Commit Graph

Select branches

Hide Pull Requests

experiment/dynamic-extension-modules

feat/ds4-layered-prefill

feat/flash-next-layered-prefill

feat/qwen27b-layered-prefill

feat/sched-prefill-yield

fedeizzo/hrx-integration

main

perf/qwen27b-dflash2-npu

#1001

#1003

#1006

#1007

#1008

#1009

#1010

#1011

#1012

#1013

#1014

#1016

#1017

#1021

#1022

#1023

#1026

#1027

#1027/merge

#1029

#1031

#1032

#1033

#1034

#1035

#1036

#1036/merge

#1037

#1038

#1039

#1039/merge

#1041

#1042

#1042/merge

#1043

#1044

#1045

#1046

#1047

#1048

#1049

#1050

#1050/merge

#1052

#1054

#1055

#1059

#1061

#1062

#1063

#1064

#1065

#1068

#1068/merge

#1069

#1070

#1071

#1072

#1073

#1074

#1076

#1076/merge

#1077

#1077/merge

#1079

#1079/merge

#1080

#1080/merge

#183

#185

#186

#198

#199

#200

#209

#210

#212

#213

#218

#219

#220

#221

#223

#225

#226

#227

#228

#229

#230

#231

#234

#240

#241

#242

#243

#244

#246

#247

#249

#250

#251

#252

#255

#257

#269

#274

#276

#279

#280

#281

#282

#283

#284

#286

#288

#289

#292

#293

#294

#295

#296

#297

#299

#301

#305

#306

#307

#308

#309

#312

#314

#317

#319

#320

#321

#322

#323

#324

#327

#330

#332

#334

#342

#344

#348

#350

#351

#352

#353

#355

#356

#358

#359

#360

#361

#362

#363

#365

#366

#367

#369

#372

#373

#377

#378

#379

#380

#381

#382

#384

#385

#386

#389

#390

#391

#392

#393

#394

#396

#397

#398

#400

#402

#403

#404

#405

#406

#407

#408

#409

#412

#413

#415

#416

#420

#421

#422

#423

#424

#425

#426

#428

#432

#434

#435

#436

#437

#438

#440

#441

#443

#444

#445

#446

#447

#448

#449

#453

#458

#459

#460

#463

#464

#466

#467

#468

#469

#470

#471

#472

#473

#476

#477

#479

#480

#481

#482

#483

#484

#485

#488

#489

#490

#491

#493

#494

#496

#498

#499

#500

#501

#503

#504

#505

#506

#507

#508

#509

#510

#511

#512

#513

#515

#517

#518

#519

#520

#521

#523

#524

#525

#526

#527

#528

#529

#530

#531

#532

#533

#534

#535

#536

#539

#540

#541

#543

#544

#545

#546

#547

#548

#549

#550

#551

#552

#553

#554

#555

#556

#557

#561

#563

#565

#566

#567

#568

#568/merge

#569

#570

#571

#572

#573

#574

#577

#578

#579

#582

#583

#585

#586

#589

#590

#591

#592

#596

#602

#603

#605

#606

#607

#608

#609

#610

#613

#614

#615

#616

#618

#621

#622

#623

#624

#625

#626

#629

#630

#633

#635

#636

#637

#638

#639

#640

#643

#644

#645

#646

#647

#649

#650

#651

#653

#654

#655

#656

#657

#658

#659

#660

#661

#663

#664

#665

#667

#668

#669

#672

#674

#675

#676

#677

#679

#680

#681

#682

#682/merge

#683

#686

#688

#690

#691

#692

#694

#695

#697

#698

#699

#700

#702

#703

#704

#705

#707

#708

#709

#710

#711

#712

#713

#714

#722

#729

#730

#731

#732

#733

#734

#735

#736

#738

#739

#740

#741

#742

#743

#744

#746

#747

#748

#749

#750

#751

#752

#754

#755

#756

#757

#758

#759

#765

#766

#767

#768

#769

#770

#773

#774

#775

#776

#777

#778

#779

#782

#783

#784

#785

#787

#788

#790

#791

#792

#793

#794

#795

#796

#799

#800

#802

#805

#806

#808

#810

#811

#813

#814

#815

#816

#817

#818

#819

#821

#822

#823

#826

#827

#828

#830

#831

#832

#833

#834

#836

#839

#840

#841

#842

#844

#845

#846

#847

#849

#850

#851

#854

#855

#856

#857

#858

#859

#860

#861

#861/merge

#862

#866

#868

#869

#870

#871

#872

#874

#875

#876

#877

#878

#879

#88

#880

#881

#882

#883

#884

#885

#885/merge

#886

#887

#888

#890

#891

#892

#893

#893/merge

#894

#895

#897

#899

#900

#900/merge

#902

#903

#904

#904/merge

#905

#907

#908

#910

#910/merge

#911

#912

#912/merge

#913

#914

#915

#919

#922

#923

#924

#927

#929

#930

#931

#933

#934

#934/merge

#935

#936

#938

#939

#940

#941

#942

#943

#943/merge

#944

#944/merge

#945

#945/merge

#946

#947

#950

#951

#953

#954

#955

#956

#958

#958/merge

#959

#961

#962

#963

#964

#965

#966

#967

#968

#969

#969/merge

#981

#983

#984

#985

#988

#989

#990

#991

#992

#993

#995

#996

#998

#998/merge

#999

research/cache-redesign-2026-10-08

v0.1.0

v0.1.1

v0.10.0

v0.11.0

v0.12.0

v0.13.0

v0.14.0

v0.15.0

v0.15.1

v0.16.0

v0.17.0

v0.17.1

v0.18.0

v0.19.0

v0.2.0

v0.20.0

v0.21.0

v0.22.0

v0.23.0

v0.24.0

v0.25.0

v0.26.0

v0.27.0

v0.28.0

v0.28.1

v0.29.0

v0.3.0

v0.30.0

v0.31.0

v0.4.0

v0.5.0

v0.6.0

v0.7.0

v0.7.1

v0.8.0

v0.8.1

v0.9.0

v0.9.1

Mono Color

• 9c9a99b01e feat(ds4): batch decodes between layer groups in their own workspace, DSpark included feat/ds4-layered-prefill Francesco Bozzo 2026-11-07 19:22:36 +01:00
• 80e7eca557 feat(ext): experiment with extensions built as loadable modules experiment/dynamic-extension-modules Federico Izzo 2026-11-07 15:41:47 +01:00
• ffb0fbff12 fix(qwen38-flash-next): keep the deferred n-gram prefetch across a decode in the layer hook feat/flash-next-layered-prefill Francesco Bozzo 2026-11-05 23:34:50 +01:00
• a59d56c613 feat(qwen27b): split prompt passes by layers so other sessions keep decoding feat/qwen27b-layered-prefill Francesco Bozzo 2026-11-07 16:40:54 +01:00
• 12b09d67f7 feat(ds4): split prompt passes by layers so other sessions keep decoding Francesco Bozzo 2026-11-07 16:41:03 +01:00
• 78fedd5fef feat(qwen38-flash-next): batch every decoding session in each round between layer groups Francesco Bozzo 2026-11-05 23:02:23 +01:00
• 80dc415c16 feat(qwen38-flash-next): split prompt passes by layers so other sessions keep decoding Francesco Bozzo 2026-11-05 22:13:50 +01:00
• 2492ec7b61 fix(sched): let a prompt one pass finishes go before longer peers beside decoders feat/sched-prefill-yield Francesco Bozzo 2026-11-07 18:27:43 +01:00
• a994fc67be fix(serve): step every decoder once in each round between layer groups Francesco Bozzo 2026-11-07 06:28:48 +01:00
• ae136a9a49 fix(serve): resume decoders whose capture finished between layer groups Francesco Bozzo 2026-11-07 06:13:56 +01:00
• d36c5c6b13 fix(serve): leave decode rounds between layer groups out of the prompt's prefill time Francesco Bozzo 2026-11-06 23:52:03 +01:00
• 04482594e1 fix(sched): size layer groups by the pass's own tokens and let the conformance check batch decoders Francesco Bozzo 2026-11-05 22:58:19 +01:00
• ed55c89a42 test(ext): cover yields around cancellation and captures, and allow stall boundaries Francesco Bozzo 2026-11-05 22:35:42 +01:00
• 973b5fe480 feat(ext): let text runners split prompt passes by layers so other sessions keep decoding Francesco Bozzo 2026-11-05 21:08:25 +01:00
• 4938016878 feat(ext): let text providers declare their speculation defaults (#1047) main Francesco MocaandClaude Opus 5.5 2026-11-07 15:51:26 +01:00
• f6dc1ded7e build(ext): require extension pkg-config packages only for selected extensions (#1074) Federico Izzo 2026-11-07 14:55:41 +01:00
• e1201933ca feat(ext): open GGUF tensors core cannot size for providers that accept them (#966) Thiago Oliveira SantosandFederico Izzo 2026-11-07 10:21:56 -03:00
• fb619ba141 feat(ext): let text runners see the whole prompt end during prefill (#1069) 3 people 2026-11-07 14:02:29 +01:00
• aac3f843f9 feat(ext): add optional text hidden-state encoding to text runners (#1073) Federico IzzoandD3rPapts 2026-11-07 13:48:34 +01:00
• 832ecdbbd3 feat(ext): deliver declared provider options to text runners and commands (#1046) 3 people 2026-11-07 13:37:21 +01:00
• 80786f73d9 perf(qwen38-flash-next): overlap the shared expert with routed experts on decode (#1048) Vincent Kelleher 2026-11-07 13:36:58 +01:00
• c229292182 perf(image): decode PNG with new Rust png crate (#913) Shawn WilsherandFederico Izzo 2026-11-07 03:42:31 -08:00
• 8f2dc4e03f ci: diff the extension API against the merged main parent (#1072) Federico Izzo 2026-11-07 12:39:42 +01:00
• 61c0de89bd test(ds4): gate ds4.target on the official continuations and fix ds4.serving and ds4.dspark (#1070) Shawn Wilsher 2026-11-07 03:02:26 -08:00
• 794972d6a3 test(functional): run DeepSeek-V4-Flash in Devlab with a curated suite set (#1044) Shawn Wilsher 2026-11-07 02:06:21 -08:00
• f07fd2fefc fix(sampling): end a reasoning budget when the constraint grammar leaves reasoning (#1064) Shawn Wilsher 2026-11-06 13:59:56 -08:00
• 0e2cab6cc2 perf(ds4): faster DeepSeek V4 Flash decode at long context (#1062) Shawn WilsherandFrancesco Bozzo 2026-11-06 12:27:11 -08:00
• 800eb68d12 fix(ds4): compare GGUF float metadata with a relative tolerance (#1065) Shawn Wilsher 2026-11-06 12:23:50 -08:00
• 0d12f7b7f2 fix(serve): refuse a RAM cache the host cannot hold with the remedies instead of bad_alloc (#1063) 3 people 2026-11-06 20:03:53 +01:00
• 5d2bf02f6f test: order NULL-stream fills and uploads ahead of non-blocking stream launches (#1055) piotrrak 2026-11-06 17:51:27 +01:00
• de29d6a7b2 fix(json): accept annotation keywords such as $schema in strict schemas (#1059) Francesco Bozzo 2026-11-06 16:07:12 +01:00
• b7cf081da2 fix(qwen38_flash_next): decline checkpoints with unsupported expert formats in the probe (#1041) storm1erandFrancesco Bozzo 2026-11-06 14:09:11 +01:00
• 0f251d4a85 fix(json): accept general type unions in strict tool schemas (#1052) Francesco Bozzo 2026-11-06 13:31:09 +01:00
• 211dba49fa fix(ds4): build the Q8 to F16 weight cache in one pass so output never depends on free memory (#1045) Shawn Wilsher 2026-11-06 01:44:05 -08:00
• 1175eebeb8 fix(serve): keep a tool opener's leading separator out of required structured replies (#1043) Shawn Wilsher 2026-11-06 00:50:26 -08:00
• 6cc8a42ca8 refactor(sched): move the text scheduler's work-unit policy into src/sched (#1037) Francesco Bozzo 2026-11-05 21:07:30 +01:00
• fd6e228f10 ci(release): wait for the current PR validation check (#1038) Francesco Bozzo 2026-11-05 21:02:44 +01:00
• be7bd3be30 chore(main): release 0.31.0 (#1035) v0.31.0 gufo-release-bot[bot] 2026-11-05 20:44:48 +01:00
• ce61761085 fix(qwen): prepare the vision sidecar before serving again (#1034) Sami SilmaniandClaude Opus 5.5 2026-11-05 20:10:41 +01:00
• 396aa8fa90 feat(ext): let shipped and unshipped extensions share provider options (#1033) Federico Izzo 2026-11-05 19:35:11 +01:00
• b2e132b581 chore(qwen38-flash-next): move to the Unsloth Dynamic 3.1 UD-Q4_K_XL weights (#1031) Francesco Bozzo 2026-11-05 19:06:05 +01:00
• 55c21af075 feat(ext): add public link targets for extension executables (#1032) Federico Izzo 2026-11-05 18:41:55 +01:00
• 2c97b72d86 feat(ext): let text providers register native tool dialects (#1013) Francesco MocaandClaude Opus 5.5 2026-11-05 14:58:13 +01:00
• d02a1d2d5f feat(ext): keep indivisible prompt spans whole across prefill stops and checkpoints (#1011) 3 people 2026-11-05 13:56:53 +01:00
• d86d4e7cd2 ci(release): run release automation only on manual dispatch (#1029) Francesco Bozzo 2026-11-05 13:08:56 +01:00
• 2728aa7e22 feat(ext): let text providers declare their reasoning delimiters (#1008) 3 people 2026-11-05 13:01:15 +01:00
• 555f859fc3 feat(cache): add window-row components for sliding-window attention (#1010) 3 people 2026-11-05 13:00:19 +01:00
• 311531543e feat(ext): let text providers supply sampling defaults and restore them in prompt and chat (#1009) 3 people 2026-11-05 11:26:39 +01:00
• 65fe68a9f3 refactor(ext): make the component cache the only continuation contract (#1026) Federico Izzo 2026-11-05 10:11:46 +01:00
• 409cb6ff0e test(functional): allow the aligned stable boundary in pressure retries (#1022) Francesco Bozzo 2026-11-04 23:21:34 +01:00
• 118b7c63f8 fix(gguf): open tensorless split shards that end without data padding (#1021) Francesco Bozzo 2026-11-04 23:07:55 +01:00
• f801b8cac0 chore(main): release 0.30.0 (#1017) v0.30.0 gufo-release-bot[bot] 2026-11-04 23:00:10 +01:00
• 186a507da7 fix(qwen38-flash-next): end capacity-limited prefill steps on the selection block (#1016) Francesco Bozzo 2026-11-04 21:50:50 +01:00
• 4fc4a2236b feat(cache): remove the legacy continuation cache (#1014) Federico Izzo 2026-11-04 20:16:09 +01:00
• da04ba521b fix(json): keep tool argument schemas beside arrays without items (#1012) Francesco Moca 2026-11-04 19:25:02 +01:00
• 6c1ac518cc test(cache): run serving tests on the component cache (#1007) Federico Izzo 2026-11-04 18:13:16 +01:00
• 6667372738 perf(qwen38-flash-next): capture prompt checkpoints inside prefill passes without splitting them (#963) 3 people 2026-11-04 16:33:22 +01:00
• 346db1d075 feat(ext): add a public link target for the component cache (#1006) Federico Izzo 2026-11-04 16:30:40 +01:00
• b29229e349 feat(cache): migrate DeepSeek V4 Flash to the component cache (#1003) Federico Izzo 2026-11-04 16:09:24 +01:00
• 883f67b609 chore(main): release 0.29.0 (#1001) v0.29.0 gufo-release-bot[bot] 2026-11-04 15:53:06 +01:00
• e28fef40aa feat(qwen38-flash-next): scale YaRN with the context a session holds (#905) 3 people 2026-11-04 13:20:27 +01:00
• 6905be287e fix(serve): mark realtime responses active before announcing them (#999) Francesco Bozzo 2026-11-04 11:37:40 +01:00
• 18c89e3ba1 fix(cache): align Flash-Next prefill and keep disk readahead progressing (#877) 3 people 2026-11-04 10:49:34 +01:00
• 1c75887ba3 perf(qwen): specialize the K-quant warp dot and run K-quant GEMV one row per warp (#995) Sami SilmaniandClaude Opus 5.5 2026-11-04 02:01:34 +01:00
• d2b01f24a9 perf(qwen27b): verify long prompt-lookup copies in one wide pass (#992) Lukáš HanákandClaude Opus 5.5 2026-11-03 22:14:52 +01:00
• 805c942155 fix(qwen38-flash-next): count the draft block's hidden row beside wide copy rows in session bytes (#993) Lukáš HanákandClaude Opus 5.5 2026-11-03 22:13:36 +01:00
• 56ed8f4d23 perf(qwen): step IQ4 small-batch K-quant tokens so ROCm 10.1 does not spill (#990) Lukáš Hanák 2026-11-03 17:45:58 +01:00
• 49eeada660 perf(qwen27b): draft DFlash2 through the shared prefix-plus-table output head (#988) Francesco Bozzo 2026-11-03 14:30:05 +01:00
• 159f2b8464 feat(tts): declare voice and reference conditioning capabilities independently of Qwen variants (#984) rikaaa0928 2026-11-03 19:36:34 +08:00
• 6b583876e9 fix(tts): use provider output sample rates in streaming metadata and metrics (#983) rikaaa0928 2026-11-03 18:37:16 +08:00
• 108ade2718 ci!: make nix flake check the canonical hosted gate (#981) Federico Izzo 2026-11-03 10:44:06 +01:00
• 67dd6354c7 fix(qwen38-flash-next): widen BF16 HC inject rows to F32 and keep the mixer's F32 rows beside a separate inject projection (#950) Lukáš HanákandClaude Opus 5.5 2026-11-03 08:38:41 +01:00
• 96c9a0f535 perf(qwen38-flash-next): verify long prompt-lookup copies in one wide pass (#961) 3 people 2026-11-03 08:11:21 +01:00
• 9bd24d26b0 feat(serve): drop API-key auth, accept model aliases in every modality, answer GET / with health (#957) (#967) Federico Izzo 2026-11-02 21:19:16 +01:00
• 9b3b1f4844 fix(serve): ignore unknown request fields on embeddings, audio, image and video routes (#968) Francesco Bozzo 2026-11-02 19:56:43 +01:00
• 8ab834a590 fix(cache): refuse a RAM-cache segment only when registering it cuts the heap below the reserve (#903) Shawn WilsherandFrancesco Bozzo 2026-11-02 08:37:43 -08:00
• cefc651ee9 test(functional): accept record-capacity retention refusals for unchanged retries (#965) Francesco Bozzo 2026-11-02 17:28:45 +01:00
• b911c150e9 fix(serve): report sizes and fitting sessions when request states exceed capacity (#964) Francesco Bozzo 2026-11-02 16:31:22 +01:00
• 773a164192 test(functional): run cache-stalls at 131072 context to fit the devlab runner (#956) Francesco Bozzo 2026-11-02 12:26:05 +01:00
• 9be1f725d8 docs(server): document llama-swap readiness configuration (#941) vmlinuzx 2026-11-02 05:20:54 -06:00
• abb6acabce fix(qwen38-flash-next): keep cached requests alive without snapshot device headroom and return to 4,096-row prefill passes (#954) Francesco Bozzo 2026-11-02 11:36:50 +01:00
• c32136f87c docs(extensions): mark cards 27-29 won't do (#955) Federico Izzo 2026-11-02 10:29:20 +01:00
• 6a027439d7 fix(qwen38_flash_next): restore serving concurrency and GPU tests (#953) Federico Izzo 2026-11-02 09:56:36 +01:00
• 276dd180c7 build(extensions): add distribution recipes and conformance policy (#951) Federico Izzo 2026-11-02 09:22:00 +01:00
• 6bbb6bb6bd feat(extensions): generate Nix serving options from metadata (#947) Federico Izzo 2026-11-02 08:20:55 +01:00
• b6336a3457 fix(ds4): default the served context to the GGUF native length (#946) Francesco Bozzo 2026-11-02 08:06:35 +01:00
• 6246d77a22 fix(ds4): keep caller text out of DeepSeek control tokens (#691) Jonathan Yates 2026-11-01 22:22:05 -08:00
• 13f6456628 feat(extensions): add realtime speech, speak and decision contracts (#942) Federico Izzo 2026-11-02 06:43:26 +01:00
• 05f0a9a0f7 feat(image): support Qwen Image 2.1 Turbo generation and editing (#784) Julius BrockelmannandFrancesco Bozzo 2026-11-02 01:22:42 +01:00
• f25f4744a6 feat(extensions): derive API revisions from permanent change records (#940) Federico Izzo 2026-11-02 00:14:09 +01:00
• 7b9122f279 docs(extensions): record Card 24 external LFM3 text provider conformance (#938) Federico Izzo 2026-11-01 23:48:31 +01:00
• 9d1e7c8ef7 feat(extensions): support external extension source trees (Card 23) (#936) Federico Izzo 2026-11-01 23:45:39 +01:00
• eddcfea65b feat(embedding): select the EmbeddingGemma 2 video frame budget per request (#888) jabataandClaude Opus 5.5 2026-11-01 13:30:43 -08:00
• eb313910a5 fix(windows): keep same-named HIP sources from swapping device code (#914) ThomasandClaude Opus 5.5 2026-11-01 16:26:49 -05:00
• 385d9a7f83 feat(extensions): migrate text generation models (Cards 19-22) (#935) Federico Izzo 2026-11-01 21:54:24 +01:00
• 6ab4569ba1 feat(serve): report completion_tokens_details.reasoning_tokens on chat completions (#915) Daiki Noda 2026-11-02 05:37:18 +09:00
• bdf800de47 perf(ds4): faster DeepSeek V4 Flash decode, top-p sampling, and unaligned resumed prefill (#897) Shawn Wilsher 2026-11-01 11:38:14 -08:00
• a6c53a6273 perf(qwen38-flash-next): fuse the prefill DeltaNet norms, gates, recurrence and epilogue (#895) Lukáš HanákandClaude Opus 5.5 2026-11-01 20:05:16 +01:00
• 8f5aece0c2 feat(extensions): migrate speech, audio, image, and video models to extensions (#931) Federico Izzo 2026-11-01 18:42:12 +01:00
• eb1d52e68a perf(qwen38-flash-next): run a decode batch's HC combine and Q8_0 mixer in one launch (#894) Lukáš HanákandClaude Opus 5.5 2026-11-01 18:14:19 +01:00