tetherto · iancris · Aug 18, 2026 · Aug 14, 2026 · Aug 14, 2026 · Aug 14, 2026
@@ -1,5 +1,12 @@
 # Changelog
 
+## [0.33.0] - 2026-08-17
+
+### Changed
+
+- `qvac-fabric` dependency bumped `10069.0.0` -> `10069.1.0` (VisionPsy Nano
+  support and its Flash preprocessing rule; no API change for this package).
+
 ## [0.32.0] - 2026-08-10
 
 ### Changed

@@ -1,6 +1,6 @@
 {
   "name": "@qvac/embed-llamacpp",
-  "version": "0.32.0",
+  "version": "0.33.0",
   "description": "bert addon for qvac",
   "addon": true,
   "engines": {

@@ -6,7 +6,7 @@
     },
     {
       "name": "qvac-fabric",
-      "version>=": "10069.0.0"
+      "version>=": "10069.1.0"
     },
     {
       "name": "qvac-lib-inference-addon-cpp",

@@ -1,5 +1,12 @@
 # Changelog
 
+## [0.5.0] - 2026-08-17
+
+### Changed
+
+- `qvac-fabric` dependency bumped `10069.0.0` -> `10069.1.0` (VisionPsy Nano
+  support and its Flash preprocessing rule; no API change for this package).
+
 ## [0.4.0] - 2026-08-10
 
 ### Changed

@@ -1,6 +1,6 @@
 {
   "name": "@qvac/fabric",
-  "version": "0.4.0",
+  "version": "0.5.0",
   "description": "Shared bare addon hosting the qvac-fabric (forked llama.cpp + ggml) runtime for QVAC inference addons",
   "addon": true,
   "engines": {

@@ -6,7 +6,7 @@
     },
     {
       "name": "qvac-fabric",
-      "version>=": "10069.0.0"
+      "version>=": "10069.1.0"
     },
     {
       "name": "qvac-lint-cpp",

@@ -1,5 +1,20 @@
 # Changelog
 
+## [0.44.0] - 2026-08-17
+
+### Added
+
+- `image_no_upscale` in the addon load config — an idefics3-style preprocessing
+  override forwarded to the vision context, accepting `"on"` or `"off"`. Left
+  unset, the model's own GGUF value is used unchanged. This is what separates the
+  VisionPsy Flash checkpoint from the base one, whose mmprojs are otherwise
+  indistinguishable: a Flash checkpoint loaded without it silently runs base
+  preprocessing, which changes the image token count and so moves both accuracy
+  and encode time.
+- `qvac-fabric` dependency bumped `10069.0.0` -> `10069.1.0` (VisionPsy Nano
+  support and its Flash preprocessing rule), which is what supplies
+  `image_no_upscale` on `common_params` and `mtmd_context_params`.
+
 ## [0.43.0] - 2026-08-14
 
 This release removes the Qwen3-only dynamic tools feature behind

@@ -36,6 +36,31 @@ static void handleImageTileMode(common_params& params, const std::string& raw) {
   }
 }
 
+// Selects the idefics3-style no-upscale preprocessing rule. Tri-state on the
+// fabric side, but the config map can only say "the caller set something", so
+// this handler only ever writes 0 or 1; leaving the key out keeps fabric's -1
+// model default. Needed because the VisionPsy base and Flash mmprojs declare
+// identical vision hparams, so without it a Flash model silently runs base
+// preprocessing.
+static void
+handleImageNoUpscale(common_params& params, const std::string& raw) {
+  std::string val = raw;
+  std::transform(val.begin(), val.end(), val.begin(), ::tolower);
+  if (val == "1" || val == "on" || val == "true") {
+    params.image_no_upscale = 1;
+  } else if (val == "0" || val == "off" || val == "false") {
+    params.image_no_upscale = 0;
+  } else {
+    throw qvac_errors::StatusError(
+        errors::ADDON_ID,
+        qvac_errors::general_error::toString(
+            qvac_errors::general_error::InvalidArgument),
+        string_format(
+            "image-no-upscale must be 0/off/false or 1/on/true, got: %s",
+            raw.c_str()));
+  }
+}
+
 static void
 handleImageMaxTokens(common_params& params, const std::string& raw) {
   try {
@@ -73,6 +98,8 @@ const LoadConfigHandlerList LOAD_CONFIG_HANDLERS = {
     {"image_max_tokens", handleImageMaxTokens},
     {"image-min-tokens", handleImageMinTokens},
     {"image_min_tokens", handleImageMinTokens},
+    {"image-no-upscale", handleImageNoUpscale},
+    {"image_no_upscale", handleImageNoUpscale},
 };
 
 void applyLoadConfigHandlers(

@@ -204,6 +204,7 @@ void MtmdLlmContext::initVisionContext() {
   mparams.print_timings = true;
   mparams.n_threads = params_.cpuparams.n_threads;
   mparams.image_tile_mode = params_.image_tile_mode;
+  mparams.image_no_upscale = params_.image_no_upscale;
   // Forward the per-image token budget to the vision encoder. These were
   // previously dropped: the addon parsed image_min/max_tokens into
   // common_params but never copied them into mtmd_context_params, so a

@@ -1,6 +1,6 @@
 {
   "name": "@qvac/llm-llamacpp",
-  "version": "0.43.0",
+  "version": "0.44.0",
   "description": "llama addon for qvac",
   "addon": true,
   "scripts": {

@@ -46,21 +46,66 @@ const noGpu = String(noGpuEnv || '').toLowerCase() === 'true'
 // CPU-only platforms (no GPU inference path today)
 const useCpu = isDarwinX64 || isLinuxArm64
 
-// The default VLM pair for every image test that does not pass its own config.
+// Read a string env var. Bare doesn't define `process` as a global at
+// module-init time, so try bare-os first and fall back behind a `typeof`
+// guard; `_envInt` further down is the integer twin of this. Going through
+// os.getEnv() is what makes a variable settable on Device Farm, which has no
+// `env:` block and injects values via os.setEnv() from qvacPerfConfig.txt.
+function _envStr(key) {
+  if (typeof os.getEnv === 'function') return os.getEnv(key) || ''
+  if (typeof process !== 'undefined' && process.env) return process.env[key] || ''
+  return ''
+}
+
+// The VLM pairs available to every image test that does not pass its own
+// config. Both sit inside ONE literal deliberately: validate-mobile-manifest.js
+// reads the prestage-set marker with a regex plus brace matching and takes the
+// first `{` after it, so a ternary here would silently drop the second pair
+// from the expected set. Keeping both inside makes the set their union, and
+// each consuming test opts the pair it isn't using out with prestage-ignore.
+//
+// No downloadUrl on these: ensureModel() resolves url + sha256 + bytes from
+// models.manifest.json by modelName and ignores any URL passed alongside it.
 // prestage-set: multimodal-default
-const MULTIMODAL_MODEL_CONFIG = {
-  llmModel: {
-    modelName: 'SmolVLM2-500M-Video-Instruct-Q8_0.gguf',
-    downloadUrl:
-      'https://huggingface.co/ggml-org/SmolVLM2-500M-Video-Instruct-GGUF/resolve/main/SmolVLM2-500M-Video-Instruct-Q8_0.gguf'
+const MULTIMODAL_MODEL_CONFIGS = {
+  smolvlm2: {
+    llmModel: { modelName: 'SmolVLM2-500M-Video-Instruct-Q8_0.gguf' },
+    projModel: { modelName: 'mmproj-SmolVLM2-500M-Video-Instruct-Q8_0.gguf' },
+    ctx_size: '2048',
+    batchCtxSize: '4096'
   },
-  projModel: {
-    modelName: 'mmproj-SmolVLM2-500M-Video-Instruct-Q8_0.gguf',
-    downloadUrl:
-      'https://huggingface.co/ggml-org/SmolVLM2-500M-Video-Instruct-GGUF/resolve/main/mmproj-SmolVLM2-500M-Video-Instruct-Q8_0.gguf'
-  },
-  ctx_size: '2048'
+  // VisionPsy Nano, base checkpoint. Needs no image-no-upscale key: base
+  // preprocessing is what its GGUF already declares. Only loadable against a
+  // fabric carrying the VisionPsy projector alias (qvac-fabric-llm.cpp#205),
+  // which on this branch comes from the vcpkg-overlays/ports/qvac-fabric pin.
+  visionpsy: {
+    llmModel: { modelName: 'visionpsy-nano-460m-q8_0.gguf' },
+    projModel: { modelName: 'mmproj-visionpsy-nano-460m-q8.gguf' },
+    ctx_size: '4096',
+    batchCtxSize: '8192'
+  }
+}
+
+// QVAC_VLM_MODEL selects the pair for every test that takes the default, so the
+// same assertions can run against either model without forking them.
+//
+// VisionPsy Nano base is the default: it is the model this suite is here to
+// cover, and it exercises real idefics3 slicing. SmolVLM2 stays available as
+// QVAC_VLM_MODEL=smolvlm2 for an A/B, but note its mmproj declares no
+// clip.vision.preproc_image_size, so fabric falls back to an overview-only
+// encode — ~64 image tokens regardless of image size, against VisionPsy's ~862.
+// It is the cheaper baseline, not the equivalent one.
+//
+// A typo throws rather than falling back, because silently measuring the wrong
+// model is the worse failure — same rule as QVAC_QWEN35_MTMD_SIZE.
+const VLM_MODEL = (_envStr('QVAC_VLM_MODEL') || 'visionpsy').toLowerCase()
+if (!MULTIMODAL_MODEL_CONFIGS[VLM_MODEL]) {
+  throw new Error(
+    `QVAC_VLM_MODEL must be one of ${Object.keys(MULTIMODAL_MODEL_CONFIGS).join(', ')} ` +
+      `(got "${VLM_MODEL}")`
+  )
 }
+const MULTIMODAL_MODEL_CONFIG = MULTIMODAL_MODEL_CONFIGS[VLM_MODEL]
 
 // Opt-in larger VLM pair — only tests that import LARGE_MULTIMODAL_CONFIG and
 // pass it to setupMultimodalInference() load these.

@@ -9,6 +9,8 @@ const LlmLlamacpp = require('../../index.js')
 const { ensureModel, safeTest, getMediaPath } = require('./utils')
 const { attachSpecLogger } = require('./spec-logger')
 // prestage-uses: multimodal-default — MULTIMODAL_MODEL_CONFIG, loaded via ensureModel() below
+// prestage-ignore: SmolVLM2-500M-Video-Instruct-Q8_0.gguf — opt-in via QVAC_VLM_MODEL=smolvlm2 only
+// prestage-ignore: mmproj-SmolVLM2-500M-Video-Instruct-Q8_0.gguf — opt-in via QVAC_VLM_MODEL=smolvlm2 only
 const { MULTIMODAL_MODEL_CONFIG } = require('./_image-common.js')
 
 const platform = os.platform()
@@ -82,6 +84,13 @@ const CASES = [
   {
     id: 'count-fingers',
     user: 'How many fingers are on one typical human hand? Answer with one word.',
+    // Workaround, not a fix: VisionPsy answers this wrong, so we ask a wording it
+    // gets right. Defensible only because this test covers batch scheduling, not
+    // answer quality. "one" and "typical" are what break it, and they are exactly
+    // what Llama-3.2-1B needs to avoid answering "Fifty", so the two paths cannot
+    // share one string. Greedy, so it is the same every run. Re-measure both
+    // models before editing either wording; full table in the commit.
+    vlmUser: 'How many fingers are on a human hand? Answer with one word.',
     expected: ['five', '5', 'ten', '10']
   },
   {
@@ -91,9 +100,15 @@ const CASES = [
   },
   { id: 'story-canyon', story: true, expected: ['canyon'] },
   {
+    // Keep this open-ended. Offering the options instead ("yellow, green, or
+    // purple?") made it worse, not better: Llama-3.2-1B picked "Green" off the
+    // list and broke a test that had been passing. A weak model will take a
+    // distractor when one is handed to it.
+    // "sand" covers VisionPsy, which answers "sandstone" — a yellow-brown shade,
+    // and a fair reading of the question rather than a wrong colour.
     id: 'primary-yellow',
     user: 'What primary color is the sun often drawn as? Answer with one word.',
-    expected: ['yellow', 'orange', 'red']
+    expected: ['yellow', 'orange', 'red', 'sand']
   },
   { id: 'story-saffron', story: true, expected: ['saffron'] }
 ]
@@ -150,7 +165,12 @@ const IMAGE_CASES = [
       'photo',
       'photograph',
       'picture',
-      'news'
+      'news',
+      // Masthead rather than headline. SmolVLM2 reads the banner headline
+      // ("STORM."); VisionPsy names the publication ("New York Times"). Both are
+      // true readings of the page, and "news" does not match "new york times".
+      'times',
+      'york'
     ]
   }
 ]
@@ -170,8 +190,19 @@ function normalizeText(text) {
     .trim()
 }
 
+// Drop a leading reasoning trace before matching. Some VLMs (VisionPsy Nano)
+// open a `<think>` block even under a one-word system prompt and with a chat
+// template that has no thinking branch, so the answer sits after it. A block
+// left unterminated by the token budget strips to empty, which fails loudly
+// rather than matching on the reasoning text.
+function stripReasoning(text) {
+  const s = String(text || '')
+  const closed = s.replace(/<think>[\s\S]*?<\/think>/g, ' ')
+  return closed.replace(/<think>[\s\S]*$/, ' ')
+}
+
 function containsExpectedWord(text, expectedOptions) {
-  const normalized = normalizeText(text)
+  const normalized = normalizeText(stripReasoning(text))
   const options = Array.isArray(expectedOptions) ? expectedOptions : [expectedOptions]
   return options.some((option) => normalized.includes(option))
 }
@@ -236,10 +267,18 @@ function buildVlmBatchItem(item) {
   return {
     id: item.id,
     prompt: [
-      { role: 'system', content: 'Answer with one word only.' },
-      { role: 'user', content: item.user }
+      // "Do not explain" is aimed at VisionPsy, which opens a <think> trace even
+      // under a one-word instruction and with a chat template that has no
+      // thinking branch. At predict 64 the trace was still unterminated, so the
+      // answer never arrived and stripReasoning() correctly reduced it to empty.
+      { role: 'system', content: 'Answer with one word only. Do not explain or think first.' },
+      // vlmUser overrides user for the VLM pair only; see count-fingers.
+      { role: 'user', content: item.vlmUser || item.user }
     ],
-    runOptions: { generationParams: { predict: 16 } }
+    // 128, not 16. A reasoning model spends a 16-token budget restating the
+    // question, and 64 was still short of closing the trace. Models that answer
+    // in one word stop at their EOG token, so this costs them nothing.
+    runOptions: { generationParams: { predict: 128 } }
   }
 }
 
@@ -285,12 +324,17 @@ async function setupMultimodalBatchModel(t, configOverrides = {}) {
   const modelPath = path.join(dirPath, modelName)
   const projModelPath = path.join(dirPath, projModelName)
 
-  // ctx_size 4096 gives each of the 4 parallel slots ~1024 tokens — enough for
-  // SmolVLM2-500M vision tokens (~256 per image) + prompt + output.
+  // Sized so each of the 4 parallel slots holds one image plus prompt and
+  // output. The per-image cost is model-specific, so the value travels with the
+  // model rather than being hardcoded: SmolVLM2-500M emits ~256 vision tokens
+  // per image, so 4096 leaves each slot ~1024. VisionPsy Nano caps its long
+  // side at 2048 and slices at 512, so both images used here become a 13-crop
+  // grid at ~858 tokens — four of those would need ~4000 of 4096 before any
+  // output, hence 8192 for that pair.
   const config = {
     device: useCpu ? 'cpu' : 'gpu',
     gpu_layers: '99',
-    ctx_size: '4096',
+    ctx_size: MULTIMODAL_MODEL_CONFIG.batchCtxSize,
     temp: '0',
     top_p: '1',
     top_k: '1',

@@ -9,6 +9,8 @@ const test = require('brittle')
 const fs = require('bare-fs')
 // prestage-uses: multimodal-default — setupMultimodalInference() default in _image-common.js
 // prestage-uses: multimodal-large — LARGE_MULTIMODAL_CONFIG, passed explicitly below
+// prestage-ignore: SmolVLM2-500M-Video-Instruct-Q8_0.gguf — opt-in via QVAC_VLM_MODEL=smolvlm2 only
+// prestage-ignore: mmproj-SmolVLM2-500M-Video-Instruct-Q8_0.gguf — opt-in via QVAC_VLM_MODEL=smolvlm2 only
 const {
   DEVICE_CONFIGS,
   LARGE_MULTIMODAL_CONFIG,

@@ -4,6 +4,8 @@
 // each image in its own group. See _image-common.js for details.
 
 // prestage-uses: multimodal-default — setupMultimodalInference() default in _image-common.js
+// prestage-ignore: SmolVLM2-500M-Video-Instruct-Q8_0.gguf — opt-in via QVAC_VLM_MODEL=smolvlm2 only
+// prestage-ignore: mmproj-SmolVLM2-500M-Video-Instruct-Q8_0.gguf — opt-in via QVAC_VLM_MODEL=smolvlm2 only
 const { runPerImageBackendTests } = require('./_image-common.js')
 
 runPerImageBackendTests({

@@ -6,6 +6,8 @@
 // from earlier iterations even when the final run OOMs.
 
 // prestage-uses: multimodal-default — setupMultimodalInference() default in _image-common.js
+// prestage-ignore: SmolVLM2-500M-Video-Instruct-Q8_0.gguf — opt-in via QVAC_VLM_MODEL=smolvlm2 only
+// prestage-ignore: mmproj-SmolVLM2-500M-Video-Instruct-Q8_0.gguf — opt-in via QVAC_VLM_MODEL=smolvlm2 only
 const { runPerImageBackendTests } = require('./_image-common.js')
 
 runPerImageBackendTests({

@@ -18,6 +18,8 @@
 const test = require('brittle')
 const fs = require('bare-fs')
 // prestage-uses: multimodal-default — setupMultimodalInference() default in _image-common.js
+// prestage-ignore: SmolVLM2-500M-Video-Instruct-Q8_0.gguf — opt-in via QVAC_VLM_MODEL=smolvlm2 only
+// prestage-ignore: mmproj-SmolVLM2-500M-Video-Instruct-Q8_0.gguf — opt-in via QVAC_VLM_MODEL=smolvlm2 only
 const {
   DEVICE_CONFIGS,
   TEST_CONSTANTS,