@qvac/sdk 0.19.1 → 0.20.1

package/NOTICE

@@ -73,12 +73,16 @@ Third-Party Model Licenses
     https://github.com/tetherto/qvac/releases/download/bci-test-assets-v0.1.0/bci-embedder.bin
   cosyvoice3-campplus-f32
     https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512
+  cosyvoice3-flow-bf16
+    https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512
   cosyvoice3-flow-f32
     https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512
   cosyvoice3-hift-f32
     https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512
   cosyvoice3-llm-q8_0
     https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512
+  cosyvoice3-llm-q8_0-fused
+    https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512
   cosyvoice3-s3tok-f16
     https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512
   craft_mlt_25k
@@ -207,6 +211,18 @@ Third-Party Model Licenses
     https://huggingface.co/HuggingFaceVLA/smolvla_libero
   SmolVLM2-500M-Video-Instruct-GGUF
     https://huggingface.co/ggml-org/SmolVLM2-500M-Video-Instruct-GGUF
+  TranslatePsy-AfriSLM-0.8B-Q4-GGUF
+    https://huggingface.co/qvac/TranslatePsy-AfriSLM-0.8B-Q4-GGUF
+  TranslatePsy-AfriSLM-0.8B-Q8-GGUF
+    https://huggingface.co/qvac/TranslatePsy-AfriSLM-0.8B-Q8-GGUF
+  TranslatePsy-AfriSLM-2B-Q4-GGUF
+    https://huggingface.co/qvac/TranslatePsy-AfriSLM-2B-Q4-GGUF
+  TranslatePsy-AfriSLM-2B-Q8-GGUF
+    https://huggingface.co/qvac/TranslatePsy-AfriSLM-2B-Q8-GGUF
+  TranslatePsy-AfriSLM-4B-Q4-GGUF
+    https://huggingface.co/qvac/TranslatePsy-AfriSLM-4B-Q4-GGUF
+  TranslatePsy-AfriSLM-4B-Q8-GGUF
+    https://huggingface.co/qvac/TranslatePsy-AfriSLM-4B-Q8-GGUF
   umt5-xxl-enc-q8_0
     https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged
   unlimited-ocr-gguf
@@ -318,8 +334,12 @@ Third-Party Model Licenses
 
 --- mit (MIT License) ---
 
+  acestep-5Hz-lm-0.6B-BF16
+    https://huggingface.co/ACE-Step/acestep-5Hz-lm-0.6B
   acestep-5Hz-lm-0.6B-Q8_0
     https://huggingface.co/ACE-Step/acestep-5Hz-lm-0.6B
+  acestep-v15-base-Q4_K_M
+    https://huggingface.co/ACE-Step/acestep-v15-base
   acestep-v15-sft-Q8_0
     https://huggingface.co/ACE-Step/acestep-v15-sft
   acestep-v15-turbo-Q4_K_M
@@ -608,6 +628,12 @@ Third-Party Model Licenses
     https://huggingface.co/nvidia/GR00T-N1.7-LIBERO
   groot-q8_vf16
     https://huggingface.co/nvidia/GR00T-N1.7-3B
+  nemotron-3.5-asr-streaming-0.6b.f16
+    https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b
+  nemotron-3.5-asr-streaming-0.6b.q4_0
+    https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b
+  nemotron-3.5-asr-streaming-0.6b.q8_0
+    https://huggingface.co/nvidia/nemotron-3.5-asr-streaming-0.6b
   parakeet-unified-en-0.6b.f16
     https://huggingface.co/nvidia/parakeet-unified-en-0.6b
   parakeet-unified-en-0.6b.q4_0
@@ -663,54 +689,60 @@ JavaScript Dependencies
 
 --- apache-2.0 (Apache License 2.0) ---
 
-  @hyperswarm/secret-stream@6.9.1
+  @hyperswarm/secret-stream@6.9.2
     https://github.com/holepunchto/hyperswarm-secret-stream
-  @qvac/asr-ggml@0.3.3
+  @qvac/asr-ggml@0.5.3
     https://github.com/tetherto/qvac
-  @qvac/audiogen-ggml@0.3.3
+  @qvac/asr-ggml-darwin-arm64@0.5.3
     https://github.com/tetherto/qvac
-  @qvac/bci-whispercpp@0.8.2
+  @qvac/audiogen-ggml@0.4.1
     https://github.com/tetherto/qvac
-  @qvac/classification-ggml@0.23.0
+  @qvac/audiogen-ggml-darwin-arm64@0.4.1
     https://github.com/tetherto/qvac
-  @qvac/decoder-audio@0.5.0
+  @qvac/bci-whispercpp@0.9.1
     https://github.com/tetherto/qvac
-  @qvac/diagnostics@0.1.2
+  @qvac/classification-ggml@0.26.3
     https://github.com/tetherto/qvac
-  @qvac/diffusion-cpp@0.21.0
+  @qvac/decoder-audio@0.6.0
     https://github.com/tetherto/qvac
-  @qvac/embed-llamacpp@0.37.0
+  @qvac/diffusion-cpp@0.24.0
     https://github.com/tetherto/qvac
-  @qvac/error@0.1.1
-  @qvac/fabric@0.10.0
+  @qvac/embed-llamacpp@0.41.3
     https://github.com/tetherto/qvac
-  @qvac/infer-base@0.4.2
+  @qvac/error@0.1.1
+  @qvac/fabric@0.16.1
     https://github.com/tetherto/qvac
   @qvac/infer-base@0.6.2
     https://github.com/tetherto/qvac
-  @qvac/inference@0.19.1
+  @qvac/inference@0.20.0
     https://github.com/tetherto/qvac
   @qvac/langdetect-text@0.1.2
     https://github.com/tetherto/qvac
-  @qvac/llm-llamacpp@0.49.1
+  @qvac/llm-llamacpp@0.53.2
     https://github.com/tetherto/qvac
   @qvac/logging@0.1.1
     https://github.com/tetherto/qvac
-  @qvac/ocr-ggml@0.21.0
+  @qvac/model-fit@0.12.2
+    https://github.com/tetherto/qvac
+  @qvac/ocr-ggml@0.24.3
+    https://github.com/tetherto/qvac
+  @qvac/rag@0.8.1
     https://github.com/tetherto/qvac
   @qvac/registry-client@0.6.1
     https://github.com/tetherto/qvac
   @qvac/registry-schema@0.3.0
     https://github.com/tetherto/qvac
-  @qvac/translation-nmtcpp@0.13.0
+  @qvac/translation-nmtcpp@0.16.4
     https://github.com/tetherto/qvac
-  @qvac/tts-ggml@0.8.1
+  @qvac/tts-ggml@0.9.2
     https://github.com/tetherto/qvac
-  @qvac/vla-ggml@0.24.0
+  @qvac/tts-ggml-darwin-arm64@0.9.2
+    https://github.com/tetherto/qvac
+  @qvac/vla-ggml@0.27.3
     https://github.com/tetherto/qvac
   adaptive-timeout@1.0.1
     https://github.com/holepunchto/adaptive-timeout
-  b4a@1.8.1
+  b4a@1.9.0
     https://github.com/holepunchto/b4a
   bare-abort@2.0.13
     https://github.com/holepunchto/bare-abort
@@ -728,7 +760,7 @@ JavaScript Dependencies
     https://github.com/holepunchto/bare-bundle
   bare-bundle-id@1.0.2
     https://github.com/holepunchto/bare-bundle-id
-  bare-cov@1.2.2
+  bare-cov@1.2.3
     https://github.com/holepunchto/bare-cov
   bare-cpu-info@0.1.1
     https://github.com/holepunchto/bare-cpu-info
@@ -736,7 +768,7 @@ JavaScript Dependencies
     https://github.com/holepunchto/bare-crypto
   bare-debug-log@2.0.0
     https://github.com/holepunchto/bare-debug-log
-  bare-dns@2.2.0
+  bare-dns@2.2.1
     https://github.com/holepunchto/bare-dns
   bare-encoding@1.0.3
     https://github.com/holepunchto/bare-encoding
@@ -744,9 +776,9 @@ JavaScript Dependencies
     https://github.com/holepunchto/bare-env
   bare-events@2.9.2
     https://github.com/holepunchto/bare-events
-  bare-fetch@3.2.1
+  bare-fetch@3.4.0
     https://github.com/holepunchto/bare-fetch
-  bare-ffmpeg@1.5.0
+  bare-ffmpeg@1.6.0
     https://github.com/holepunchto/bare-ffmpeg
   bare-form-data@1.2.2
     https://github.com/holepunchto/bare-form-data
@@ -760,7 +792,7 @@ JavaScript Dependencies
     https://github.com/holepunchto/bare-hrtime
   bare-http-parser@2.1.4
     https://github.com/holepunchto/bare-http-parser
-  bare-http1@4.6.1
+  bare-http1@4.6.2
     https://github.com/holepunchto/bare-http1
   bare-https@3.1.0
     https://github.com/holepunchto/bare-https
@@ -768,19 +800,15 @@ JavaScript Dependencies
     https://github.com/holepunchto/bare-inspect
   bare-inspector@6.1.0
     https://github.com/holepunchto/bare-inspector
-  bare-lief@0.2.7
-    https://github.com/holepunchto/bare-lief
-  bare-link@3.3.0
-    https://github.com/holepunchto/bare-link
   bare-mime@1.0.0
     https://github.com/holepunchto/bare-mime
   bare-module@6.4.0
     https://github.com/holepunchto/bare-module
-  bare-module-lexer@1.6.6
+  bare-module-lexer@1.6.7
     https://github.com/holepunchto/bare-module-lexer
   bare-module-resolve@1.12.5
     https://github.com/holepunchto/bare-module-resolve
-  bare-module-traverse@2.5.5
+  bare-module-traverse@2.5.6
     https://github.com/holepunchto/bare-module-traverse
   bare-net@2.3.3
     https://github.com/holepunchto/bare-net
@@ -800,12 +828,10 @@ JavaScript Dependencies
     https://github.com/holepunchto/bare-process
   bare-rpc@1.3.8
     https://github.com/holepunchto/bare-rpc
-  bare-runtime@1.32.0
+  bare-runtime@1.33.4
     https://github.com/holepunchto/bare-runtime
-  bare-runtime-darwin-arm64@1.32.0
+  bare-runtime-darwin-arm64@1.33.4
     https://github.com/holepunchto/bare-runtime
-  bare-semver@1.1.0
-    https://github.com/holepunchto/bare-semver
   bare-signals@4.2.0
     https://github.com/holepunchto/bare-signals
   bare-signals@5.0.0
@@ -814,13 +840,11 @@ JavaScript Dependencies
     https://github.com/holepunchto/bare-stdio
   bare-stream@2.13.4
     https://github.com/holepunchto/bare-stream
-  bare-structured-clone@1.6.0
-    https://github.com/holepunchto/bare-structured-clone
   bare-stylize@0.0.1
     https://github.com/holepunchto/bare-stylize
   bare-subprocess@5.2.3
     https://github.com/holepunchto/bare-subprocess
-  bare-subprocess@6.1.0
+  bare-subprocess@6.2.0
     https://github.com/holepunchto/bare-subprocess
   bare-tcp@2.6.1
     https://github.com/holepunchto/bare-tcp
@@ -828,16 +852,17 @@ JavaScript Dependencies
     https://github.com/holepunchto/bare-tls
   bare-tty@5.2.1
     https://github.com/holepunchto/bare-tty
-  bare-type@1.1.1
+  bare-type@1.4.0
     https://github.com/holepunchto/bare-type
-  bare-type-stripper@0.1.6
+  bare-type-stripper@0.1.7
     https://github.com/holepunchto/bare-type-stripper
   bare-url@2.5.4
     https://github.com/holepunchto/bare-url
   bare-utils@1.6.0
     https://github.com/holepunchto/bare-utils
-  bare-v8-to-istanbul@1.0.2
-  bare-ws@3.1.0
+  bare-v8-to-istanbul@1.0.4
+    https://github.com/holepunchto/bare-v8-to-istanbul
+  bare-ws@3.2.0
     https://github.com/holepunchto/bare-ws
   bare-zlib@1.4.1
     https://github.com/holepunchto/bare-zlib
@@ -869,7 +894,7 @@ JavaScript Dependencies
     https://github.com/holepunchto/hypercore-id-encoding
   hypercore-stats@2.5.1
     https://github.com/holepunchto/hypercore-stats
-  hypercore-storage@3.2.1
+  hypercore-storage@3.3.1
     https://github.com/holepunchto/hypercore-storage
   hyperdb@6.9.0
     https://github.com/holepunchto/hyperdb
@@ -879,9 +904,9 @@ JavaScript Dependencies
     https://github.com/holepunchto/hyperdht-stats
   hyperdispatch@1.6.0
     https://github.com/holepunchto/hyperdispatch
-  hyperdrive@13.3.3
+  hyperdrive@13.3.4
     https://github.com/holepunchto/hyperdrive
-  hyperschema@1.24.0
+  hyperschema@1.26.2
     https://github.com/holepunchto/hyperschema
   hyperswarm-stats@1.3.1
     https://github.com/holepunchto/hyperswarm-stats
@@ -907,11 +932,11 @@ JavaScript Dependencies
     https://github.com/holepunchto/refcounter
   require-addon@1.3.0
     https://github.com/holepunchto/require-addon
-  require-asset@1.2.2
+  require-asset@1.2.4
     https://github.com/holepunchto/require-asset
   resource-on-exit@1.0.0
     https://github.com/holepunchto/bare-teardown
-  rocksdb-native@3.17.4
+  rocksdb-native@3.18.1
     https://github.com/holepunchto/rocksdb-native
   scope-lock@1.2.4
     https://github.com/holepunchto/scope-lock
@@ -921,7 +946,7 @@ JavaScript Dependencies
     https://github.com/holepunchto/sub-encoder
   text-decoder@1.2.7
     https://github.com/holepunchto/text-decoder
-  udx-native@1.21.1
+  udx-native@1.21.3
     https://github.com/holepunchto/udx-native
   unslab@1.3.0
     https://github.com/holepunchto/unslab
@@ -949,10 +974,6 @@ JavaScript Dependencies
 
 --- mit (MIT License) ---
 
-  @jridgewell/resolve-uri@3.1.2
-    https://github.com/jridgewell/resolve-uri
-  @jridgewell/sourcemap-codec@1.6.0
-    https://github.com/jridgewell/sourcemaps
   @jridgewell/trace-mapping@0.3.31
     https://github.com/jridgewell/sourcemaps
   @types/istanbul-lib-coverage@2.0.6
@@ -967,7 +988,7 @@ JavaScript Dependencies
     https://github.com/mafintosh/codecs
   convert-source-map@2.0.0
     https://github.com/thlorenz/convert-source-map
-  corestore@7.12.4
+  corestore@7.12.5
     https://github.com/holepunchto/corestore
   debounceify@1.1.0
     https://github.com/mafintosh/debounceify
@@ -987,13 +1008,13 @@ JavaScript Dependencies
     https://github.com/mafintosh/generate-string
   hyperbee@2.27.3
     https://github.com/holepunchto/hyperbee
-  hypercore@11.35.5
+  hypercore@11.36.1
     https://github.com/holepunchto/hypercore
   hypercore-crypto@3.7.0
     https://github.com/mafintosh/hypercore-crypto
   hyperdht@6.34.0
     https://github.com/holepunchto/hyperdht
-  hyperswarm@4.17.1
+  hyperswarm@4.17.2
     https://github.com/holepunchto/hyperswarm
   is-options@1.0.2
     https://github.com/mafintosh/is-options
@@ -1001,6 +1022,8 @@ JavaScript Dependencies
     https://github.com/mikolalysenko/is-property
   kademlia-routing-table@1.0.6
     https://github.com/mafintosh/kademlia-routing-table
+  llm-splitter@0.2.0
+    https://github.com/nearform/llm-splitter
   mutexify@1.4.0
     https://github.com/mafintosh/mutexify
   nat-sampler@1.0.1
@@ -1071,6 +1094,6 @@ JavaScript Dependencies
     https://github.com/mafintosh/xache
   z32@1.1.0
     https://github.com/mafintosh/z32
-  zod@4.6.1
+  zod@4.6.5
     https://github.com/colinhacks/zod
 

package/README.md

@@ -144,7 +144,9 @@ initialized, the event omits the resource block.
 
 Use `assessModelFit` to check, before downloading anything, whether models are
 likely to fit in this device's memory. It reads generated catalog metadata plus a
-fresh memory sample — no weights, no load, no native probe:
+fresh memory sample, and for a single candidate the registry's weightless
+description of the artifact, so the engine's own fitter can answer — no weights,
+no load:
 
 ```ts
 import { assessModelFit, QWEN3_8B_INST_Q4_K_M } from '@qvac/sdk'

package/dist/examples/abot-world.js

@@ -1,4 +1,4 @@
-import { cancel, loadModel, unloadModel, worldCreateScene, worldStep, ABOT_WORLD_0_5B_Q8_0, ABOT_WORLD_0_5B_LF_VAE, ABOT_WORLD_0_5B_LF_VAE_F16, UMT5_XXL_ENC_Q8_0 } from '@qvac/sdk';
+import { cancel, loadModel, unloadModel, worldCreateScene, worldStep, ABOT_WORLD_0_5B_Q8_0, ABOT_WORLD_0_5B_LF_TAEHV_VAE, ABOT_WORLD_0_5B_LF_WAN_VAE, UMT5_XXL_ENC_Q8_0 } from '@qvac/sdk';
 import fs from 'fs';
 import path from 'path';
 // ABot-World: build a world from a single image, then walk it. Each step
@@ -38,9 +38,9 @@ try {
             //   vaeModelSrc   — the full-precision Wan2.2 VAE. ENCODES the first frame
             //     during worldCreateScene, and is not touched again after that.
             // Swapping them loads at first and fails once the walk starts.
-            taehvModelSrc: ABOT_WORLD_0_5B_LF_VAE,
+            taehvModelSrc: ABOT_WORLD_0_5B_LF_TAEHV_VAE,
             t5XxlModelSrc: UMT5_XXL_ENC_Q8_0,
-            vaeModelSrc: ABOT_WORLD_0_5B_LF_VAE_F16,
+            vaeModelSrc: ABOT_WORLD_0_5B_LF_WAN_VAE,
             world: {
                 seed: 42,
                 // ~3.7x fewer frame-passes per block. Without it, block times ramp from

package/dist/examples/advisory-model-fit.d.ts

@@ -0,0 +1,59 @@
+/**
+ * Advisory llama.cpp fit check (QVAC-22629).
+ *
+ * Before a completion or embedding load, the SDK runs `@qvac/model-fit` in one
+ * disposable Bare child and projects whether the exact configuration it is
+ * about to load will fit in device memory.
+ *
+ * The result is ADVISORY. It never blocks a load. `does-not-fit` is logged and
+ * the ordinary load path runs unchanged. Crashes, timeouts, malformed
+ * responses, unsupported configurations, and internal errors all resolve to
+ * "no evidence" and are equally non-blocking. Nothing consumes the verdict
+ * yet — this PR only produces it.
+ *
+ * The verdict is emitted on the SDK server log stream, not to stdout, so this
+ * example subscribes to `loggingStream({ id: SDK_LOG_ID })` and reprints the
+ * `[advisory-fit:…]` lines.
+ *
+ * ---------------------------------------------------------------------------
+ * What fits on this machine (Apple M4 Pro, 24 GiB unified memory)
+ * ---------------------------------------------------------------------------
+ *
+ * Measured with `@qvac/model-fit@0.8.0` (the first release carrying the
+ * qvac-fabric#214 memory-reporting fix) at the default 1024 MiB margin. The
+ * fitter budgets against what the machine can actually keep resident
+ * (total − wired − compressor: 17.4 GiB on this machine at idle), not the raw
+ * RAM figure.
+ *
+ *   PROJECTED TO FIT — all layers on GPU
+ *     Qwen3.5 0.8B  Q4_K_M   0.5 GiB  @   4k ctx
+ *     gpt-oss-20B   Q4_K_M  10.8 GiB  @  32k ctx
+ *     gte-large     fp16     0.6 GiB  (embedding, context pinned to 512)
+ *
+ *   PROJECTED NOT TO FIT — try any of these to see a `does-not-fit` verdict
+ *     gpt-oss-20B   Q4_K_M  10.8 GiB  @ 128k ctx with an f32 KV cache ← below
+ *     Gemma 4 31B   Q4_K_M  18.3 GiB  @   1k ctx   ← the weights alone do not
+ *                                                    fit with gpu_layers: 99
+ *
+ * gpt-oss-20B at 128k context FITS with the default KV cache and DOES NOT FIT
+ * once `cache-type-k`/`cache-type-v` are set to `f32`. Same model, same
+ * context, same machine — the verdict tracks the configuration, not the file
+ * size.
+ *
+ * Two boundaries worth understanding when reading verdicts:
+ *
+ * 1. The verdict answers for a PLACEMENT, not a model. The SDK's schema
+ *    default is `gpu_layers: 99` ("everything on the GPU"), so `does-not-fit`
+ *    means "not at this placement". Omit `gpu_layers` and the fitter is free
+ *    to move layers to the CPU side instead.
+ * 2. A `does-not-fit` configuration can still RUN on macOS when the OS
+ *    compresses and pages hard enough — and a `fits` configuration right at
+ *    the boundary can still fail at first decode under memory pressure.
+ *    Prediction cannot separate those cases from a snapshot; the addon-side
+ *    probe decode (QVAC-24114) is the runtime check that catches the
+ *    remainder. The verdict here is the honest working-set budget, and the
+ *    measured failure modes punish over-committing, so treat `does-not-fit`
+ *    as "expect degradation or decode failure", not "the load will error".
+ */
+export {};
+//# sourceMappingURL=advisory-model-fit.d.ts.map

package/dist/examples/advisory-model-fit.js

@@ -0,0 +1,156 @@
+/**
+ * Advisory llama.cpp fit check (QVAC-22629).
+ *
+ * Before a completion or embedding load, the SDK runs `@qvac/model-fit` in one
+ * disposable Bare child and projects whether the exact configuration it is
+ * about to load will fit in device memory.
+ *
+ * The result is ADVISORY. It never blocks a load. `does-not-fit` is logged and
+ * the ordinary load path runs unchanged. Crashes, timeouts, malformed
+ * responses, unsupported configurations, and internal errors all resolve to
+ * "no evidence" and are equally non-blocking. Nothing consumes the verdict
+ * yet — this PR only produces it.
+ *
+ * The verdict is emitted on the SDK server log stream, not to stdout, so this
+ * example subscribes to `loggingStream({ id: SDK_LOG_ID })` and reprints the
+ * `[advisory-fit:…]` lines.
+ *
+ * ---------------------------------------------------------------------------
+ * What fits on this machine (Apple M4 Pro, 24 GiB unified memory)
+ * ---------------------------------------------------------------------------
+ *
+ * Measured with `@qvac/model-fit@0.8.0` (the first release carrying the
+ * qvac-fabric#214 memory-reporting fix) at the default 1024 MiB margin. The
+ * fitter budgets against what the machine can actually keep resident
+ * (total − wired − compressor: 17.4 GiB on this machine at idle), not the raw
+ * RAM figure.
+ *
+ *   PROJECTED TO FIT — all layers on GPU
+ *     Qwen3.5 0.8B  Q4_K_M   0.5 GiB  @   4k ctx
+ *     gpt-oss-20B   Q4_K_M  10.8 GiB  @  32k ctx
+ *     gte-large     fp16     0.6 GiB  (embedding, context pinned to 512)
+ *
+ *   PROJECTED NOT TO FIT — try any of these to see a `does-not-fit` verdict
+ *     gpt-oss-20B   Q4_K_M  10.8 GiB  @ 128k ctx with an f32 KV cache ← below
+ *     Gemma 4 31B   Q4_K_M  18.3 GiB  @   1k ctx   ← the weights alone do not
+ *                                                    fit with gpu_layers: 99
+ *
+ * gpt-oss-20B at 128k context FITS with the default KV cache and DOES NOT FIT
+ * once `cache-type-k`/`cache-type-v` are set to `f32`. Same model, same
+ * context, same machine — the verdict tracks the configuration, not the file
+ * size.
+ *
+ * Two boundaries worth understanding when reading verdicts:
+ *
+ * 1. The verdict answers for a PLACEMENT, not a model. The SDK's schema
+ *    default is `gpu_layers: 99` ("everything on the GPU"), so `does-not-fit`
+ *    means "not at this placement". Omit `gpu_layers` and the fitter is free
+ *    to move layers to the CPU side instead.
+ * 2. A `does-not-fit` configuration can still RUN on macOS when the OS
+ *    compresses and pages hard enough — and a `fits` configuration right at
+ *    the boundary can still fail at first decode under memory pressure.
+ *    Prediction cannot separate those cases from a snapshot; the addon-side
+ *    probe decode (QVAC-24114) is the runtime check that catches the
+ *    remainder. The verdict here is the honest working-set budget, and the
+ *    measured failure modes punish over-committing, so treat `does-not-fit`
+ *    as "expect degradation or decode failure", not "the load will error".
+ */
+import { completion, loadModel, unloadModel, loggingStream, SDK_LOG_ID, QWEN3_5_0_8B_MULTIMODAL_Q4_K_M, GPT_OSS_20B_INST_Q4_K_M } from '@qvac/sdk';
+// The oversized load below is expected to be reported as `does-not-fit` and
+// then attempted anyway, because the check is advisory. It really does try to
+// allocate ~11 GiB, so it stays opt-in.
+const ATTEMPT_OVERSIZED = process.env['QVAC_FIT_DEMO_ATTEMPT_OVERSIZED'] === '1';
+// Reprint the worker's advisory verdicts. They arrive on the SDK server log
+// stream; everything else on that stream is filtered out to keep this readable.
+//
+// Called once per phase rather than once for the process: a subscription
+// currently stops delivering after any `unloadModel`, so a single one would go
+// silent before the second verdict. Resubscribing after the unload works.
+function watchVerdicts() {
+    void (async () => {
+        for await (const log of loggingStream({ id: SDK_LOG_ID })) {
+            if (log.message.includes('[advisory-fit:')) {
+                console.log(`▸ [${log.level.toUpperCase()}] ${log.message}`);
+            }
+        }
+    })().catch(() => {
+        // Stream terminated — normal on shutdown.
+    });
+}
+watchVerdicts();
+try {
+    // 1. A load the fitter projects to fit. The verdict carries the plan it
+    //    projected: resolved context, offloaded layers, and GPU device count.
+    console.log('▸ Loading Qwen3.5 0.8B @ 4k — expected verdict: projected to fit');
+    const smallModelId = await loadModel({
+        modelSrc: QWEN3_5_0_8B_MULTIMODAL_Q4_K_M,
+        modelConfig: { ctx_size: 4096 }
+    });
+    console.log(`▸ Loaded ${smallModelId}\n`);
+    const result = completion({
+        modelId: smallModelId,
+        history: [{ role: 'user', content: 'Say hello in five words.' }],
+        stream: false,
+        generationParams: { predict: 48 }
+    });
+    const final = await result.final;
+    console.log(`▸ Completion still works normally: ${final.contentText.trim().slice(0, 120)}\n`);
+    // Unloaded before the next phase, so the second verdict is measured on an
+    // idle machine and stays comparable to the fixture tables above. Leaving it
+    // loaded would shift the verdict: the check reserves resident weight bytes
+    // through the fit margin, and since model-fit 0.8.0 the fit child also sees
+    // system-wide wired memory.
+    await unloadModel({ modelId: smallModelId, clearStorage: false });
+    watchVerdicts();
+    // 2. A load the fitter projects NOT to fit. The point of this example is that
+    //    the SDK reports the verdict and then loads anyway — the check is
+    //    evidence, not admission control.
+    if (!ATTEMPT_OVERSIZED) {
+        console.log('▸ Skipping the oversized gpt-oss-20B load.');
+        console.log('▸ Set QVAC_FIT_DEMO_ATTEMPT_OVERSIZED=1 to let it run and watch the');
+        console.log('  load proceed past a `does-not-fit` verdict (allocates ~11 GiB).');
+    }
+    else {
+        console.log('▸ Loading gpt-oss-20B @ 128k with an f32 KV cache');
+        console.log('▸ Expected verdict: projected NOT to fit');
+        console.log('▸ The load is attempted regardless. That is the fail-open contract:');
+        console.log('  the verdict is evidence, not admission control.\n');
+        const bigModelId = await loadModel({
+            modelSrc: GPT_OSS_20B_INST_Q4_K_M,
+            modelConfig: {
+                ctx_size: 131072,
+                'cache-type-k': 'f32',
+                'cache-type-v': 'f32'
+            }
+        });
+        console.log(`▸ Load returned ${bigModelId} — the advisory verdict did not block it`);
+        // Loading is not the same as being usable: a model can load and then fail
+        // at decode time. Gemma 4 31B does exactly that on this machine. So run a
+        // real completion and report throughput rather than trusting the load.
+        try {
+            const check = completion({
+                modelId: bigModelId,
+                history: [{ role: 'user', content: 'Name three colours. Answer briefly.' }],
+                stream: false,
+                generationParams: { predict: 40 }
+            });
+            const checkFinal = await check.final;
+            console.log(`▸ ...and it actually runs: ${checkFinal.stats?.tokensPerSecond?.toFixed(1) ?? '?'} tok/s ` +
+                `— the OS compressed and paged its way past the working-set budget`);
+        }
+        catch (inferenceError) {
+            console.log(`▸ ...but it cannot run: ${inferenceError instanceof Error ? inferenceError.message : String(inferenceError)}`);
+            console.log('▸ The verdict was right, and `loadModel` succeeding did not mean usable.');
+        }
+        await unloadModel({ modelId: bigModelId, clearStorage: false });
+    }
+}
+catch (error) {
+    // A failing load here is the native loader's own error, not the fit check.
+    // The check never throws and never converts a verdict into a load failure.
+    console.error('✖', error instanceof Error ? error.message : error);
+    process.exitCode = 1;
+}
+// The log subscription is an open stream and would otherwise keep the process
+// alive after the work is done.
+process.exit(process.exitCode ?? 0);

package/dist/examples/asr/parakeet-nemotron-filesystem.d.ts

@@ -0,0 +1,2 @@
+export {};
+//# sourceMappingURL=parakeet-nemotron-filesystem.d.ts.map