mlx-bun 0.4.0 → 1.0.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +1453 -88
- package/THIRD_PARTY_NOTICES.md +882 -0
- package/bin/mlx-bun.convert.mjs +4 -0
- package/bin/mlx-bun.fuse.mjs +4 -0
- package/bin/mlx-bun.generate.mjs +4 -0
- package/bin/mlx-bun.lora.mjs +4 -0
- package/bin/mlx-bun.mjs +13 -21
- package/bin/mlx-bun.server.mjs +4 -0
- package/bin/mlx-bun.upload.mjs +4 -0
- package/dist/web/app.js +9584 -0
- package/package.json +50 -53
- package/src/cli/ane-install.ts +76 -0
- package/src/cli/args.ts +179 -0
- package/src/cli/engine-entry.ts +68 -0
- package/src/cli/inference.ts +214 -0
- package/src/cli/job-entry.ts +97 -0
- package/src/cli/library-host.ts +117 -0
- package/src/cli/main.ts +57 -0
- package/src/cli/memory-engine.ts +163 -0
- package/src/cli/memory.ts +36 -0
- package/src/cli/mlx-lm-aliases.ts +309 -0
- package/src/cli/model-choice.ts +33 -0
- package/src/cli/model-selection.ts +84 -0
- package/src/cli/module-host.ts +11 -0
- package/src/cli/module-verbs.ts +50 -0
- package/src/cli/numerical-policy.ts +12 -0
- package/src/cli/perplexity.ts +89 -0
- package/src/cli/serve-host.ts +423 -0
- package/src/cli/serve-isolated.ts +217 -0
- package/src/cli/serve-options.ts +106 -0
- package/src/cli/serve-state.ts +205 -0
- package/src/cli/serve.ts +500 -0
- package/src/cli/served-model-host.ts +102 -0
- package/src/cli/server-entry.ts +148 -0
- package/src/cli/serving-unit.ts +177 -0
- package/src/{tui.ts → cli/terminal.ts} +13 -8
- package/src/cli/worker-adapters.ts +34 -0
- package/src/cli/worker-entry.ts +272 -0
- package/src/cli/worker-unit.ts +179 -0
- package/src/engine/adapter-operation.ts +26 -0
- package/src/engine/cache-services.ts +114 -0
- package/src/engine/completion.ts +39 -0
- package/src/engine/generation-gateway.ts +358 -0
- package/src/engine/index.ts +59 -0
- package/src/engine/kv-budget.ts +120 -0
- package/src/engine/media-preparation.ts +232 -0
- package/src/engine/model-binding.ts +24 -0
- package/src/engine/model-host.ts +513 -0
- package/src/engine/model-serving.ts +89 -0
- package/src/{serve → engine}/preparation.ts +10 -8
- package/src/{serve → engine}/session-completion-engine.ts +6 -6
- package/src/engine/telemetry.ts +109 -0
- package/src/hub/downloads.ts +116 -0
- package/src/jobs/db.ts +16 -7
- package/src/jobs/events.ts +29 -0
- package/src/jobs/executable.ts +5 -0
- package/src/jobs/host.ts +51 -0
- package/src/jobs/in-process.ts +55 -0
- package/src/{contracts/jobs.ts → jobs/protocol.ts} +7 -9
- package/src/jobs/runner.ts +150 -231
- package/src/jobs/service.ts +75 -0
- package/src/jobs/sse.ts +59 -23
- package/src/jobs/worker-memory.ts +46 -0
- package/src/jobs/worker-process.ts +163 -0
- package/src/jobs/worker-supervisor.ts +258 -0
- package/src/modules.ts +58 -0
- package/src/publishing/catalog-hub.ts +18 -0
- package/src/publishing/credentials.ts +31 -0
- package/src/publishing/upload.ts +37 -0
- package/src/residency/leased-adapters.ts +21 -0
- package/src/residency/local-records.ts +16 -0
- package/src/residency/model-residency.ts +405 -0
- package/src/residency/record-index.ts +28 -0
- package/src/residency/resident-estimate.ts +50 -0
- package/src/{anthropic.ts → server/anthropic.ts} +3 -81
- package/src/server/cache-routes.ts +36 -0
- package/src/{serve → server}/chat-request.ts +12 -10
- package/src/{serve → server}/chat-stage.ts +27 -22
- package/src/{client.ts → server/client.ts} +28 -2
- package/src/{serve → server}/completion-executor.ts +42 -42
- package/src/server/decision-body.ts +55 -0
- package/src/server/decision-routes.ts +134 -0
- package/src/server/decision-wire.ts +228 -0
- package/src/{serve → server}/discovery-routes.ts +42 -83
- package/src/server/generated-token-history.ts +57 -0
- package/src/{serve → server}/http.ts +25 -13
- package/src/{serve → server}/inference-request.ts +5 -5
- package/src/server/job-routes.ts +20 -0
- package/src/server/media-prompt.ts +63 -0
- package/src/server/memory-completion-client.ts +46 -0
- package/src/server/model-routes.ts +192 -0
- package/src/{serve → server}/openai-wire.ts +3 -2
- package/src/{serve → server}/pipeline.ts +1 -1
- package/src/server/prompt-builder.ts +13 -0
- package/src/server/prompt-contracts.ts +17 -0
- package/src/server/proxy-routes.ts +187 -0
- package/src/server/publishing-routes.ts +41 -0
- package/src/{serve → server}/request-plan.ts +8 -8
- package/src/{serve → server}/request-prep.ts +82 -68
- package/src/{serve → server}/responses-client.ts +30 -13
- package/src/{responses.ts → server/responses.ts} +12 -93
- package/src/server/routes.ts +205 -0
- package/src/server/start.ts +72 -0
- package/src/server/status-routes.ts +108 -0
- package/src/{serve → server}/text-completion-stage.ts +25 -4
- package/src/server/text-prompt.ts +12 -0
- package/src/{serve → server}/token-streams.ts +36 -38
- package/src/server/worker-routes.ts +311 -0
- package/src/storage/paths.ts +61 -0
- package/src/web/assets.ts +37 -0
- package/src/web/{src → browser}/api.ts +5 -23
- package/src/web/browser/chat-host.ts +17 -0
- package/src/web/browser/installed-panels.ts +6 -0
- package/src/web/browser/main.ts +55 -0
- package/src/web/browser/markdown.ts +279 -0
- package/src/web/browser/memory-panel.ts +9 -0
- package/src/web/browser/palette.ts +64 -0
- package/src/web/browser/protocol.ts +13 -0
- package/src/web/browser/shell.ts +467 -0
- package/src/web/{src → browser}/status.ts +32 -36
- package/src/web/build.ts +51 -0
- package/src/web/public/app.html +1054 -0
- package/src/web/{sw.js → public/sw.js} +4 -3
- package/src/web/{vendor → public/vendor}/hljs-theme.css +2 -2
- package/src/web/{vendor → public/vendor}/hljs.js +2 -2
- package/THIRD_PARTY_LICENSES.md +0 -51
- package/docs/README.md +0 -94
- package/docs/archive/hlg-sampling.md +0 -350
- package/docs/archive/investigations/audio-input-plan.md +0 -356
- package/docs/archive/investigations/chunk-finetune-experiment.md +0 -84
- package/docs/archive/investigations/colibri-glm52-port.md +0 -1486
- package/docs/archive/investigations/colibri-oracle-pin.md +0 -1284
- package/docs/archive/investigations/curve-bisector-routing.md +0 -123
- package/docs/archive/investigations/decode-roofline-lookagain.md +0 -348
- package/docs/archive/investigations/diffusion-gemma-port.md +0 -602
- package/docs/archive/investigations/dspark-handoff.md +0 -224
- package/docs/archive/investigations/expert-offload-single-user-moe.md +0 -292
- package/docs/archive/investigations/faithful-l1-consolidation.md +0 -310
- package/docs/archive/investigations/hlg-sampling-investigation.md +0 -970
- package/docs/archive/investigations/indexshare-performance-spike.md +0 -357
- package/docs/archive/investigations/kernel-perf-review-2026-07.md +0 -171
- package/docs/archive/investigations/lab-build-journal.md +0 -157
- package/docs/archive/investigations/minicpm5-decode-megakernel.md +0 -489
- package/docs/archive/investigations/openjarvis-fit.md +0 -198
- package/docs/archive/investigations/openwiki-evaluation.md +0 -249
- package/docs/archive/investigations/optimization_plan.md +0 -346
- package/docs/archive/investigations/orpo-base-uf-experiment-and-directions.md +0 -120
- package/docs/archive/investigations/orpo-flash-cce-pin-leak.md +0 -167
- package/docs/archive/investigations/orpo-uf-testing-handoff.md +0 -126
- package/docs/archive/investigations/pareto-specialized-runtime-findings.md +0 -92
- package/docs/archive/investigations/pi-builtin-investigation.md +0 -308
- package/docs/archive/investigations/pre-colibri-stabilization.md +0 -329
- package/docs/archive/investigations/qwen38-27b-campaign-log.md +0 -1092
- package/docs/archive/investigations/segmented-backward-handoff.md +0 -148
- package/docs/archive/investigations/starter-model-port-handoff.md +0 -342
- package/docs/archive/investigations/steel-flash-cce-handoff.md +0 -489
- package/docs/archive/investigations/the-dreaming-handoff.md +0 -141
- package/docs/archive/investigations/trainer-validation-experiment.md +0 -173
- package/docs/archive/planning/chat-ui-vision-memory-first.md +0 -199
- package/docs/archive/planning/chat-ui-vision.md +0 -50
- package/docs/archive/planning/chunk-task-roadmap.md +0 -77
- package/docs/archive/planning/curve-sampler-research-plan.md +0 -331
- package/docs/archive/planning/journal.md +0 -316
- package/docs/archive/planning/memory-docs-and-dag-plan.md +0 -194
- package/docs/archive/planning/pi-integration-review.md +0 -178
- package/docs/archive/planning/release-notes-v0.0.10.md +0 -57
- package/docs/archive/planning/release-notes-v0.0.11.md +0 -239
- package/docs/archive/planning/release-notes-v0.0.12.md +0 -80
- package/docs/archive/planning/release-notes-v0.0.13.md +0 -77
- package/docs/archive/planning/release-notes-v0.0.9.md +0 -122
- package/docs/archive/planning/release-notes-v0.2.0.md +0 -49
- package/docs/archive/planning/release-notes-v0.3.0.md +0 -35
- package/docs/archive/planning/web-ui-pass-plan.md +0 -213
- package/docs/archive/planning/website-readme-pass-plan.md +0 -224
- package/docs/design/batching.md +0 -1841
- package/docs/design/decode-speed-program.md +0 -2349
- package/docs/design/dreaming-nightly-pipeline.md +0 -684
- package/docs/design/generic-model-support.md +0 -705
- package/docs/design/kv-cache.md +0 -824
- package/docs/design/orpo-training.md +0 -680
- package/docs/design/speculative-decoding.md +0 -1990
- package/docs/design/turboquant.md +0 -2212
- package/docs/design/unified-engine-frontier-plan.md +0 -1880
- package/docs/design/web-chat-redesign.md +0 -695
- package/docs/planning/PRODUCT_ROADMAP.md +0 -594
- package/docs/planning/ResearchTopics.md +0 -119
- package/docs/planning/release-notes-v0.4.0.md +0 -45
- package/docs/planning/video-series.md +0 -421
- package/docs/reference/benchmarks.md +0 -3613
- package/docs/reference/cli.md +0 -585
- package/docs/reference/distribution.md +0 -355
- package/docs/reference/environment.md +0 -398
- package/docs/reference/glossary.md +0 -39
- package/docs/reference/library-api.md +0 -674
- package/docs/reference/memory.md +0 -260
- package/docs/reference/models.md +0 -334
- package/docs/reference/server-api.md +0 -1775
- package/docs/reference/server-config.md +0 -846
- package/docs/reference/training.md +0 -462
- package/docs/reference/troubleshooting.md +0 -158
- package/src/admission-plan.ts +0 -18
- package/src/audio/conformer.ts +0 -647
- package/src/audio/decode.ts +0 -133
- package/src/audio/features.ts +0 -202
- package/src/audio/transcode.ts +0 -67
- package/src/backends/mlx/assistant-target.ts +0 -46
- package/src/backends/mlx/autoregressive-draft-rows.ts +0 -120
- package/src/backends/mlx/autoregressive.ts +0 -108
- package/src/backends/mlx/batch-group.ts +0 -1790
- package/src/backends/mlx/cache-attachment.ts +0 -67
- package/src/backends/mlx/cache-layout.ts +0 -31
- package/src/backends/mlx/checkpoint-state.ts +0 -53
- package/src/backends/mlx/continuation-persistence.ts +0 -106
- package/src/backends/mlx/continuation-request.ts +0 -43
- package/src/backends/mlx/continuation.ts +0 -77
- package/src/backends/mlx/diffusion.ts +0 -39
- package/src/backends/mlx/draft-target.ts +0 -32
- package/src/backends/mlx/evaluation.ts +0 -105
- package/src/backends/mlx/fill-append.ts +0 -45
- package/src/backends/mlx/full-kv-row-append.ts +0 -32
- package/src/backends/mlx/full-kv-row-donor.ts +0 -54
- package/src/backends/mlx/full-transitioning-kv-rows.ts +0 -65
- package/src/backends/mlx/gateway-binding.ts +0 -130
- package/src/backends/mlx/graph.ts +0 -38
- package/src/backends/mlx/kv-codec.ts +0 -9
- package/src/backends/mlx/kv-maintenance.ts +0 -133
- package/src/backends/mlx/kv-tensor-rows.ts +0 -212
- package/src/backends/mlx/memory-client.ts +0 -189
- package/src/backends/mlx/methods.ts +0 -90
- package/src/backends/mlx/model-prompt.ts +0 -164
- package/src/backends/mlx/model-serving.ts +0 -79
- package/src/backends/mlx/prefill-cohort.ts +0 -89
- package/src/backends/mlx/prefill-rows.ts +0 -200
- package/src/backends/mlx/prefill.ts +0 -33
- package/src/backends/mlx/projected-context-rows.ts +0 -72
- package/src/backends/mlx/request-policy.ts +0 -20
- package/src/backends/mlx/request-state-policy.ts +0 -41
- package/src/backends/mlx/rollback.ts +0 -89
- package/src/backends/mlx/row-sampling.ts +0 -49
- package/src/backends/mlx/serial-executor.ts +0 -242
- package/src/backends/mlx/speculative-group.ts +0 -359
- package/src/backends/mlx/speculative-round.ts +0 -161
- package/src/backends/mlx/speculative.ts +0 -114
- package/src/backends/mlx/state-rows.ts +0 -104
- package/src/backends/mlx/state-views.ts +0 -46
- package/src/backends/mlx/target-layout-capability.ts +0 -35
- package/src/backends/mlx/text-engine.ts +0 -73
- package/src/backends/mlx/transitioning-kv-rows.ts +0 -165
- package/src/chat-template.ts +0 -382
- package/src/cli.ts +0 -3168
- package/src/config.ts +0 -399
- package/src/contracts/completion.ts +0 -16
- package/src/contracts/execution.ts +0 -60
- package/src/contracts/generation.ts +0 -103
- package/src/contracts/host.ts +0 -7
- package/src/contracts/pi-provider.ts +0 -30
- package/src/contracts/pi.ts +0 -259
- package/src/contracts/prefix-cache.ts +0 -23
- package/src/contracts/resources.ts +0 -25
- package/src/contracts/sampling.ts +0 -23
- package/src/contracts/scheduling.ts +0 -45
- package/src/contracts/tsconfig.json +0 -14
- package/src/dataset/generators.ts +0 -1087
- package/src/dataset/index.ts +0 -12
- package/src/dataset/job.ts +0 -81
- package/src/dataset/llm.ts +0 -71
- package/src/dataset/registry.ts +0 -528
- package/src/diffusion/diffusion-generate.ts +0 -507
- package/src/download.ts +0 -505
- package/src/embed.ts +0 -77
- package/src/engine/admission.ts +0 -59
- package/src/engine/cancellation.ts +0 -38
- package/src/engine/engine.ts +0 -56
- package/src/engine/execution-coordinator.ts +0 -70
- package/src/engine/execution-plan.ts +0 -53
- package/src/engine/resources.ts +0 -53
- package/src/engine/scheduler.ts +0 -72
- package/src/engine/session.ts +0 -247
- package/src/eval/README.md +0 -130
- package/src/eval/ifeval.ts +0 -81
- package/src/eval/kl-prompts.ts +0 -29
- package/src/eval/kl.ts +0 -590
- package/src/eval/perplexity.ts +0 -164
- package/src/eval/quality-db.ts +0 -92
- package/src/eval/reference-logits.ts +0 -207
- package/src/eval/runner.ts +0 -124
- package/src/eval/score.ts +0 -49
- package/src/eval/tasks/bfcl.ts +0 -535
- package/src/eval/tasks/gsm8k.ts +0 -98
- package/src/eval/tasks/hashhop.ts +0 -139
- package/src/eval/tasks/humaneval.ts +0 -275
- package/src/eval/tasks/ifeval.ts +0 -525
- package/src/eval/tasks/mmlu.ts +0 -157
- package/src/eval/ultrafeedback.ts +0 -60
- package/src/evaldb.ts +0 -199
- package/src/expert-io.ts +0 -393
- package/src/expert-offload-build.ts +0 -77
- package/src/expert-offload.ts +0 -76
- package/src/expert-residency.ts +0 -1134
- package/src/expert-usage.ts +0 -415
- package/src/fill/echo-index.ts +0 -188
- package/src/fill/fill-session.ts +0 -452
- package/src/fill/proposal.ts +0 -68
- package/src/fill/schema-rows.ts +0 -506
- package/src/fill/tool-boundary.ts +0 -103
- package/src/fit.ts +0 -249
- package/src/generate.ts +0 -1346
- package/src/grammar.ts +0 -591
- package/src/harness-pi.ts +0 -184
- package/src/hf-push.ts +0 -628
- package/src/hub-rest.ts +0 -242
- package/src/index.ts +0 -63
- package/src/inference/denoising.ts +0 -17
- package/src/inference/draft-acceptance.ts +0 -41
- package/src/inference/draft-output.ts +0 -49
- package/src/inference/graph.ts +0 -33
- package/src/inference/prefill.ts +0 -39
- package/src/inference/rollback.ts +0 -7
- package/src/jobs/index.ts +0 -31
- package/src/jobs/job-entry.ts +0 -107
- package/src/jobs/task-client.ts +0 -19
- package/src/jobs/types.ts +0 -2
- package/src/kv-scheme.ts +0 -257
- package/src/kv-store.ts +0 -1099
- package/src/lab/README.md +0 -18
- package/src/lab/curve/curve-designer.html +0 -303
- package/src/lab/curve/curve-sampler.ts +0 -122
- package/src/lab/expert-trace/expert-trace.ts +0 -69
- package/src/lab/paged-kv/paged-kv-rows.ts +0 -83
- package/src/lab/paged-kv/paged-kv.ts +0 -270
- package/src/library.ts +0 -40
- package/src/lora.ts +0 -408
- package/src/md-modules.d.ts +0 -49
- package/src/media-fetch.ts +0 -442
- package/src/memory/article.ts +0 -558
- package/src/memory/chunk-validation.ts +0 -150
- package/src/memory/chunk.ts +0 -318
- package/src/memory/cluster.ts +0 -351
- package/src/memory/coverage.ts +0 -134
- package/src/memory/crosslink.ts +0 -0
- package/src/memory/db.ts +0 -253
- package/src/memory/entity.ts +0 -203
- package/src/memory/gate.ts +0 -187
- package/src/memory/history.ts +0 -152
- package/src/memory/ingest.ts +0 -335
- package/src/memory/model.ts +0 -183
- package/src/memory/normalize.ts +0 -671
- package/src/memory/parse.ts +0 -22
- package/src/memory/pipeline.ts +0 -275
- package/src/memory/prompts.ts +0 -65
- package/src/memory/query.ts +0 -334
- package/src/memory/reconcile.ts +0 -734
- package/src/memory/redirect.ts +0 -108
- package/src/memory/reindex.ts +0 -325
- package/src/memory/resolve.ts +0 -307
- package/src/memory/rest.ts +0 -245
- package/src/memory/route.ts +0 -272
- package/src/memory/schedule.ts +0 -165
- package/src/memory/stages.ts +0 -777
- package/src/memory/synthesize.ts +0 -1638
- package/src/memory/tools.ts +0 -749
- package/src/memory/vault.ts +0 -1194
- package/src/memory/wikify.ts +0 -1064
- package/src/mlx/array.ts +0 -388
- package/src/mlx/autograd.ts +0 -304
- package/src/mlx/checkpoint.ts +0 -149
- package/src/mlx/compile.ts +0 -168
- package/src/mlx/custom-vjp.ts +0 -188
- package/src/mlx/ffi.ts +0 -512
- package/src/mlx/materialize.ts +0 -23
- package/src/mlx/metal-kernel.ts +0 -159
- package/src/mlx/ops.ts +0 -1021
- package/src/mlx/quantized-rows.ts +0 -36
- package/src/mlx/token-bitmask.ts +0 -52
- package/src/mlx/turboquant-kv-decode.ts +0 -82
- package/src/mlx/turboquant-kv-inverse.ts +0 -122
- package/src/mlx/turboquant-ops.ts +0 -861
- package/src/mlx/turboquant-tables.ts +0 -237
- package/src/mmap.ts +0 -90
- package/src/model/artifact-identity.ts +0 -73
- package/src/model/batched-kv.ts +0 -235
- package/src/model/batched-mask.ts +0 -279
- package/src/model/batched-quant.ts +0 -212
- package/src/model/batched-quantized-kv.ts +0 -144
- package/src/model/batched-rotating-quant.ts +0 -353
- package/src/model/batched-rotating-state.ts +0 -149
- package/src/model/batched-rotating.ts +0 -338
- package/src/model/batched-row-storage.ts +0 -179
- package/src/model/batched-ssm.ts +0 -52
- package/src/model/batched-turboquant-kv.ts +0 -93
- package/src/model/compiled-decode.ts +0 -712
- package/src/model/decoded-kv-donor.ts +0 -12
- package/src/model/delayed-quantized-kv.ts +0 -72
- package/src/model/delayed-rotating-quantized-kv.ts +0 -93
- package/src/model/delayed-turboquant-kv.ts +0 -55
- package/src/model/diffusion-gemma.ts +0 -1026
- package/src/model/factory.ts +0 -189
- package/src/model/fingerprint.ts +0 -33
- package/src/model/flash-attention.ts +0 -694
- package/src/model/full-prefill-padding.ts +0 -56
- package/src/model/full-prefill-row.ts +0 -105
- package/src/model/gemma4-base.ts +0 -2351
- package/src/model/gemma4.ts +0 -1315
- package/src/model/generated/gemma4-12b.ts +0 -429
- package/src/model/generated/gemma4-26b.ts +0 -448
- package/src/model/generated/gemma4-e4b.ts +0 -717
- package/src/model/generated/index.ts +0 -18
- package/src/model/glm52-atlas.ts +0 -511
- package/src/model/glm52-cache.ts +0 -792
- package/src/model/glm52-config.ts +0 -262
- package/src/model/glm52-container.ts +0 -305
- package/src/model/glm52-coupling.ts +0 -388
- package/src/model/glm52-dsa.ts +0 -606
- package/src/model/glm52-expert-layout.ts +0 -253
- package/src/model/glm52-layout.ts +0 -152
- package/src/model/glm52-memory.ts +0 -500
- package/src/model/glm52-mla.ts +0 -792
- package/src/model/glm52-moe.ts +0 -223
- package/src/model/glm52-pilot.ts +0 -537
- package/src/model/glm52-reference.ts +0 -445
- package/src/model/glm52-residency.ts +0 -557
- package/src/model/glm52-resident-weights.ts +0 -135
- package/src/model/glm52-streamed-experts.ts +0 -453
- package/src/model/glm52-streamed-metal.ts +0 -610
- package/src/model/glm52-weights.ts +0 -503
- package/src/model/glm52.ts +0 -1132
- package/src/model/implementation.ts +0 -69
- package/src/model/kv-attention-view.ts +0 -67
- package/src/model/minicpm5.ts +0 -376
- package/src/model/profile.ts +0 -443
- package/src/model/qwen-append.ts +0 -12
- package/src/model/qwen-conv.ts +0 -52
- package/src/model/qwen3-delta-state.ts +0 -74
- package/src/model/qwen3-delta.ts +0 -451
- package/src/model/qwen3-moe.ts +0 -404
- package/src/model/qwen3-mrope.ts +0 -252
- package/src/model/qwen3.ts +0 -277
- package/src/model/qwen3_5-checkpoint.ts +0 -164
- package/src/model/qwen3_5.ts +0 -819
- package/src/model/rotating-kv-donor.ts +0 -35
- package/src/model/rotating-kv-layout.ts +0 -262
- package/src/model/rotating-row-transaction.ts +0 -35
- package/src/model/speculative-rotating-kv.ts +0 -101
- package/src/model/ssm-prefill-padding.ts +0 -39
- package/src/model/support.ts +0 -102
- package/src/model/trellis-balanced-scatter.ts +0 -51
- package/src/model/trellis-linear.ts +0 -672
- package/src/model/trellis-shared-m.ts +0 -151
- package/src/model/trellis-shared-scatter.ts +0 -144
- package/src/model/trellis-splitk-prefill.ts +0 -158
- package/src/model/trellis-tiled-prefill.ts +0 -136
- package/src/model/trellis-vector-expand.ts +0 -62
- package/src/model/trellis-wide-prefill.ts +0 -94
- package/src/model/turboquant-codec.ts +0 -83
- package/src/model/universal/archs.ts +0 -332
- package/src/model/universal/dense.ts +0 -612
- package/src/model/universal/modules.ts +0 -170
- package/src/model/universal/rope.ts +0 -381
- package/src/native/expert_io.c +0 -523
- package/src/native/frame_extract.swift +0 -140
- package/src/native-pack.ts +0 -162
- package/src/pi-provider.ts +0 -53
- package/src/pi-session.ts +0 -81
- package/src/pi-terminal.ts +0 -274
- package/src/pi-web.ts +0 -1932
- package/src/preflight.ts +0 -158
- package/src/prompt-cache.ts +0 -460
- package/src/quantize/allocator.ts +0 -627
- package/src/quantize/atomic-output.ts +0 -38
- package/src/quantize/calibration.ts +0 -181
- package/src/quantize/config-writer.ts +0 -145
- package/src/quantize/index.ts +0 -84
- package/src/quantize/job.ts +0 -162
- package/src/quantize/quantizer.ts +0 -666
- package/src/quantize/rotate.ts +0 -603
- package/src/quantize/safetensors-writer.ts +0 -249
- package/src/quantize/sensitivity.ts +0 -216
- package/src/quantize/trellis.ts +0 -586
- package/src/quantize/weight-transform.ts +0 -292
- package/src/registry.ts +0 -602
- package/src/runtime-config.ts +0 -81
- package/src/safetensors.ts +0 -158
- package/src/sampler.ts +0 -1011
- package/src/serve/admin-routes.ts +0 -184
- package/src/serve/assets/curve-terrain.html +0 -105
- package/src/serve/assets/training-inference-map.html +0 -575
- package/src/serve/aux-routes.ts +0 -132
- package/src/serve/batch-scheduler.ts +0 -8
- package/src/serve/checkpoint-identity.ts +0 -64
- package/src/serve/generation-gateway.ts +0 -496
- package/src/serve/isolate.ts +0 -609
- package/src/serve/kv-budget.ts +0 -14
- package/src/serve/lab-routes.ts +0 -341
- package/src/serve/lane-registry.ts +0 -45
- package/src/serve/model-admin-routes.ts +0 -155
- package/src/serve/model-binding.ts +0 -43
- package/src/serve/model-host.ts +0 -511
- package/src/serve/parent-application.ts +0 -56
- package/src/serve/prompt-response-trace.ts +0 -147
- package/src/serve/response-lease.ts +0 -24
- package/src/serve/session-search.ts +0 -326
- package/src/serve/static-routes.ts +0 -101
- package/src/server.ts +0 -1314
- package/src/spec/assistant-rows.ts +0 -92
- package/src/spec/assistant-source.ts +0 -126
- package/src/spec/cache-identity.ts +0 -22
- package/src/spec/deepspec-rows.ts +0 -29
- package/src/spec/deepspec-source.ts +0 -145
- package/src/spec/dflash-source.ts +0 -149
- package/src/spec/drafter.ts +0 -304
- package/src/spec/dspark/ab-stats.ts +0 -136
- package/src/spec/dspark/calibration.ts +0 -112
- package/src/spec/dspark/data-dflash.ts +0 -156
- package/src/spec/dspark/data.ts +0 -201
- package/src/spec/dspark/deepspec-module.ts +0 -776
- package/src/spec/dspark/generate-dflash.ts +0 -245
- package/src/spec/dspark/generate.ts +0 -290
- package/src/spec/dspark/loader.ts +0 -23
- package/src/spec/dspark/loss.ts +0 -213
- package/src/spec/dspark/module-dflash.ts +0 -560
- package/src/spec/dspark/module.ts +0 -540
- package/src/spec/dspark/quantize-drafter.ts +0 -91
- package/src/spec/dspark/sample.ts +0 -112
- package/src/spec/generate.ts +0 -276
- package/src/spec/glm52-mtp-graph.ts +0 -29
- package/src/spec/glm52-mtp-rows.ts +0 -132
- package/src/spec/glm52-mtp-source.ts +0 -317
- package/src/spec/glm52-mtp-state.ts +0 -33
- package/src/spec/ngram-source.ts +0 -204
- package/src/spec/projected-draft-rows.ts +0 -73
- package/src/spec/qwen-mtp-module.ts +0 -171
- package/src/spec/qwen-mtp-rows.ts +0 -219
- package/src/spec/qwen-mtp-source.ts +0 -383
- package/src/spec/qwen-mtp-state.ts +0 -30
- package/src/spec/serve-loop.ts +0 -585
- package/src/spec/source.ts +0 -215
- package/src/spec/two-model-rows.ts +0 -61
- package/src/spec/two-model.ts +0 -113
- package/src/ssd-cache.ts +0 -372
- package/src/ssd-durability.ts +0 -211
- package/src/tokenizer.ts +0 -335
- package/src/tool-approvals.ts +0 -131
- package/src/tool-call.ts +0 -695
- package/src/train/dataset.ts +0 -484
- package/src/train/diffusion-lora.ts +0 -216
- package/src/train/export.ts +0 -32
- package/src/train/flash-cce.ts +0 -1234
- package/src/train/forward.ts +0 -266
- package/src/train/fuse.ts +0 -243
- package/src/train/index.ts +0 -46
- package/src/train/job.ts +0 -205
- package/src/train/lora-params.ts +0 -254
- package/src/train/loss.ts +0 -1591
- package/src/train/merge.ts +0 -195
- package/src/train/optimizer.ts +0 -183
- package/src/train/prefix-shared.ts +0 -490
- package/src/train/rank.ts +0 -137
- package/src/train/segmented.ts +0 -1828
- package/src/train/steel-qmm-header.ts +0 -1461
- package/src/train/trainer.ts +0 -1221
- package/src/train/watch.ts +0 -542
- package/src/vision/diffusion-vision.ts +0 -620
- package/src/vision/embedder.ts +0 -168
- package/src/vision/preprocess.ts +0 -247
- package/src/vision/prompt.ts +0 -373
- package/src/vision/qwen3vl-preprocess.ts +0 -409
- package/src/vision/qwen3vl-prompt.ts +0 -196
- package/src/vision/qwen3vl-tower.ts +0 -555
- package/src/vision/siglip.ts +0 -599
- package/src/vision/video-frames.ts +0 -148
- package/src/web/app.html +0 -2143
- package/src/web/app.js +0 -5703
- package/src/web/skills/memory/SKILL.md +0 -120
- package/src/web/skills/web-research/SKILL.md +0 -46
- package/src/web/skills.ts +0 -97
- package/src/web/src/adapters-panel.ts +0 -288
- package/src/web/src/assistant.ts +0 -451
- package/src/web/src/chat.ts +0 -1057
- package/src/web/src/composer.ts +0 -959
- package/src/web/src/dataset.ts +0 -132
- package/src/web/src/finetune.ts +0 -267
- package/src/web/src/hub.ts +0 -428
- package/src/web/src/main.ts +0 -98
- package/src/web/src/markdown.ts +0 -0
- package/src/web/src/memory-panel.ts +0 -572
- package/src/web/src/model-picker.ts +0 -159
- package/src/web/src/palette.ts +0 -346
- package/src/web/src/protocol.ts +0 -26
- package/src/web/src/quantize.ts +0 -167
- package/src/web/src/rag.ts +0 -274
- package/src/web/src/sessions.ts +0 -469
- package/src/web/src/shell.ts +0 -804
- package/src/web/src/ui-catalog.ts +0 -78
- package/src/web/vendor/README.md +0 -61
- package/src/web-assets.ts +0 -47
- package/src/web-tools.ts +0 -478
- package/src/weights.ts +0 -255
- /package/src/{serve → server}/completion-sink.ts +0 -0
- /package/src/web/{icon.svg → public/icon.svg} +0 -0
- /package/src/web/{manifest.webmanifest → public/manifest.webmanifest} +0 -0
- /package/src/web/{vendor → public/vendor}/hljs-LICENSE +0 -0
|
@@ -1,148 +0,0 @@
|
|
|
1
|
-
// Video file → RGB frames via the AVFoundation sidecar (PLAN 14w) — the
|
|
2
|
-
// video twin of the audio path's afconvert doctrine: macOS-native codecs
|
|
3
|
-
// (H.264/HEVC/ProRes/…), zero vendored decoders. The sidecar
|
|
4
|
-
// (src/native/frame_extract.swift, shipped as `mlx-bun-frame-extract` beside
|
|
5
|
-
// the binary and in the native pack) samples frames at fps=2 with
|
|
6
|
-
// zero-tolerance timestamps — mirroring mlx-vlm's do_sample_frames — and
|
|
7
|
-
// writes PNGs that decode through the SAME fast-png path as image inputs.
|
|
8
|
-
//
|
|
9
|
-
// Resolution order (afconvert needs none of this because /usr/bin ships it;
|
|
10
|
-
// our extractor is the missing CLI face on Apple's decoder):
|
|
11
|
-
// MLX_BUN_FRAME_EXTRACT env → beside the executable (release bundle) →
|
|
12
|
-
// the native-pack cache → dist-native (dev tree) → compile-on-demand from
|
|
13
|
-
// src/native (from-source installs with Xcode CLT; cached in dist-native).
|
|
14
|
-
|
|
15
|
-
import { existsSync, mkdirSync } from "node:fs";
|
|
16
|
-
import { rm } from "node:fs/promises";
|
|
17
|
-
import { randomUUID } from "node:crypto";
|
|
18
|
-
import { tmpdir } from "node:os";
|
|
19
|
-
import { dirname, join } from "node:path";
|
|
20
|
-
import { nativePackDir } from "../native-pack";
|
|
21
|
-
import { decodeImage, type RGBImage } from "./preprocess";
|
|
22
|
-
import { runtimeValue } from "../runtime-config";
|
|
23
|
-
|
|
24
|
-
export const VIDEO_SAMPLE_FPS = 2;
|
|
25
|
-
export const VIDEO_MAX_FRAMES = 768; // Qwen3VLVideoProcessor max_frames
|
|
26
|
-
|
|
27
|
-
const BIN = "mlx-bun-frame-extract";
|
|
28
|
-
|
|
29
|
-
/** Memoize the PROMISE, not the result: a second video request arriving
|
|
30
|
-
* while the first compile is in flight must await the same compile, not
|
|
31
|
-
* observe a half-set cache and spuriously report the sidecar missing. */
|
|
32
|
-
let devCompile: Promise<string | null> | null = null;
|
|
33
|
-
|
|
34
|
-
/** Compile the sidecar from the repo tree (dev/from-source convenience —
|
|
35
|
-
* needs swiftc from the Xcode CLT). Null when the source or toolchain is
|
|
36
|
-
* unavailable. */
|
|
37
|
-
function compileFromSource(): Promise<string | null> {
|
|
38
|
-
devCompile ??= (async () => {
|
|
39
|
-
const src = join(import.meta.dir, "..", "native", "frame_extract.swift");
|
|
40
|
-
if (!existsSync(src)) return null;
|
|
41
|
-
const out = join(import.meta.dir, "..", "..", "dist-native", BIN);
|
|
42
|
-
try {
|
|
43
|
-
mkdirSync(dirname(out), { recursive: true });
|
|
44
|
-
const proc = Bun.spawn(["swiftc", "-O", src, "-o", out], {
|
|
45
|
-
stdout: "ignore", stderr: "pipe",
|
|
46
|
-
});
|
|
47
|
-
if ((await proc.exited) === 0 && existsSync(out)) return out;
|
|
48
|
-
} catch {
|
|
49
|
-
// no swiftc — resolver returns null and serving 400s with the reason
|
|
50
|
-
}
|
|
51
|
-
return null;
|
|
52
|
-
})();
|
|
53
|
-
return devCompile;
|
|
54
|
-
}
|
|
55
|
-
|
|
56
|
-
/** Locate (or dev-compile) the extractor; null ⟹ video input unavailable
|
|
57
|
-
* on this install (surfaced as a clear 400, never a crash). */
|
|
58
|
-
export async function resolveFrameExtract(): Promise<string | null> {
|
|
59
|
-
const explicit = runtimeValue("MLX_BUN_FRAME_EXTRACT");
|
|
60
|
-
if (explicit) return explicit;
|
|
61
|
-
const candidates = [
|
|
62
|
-
join(dirname(process.execPath), BIN),
|
|
63
|
-
join(nativePackDir(), BIN),
|
|
64
|
-
join(import.meta.dir, "..", "..", "dist-native", BIN),
|
|
65
|
-
];
|
|
66
|
-
for (const c of candidates) if (existsSync(c)) return c;
|
|
67
|
-
return compileFromSource();
|
|
68
|
-
}
|
|
69
|
-
|
|
70
|
-
/** Decode a video container to sampled RGB frames. Throws with the
|
|
71
|
-
* sidecar's diagnostic on failure (the serve layer surfaces it as a 400,
|
|
72
|
-
* the afconvert pattern — timeout kill included). */
|
|
73
|
-
export async function extractVideoFrames(
|
|
74
|
-
bytes: Uint8Array,
|
|
75
|
-
opts: { fps?: number; maxFrames?: number } = {},
|
|
76
|
-
): Promise<RGBImage[]> {
|
|
77
|
-
const bin = await resolveFrameExtract();
|
|
78
|
-
if (!bin)
|
|
79
|
-
throw new Error(
|
|
80
|
-
"video input needs the frame-extraction sidecar (mlx-bun-frame-extract) " +
|
|
81
|
-
"— ships beside the release binary and in the native pack; from a " +
|
|
82
|
-
"source tree it compiles on first use via the Xcode CLT (swiftc)",
|
|
83
|
-
);
|
|
84
|
-
const base = join(tmpdir(), `mlx-bun-video-${randomUUID()}`);
|
|
85
|
-
// AVFoundation infers the container from the path EXTENSION — sniff the
|
|
86
|
-
// ISO-BMFF `ftyp` box (mp4/mov share it; the `qt ` major brand is
|
|
87
|
-
// QuickTime) so the temp file opens. Unknown magic defaults to .mp4.
|
|
88
|
-
const isFtyp = bytes.length >= 12 &&
|
|
89
|
-
bytes[4] === 0x66 && bytes[5] === 0x74 && bytes[6] === 0x79 && bytes[7] === 0x70;
|
|
90
|
-
const brand = isFtyp
|
|
91
|
-
? String.fromCharCode(bytes[8]!, bytes[9]!, bytes[10]!, bytes[11]!)
|
|
92
|
-
: "";
|
|
93
|
-
const ext = brand.startsWith("qt") ? "mov" : "mp4";
|
|
94
|
-
const inPath = `${base}.${ext}`;
|
|
95
|
-
const outDir = `${base}.frames`;
|
|
96
|
-
try {
|
|
97
|
-
// Inside the try so a part-way write failure (ENOSPC on a large clip)
|
|
98
|
-
// still hits the finally's cleanup instead of stranding the temp file.
|
|
99
|
-
await Bun.write(inPath, bytes);
|
|
100
|
-
const proc = Bun.spawn(
|
|
101
|
-
[
|
|
102
|
-
bin, inPath, outDir,
|
|
103
|
-
String(opts.fps ?? VIDEO_SAMPLE_FPS),
|
|
104
|
-
String(opts.maxFrames ?? VIDEO_MAX_FRAMES),
|
|
105
|
-
],
|
|
106
|
-
{ stdout: "pipe", stderr: "pipe" },
|
|
107
|
-
);
|
|
108
|
-
// Decoding is fast (hardware) but long clips take real time; a hung
|
|
109
|
-
// decoder must not pin the request forever — kill → nonzero → 400.
|
|
110
|
-
const killer = setTimeout(() => proc.kill(), 120_000);
|
|
111
|
-
const exit = await proc.exited;
|
|
112
|
-
clearTimeout(killer);
|
|
113
|
-
if (exit !== 0) {
|
|
114
|
-
const err = (await new Response(proc.stderr).text()).trim();
|
|
115
|
-
throw new Error(
|
|
116
|
-
`video decode failed (frame-extract exit ${exit}): ` +
|
|
117
|
-
`${err.split("\n").at(-1) || "unknown error"} — supported containers: ` +
|
|
118
|
-
`whatever AVFoundation decodes on this macOS (mp4/mov H.264/HEVC/ProRes/…)`,
|
|
119
|
-
);
|
|
120
|
-
}
|
|
121
|
-
const head = (await new Response(proc.stdout).text()).trim();
|
|
122
|
-
const n = Number(/^N (\d+)/.exec(head)?.[1] ?? 0);
|
|
123
|
-
if (!n) throw new Error(`video decode produced no frames (${head})`);
|
|
124
|
-
// Aggregate decoded-RGB budget (belt over the sidecar's 1024px
|
|
125
|
-
// longest-edge cap): a valid low-bitrate clip must not balloon into
|
|
126
|
-
// gigabytes of frame buffers before preprocessing.
|
|
127
|
-
const w = Number(/W (\d+)/.exec(head)?.[1] ?? 0);
|
|
128
|
-
const hh = Number(/H (\d+)/.exec(head)?.[1] ?? 0);
|
|
129
|
-
const DECODED_BUDGET = 1536 * 1024 * 1024; // 1.5 GiB
|
|
130
|
-
if (n * w * hh * 3 > DECODED_BUDGET)
|
|
131
|
-
throw new Error(
|
|
132
|
-
`video decodes to ${(n * w * hh * 3 / 1024 / 1024).toFixed(0)} MB of ` +
|
|
133
|
-
`frames (${n} × ${w}x${hh}) — over the ${DECODED_BUDGET / 1024 / 1024} MB ` +
|
|
134
|
-
`budget; send a shorter or smaller clip`,
|
|
135
|
-
);
|
|
136
|
-
const frames: RGBImage[] = [];
|
|
137
|
-
for (let i = 0; i < n; i++) {
|
|
138
|
-
const f = join(outDir, `frame-${String(i).padStart(4, "0")}.png`);
|
|
139
|
-
frames.push(await decodeImage(new Uint8Array(await Bun.file(f).arrayBuffer())));
|
|
140
|
-
}
|
|
141
|
-
return frames;
|
|
142
|
-
} finally {
|
|
143
|
-
await Promise.allSettled([
|
|
144
|
-
rm(inPath, { force: true }),
|
|
145
|
-
rm(outDir, { recursive: true, force: true }),
|
|
146
|
-
]);
|
|
147
|
-
}
|
|
148
|
-
}
|