QVAC-22629 feat: run an advisory llama.cpp fit check before loadModel

packages/inference/NOTICE

@@ -63,14 +63,24 @@ Third-Party Model Licenses
 
   abot-world-0-5b-lf-dit-q8_0
     https://huggingface.co/acvlab/ABot-World-0-5B-LF
+  audio8-codec-decoder-f16
+    https://huggingface.co/Audio8/Audio8-TTS-Preview-0.6b
+  audio8-codec-encoder-f16
+    https://huggingface.co/Audio8/Audio8-TTS-Preview-0.6b
+  audio8-lm-f16
+    https://huggingface.co/Audio8/Audio8-TTS-Preview-0.6b
   bci-embedder
     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-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-s3tok-f16
+    https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512
   craft_mlt_25k
     https://www.jaided.ai/easyocr/modelhub/
   crnn_mobilenet_v3_small
@@ -179,6 +189,8 @@ Third-Party Model Licenses
     https://huggingface.co/unsloth/Qwen3.6-27B-GGUF
   Qwen3.6-35B-A3B-GGUF
     https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF
+  Qwen3.8-27B-GGUF
+    https://huggingface.co/unsloth/Qwen3.8-27B-GGUF
   ru-base-ggml-model-f16
     https://huggingface.co/CheeLi03/whisper-base-rus-8
   ru-tiny-ggml-model-f16
@@ -199,6 +211,10 @@ Third-Party Model Licenses
     https://huggingface.co/Comfy-Org/Wan_2.1_ComfyUI_repackaged
   unlimited-ocr-gguf
     https://huggingface.co/vimalnakrani/unlimited-ocr-gguf
+  VisionPsy-Nano-460M-Flash-GGUFs
+    https://huggingface.co/qvac/VisionPsy-Nano-460M-Flash-GGUFs
+  VisionPsy-Nano-460M-GGUFs
+    https://huggingface.co/qvac/VisionPsy-Nano-460M-GGUFs
   voice-en
     https://huggingface.co/FunAudioLLM/Fun-CosyVoice3-0.5B-2512
   voice-zh-2
@@ -348,6 +364,12 @@ Third-Party Model Licenses
     https://huggingface.co/ChristianAzinn/gte-large-gguf
   gte-large-gguf
     https://huggingface.co/ChristianAzinn/gte-large-gguf
+  indic-conformer-ctc.f16
+    https://huggingface.co/ai4bharat/indic-conformer-600m-multilingual
+  indic-conformer-ctc.q4_0
+    https://huggingface.co/ai4bharat/indic-conformer-600m-multilingual
+  indic-conformer-ctc.q8_0
+    https://huggingface.co/ai4bharat/indic-conformer-600m-multilingual
   lavasr-denoiser-f16
     https://huggingface.co/spaces/YatharthS
   lavasr-enhancer-f16
@@ -646,7 +668,9 @@ JavaScript Dependencies
   @qvac/error@0.1.1
   @qvac/logging@0.1.1
     https://github.com/tetherto/qvac
-  @qvac/rag@0.6.4
+  @qvac/model-fit@0.8.0
+    https://github.com/tetherto/qvac
+  @qvac/rag@0.8.0
     https://github.com/tetherto/qvac
   @qvac/registry-client@0.6.1
     https://github.com/tetherto/qvac
@@ -672,23 +696,25 @@ JavaScript Dependencies
     https://github.com/holepunchto/bare-cpu-info
   bare-crypto@1.15.3
     https://github.com/holepunchto/bare-crypto
-  bare-dns@2.1.4
+  bare-dns@2.2.0
     https://github.com/holepunchto/bare-dns
   bare-env@3.0.1
     https://github.com/holepunchto/bare-env
-  bare-events@2.9.1
+  bare-events@2.9.2
     https://github.com/holepunchto/bare-events
+  bare-fetch@3.2.0
+    https://github.com/holepunchto/bare-fetch
   bare-form-data@1.2.2
     https://github.com/holepunchto/bare-form-data
-  bare-fs@4.8.0
+  bare-fs@4.8.1
     https://github.com/holepunchto/bare-fs
   bare-gpu-info@0.1.1
     https://github.com/holepunchto/bare-gpu-info
   bare-hrtime@2.1.1
     https://github.com/holepunchto/bare-hrtime
-  bare-http-parser@1.1.5
+  bare-http-parser@2.1.4
     https://github.com/holepunchto/bare-http-parser
-  bare-http1@4.5.8
+  bare-http1@4.6.1
     https://github.com/holepunchto/bare-http1
   bare-https@3.0.0
     https://github.com/holepunchto/bare-https
@@ -714,17 +740,25 @@ JavaScript Dependencies
     https://github.com/holepunchto/bare-process
   bare-rpc@1.3.8
     https://github.com/holepunchto/bare-rpc
+  bare-runtime@1.31.0
+    https://github.com/holepunchto/bare-runtime
+  bare-runtime-darwin-arm64@1.31.0
+    https://github.com/holepunchto/bare-runtime
   bare-semver@1.1.0
     https://github.com/holepunchto/bare-semver
   bare-signals@5.0.0
     https://github.com/holepunchto/bare-signals
   bare-stdio@1.0.3
     https://github.com/holepunchto/bare-stdio
-  bare-stream@2.13.3
+  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-subprocess@6.1.0
+    https://github.com/holepunchto/bare-subprocess
   bare-tcp@2.5.4
     https://github.com/holepunchto/bare-tcp
-  bare-tls@3.1.8
+  bare-tls@3.1.9
     https://github.com/holepunchto/bare-tls
   bare-tty@5.1.2
     https://github.com/holepunchto/bare-tty
@@ -736,7 +770,7 @@ JavaScript Dependencies
     https://github.com/holepunchto/bare-zlib
   blind-relay@1.6.1
     https://github.com/holepunchto/blind-relay
-  compact-encoding@3.3.0
+  compact-encoding@3.3.2
     https://github.com/holepunchto/compact-encoding
   compact-encoding-bitfield@1.1.0
     https://github.com/compact-encoding/compact-encoding-bitfield
@@ -760,7 +794,7 @@ JavaScript Dependencies
     https://github.com/holepunchto/hypercore-stats
   hypercore-storage@3.2.1
     https://github.com/holepunchto/hypercore-storage
-  hyperdb@6.8.0
+  hyperdb@6.9.0
     https://github.com/holepunchto/hyperdb
   hyperdht-address@1.1.1
     https://github.com/holepunchto/hyperdht-address
@@ -780,7 +814,7 @@ JavaScript Dependencies
     https://github.com/holepunchto/mirror-drive
   noise-handshake@4.2.0
     https://github.com/holepunchto/noise-handshake
-  paparam@1.12.0
+  paparam@1.13.0
     https://github.com/holepunchto/paparam
   passive-core-watcher@1.0.1
     https://github.com/holepunchto/passive-core-watcher
@@ -796,6 +830,8 @@ JavaScript Dependencies
     https://github.com/holepunchto/refcounter
   require-addon@1.2.0
     https://github.com/holepunchto/require-addon
+  require-asset@1.2.2
+    https://github.com/holepunchto/require-asset
   resource-on-exit@1.0.0
     https://github.com/holepunchto/bare-teardown
   rocksdb-native@3.17.4
@@ -808,7 +844,7 @@ JavaScript Dependencies
     https://github.com/holepunchto/sub-encoder
   text-decoder@1.2.7
     https://github.com/holepunchto/text-decoder
-  udx-native@1.21.0
+  udx-native@1.21.1
     https://github.com/holepunchto/udx-native
   unslab@1.3.0
     https://github.com/holepunchto/unslab
@@ -840,7 +876,7 @@ JavaScript Dependencies
     https://github.com/mafintosh/bogon
   codecs@3.1.0
     https://github.com/mafintosh/codecs
-  corestore@7.12.0
+  corestore@7.12.2
     https://github.com/holepunchto/corestore
   debounceify@1.1.0
     https://github.com/mafintosh/debounceify
@@ -858,11 +894,11 @@ JavaScript Dependencies
     https://github.com/mafintosh/generate-string
   hyperbee@2.27.3
     https://github.com/holepunchto/hyperbee
-  hypercore@11.35.1
+  hypercore@11.35.2
     https://github.com/holepunchto/hypercore
   hypercore-crypto@3.7.0
     https://github.com/mafintosh/hypercore-crypto
-  hyperdht@6.33.1
+  hyperdht@6.33.2
     https://github.com/holepunchto/hyperdht
   hyperswarm@4.17.0
     https://github.com/holepunchto/hyperswarm
@@ -910,9 +946,9 @@ JavaScript Dependencies
     https://github.com/holepunchto/sodium-universal
   speedometer@1.1.0
     https://github.com/mafintosh/speedometer
-  streamx@2.28.0
+  streamx@2.28.1
     https://github.com/mafintosh/streamx
-  tar-stream@3.2.0
+  tar-stream@3.2.1
     https://github.com/mafintosh/tar-stream
   teex@1.0.1
     https://github.com/mafintosh/teex

packages/inference/package.json

@@ -163,6 +163,7 @@
   "dependencies": {
     "@qvac/error": "^0.1.1",
     "@qvac/logging": "^0.1.1",
+    "@qvac/model-fit": "^0.8.0",
     "@qvac/rag": "^0.8.0",
     "@qvac/registry-client": "^0.6.1",
     "bare-abort-controller": "^1.1.2",
@@ -177,6 +178,7 @@
     "bare-os": "^3.9.3",
     "bare-path": "^3.1.1",
     "bare-rpc": "^1.3.8",
+    "bare-runtime": "^1.24.2",
     "bare-stream": "^2.13.3",
     "bare-url": "^2.4.5",
     "bare-zlib": "^1.4.0",
@@ -187,6 +189,7 @@
     "hyperswarm": "^4.17.0",
     "semver": "^7.8.5",
     "tar-stream": "^3.2.0",
+    "which-runtime": "^1.2.1",
     "zod": "^4.4.3"
   },
   "peerDependencies": {

packages/inference/src/plugins/builtin/llamacpp-embedding/plugin.ts

@@ -1,4 +1,4 @@
-import EmbedLlamacpp, { IdMapIndex, type GGMLConfig } from '@qvac/embed-llamacpp'
+import EmbedLlamacpp, { IdMapIndex } from '@qvac/embed-llamacpp'
 import embedAddonLogging from '@qvac/embed-llamacpp/addonLogging'
 import type { TurboVecIndexProvider } from '@qvac/rag'
 import {
@@ -19,6 +19,7 @@ import { embed } from '@/plugins/ops/embed'
 import { forwardModelExecution } from '@/profiling/model-execution'
 import { isMobile } from '@/runtime/state'
 import { stripMultiGpuKeys } from '@/utils/multi-gpu-mobile'
+import { transformEmbedConfig } from '@/plugins/builtin/llamacpp-embedding/transform'
 
 const turbovecIndexProvider: TurboVecIndexProvider = {
   create(options) {
@@ -29,53 +30,6 @@ const turbovecIndexProvider: TurboVecIndexProvider = {
   }
 }
 
-function transformEmbedConfig(embedConfig: EmbedConfig): GGMLConfig {
-  const config: GGMLConfig = {
-    device: embedConfig.device as 'gpu' | 'cpu',
-    gpu_layers: `${embedConfig.gpuLayers}` as `${number}`,
-    batch_size: `${embedConfig.batchSize}` as `${number}`
-  }
-
-  if (embedConfig.flashAttention) {
-    config.flash_attn = embedConfig.flashAttention
-  }
-
-  if (embedConfig.pooling) {
-    config.pooling = embedConfig.pooling
-  }
-
-  if (embedConfig.attention) {
-    config.attention = embedConfig.attention
-  }
-
-  if (typeof embedConfig.embdNormalize === 'number') {
-    config.embd_normalize = `${embedConfig.embdNormalize}`
-  }
-
-  if (embedConfig.mainGpu !== undefined) {
-    config['main-gpu'] =
-      typeof embedConfig.mainGpu === 'number' ? `${embedConfig.mainGpu}` : embedConfig.mainGpu
-  }
-
-  if (embedConfig.splitMode) {
-    config['split-mode'] = embedConfig.splitMode
-  }
-
-  if (embedConfig.tensorSplit) {
-    config['tensor-split'] = embedConfig.tensorSplit
-  }
-
-  if (typeof embedConfig.verbosity === 'number') {
-    config.verbosity = `${embedConfig.verbosity}`
-  }
-
-  if (embedConfig.openclCacheDir) {
-    config.openclCacheDir = embedConfig.openclCacheDir
-  }
-
-  return config
-}
-
 function createEmbeddingsModel(modelId: string, modelPath: string, embedConfig: EmbedConfig) {
   const logger = createStreamLogger(modelId, ModelType.llamacppEmbedding)
   registerAddonLogger(modelId, ModelType.llamacppEmbedding, logger)

packages/inference/src/plugins/builtin/llamacpp-embedding/transform.ts

@@ -0,0 +1,55 @@
+import type { GGMLConfig } from '@qvac/embed-llamacpp'
+import { type EmbedConfig } from '@/schemas/index'
+
+/**
+ * Converts an EmbedConfig into the flat string-keyed map the C++ addon expects.
+ *
+ * Extracted from the plugin so the advisory fit check can build its request
+ * from the same transform the real embedding load uses.
+ */
+export function transformEmbedConfig(embedConfig: EmbedConfig): GGMLConfig {
+  const config: GGMLConfig = {
+    device: embedConfig.device as 'gpu' | 'cpu',
+    gpu_layers: `${embedConfig.gpuLayers}` as `${number}`,
+    batch_size: `${embedConfig.batchSize}` as `${number}`
+  }
+
+  if (embedConfig.flashAttention) {
+    config.flash_attn = embedConfig.flashAttention
+  }
+
+  if (embedConfig.pooling) {
+    config.pooling = embedConfig.pooling
+  }
+
+  if (embedConfig.attention) {
+    config.attention = embedConfig.attention
+  }
+
+  if (typeof embedConfig.embdNormalize === 'number') {
+    config.embd_normalize = `${embedConfig.embdNormalize}`
+  }
+
+  if (embedConfig.mainGpu !== undefined) {
+    config['main-gpu'] =
+      typeof embedConfig.mainGpu === 'number' ? `${embedConfig.mainGpu}` : embedConfig.mainGpu
+  }
+
+  if (embedConfig.splitMode) {
+    config['split-mode'] = embedConfig.splitMode
+  }
+
+  if (embedConfig.tensorSplit) {
+    config['tensor-split'] = embedConfig.tensorSplit
+  }
+
+  if (typeof embedConfig.verbosity === 'number') {
+    config.verbosity = `${embedConfig.verbosity}`
+  }
+
+  if (embedConfig.openclCacheDir) {
+    config.openclCacheDir = embedConfig.openclCacheDir
+  }
+
+  return config
+}

packages/inference/src/plugins/ops/load-model.ts

@@ -21,6 +21,7 @@ import {
   ModelFileLocateFailedError
 } from '@/errors/index'
 import { getPlugin } from '@/plugins/index'
+import { runAdvisoryFitCheck } from '@/resources/model-fit/native-probe/advisory-fit'
 import { promises as fsPromises } from 'bare-fs'
 import path from 'bare-path'
 import { getEngineLogger } from '@/logging/index'
@@ -93,6 +94,21 @@ export async function loadModel(
     }
   }
 
+  // Advisory: every outcome — including a projected insufficiency — continues
+  // to the ordinary load below. Runs after config resolution and path
+  // validation so it sees the same state the real load uses, and before
+  // `createModel()` so it never competes with the native load for device
+  // memory. The outcome is stored on the registry entry for internal use;
+  // it is deliberately not exposed on any public API yet.
+  const fitProbe = await runAdvisoryFitCheck({
+    modelId,
+    modelType: modelType as CanonicalModelType,
+    modelPath,
+    modelConfig,
+    artifacts,
+    isShardedModel
+  })
+
   logger.info(`${modelType}: Loading model ${modelId}...`)
   startLogBuffering(modelId)
 
@@ -118,7 +134,8 @@ export async function loadModel(
       path: modelPath,
       config: modelConfig,
       modelType: modelType as CanonicalModelType,
-      name: modelName
+      name: modelName,
+      fitProbe
     })
 
     const loadResult: LoadModelResult =