326
- "s390x"
327
- ],
328
- "dev": true,
329
- "license": "MIT",
330
- "optional": true,
331
- "os": [
332
- "linux"
333
- ],
334
- "engines": {
335
- "node": ">=18"
336
- }
337
- },
338
- "node_modules/@esbuild/linux-x64": {
339
- "version": "0.25.12",
340
- "resolved": "https://registry.npmjs.org/@esbuild/linux-x64/-/linux-x64-0.25.12.tgz",
341
- "integrity": "sha512-uqZMTLr/zR/ed4jIGnwSLkaHmPjOjJvnm6TVVitAa08SLS9Z0VM8wIRx7gWbJB5/J54YuIMInDquWyYvQLZkgw==",
342
- "cpu": [
343
- "x64"
344
- ],
345
- "dev": true,
346
- "license": "MIT",
347
- "optional": true,
348
- "os": [
349
- "linux"
350
- ],
351
- "engines": {
352
- "node": ">=18"
353
- }
354
- },
355
- "node_modules/@esbuild/netbsd-arm64": {
356
- "version": "0.25.12",
357
- "resolved": "https://registry.npmjs.org/@esbuild/netbsd-arm64/-/netbsd-arm64-0.25.12.tgz",
358
- "integrity": "sha512-xXwcTq4GhRM7J9A8Gv5boanHhRa/Q9KLVmcyXHCTaM4wKfIpWkdXiMog/KsnxzJ0A1+nD+zoecuzqPmCRyBGjg==",
359
- "cpu": [
360
- "arm64"
361
- ],
362
- "dev": true,
363
- "license": "MIT",
364
- "optional": true,
365
- "os": [
366
- "netbsd"
367
- ],
368
- "engines": {
369
- "node": ">=18"
370
- }
371
- },
372
- "node_modules/@esbuild/netbsd-x64": {
373
- "version": "0.25.12",
374
- "resolved": "https://registry.npmjs.org/@esbuild/netbsd-x64/-/netbsd-x64-0.25.12.tgz",
375
- "integrity": "sha512-Ld5pTlzPy3YwGec4OuHh1aCVCRvOXdH8DgRjfDy/oumVovmuSzWfnSJg+VtakB9Cm0gxNO9BzWkj6mtO1FMXkQ==",
376
- "cpu": [
377
- "x64"
378
- ],
379
- "dev": true,
380
- "license": "MIT",
381
- "optional": true,
382
- "os": [
383
- "netbsd"
384
- ],
385
- "engines": {
386
- "node": ">=18"
387
- }
388
- },
389
- "node_modules/@esbuild/openbsd-arm64": {
390
- "version": "0.25.12",
391
- "resolved": "https://registry.npmjs.org/@esbuild/openbsd-arm64/-/openbsd-arm64-0.25.12.tgz",
392
- "integrity": "sha512-fF96T6KsBo/pkQI950FARU9apGNTSlZGsv1jZBAlcLL1MLjLNIWPBkj5NlSz8aAzYKg+eNqknrUJ24QBybeR5A==",
393
- "cpu": [
394
- "arm64"
395
- ],
396
- "dev": true,
397
- "license": "MIT",
398
- "optional": true,
399
- "os": [
400
- "openbsd"
401
- ],
402
- "engines": {
403
- "node": ">=18"
404
- }
405
- },
406
- "node_modules/@esbuild/openbsd-x64": {
407
- "version": "0.25.12",
408
- "resolved": "https://registry.npmjs.org/@esbuild/openbsd-x64/-/openbsd-x64-0.25.12.tgz",
409
- "integrity": "sha512-MZyXUkZHjQxUvzK7rN8DJ3SRmrVrke8ZyRusHlP+kuwqTcfWLyqMOE3sScPPyeIXN/mDJIfGXvcMqCgYKekoQw==",
410
- "cpu": [
411
- "x64"
412
- ],
413
- "dev": true,
414
- "license": "MIT",
415
- "optional": true,
416
- "os": [
417
- "openbsd"
418
- ],
419
- "engines": {
420
- "node": ">=18"
421
- }
422
- },
423
- "node_modules/@esbuild/openharmony-arm64": {
424
- "version": "0.25.12",
425
- "resolved": "https://registry.npmjs.org/@esbuild/openharmony-arm64/-/openharmony-arm64-0.25.12.tgz",
426
- "integrity": "sha512-rm0YWsqUSRrjncSXGA7Zv78Nbnw4XL6/dzr20cyrQf7ZmRcsovpcRBdhD43Nuk3y7XIoW2OxMVvwuRvk9XdASg==",
427
- "cpu": [
428
- "arm64"
429
- ],
430
- "dev": true,
431
- "license": "MIT",
432
- "optional": true,
433
- "os": [
434
- "openharmony"
435
- ],
436
- "engines": {
437
- "node": ">=18"
438
- }
439
- },
440
- "node_modules/@esbuild/sunos-x64": {
441
- "version": "0.25.12",
442
- "resolved": "https://registry.npmjs.org/@esbuild/sunos-x64/-/sunos-x64-0.25.12.tgz",
443
- "integrity": "sha512-3wGSCDyuTHQUzt0nV7bocDy72r2lI33QL3gkDNGkod22EsYl04sMf0qLb8luNKTOmgF/eDEDP5BFNwoBKH441w==",
444
- "cpu": [
445
- "x64"
446
- ],
447
- "dev": true,
448
- "license": "MIT",
449
- "optional": true,
450
- "os": [
451
- "sunos"
452
- ],
453
- "engines": {
454
- "node": ">=18"
455
- }
456
- },
457
- "node_modules/@esbuild/win32-arm64": {
458
- "version": "0.25.12",
459
- "resolved": "https://registry.npmjs.org/@esbuild/win32-arm64/-/win32-arm64-0.25.12.tgz",
460
- "integrity": "sha512-rMmLrur64A7+DKlnSuwqUdRKyd3UE7oPJZmnljqEptesKM8wx9J8gx5u0+9Pq0fQQW8vqeKebwNXdfOyP+8Bsg==",
461
- "cpu": [
462
- "arm64"
463
- ],
464
- "dev": true,
465
- "license": "MIT",
466
- "optional": true,
467
- "os": [
468
- "win32"
469
- ],
470
- "engines": {
471
- "node": ">=18"
472
- }
473
- },
474
- "node_modules/@esbuild/win32-ia32": {
475
- "version": "0.25.12",
476
- "resolved": "https://registry.npmjs.org/@esbuild/win32-ia32/-/win32-ia32-0.25.12.tgz",
477
- "integrity": "sha512-HkqnmmBoCbCwxUKKNPBixiWDGCpQGVsrQfJoVGYLPT41XWF8lHuE5N6WhVia2n4o5QK5M4tYr21827fNhi4byQ==",
478
- "cpu": [
479
- "ia32"
480
- ],
481
- "dev": true,
482
- "license": "MIT",
483
- "optional": true,
484
- "os": [
485
- "win32"
486
- ],
487
- "engines": {
488
- "node": ">=18"
489
- }
490
- },
491
- "node_modules/@esbuild/win32-x64": {
492
- "version": "0.25.12",
493
- "resolved": "https://registry.npmjs.org/@esbuild/win32-x64/-/win32-x64-0.25.12.tgz",
494
- "integrity": "sha512-alJC0uCZpTFrSL0CCDjcgleBXPnCrEAhTBILpeAp7M/OFgoqtAetfBzX0xM00MUsVVPpVjlPuMbREqnZCXaTnA==",
495
- "cpu": [
496
- "x64"
497
- ],
498
- "dev": true,
499
- "license": "MIT",
500
- "optional": true,
501
- "os": [
502
- "win32"
503
- ],
504
- "engines": {
505
- "node": ">=18"
506
- }
507
- },
508
- "node_modules/@jridgewell/sourcemap-codec": {
509
- "version": "1.6.0",
510
- "resolved": "https://registry.npmjs.org/@jridgewell/sourcemap-codec/-/sourcemap-codec-1.6.0.tgz",
511
- "integrity": "sha512-T7jf+5zgsZHwNJ4lvQ7/aezbyk0nNX+zJVWpmHA7VYsEx7a7qr5Rg5IbtJFqkgze5Y2sruq1RUY8Q837Od7iFw==",
512
- "license": "MIT"
513
- },
514
- "node_modules/@napi-rs/lzma-linux-x64-gnu": {
515
- "version": "1.5.1",
516
- "resolved": "https://registry.npmjs.org/@napi-rs/lzma-linux-x64-gnu/-/lzma-linux-x64-gnu-1.5.1.tgz",
517
- "integrity": "sha512-oTXEIha4SsuXdTA4Iyskj0kpdx2yVXdhd75c2v3xGrHFfVMsbhTPZU/nMPL4sWKo4pBHm3aucLaqGlF696dTyQ==",
518
- "cpu": [
519
- "x64"
520
- ],
521
- "dev": true,
522
- "license": "MIT",
523
- "optional": true,
524
- "os": [
525
- "linux"
526
- ],
527
- "engines": {
528
- "node": "^22.20 || ^24.12 || >=25"
529
- }
530
- },
531
- "node_modules/@rollup/rollup-android-arm-eabi": {
532
- "version": "4.63.4",
533
- "resolved": "https://registry.npmjs.org/@rollup/rollup-android-arm-eabi/-/rollup-android-arm-eabi-4.63.4.tgz",
534
- "integrity": "sha512-I+BSHzTAhKN2n7ZwGZsegGcZjDpLqFOMAtJz/u6uFGe0pUFbq56dEHjqJV/ZUdRJtNXNxA+hREUatZBvMR3Oiw==",
535
- "cpu": [
536
- "arm"
537
- ],
538
- "dev": true,
539
- "license": "MIT",
540
- "optional": true,
541
- "os": [
542
- "android"
543
- ]
544
- },
545
- "node_modules/@rollup/rollup-android-arm64": {
546
- "version": "4.63.4",
547
- "resolved": "https://registry.npmjs.org/@rollup/rollup-android-arm64/-/rollup-android-arm64-4.63.4.tgz",
548
- "integrity": "sha512-pu3BdjS2LtEzRu2elmGzS3fIeWSZy4BMDIaLNwjorO76+k2d0LMluijhsDx3KQyQBQ/lLUZCQA9/s6csvUfuhw==",
549
- "cpu": [
550
- "arm64"
551
- ],
552
- "dev": true,
553
- "license": "MIT",
554
- "optional": true,
555
- "os": [
556
- "android"
557
- ]
558
- },
559
- "node_modules/@rollup/rollup-darwin-arm64": {
560
- "version": "4.63.4",
561
- "resolved": "https://registry.npmjs.org/@rollup/rollup-darwin-arm64/-/rollup-darwin-arm64-4.63.4.tgz",
562
- "integrity": "sha512-xfSrj9MHnWK9GaSqT9U0ImHtH/N8WZlHLx4cZHiuLcqs640hvZ3hLPd5UR2AZS57FaE8HrRUSpltbZdWRxHiDA==",
563
- "cpu": [
564
- "arm64"
565
- ],
566
- "dev": true,
567
- "license": "MIT",
568
- "optional": true,
569
- "os": [
570
- "darwin"
571
- ]
572
- },
573
- "node_modules/@rollup/rollup-darwin-x64": {
574
- "version": "4.63.4",
575
- "resolved": "https://registry.npmjs.org/@rollup/rollup-darwin-x64/-/rollup-darwin-x64-4.63.4.tgz",
576
- "integrity": "sha512-bqU99PLJb/dqb3S0GIMdeuyAEETSUgZBoqXYd3Sd+WCsV+MmPhnN6JrotWyir31+QgH7EvvE5/mwGJlEoci8Fw==",
577
- "cpu": [
578
- "x64"
579
- ],
580
- "dev": true,
581
- "license": "MIT",
582
- "optional": true,
583
- "os": [
584
- "darwin"
585
- ]
586
- },
587
- "node_modules/@rollup/rollup-freebsd-arm64": {
588
- "version": "4.63.4",
589
- "resolved": "https://registry.npmjs.org/@rollup/rollup-freebsd-arm64/-/rollup-freebsd-arm64-4.63.4.tgz",
590
- "integrity": "sha512-JinsFZ5G40oXQb+sUuiA5x689vhr6dDYK0H0NL+rwKdL6CqnmYN8PE4ZwfRSoIjrCxqTQG/SLfTtSvHeGxoVlw==",
591
- "cpu": [
592
- "arm64"
593
- ],
594
- "dev": true,
595
- "license": "MIT",
596
- "optional": true,
597
- "os": [
598
- "freebsd"
599
- ]
600
- },
601
- "node_modules/@rollup/rollup-freebsd-x64": {
602
- "version": "4.63.4",
603
- "resolved": "https://registry.npmjs.org/@rollup/rollup-freebsd-x64/-/rollup-freebsd-x64-4.63.4.tgz",
604
- "integrity": "sha512-GAdA4UxpiNm27cLHr2GqXBpAD0x9FqwYBY7/YSP0Ss0/PNi4k8gbviqpIpYbVSRBaS2ZcegXEzgTQMbRNCwxCw==",
605
- "cpu": [
606
- "x64"
607
- ],
608
- "dev": true,
609
- "license": "MIT",
610
- "optional": true,
611
- "os": [
612
- "freebsd"
613
- ]
614
- },
615
- "node_modules/@rollup/rollup-linux-arm-gnueabihf": {
616
- "version": "4.63.4",
617
- "resolved": "https://registry.npmjs.org/@rollup/rollup-linux-arm-gnueabihf/-/rollup-linux-arm-gnueabihf-4.63.4.tgz",
618
- "integrity": "sha512-qDd6NoA1znaLjp4jR5U/KWCdLAKDJNB8W9ChbbDaKbo0xA+Atln5HK6LFCZ4oJQpemtRZA288DCirFRjrspptw==",
619
- "cpu": [
620
- "arm"
621
- ],
622
- "dev": true,
623
- "license": "MIT",
624
- "optional": true,
625
- "os": [
626
- "linux"
627
- ]
628
- },
629
- "node_modules/@rollup/rollup-linux-arm-musleabihf": {
630
- "version": "4.63.4",
631
- "resolved": "https://registry.npmjs.org/@rollup/rollup-linux-arm-musleabihf/-/rollup-linux-arm-musleabihf-4.63.4.tgz",
632
- "integrity": "sha512-WtB5Tz5KTNINb8ZA+8sQ7bmjuS1JrRT7YverYIhUGdWWDlpzVWmIwuZE+jidkEXUn1l0zrEkaIMa8dHF3NGcsA==",
633
- "cpu": [
634
- "arm"
635
- ],
636
- "dev": true,
637
- "license": "MIT",
638
- "optional": true,
639
- "os": [
640
- "linux"
641
- ]
642
- },
643
- "node_modules/@rollup/rollup-linux-arm64-gnu": {
644
- "version": "4.63.4",
645
- "resolved": "https://registry.npmjs.org/@rollup/rollup-linux-arm64-gnu/-/rollup-linux-arm64-gnu-4.63.4.tgz",
646
- "integrity": "sha512-VcQ3L1tjnkKzWjryAVaFhHEWcqOfICX9uxVVoDzm2t0DpgKRHd2zOpVrJc0xsWeBZcBFyYROCIBdyR/fS174pg==",
647
- "cpu": [
648
- "arm64"
649
- ],
650
- "dev": true,
651
- "license": "MIT",
652
- "optional": true,
653
- "os": [
654
- "linux"
655
- ]
656
- },
657
- "node_modules/@rollup/rollup-linux-arm64-musl": {
658
- "version": "4.63.4",
659
- "resolved": "https://registry.npmjs.org/@rollup/rollup-linux-arm64-musl/-/rollup-linux-arm64-musl-4.63.4.tgz",
660
- "integrity": "sha512-6+ZQX6P5s0cMDN2Ypb8Lbm2+/sZYmZjdaYny992ujUU9UKi/4CWoJWsl1pNvjWJHNHGK51m+jKGLlh1ylb2ifQ==",
661
- "cpu": [
662
- "arm64"
663
- ],
664
- "dev": true,
665
- "license": "MIT",
666
- "optional": true,
667
- "os": [
668
- "linux"
669
- ]
670
- },
671
- "node_modules/@rollup/rollup-linux-loong64-gnu": {
672
- "version": "4.63.4",
673
- "resolved": "https://registry.npmjs.org/@rollup/rollup-linux-loong64-gnu/-/rollup-linux-loong64-gnu-4.63.4.tgz",
674
- "integrity": "sha512-D72ZnvkFkBXOfzMMQLcwfPLyGkKb7HZ9/mf97B7v6/P5Lbv4oFOtSY/uHbS8lH6uKUOxoKiuokdb50XZSzzbJw==",
675
- "cpu": [
676
- "loong64"
677
- ],
678
- "dev": true,
679
- "license": "MIT",
680
- "optional": true,
681
- "os": [
682
- "linux"
683
- ]
684
- },
685
- "node_modules/@rollup/rollup-linux-loong64-musl": {
686
- "version": "4.63.4",
687
- "resolved": "https://registry.npmjs.org/@rollup/rollup-linux-loong64-musl/-/rollup-linux-loong64-musl-4.63.4.tgz",
688
- "integrity": "sha512-piU6BxeqA3O9KSu3kRCIQQtNqFFaTu21SEV4FwaRZowpnj3bLaWPZHw+xFqCs0XlJ+aOH3PTRWGoglH+mKA/OA==",
689
- "cpu": [
690
- "loong64"
691
- ],
692
- "dev": true,
693
- "license": "MIT",
694
- "optional": true,
695
- "os": [
696
- "linux"
697
- ]
698
- },
699
- "node_modules/@rollup/rollup-linux-ppc64-gnu": {
700
- "version": "4.63.4",
701
- "resolved": "https://registry.npmjs.org/@rollup/rollup-linux-ppc64-gnu/-/rollup-linux-ppc64-gnu-4.63.4.tgz",
702
- "integrity": "sha512-/5PGpHwqt2EEEOUs1XwzubE/ucr0dWDQ+to3zqi4Ds7EWpwtQ79wXc4JBoxqj/OwpawTsKWzJxHfSuBOq3DrWA==",
703
- "cpu": [
704
- "ppc64"
705
- ],
706
- "dev": true,
707
- "license": "MIT",
708
- "optional": true,
709
- "os": [
710
- "linux"
711
- ]
712
- },
713
- "node_modules/@rollup/rollup-linux-ppc64-musl": {
714
- "version": "4.63.4",
715
- "resolved": "https://registry.npmjs.org/@rollup/rollup-linux-ppc64-musl/-/rollup-linux-ppc64-musl-4.63.4.tgz",
716
- "integrity": "sha512-cX3beZDLWt7G2oJF+nhChiT+qtaihs+S2xi7ziGmVB+2pwPng6D0Ed0HmElQOgv2UsUmSJJLGwpBao/3TDx3VA==",
717
- "cpu": [
718
- "ppc64"
719
- ],
720
- "dev": true,
721
- "license": "MIT",
722
- "optional": true,
723
- "os": [
724
- "linux"
725
- ]
726
- },
727
- "node_modules/@rollup/rollup-linux-riscv64-gnu": {
728
- "version": "4.63.4",
729
- "resolved": "https://registry.npmjs.org/@rollup/rollup-linux-riscv64-gnu/-/rollup-linux-riscv64-gnu-4.63.4.tgz",
730
- "integrity": "sha512-1uz2mGWHyptR7DgHHrlbdRAjXK7v7elGZ9lMja910/RP+ZYbX6xAmCiU9UZSX4hqmgtHMv6lr5l3kq1HIOpcag==",
731
- "cpu": [
732
- "riscv64"
733
- ],
734
- "dev": true,
735
- "license": "MIT",
736
- "optional": true,
737
- "os": [
738
- "linux"
739
- ]
740
- },
741
- "node_modules/@rollup/rollup-linux-riscv64-musl": {
742
- "version": "4.63.4",
743
- "resolved": "https://registry.npmjs.org/@rollup/rollup-linux-riscv64-musl/-/rollup-linux-riscv64-musl-4.63.4.tgz",
744
- "integrity": "sha512-nLS8topojxyz7SRpKR2IODRpQ0XPZ+xaOXvT3+hqK/Uy8Lo5HFgkkIBiIrCu5tL5YqzTvgovGw55PwpahTAGig==",
745
- "cpu": [
746
- "riscv64"
747
- ],
748
- "dev": true,
749
- "license": "MIT",
750
- "optional": true,
751
- "os": [
752
- "linux"
753
- ]
754
- },
755
- "node_modules/@rollup/rollup-linux-s390x-gnu": {
756
- "version": "4.63.4",
757
- "resolved": "https://registry.npmjs.org/@rollup/rollup-linux-s390x-gnu/-/rollup-linux-s390x-gnu-4.63.4.tgz",
758
- "integrity": "sha512-gs7DRKotr3l3q+jGPQBjH0ng1FjlEDm5ueQrkw5JtQvtLyEIcLASqAEaor56BhkKRzk+IcQzrcanBdb/bBQn8g==",
759
- "cpu": [
760
- "s390x"
761
- ],
762
- "dev": true,
763
- "license": "MIT",
764
- "optional": true,
765
- "os": [
766
- "linux"
767
- ]
768
- },
769
- "node_modules/@rollup/rollup-linux-x64-gnu": {
770
- "version": "4.63.4",
771
- "resolved": "https://registry.npmjs.org/@rollup/rollup-linux-x64-gnu/-/rollup-linux-x64-gnu-4.63.4.tgz",
772
- "integrity": "sha512-791ET7W17NnScOZM7h4dX5hYspxE28htPFsb1awY/NRR8+PRNkS53e475rDdxXXDrP+kwnCcNWg9CX5ztn/Aqw==",
773
- "cpu": [
774
- "x64"
775
- ],
776
- "dev": true,
777
- "license": "MIT",
778
- "optional": true,
779
- "os": [
780
- "linux"
781
- ]
782
- },
783
- "node_modules/@rollup/rollup-linux-x64-musl": {
784
- "version": "4.63.4",
785
- "resolved": "https://registry.npmjs.org/@rollup/rollup-linux-x64-musl/-/rollup-linux-x64-musl-4.63.4.tgz",
786
- "integrity": "sha512-iwZQRcmj7g88g3tzefIrQY7qvmuA/cfYwhrDtTBhsmukO4U2huVO5W+86XacUMRvdSFVAc6kZUZy21JaRwiB9w==",
787
- "cpu": [
788
- "x64"
789
- ],
790
- "dev": true,
791
- "license": "MIT",
792
- "optional": true,
793
- "os": [
794
- "linux"
795
- ]
796
- },
797
- "node_modules/@rollup/rollup-openbsd-x64": {
798
- "version": "4.63.4",
799
- "resolved": "https://registry.npmjs.org/@rollup/rollup-openbsd-x64/-/rollup-openbsd-x64-4.63.4.tgz",
800
- "integrity": "sha512-dVHFp9gRWrdTpnqQuGfCwd7hOQDatK1VCP2iWhLY/cGrOQs/ucFzJ6A5SRqbXX12ZDI8EUuejSM5kwg+ja7Png==",
801
- "cpu": [
802
- "x64"
803
- ],
804
- "dev": true,
805
- "license": "MIT",
806
- "optional": true,
807
- "os": [
808
- "openbsd"
809
- ]
810
- },
811
- "node_modules/@rollup/rollup-openharmony-arm64": {
812
- "version": "4.63.4",
813
- "resolved": "https://registry.npmjs.org/@rollup/rollup-openharmony-arm64/-/rollup-openharmony-arm64-4.63.4.tgz",
814
- "integrity": "sha512-t3NlauOW6gxZVVFcBEnO62Cb4wbyDFL416gTg1uFI/2tgqYQlf69FbSE115Ajre9I+c26Lk4mcmdFUsS/DGifQ==",
815
- "cpu": [
816
- "arm64"
817
- ],
818
- "dev": true,
819
- "license": "MIT",
820
- "optional": true,
821
- "os": [
822
- "openharmony"
823
- ]
824
- },
825
- "node_modules/@rollup/rollup-win32-arm64-msvc": {
826
- "version": "4.63.4",
827
- "resolved": "https://registry.npmjs.org/@rollup/rollup-win32-arm64-msvc/-/rollup-win32-arm64-msvc-4.63.4.tgz",
828
- "integrity": "sha512-xWuIaSye5FWZF8+UYtVEcHtRJDN5kN9Kfgxx3Kq8XIov9KSKbc1fiqQCm90SKrgQbUXZelbnUhnlUJmfSE7P9A==",
829
- "cpu": [
830
- "arm64"
831
- ],
832
- "dev": true,
833
- "license": "MIT",
834
- "optional": true,
835
- "os": [
836
- "win32"
837
- ]
838
- },
839
- "node_modules/@rollup/rollup-win32-ia32-msvc": {
840
- "version": "4.63.4",
841
- "resolved": "https://registry.npmjs.org/@rollup/rollup-win32-ia32-msvc/-/rollup-win32-ia32-msvc-4.63.4.tgz",
842
- "integrity": "sha512-9ALJJUOg/ZflMJepVo2PlgsGxSaxN7SQ4Z8GoZfVlarWr6r3rkHUNsd/zAio7p4YMtChSMXPionxej4Hkf6CXQ==",
843
- "cpu": [
844
- "ia32"
845
- ],
846
- "dev": true,
847
- "license": "MIT",
848
- "optional": true,
849
- "os": [
850
- "win32"
851
- ]
852
- },
853
- "node_modules/@rollup/rollup-win32-x64-gnu": {
854
- "version": "4.63.4",
855
- "resolved": "https://registry.npmjs.org/@rollup/rollup-win32-x64-gnu/-/rollup-win32-x64-gnu-4.63.4.tgz",
856
- "integrity": "sha512-blj9z5qx/Pv4WU0W1NMFDB97e0JH5ed+aZGywW8WCvp/NhWX/4PFAq5uu6Q0AebNn+Vo6KzUYDT++JzTT5ojlQ==",
857
- "cpu": [
858
- "x64"
859
- ],
860
- "dev": true,
861
- "license": "MIT",
862
- "optional": true,
863
- "os": [
864
- "win32"
865
- ]
866
- },
867
- "node_modules/@rollup/rollup-win32-x64-msvc": {
868
- "version": "4.63.4",
869
- "resolved": "https://registry.npmjs.org/@rollup/rollup-win32-x64-msvc/-/rollup-win32-x64-msvc-4.63.4.tgz",
870
- "integrity": "sha512-Erx822VRBwLa124shbj+wNXe//BOgMEctDV0m1aqTQdNO1S69DgNUCFKC1RCeZfixs1J31l6igk1ziyXErbigQ==",
871
- "cpu": [
872
- "x64"
873
- ],
874
- "dev": true,
875
- "license": "MIT",
876
- "optional": true,
877
- "os": [
878
- "win32"
879
- ]
880
- },
881
- "node_modules/@types/estree": {
882
- "version": "1.0.9",
883
- "resolved": "https://registry.npmjs.org/@types/estree/-/estree-1.0.9.tgz",
884
- "integrity": "sha512-GhdPgy1el4/ImP05X05Uw4cw2/M93BCUmnEvWZNStlCzEKME4Fkk+YpoA5OiHNQmoS7Cafb8Xa3Pya8m1Qrzeg==",
885
- "dev": true,
886
- "license": "MIT"
887
- },
888
- "node_modules/@vitejs/plugin-vue": {
889
- "version": "5.2.4",
890
- "resolved": "https://registry.npmjs.org/@vitejs/plugin-vue/-/plugin-vue-5.2.4.tgz",
891
- "integrity": "sha512-7Yx/SXSOcQq5HiiV3orevHUFn+pmMB4cgbEkDYgnkUWb0WfeQ/wa2yFv6D5ICiCQOVpjA7vYDXrC7AGO8yjDHA==",
892
- "dev": true,
893
- "license": "MIT",
894
- "engines": {
895
- "node": "^18.0.0 || >=20.0.0"
896
- },
897
- "peerDependencies": {
898
- "vite": "^5.0.0 || ^6.0.0",
899
- "vue": "^3.2.25"
900
- }
901
- },
902
- "node_modules/@vue/compiler-core": {
903
- "version": "3.5.43",
904
- "resolved": "https://registry.npmjs.org/@vue/compiler-core/-/compiler-core-3.5.43.tgz",
905
- "integrity": "sha512-zdiLhnbe1QQqgDT8xZMpNmyqZ3qlI+/Q/FHQco57Kwl/b05HhCzN6eVGN9QU9rbga4CrS0H5SYY8VGHZCt/1Hg==",
906
- "license": "MIT",
907
- "dependencies": {
908
- "@babel/parser": "^7.29.8",
909
- "@vue/shared": "3.5.43",
910
- "entities": "^7.0.1",
911
- "estree-walker": "^2.0.2",
912
- "source-map-js": "^1.2.1"
913
- }
914
- },
915
- "node_modules/@vue/compiler-dom": {
916
- "version": "3.5.43",
917
- "resolved": "https://registry.npmjs.org/@vue/compiler-dom/-/compiler-dom-3.5.43.tgz",
918
- "integrity": "sha512-PEZoAk3NQmsn/ejMzSOCyTYqwGqczrWm70PuhBKjjv1+TCoQAaO/zOqNwjV+honlNstT5ILxtc+8r8UUfj+iEQ==",
919
- "license": "MIT",
920
- "dependencies": {
921
- "@vue/compiler-core": "3.5.43",
922
- "@vue/shared": "3.5.43"
923
- }
924
- },
925
- "node_modules/@vue/compiler-sfc": {
926
- "version": "3.5.43",
927
- "resolved": "https://registry.npmjs.org/@vue/compiler-sfc/-/compiler-sfc-3.5.43.tgz",
928
- "integrity": "sha512-FCbrG3XNCRl+js3huuKx4IVHBLTvMkJhVepjbxSPu1gn4yWLaYtGNQdjJGZaMytXB6qb76qQDDmDSLy/vkmleQ==",
929
- "license": "MIT",
930
- "dependencies": {
931
- "@babel/parser": "^7.29.8",
932
- "@vue/compiler-core": "3.5.43",
933
- "@vue/compiler-dom": "3.5.43",
934
- "@vue/compiler-ssr": "3.5.43",
935
- "@vue/shared": "3.5.43",
936
- "estree-walker": "^2.0.2",
937
- "magic-string": "^0.30.21",
938
- "postcss": "^8.5.28",
939
- "source-map-js": "^1.2.1"
940
- }
941
- },
942
- "node_modules/@vue/compiler-ssr": {
943
- "version": "3.5.43",
944
- "resolved": "https://registry.npmjs.org/@vue/compiler-ssr/-/compiler-ssr-3.5.43.tgz",
945
- "integrity": "sha512-GF62orf7KiJX9RqrHNGrYBudsQGD0OhJ5nDs90O8UiDuS40+YMYomiXu6w6EuvtXuRDc3MSNis3EaaSKAVSWpg==",
946
- "license": "MIT",
947
- "dependencies": {
948
- "@vue/compiler-dom": "3.5.43",
949
- "@vue/shared": "3.5.43"
950
- }
951
- },
952
- "node_modules/@vue/devtools-api": {
953
- "version": "6.6.4",
954
- "resolved": "https://registry.npmjs.org/@vue/devtools-api/-/devtools-api-6.6.4.tgz",
955
- "integrity": "sha512-sGhTPMuXqZ1rVOk32RylztWkfXTRhuS7vgAKv0zjqk8gbsHkJ7xfFf+jbySxt7tWObEJwyKaHMikV/WGDiQm8g==",
956
- "license": "MIT"
957
- },
958
- "node_modules/@vue/reactivity": {
959
- "version": "3.5.43",
960
- "resolved": "https://registry.npmjs.org/@vue/reactivity/-/reactivity-3.5.43.tgz",
961
- "integrity": "sha512-G/c9GyOZNI2jVaaS6OX1EF1SSFSv7H0ERqNTl4+DTFMlZmB5eVAB53aLQNam/7NL2NPtaDD7RdVrzf8uJzMuOA==",
962
- "license": "MIT",
963
- "dependencies": {
964
- "@vue/shared": "3.5.43"
965
- }
966
- },
967
- "node_modules/@vue/runtime-core": {
968
- "version": "3.5.43",
969
- "resolved": "https://registry.npmjs.org/@vue/runtime-core/-/runtime-core-3.5.43.tgz",
970
- "integrity": "sha512-hU6U6VnVhBGQDpvlnnDlIB8ZGJBiOcgk2lh/0InltHiz3D8oSkluvuvY+do1G2H3+udeKFsmaBlgVYP7gXQzEw==",
971
- "license": "MIT",
972
- "dependencies": {
973
- "@vue/reactivity": "3.5.43",
974
- "@vue/shared": "3.5.43"
975
- }
976
- },
977
- "node_modules/@vue/runtime-dom": {
978
- "version": "3.5.43",
979
- "resolved": "https://registry.npmjs.org/@vue/runtime-dom/-/runtime-dom-3.5.43.tgz",
980
- "integrity": "sha512-Bb2Jc0YjjJdMt1SJmb9b2L/IWd3I8lIT9x9eS/xvvP9CiVgna0ffua74xKRmt4/uSJ+0r4iN8ex1jrqkhQGWQw==",
981
- "license": "MIT",
982
- "dependencies": {
983
- "@vue/reactivity": "3.5.43",
984
- "@vue/runtime-core": "3.5.43",
985
- "@vue/shared": "3.5.43",
986
- "csstype": "^3.2.3"
987
- }
988
- },
989
- "node_modules/@vue/server-renderer": {
990
- "version": "3.5.43",
991
- "resolved": "https://registry.npmjs.org/@vue/server-renderer/-/server-renderer-3.5.43.tgz",
992
- "integrity": "sha512-l2Ygjv9NehV94PSBxNWsAHC0j/eIIKbn92mBuWXAPGLnn6HfJ6MH5ubsd+Nk0YoZ5FRuxWI1P2VSoh+dbPQhCQ==",
993
- "license": "MIT",
994
- "dependencies": {
995
- "@vue/compiler-ssr": "3.5.43",
996
- "@vue/runtime-dom": "3.5.43",
997
- "@vue/shared": "3.5.43"
998
- }
999
- },
1000
- "node_modules/@vue/shared": {
1001
- "version": "3.5.43",
1002
- "resolved": "https://registry.npmjs.org/@vue/shared/-/shared-3.5.43.tgz",
1003
- "integrity": "sha512-uksS7YGMR5NZyr4JNq0Rp+QyLns0ueaz20KwzIPW9R0LH1Vnt4E+XUM29PNseEbf1www2gOuhuDi5AKOIXag9Q==",
1004
- "license": "MIT"
1005
- },
1006
- "node_modules/commander": {
1007
- "version": "15.0.0",
1008
- "resolved": "https://registry.npmjs.org/commander/-/commander-15.0.0.tgz",
1009
- "integrity": "sha512-z67u4ZhzCL/Tydu1lJARtEZYWbWaN7oYLHbsuzocr6y4N6WZAagG3RQ4FW61V1/0+jImpj293XfrcYnd1qxtPg==",
1010
- "license": "MIT",
1011
- "engines": {
1012
- "node": ">=22.12.0"
1013
- }
1014
- },
1015
- "node_modules/csstype": {
1016
- "version": "3.2.3",
1017
- "resolved": "https://registry.npmjs.org/csstype/-/csstype-3.2.3.tgz",
1018
- "integrity": "sha512-z1HGKcYy2xA8AGQfwrn0PAy+PB7X/GSj3UVJW9qKyn43xWa+gl5nXmU4qqLMRzWVLFC8KusUX8T/0kCiOYpAIQ==",
1019
- "license": "MIT"
1020
- },
1021
- "node_modules/entities": {
1022
- "version": "7.0.1",
1023
- "resolved": "https://registry.npmjs.org/entities/-/entities-7.0.1.tgz",
1024
- "integrity": "sha512-TWrgLOFUQTH994YUyl1yT4uyavY5nNB5muff+RtWaqNVCAK408b5ZnnbNAUEWLTCpum9w6arT70i1XdQ4UeOPA==",
1025
- "license": "BSD-2-Clause",
1026
- "engines": {
1027
- "node": ">=0.12"
1028
- },
1029
- "funding": {
1030
- "url": "https://github.com/fb55/entities?sponsor=1"
1031
- }
1032
- },
1033
- "node_modules/esbuild": {
1034
- "version": "0.25.12",
1035
- "resolved": "https://registry.npmjs.org/esbuild/-/esbuild-0.25.12.tgz",
1036
- "integrity": "sha512-bbPBYYrtZbkt6Os6FiTLCTFxvq4tt3JKall1vRwshA3fdVztsLAatFaZobhkBC8/BrPetoa0oksYoKXoG4ryJg==",
1037
- "dev": true,
1038
- "hasInstallScript": true,
1039
- "license": "MIT",
1040
- "bin": {
1041
- "esbuild": "bin/esbuild"
1042
- },
1043
- "engines": {
1044
- "node": ">=18"
1045
- },
1046
- "optionalDependencies": {
1047
- "@esbuild/aix-ppc64": "0.25.12",
1048
- "@esbuild/android-arm": "0.25.12",
1049
- "@esbuild/android-arm64": "0.25.12",
1050
- "@esbuild/android-x64": "0.25.12",
1051
- "@esbuild/darwin-arm64": "0.25.12",
1052
- "@esbuild/darwin-x64": "0.25.12",
1053
- "@esbuild/freebsd-arm64": "0.25.12",
1054
- "@esbuild/freebsd-x64": "0.25.12",
1055
- "@esbuild/linux-arm": "0.25.12",
1056
- "@esbuild/linux-arm64": "0.25.12",
1057
- "@esbuild/linux-ia32": "0.25.12",
1058
- "@esbuild/linux-loong64": "0.25.12",
1059
- "@esbuild/linux-mips64el": "0.25.12",
1060
- "@esbuild/linux-ppc64": "0.25.12",
1061
- "@esbuild/linux-riscv64": "0.25.12",
1062
- "@esbuild/linux-s390x": "0.25.12",
1063
- "@esbuild/linux-x64": "0.25.12",
1064
- "@esbuild/netbsd-arm64": "0.25.12",
1065
- "@esbuild/netbsd-x64": "0.25.12",
1066
- "@esbuild/openbsd-arm64": "0.25.12",
1067
- "@esbuild/openbsd-x64": "0.25.12",
1068
- "@esbuild/openharmony-arm64": "0.25.12",
1069
- "@esbuild/sunos-x64": "0.25.12",
1070
- "@esbuild/win32-arm64": "0.25.12",
1071
- "@esbuild/win32-ia32": "0.25.12",
1072
- "@esbuild/win32-x64": "0.25.12"
1073
- }
1074
- },
1075
- "node_modules/estree-walker": {
1076
- "version": "2.0.2",
1077
- "resolved": "https://registry.npmjs.org/estree-walker/-/estree-walker-2.0.2.tgz",
1078
- "integrity": "sha512-Rfkk/Mp/DL7JVje3u18FxFujQlTNR2q6QfMSMB7AvCBx91NGj/ba3kCfza0f6dVDbw7YlRf/nDrn7pQrCCyQ/w==",
1079
- "license": "MIT"
1080
- },
1081
- "node_modules/fdir": {
1082
- "version": "6.5.0",
1083
- "resolved": "https://registry.npmjs.org/fdir/-/fdir-6.5.0.tgz",
1084
- "integrity": "sha512-tIbYtZbucOs0BRGqPJkshJUYdL+SDH7dVM8gjy+ERp3WAUjLEFJE+02kanyHtwjWOnwrKYBiwAmM0p4kLJAnXg==",
1085
- "dev": true,
1086
- "license": "MIT",
1087
- "engines": {
1088
- "node": ">=12.0.0"
1089
- },
1090
- "peerDependencies": {
1091
- "picomatch": "^3 || ^4"
1092
- },
1093
- "peerDependenciesMeta": {
1094
- "picomatch": {
1095
- "optional": true
1096
- }
1097
- }
1098
- },
1099
- "node_modules/fsevents": {
1100
- "version": "2.3.3",
1101
- "resolved": "https://registry.npmjs.org/fsevents/-/fsevents-2.3.3.tgz",
1102
- "integrity": "sha512-5xoDfX+fL7faATnagmWPpbFtwh/R77WmMMqqHGS65C3vvB0YHrgF+B1YmZ3441tMj5n63k0212XNoJwzlhffQw==",
1103
- "dev": true,
1104
- "hasInstallScript": true,
1105
- "license": "MIT",
1106
- "optional": true,
1107
- "os": [
1108
- "darwin"
1109
- ],
1110
- "engines": {
1111
- "node": "^8.16.0 || ^10.6.0 || >=11.0.0"
1112
- }
1113
- },
1114
- "node_modules/katex": {
1115
- "version": "0.19.1",
1116
- "resolved": "https://registry.npmjs.org/katex/-/katex-0.19.1.tgz",
1117
- "integrity": "sha512-jG8uKntjSEcm2EpQOh0XturpPuuC9r3Sq5HLHdsFdWZT27E672SF19QEi+of0/YeOy15lDgRhKvr5614uRrEfA==",
1118
- "funding": [
1119
- "https://opencollective.com/katex",
1120
- "https://github.com/sponsors/katex"
1121
- ],
1122
- "license": "MIT",
1123
- "dependencies": {
1124
- "commander": "^15.0.0"
1125
- },
1126
- "bin": {
1127
- "katex": "cli.js"
1128
- }
1129
- },
1130
- "node_modules/magic-string": {
1131
- "version": "0.30.21",
1132
- "resolved": "https://registry.npmjs.org/magic-string/-/magic-string-0.30.21.tgz",
1133
- "integrity": "sha512-vd2F4YUyEXKGcLHoq+TEyCjxueSeHnFxyyjNp80yg0XV4vUhnDer/lvvlqM/arB5bXQN5K2/3oinyCRyx8T2CQ==",
1134
- "license": "MIT",
1135
- "dependencies": {
1136
- "@jridgewell/sourcemap-codec": "^1.5.5"
1137
- }
1138
- },
1139
- "node_modules/nanoid": {
1140
- "version": "3.3.19",
1141
- "resolved": "https://registry.npmjs.org/nanoid/-/nanoid-3.3.19.tgz",
1142
- "integrity": "sha512-Y2tUNy4ouw6tq5oDSKeQYGOyhkUBhNOcGV/02KC+6kd9eDGqdZd++mjMiIDilrBYvjEnCYvVtsuHCuP+okSfug==",
1143
- "funding": [
1144
- {
1145
- "type": "github",
1146
- "url": "https://github.com/sponsors/ai"
1147
- }
1148
- ],
1149
- "license": "MIT",
1150
- "bin": {
1151
- "nanoid": "bin/nanoid.cjs"
1152
- },
1153
- "engines": {
1154
- "node": "^10 || ^12 || ^13.7 || ^14 || >=15.0.1"
1155
- }
1156
- },
1157
- "node_modules/picocolors": {
1158
- "version": "1.1.1",
1159
- "resolved": "https://registry.npmjs.org/picocolors/-/picocolors-1.1.1.tgz",
1160
- "integrity": "sha512-xceH2snhtb5M9liqDsmEw56le376mTZkEX/jEb/RxNFyegNul7eNslCXP9FDj/Lcu0X8KEyMceP2ntpaHrDEVA==",
1161
- "license": "ISC"
1162
- },
1163
- "node_modules/picomatch": {
1164
- "version": "4.0.7",
1165
- "resolved": "https://registry.npmjs.org/picomatch/-/picomatch-4.0.7.tgz",
1166
- "integrity": "sha512-qcJu88Q2IWqJsDD529JKMdwGm/dvInW4HvQnRwiH9JtihJvzGOscDtHE3x1pBKeUOTysQ8kVmLnJ2kJu7yhcGA==",
1167
- "dev": true,
1168
- "license": "MIT",
1169
- "engines": {
1170
- "node": ">=12"
1171
- },
1172
- "funding": {
1173
- "url": "https://github.com/sponsors/jonschlinkert"
1174
- }
1175
- },
1176
- "node_modules/postcss": {
1177
- "version": "8.5.28",
1178
- "resolved": "https://registry.npmjs.org/postcss/-/postcss-8.5.28.tgz",
1179
- "integrity": "sha512-RRuzqDtt5Y9h3quz5hWhK+TPnsmVs6WwSU6LkJMeY4HstUEDuYTG8UJSdawMRzmzAtV+KEoG8N3Qg2qLy5vM/A==",
1180
- "funding": [
1181
- {
1182
- "type": "opencollective",
1183
- "url": "https://opencollective.com/postcss/"
1184
- },
1185
- {
1186
- "type": "tidelift",
1187
- "url": "https://tidelift.com/funding/github/npm/postcss"
1188
- },
1189
- {
1190
- "type": "github",
1191
- "url": "https://github.com/sponsors/ai"
1192
- }
1193
- ],
1194
- "license": "MIT",
1195
- "dependencies": {
1196
- "nanoid": "^3.3.18",
1197
- "picocolors": "^1.1.1",
1198
- "source-map-js": "^1.2.1"
1199
- },
1200
- "engines": {
1201
- "node": "^10 || ^12 || >=14"
1202
- }
1203
- },
1204
- "node_modules/rollup": {
1205
- "version": "4.63.4",
1206
- "resolved": "https://registry.npmjs.org/rollup/-/rollup-4.63.4.tgz",
1207
- "integrity": "sha512-4U0liVayNIoLp3GFl1FcI8561WepLnZ1rqfraGh7S9B3Ur5F9S283y8Futii7RUU2C/97tOBmBy7nYvhoiOpbQ==",
1208
- "dev": true,
1209
- "license": "MIT",
1210
- "dependencies": {
1211
- "@types/estree": "1.0.9"
1212
- },
1213
- "bin": {
1214
- "rollup": "dist/bin/rollup"
1215
- },
1216
- "engines": {
1217
- "node": ">=18.0.0",
1218
- "npm": ">=8.0.0"
1219
- },
1220
- "optionalDependencies": {
1221
- "@napi-rs/lzma-linux-x64-gnu": "1.5.1",
1222
- "@rollup/rollup-android-arm-eabi": "4.63.4",
1223
- "@rollup/rollup-android-arm64": "4.63.4",
1224
- "@rollup/rollup-darwin-arm64": "4.63.4",
1225
- "@rollup/rollup-darwin-x64": "4.63.4",
1226
- "@rollup/rollup-freebsd-arm64": "4.63.4",
1227
- "@rollup/rollup-freebsd-x64": "4.63.4",
1228
- "@rollup/rollup-linux-arm-gnueabihf": "4.63.4",
1229
- "@rollup/rollup-linux-arm-musleabihf": "4.63.4",
1230
- "@rollup/rollup-linux-arm64-gnu": "4.63.4",
1231
- "@rollup/rollup-linux-arm64-musl": "4.63.4",
1232
- "@rollup/rollup-linux-loong64-gnu": "4.63.4",
1233
- "@rollup/rollup-linux-loong64-musl": "4.63.4",
1234
- "@rollup/rollup-linux-ppc64-gnu": "4.63.4",
1235
- "@rollup/rollup-linux-ppc64-musl": "4.63.4",
1236
- "@rollup/rollup-linux-riscv64-gnu": "4.63.4",
1237
- "@rollup/rollup-linux-riscv64-musl": "4.63.4",
1238
- "@rollup/rollup-linux-s390x-gnu": "4.63.4",
1239
- "@rollup/rollup-linux-x64-gnu": "4.63.4",
1240
- "@rollup/rollup-linux-x64-musl": "4.63.4",
1241
- "@rollup/rollup-openbsd-x64": "4.63.4",
1242
- "@rollup/rollup-openharmony-arm64": "4.63.4",
1243
- "@rollup/rollup-win32-arm64-msvc": "4.63.4",
1244
- "@rollup/rollup-win32-ia32-msvc": "4.63.4",
1245
- "@rollup/rollup-win32-x64-gnu": "4.63.4",
1246
- "@rollup/rollup-win32-x64-msvc": "4.63.4",
1247
- "fsevents": "~2.3.2"
1248
- }
1249
- },
1250
- "node_modules/source-map-js": {
1251
- "version": "1.2.1",
1252
- "resolved": "https://registry.npmjs.org/source-map-js/-/source-map-js-1.2.1.tgz",
1253
- "integrity": "sha512-UXWMKhLOwVKb728IUtQPXxfYU+usdybtUrK/8uGE8CQMvrhOpwvzDBwj0QhSL7MQc7vIsISBG8VQ8+IDQxpfQA==",
1254
- "license": "BSD-3-Clause",
1255
- "engines": {
1256
- "node": ">=0.10.0"
1257
- }
1258
- },
1259
- "node_modules/tinyglobby": {
1260
- "version": "0.2.17",
1261
- "resolved": "https://registry.npmjs.org/tinyglobby/-/tinyglobby-0.2.17.tgz",
1262
- "integrity": "sha512-wXR/dYpcqKmfWpEdZjiKJOwCNFndD0DMnrW/cYjVGttEkBfVgcLFHoNrlj47mjOVic9yyNu65alsgF4NQyTa2g==",
1263
- "dev": true,
1264
- "license": "MIT",
1265
- "dependencies": {
1266
- "fdir": "^6.5.0",
1267
- "picomatch": "^4.0.4"
1268
- },
1269
- "engines": {
1270
- "node": ">=12.0.0"
1271
- },
1272
- "funding": {
1273
- "url": "https://github.com/sponsors/SuperchupuDev"
1274
- }
1275
- },
1276
- "node_modules/vite": {
1277
- "version": "6.4.3",
1278
- "resolved": "https://registry.npmjs.org/vite/-/vite-6.4.3.tgz",
1279
- "integrity": "sha512-NTKlcQjlAK7MlQoyb6LgaqHc8sso/pVyUJYWMws3jg21uTJw/LddqIFPcPqP6PzpgbIcZyKI85sFE4HBrQDA8A==",
1280
- "dev": true,
1281
- "license": "MIT",
1282
- "dependencies": {
1283
- "esbuild": "^0.25.0",
1284
- "fdir": "^6.4.4",
1285
- "picomatch": "^4.0.2",
1286
- "postcss": "^8.5.3",
1287
- "rollup": "^4.34.9",
1288
- "tinyglobby": "^0.2.13"
1289
- },
1290
- "bin": {
1291
- "vite": "bin/vite.js"
1292
- },
1293
- "engines": {
1294
- "node": "^18.0.0 || ^20.0.0 || >=22.0.0"
1295
- },
1296
- "funding": {
1297
- "url": "https://github.com/vitejs/vite?sponsor=1"
1298
- },
1299
- "optionalDependencies": {
1300
- "fsevents": "~2.3.3"
1301
- },
1302
- "peerDependencies": {
1303
- "@types/node": "^18.0.0 || ^20.0.0 || >=22.0.0",
1304
- "jiti": ">=1.21.0",
1305
- "less": "*",
1306
- "lightningcss": "^1.21.0",
1307
- "sass": "*",
1308
- "sass-embedded": "*",
1309
- "stylus": "*",
1310
- "sugarss": "*",
1311
- "terser": "^5.16.0",
1312
- "tsx": "^4.8.1",
1313
- "yaml": "^2.4.2"
1314
- },
1315
- "peerDependenciesMeta": {
1316
- "@types/node": {
1317
- "optional": true
1318
- },
1319
- "jiti": {
1320
- "optional": true
1321
- },
1322
- "less": {
1323
- "optional": true
1324
- },
1325
- "lightningcss": {
1326
- "optional": true
1327
- },
1328
- "sass": {
1329
- "optional": true
1330
- },
1331
- "sass-embedded": {
1332
- "optional": true
1333
- },
1334
- "stylus": {
1335
- "optional": true
1336
- },
1337
- "sugarss": {
1338
- "optional": true
1339
- },
1340
- "terser": {
1341
- "optional": true
1342
- },
1343
- "tsx": {
1344
- "optional": true
1345
- },
1346
- "yaml": {
1347
- "optional": true
1348
- }
1349
- }
1350
- },
1351
- "node_modules/vue": {
1352
- "version": "3.5.43",
1353
- "resolved": "https://registry.npmjs.org/vue/-/vue-3.5.43.tgz",
1354
- "integrity": "sha512-o5qZoksdnjIKvW1srZ3ab7pcDNYAerBjRe54D0LBLfRdCYFrSgBHVXokMas35czQc0//lmx4/tuY4ZNQ+Rf2Ng==",
1355
- "license": "MIT",
1356
- "dependencies": {
1357
- "@vue/compiler-dom": "3.5.43",
1358
- "@vue/compiler-sfc": "3.5.43",
1359
- "@vue/runtime-dom": "3.5.43",
1360
- "@vue/server-renderer": "3.5.43",
1361
- "@vue/shared": "3.5.43"
1362
- },
1363
- "peerDependencies": {
1364
- "typescript": "*"
1365
- },
1366
- "peerDependenciesMeta": {
1367
- "typescript": {
1368
- "optional": true
1369
- }
1370
- }
1371
- },
1372
- "node_modules/vue-router": {
1373
- "version": "4.6.4",
1374
- "resolved": "https://registry.npmjs.org/vue-router/-/vue-router-4.6.4.tgz",
1375
- "integrity": "sha512-Hz9q5sa33Yhduglwz6g9skT8OBPii+4bFn88w6J+J4MfEo4KRRpmiNG/hHHkdbRFlLBOqxN8y8gf2Fb0MTUgVg==",
1376
- "license": "MIT",
1377
- "dependencies": {
1378
- "@vue/devtools-api": "^6.6.4"
1379
- },
1380
- "funding": {
1381
- "url": "https://github.com/sponsors/posva"
1382
- },
1383
- "peerDependencies": {
1384
- "vue": "^3.5.0"
1385
- }
1386
- }
1387
- }
1388
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
website/package.json DELETED
@@ -1,36 +0,0 @@
1
- {
2
- "name": "ascendgym-ux-v2-vue",
3
- "private": true,
4
- "version": "2.0.0",
5
- "type": "module",
6
- "description": "AscendGYM leaderboard UX v2.0 — entry-based (agent × engine model), STS/PAOS metrics, v2 data contract. Aggregates every eval_summary.json in the repository as the default dataset.",
7
- "scripts": {
8
- "dev": "vite",
9
- "dev:api": "node scripts/mock-api.mjs",
10
- "build": "vite build",
11
- "prebuild": "python3 scripts/build-shape-bundle.py",
12
- "data:shapes": "python3 scripts/build-shape-bundle.py",
13
- "test:shapes": "python3 scripts/build-shape-bundle.py --check && node scripts/verify-shape-results.mjs",
14
- "preview": "vite preview --port 4174 --strictPort",
15
- "bundle": "node scripts/build-bundle.mjs",
16
- "bundle:operator": "node scripts/build-operator-bundle.mjs",
17
- "test:anchors": "node scripts/verify-anchors.mjs",
18
- "test:mapping": "node scripts/audit-field-mapping.mjs",
19
- "test:precision": "node scripts/audit-precision.mjs",
20
- "test": "npm run test:anchors && npm run test:mapping && npm run test:precision && npm run test:supplement && npm run test:pricing && npm run test:pagination && npm run test:shapes && npm run test:submission",
21
- "test:submission": "node scripts/verify-submission.mjs",
22
- "data:sync": "python3 scripts/sync-supplement.py",
23
- "test:supplement": "python3 scripts/verify-supplement.py",
24
- "test:pricing": "node scripts/verify-api-pricing.mjs",
25
- "test:pagination": "node scripts/verify-pagination.mjs"
26
- },
27
- "dependencies": {
28
- "katex": "0.19.1",
29
- "vue": "^3.5.13",
30
- "vue-router": "^4.5.0"
31
- },
32
- "devDependencies": {
33
- "@vitejs/plugin-vue": "^5.2.1",
34
- "vite": "^6.1.0"
35
- }
36
- }
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
website/public/assets/harnesses/SOURCES.md DELETED
@@ -1,9 +0,0 @@
1
- # Harness marks
2
-
3
- - `claude-code.svg`: Claude Code mark from the Simple Icons collection (CC0), https://github.com/simple-icons/simple-icons/blob/develop/icons/claudecode.svg.
4
- - `opencode.svg`: OpenCode favicon from its official repository (MIT), https://github.com/anomalyco/opencode/blob/dev/packages/ui/src/assets/favicon/favicon.svg.
5
- - Codex uses the existing OpenAI mark in `../vendors/openai.svg`; see `../vendors/SOURCES.md`.
6
-
7
- The Claude Code mark is rendered in the dashboard’s Claude orange (`#D97757`).
8
-
9
- Retrieved 2026-10-25. Marks identify the corresponding harness and do not imply endorsement.
 
 
 
 
 
 
 
 
 
 
website/public/assets/harnesses/claude-code.svg DELETED
website/public/assets/harnesses/opencode.svg DELETED
website/public/assets/logo/favicon.svg DELETED
website/public/assets/logo/lockup-primary.svg DELETED
website/public/assets/logo/mark-monochrome.svg DELETED
website/public/assets/resources/SOURCES.md DELETED
@@ -1,9 +0,0 @@
1
- # Resource logos
2
-
3
- Retrieved 25 October 2026 from the official sources:
4
-
5
- - arXiv small logomark: https://info.arxiv.org/brand/logos.html ; image: https://info.arxiv.org/brand/images/brand-logomark-primary.jpg . The image is displayed with surrounding whitespace clipped by the icon container.
6
- - Hugging Face logo: https://huggingface.co/brand ; asset: https://huggingface.co/datasets/huggingface/brand-assets/resolve/main/hf-logo.svg?download=true
7
- - GitHub mark: https://primer.style/octicons/icon/mark-github-16/ ; asset: https://raw.githubusercontent.com/primer/octicons/main/icons/mark-github-24.svg . Octicons license: https://github.com/primer/octicons/blob/main/LICENSE
8
-
9
- Brand names and logos belong to their respective owners. Resource destinations retain the existing local-report and placeholder behavior.
 
 
 
 
 
 
 
 
 
 
website/public/assets/resources/arxiv.jpg DELETED

Git LFS Details

  • SHA256: deee1e08ef17db9f06342e1188b07ec1adfe2d5e574c1712efa313b118631f98
  • Pointer size: 129 Bytes
  • Size of remote file: 3.13 kB
website/public/assets/resources/github.svg DELETED
website/public/assets/resources/huggingface.svg DELETED
website/public/assets/vendors/SOURCES.md DELETED
@@ -1,16 +0,0 @@
1
- # Vendor mark sources
2
-
3
- - OpenAI, Anthropic, Google, DeepSeek and Alibaba Cloud: Simple Icons SVG collection, retrieved 2026-07-18. Simple Icons is distributed under CC0; brand marks may remain subject to trademark rules.
4
- - KIMI: official KIMI Brand Guidelines (`moonshotai.github.io/Branding-Guide/`), K-only light-background SVG, retrieved 2026-07-18.
5
- - Z.AI: official-source mark mirrored by Wikimedia Commons (`File:Z.ai (company logo).svg`); simplified to the visible geometric paths for compact UI use.
6
- - MiniMax, Mistral AI, X (xAI / Grok), Meta, QWen, ByteDance and Baidu: Simple Icons SVG collection, retrieved 2026-09-23 (CC0).
7
-
8
- These marks are used only to identify benchmark vendors in this illustrative prototype. They do not imply endorsement.
9
-
10
- ## Rendering rules
11
-
12
- - Marks render inside a circular chip (`.vendor-chip`) with a colour ring; the
13
- registry is `src/lib/vendors.js`.
14
- - A vendor with no mark here falls back to a **monogram in the same ring**. It
15
- never falls back to another vendor's logo — misattributing an entry is a
16
- data-fidelity bug, not a cosmetic shortcut.
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
website/public/assets/vendors/alibabacloud.svg DELETED
website/public/assets/vendors/anthropic.svg DELETED
website/public/assets/vendors/baidu.svg DELETED
website/public/assets/vendors/bytedance.svg DELETED
website/public/assets/vendors/deepseek.svg DELETED
website/public/assets/vendors/google.svg DELETED
website/public/assets/vendors/kimi.svg DELETED
website/public/assets/vendors/meta.svg DELETED
website/public/assets/vendors/minimax.svg DELETED
website/public/assets/vendors/mistral.svg DELETED
website/public/assets/vendors/openai.svg DELETED