v1.3.8

Install: pip install freesolo-flash==1.3.8

PyPI: https://pypi.org/project/freesolo-flash/1.3.8/

What's Changed

• fix: fetch both console artifact names so the terminal tail is reachable by @DavidBShan in #1286
• chore: bump the verl pin to the whole-adapter lora bucket sync by @DavidBShan in #1287
• fix: put fused-expert lora on fsdp2 so peft can parametrize the expert tensor by @DavidBShan in #1288
• fix: import the preload surface above main so the weight-cache cli runs by @DavidBShan in #1281
• chore: upgrade freesolo sdk to 0.4.2 by @DavidBShan in #1290
• fix: align training charges with accepted estimates by @DavidBShan in #1282
• feat: retire small qwen models and add qwen3.8 27b by @DavidBShan in #1291
• feat: add durable hosted serving accounting by @DavidBShan in #1295
• feat: add strict managed chat parity by @DavidBShan in #1292
• feat: migrate dense grpo to fsdp2 by @DavidBShan in #1296
• chore: bump the python-deps group with 7 updates by @dependabot[bot] in #1303
• ci: bump astral-sh/setup-uv from 6.8.0 to 10.0.1 by @dependabot[bot] in #1302
• ci: bump docker/setup-buildx-action from 4.2.0 to 4.3.0 by @dependabot[bot] in #1300
• ci: bump actions/setup-python from 6.3.0 to 7.0.0 by @dependabot[bot] in #1299
• ci: bump actions/checkout from 6.1.0 to 7.0.1 by @dependabot[bot] in #1301
• refactor: organize flash source into bounded subpackages by @DavidBShan in #1289
• fix: validate grpo updates by gradient norm by @DavidBShan in #1298
• fix: harden grpo resume durability by @DavidBShan in #1304
• feat: add openai chat sampling parity by @DavidBShan in #1305
• fix: backport fp8 moe lora activation handling by @DavidBShan in #1297
• refactor: move serving profiles into deployment by @DavidBShan in #1306
• feat: add catalog-wide modal serve profiles by @DavidBShan in #1312
• feat: report training lifecycle evidence by @DavidBShan in #1314
• feat: require permanent checkpoint identities by @DavidBShan in #1313
• feat: activate hosted 27b serving and cap modal gpus per model by @DavidBShan in #1333
• fix: forward training job-scope headers to the backend authorizer by @DavidBShan in #1344
• fix: uncap hosted engine autoscaling and add one buffer container by @DavidBShan in #1345
• refactor: delete private helpers left behind by earlier refactors by @DavidBShan in #1355
• fix: pack padded opd teacher tensors before verl response slicing by @DavidBShan in #1346
• refactor: centralize failure retry authorization by @DavidBShan in #1315
• feat: add repeatable hosted inference load-testing harness by @DavidBShan in #1330
• fix: retry hosted capacity responses in smoke by @DavidBShan in #1318
• fix: reject authenticated redirects by @DavidBShan in #1308
• fix: preserve provider-authoritative zero training cost by @DavidBShan in #1323
• fix: isolate git environment credentials by @DavidBShan in #1325
• fix: authorize deployment routes before taking the per-run deploy lock by @DavidBShan in #1336
• fix: reject invalid environment source bytes by @DavidBShan in #1321
• fix: classify done marker failures by @DavidBShan in #1326
• fix: preflight serving health before adapter registration by @DavidBShan in #1332
• fix: disable packaged stream buffering by @DavidBShan in #1328
• feat: add b200 and b300 serving tiers and reject unknown gpu strings by @DavidBShan in #1347
• feat: persist hosted serving capacity telemetry and align modal admission with engine decode capacity by @DavidBShan in #1372
• fix: tear down the whole worker process group on the wall-clock cap by @DavidBShan in #1331
• refactor: make run and attempt distinct identities across the lifecycle by @DavidBShan in #1341
• fix: preserve prepared opd prompt state by @DavidBShan in #1320
• feat: add mirror tool-calling parity by @DavidBShan in #1307
• feat: gate serving promotion on exact-head streaming proof by @DavidBShan in #1334
• fix: publish score batchers transactionally by @DavidBShan in #1322
• docs: correct the parameter opener indexing cost bound by @DavidBShan in #1374
• fix: contain Vast result retrieval by @DavidBShan in #1316
• refactor: require explicit org attribution for hosted serving usage by @DavidBShan in #1343
• feat: add cancellable stream channel protocol by @DavidBShan in #1319
• feat: move all three hosted serving tiers to b200 by @DavidBShan in #1376
• feat: make the sft lora plus ratio configurable by @tomzhengy in #1390
• chore: reconcile main into dev to unblock #1392 by @DavidBShan in #1398
• ci: bump astral-sh/setup-uv from 10.0.1 to 10.3.0 by @dependabot[bot] in #1397
• build(deps): bump multidict from 6.7.1 to 6.9.1 by @dependabot[bot] in #1395
• build(deps): bump fsspec from 2026.4.0 to 2026.6.0 by @dependabot[bot] in #1394
• build(deps): bump urllib3 from 2.7.0 to 2.8.0 by @dependabot[bot] in #1388
• build(deps): bump pyjwt from 2.13.0 to 2.15.0 by @dependabot[bot] in #1387
• build(deps): bump anyio from 4.13.0 to 4.14.2 by @dependabot[bot] in #1381
• chore: bump the python-deps group across 1 directory with 13 updates by @dependabot[bot] in #1399
• ci: bump docker/build-push-action from 7.3.0 to 7.4.0 by @dependabot[bot] in #1382
• ci: bump docker/setup-buildx-action from 4.3.0 to 4.4.1 by @dependabot[bot] in #1383
• fix: rank top logprobs before truncating candidates by @DavidBShan in #1401
• fix: retain main ancestry for the dev release (merge commit required) by @DavidBShan in #1403
• fix: block permanently unverifiable billing prechecks before paid work by @DavidBShan in #1353
• fix: preserve worker results on bootstrap reentry by @tomzhengy in #1396
• fix: align Modal deployment and test pins with SDK 1.6.1 by @DavidBShan in #1404
• fix: read hosted adapter provenance under the engine's own rules by @DavidBShan in #1365
• fix: make serving promotion attributable and rollback compatible by @DavidBShan in #1400
• fix: bound concurrent deployment job admission by @DavidBShan in #1364
• fix: revalidate control-plane authorization on every request by @DavidBShan in #1356
• fix: bound serving launcher cleanup by @DavidBShan in #1329
• fix: frame hosted usage streams per choice by @DavidBShan in #1352
• fix: date usage lease renewals from the monotonic clock by @DavidBShan in #1351
• fix: reuse worker dataset loader for packaged profiling by @DavidBShan in #1354
• fix: release status report state using durable backend ordering by @DavidBShan in #1363
• Cap hosted GPU pools and remove B300 and directory file limits by @DavidBShan in #1409
• Remove LoRA+ completely by @DavidBShan in #1408
• fix(runner): reject invalid status updates by @DavidBShan in #1411
• refactor: make runtime dependencies and internal call contracts explicit by @DavidBShan in #1412
• refactor: use FSDP2 exclusively for training and resume by @DavidBShan in #1413
• Remove low-value worker and hosted serving paths by @DavidBShan in #1410
• feat: migrate Flash training and serving entirely to Modal by @DavidBShan in #1406
• build(deps): bump mcp from 1.28.1 to 1.30.0 by @dependabot[bot] in #1415
• refactor: track deployment readiness per checkpoint by @DavidBShan in #1416
• fix(serve): wait for terminal Modal containers to drain by @DavidBShan in #1418
• build(deps): bump aiohttp from 3.14.3 to 3.14.4 by @dependabot[bot] in #1436
• chore: bump modal from 1.6.1 to 1.6.2 in the python-deps group by @dependabot[bot] in #1437
• build(deps-dev): bump pillow from 12.3.0 to 13.0.0 by @dependabot[bot] in #1440
• build(deps): bump uv from 0.12.5 to 0.13.4 by @dependabot[bot] in #1439
• chore: bump pillow from 12.3.0 to 13.0.0 by @dependabot[bot] in #1438
• fix: preserve GRPO metrics under interleaved worker logs by @DavidBShan in #1417
• perf: keep packed SFT batches prefetched between epochs by @DavidBShan in #1430
• Unify hosted serving model identifiers and contract capability by @DavidBShan in #1434
• perf: bake training kernels in the correct runtime by @DavidBShan in #1421
• perf: avoid duplicate final GRPO adapter exports by @DavidBShan in #1422
• perf: omit immutable base weights from LoRA recovery checkpoints by @DavidBShan in #1431
• fix(cli): enforce deployment wait deadlines across HTTP responses by @DavidBShan in #1427
• fix: bind checkpoint verification to the serving generation by @DavidBShan in #1435
• perf: release GPUs before final checkpoint uploads by @DavidBShan in #1432
• perf: use eight artifact upload threads while preserving atomic commits by @DavidBShan in #1426
• fix(serving): require settled canary usage before promotion by @DavidBShan in #1425
• refactor: unify completion and resource cleanup policy by @DavidBShan in #1441
• refactor: make checkpoint publication results explicit by @DavidBShan in #1442
• refactor: replace worker globals with explicit attempt context by @DavidBShan in #1443
• refactor: remove billing reconciliation pipelines by @DavidBShan in #1444
• Own and bound training and recovery background work by @DavidBShan in #1414
• perf: cache worker downloads and persist CI build layers by @DavidBShan in #1424
• fix(serve): revoke interrupted checkpoint registrations by @DavidBShan in #1419
• fix: close release review cleanup and kernel cache gaps by @DavidBShan in #1446
• Simplify checkpoint deployment lifecycle to five states by @DavidBShan in #1448
• refactor: replace training patches with pinned upstream verl by @DavidBShan in #1420
• refactor: audit function names and clarify side effects by @DavidBShan in #1445
• Use permanent checkpoint identities in hosted load tests by @DavidBShan in #1449
• fix: correctness audit across training, serving and recovery by @DavidBShan in #1447
• Protect durable checkpoint sources from artifact cleanup by @DavidBShan in #1450
• Dev by @DavidBShan in #1405

Full Changelog: v1.2.111.post1...v1.3.8