@trillboards/edge-sdk 0.2.109 → 0.2.110

package/deploy/windows/setup-litertlm.ps1

@@ -211,10 +211,11 @@ function Write-LitertlmConfig {
 # /v1/models answers without ever building an engine, so it cannot tell a working
 # GPU encoder from one that fails on first use.
 function Test-LitertlmVision {
+  param([string]$InferenceModel = "$ModelId,gpu")
   try {
     $dataUri = "data:image/jpeg;base64," + $script:LitertlmProbeJpegB64
     $body = @{
-      model = "$ModelId,gpu"
+      model = $InferenceModel
       messages = @(@{ role = "user"; content = @(
         @{ type = "text"; text = "Reply with the single word: ok" },
         @{ type = "image_url"; image_url = @{ url = $dataUri } }
@@ -227,7 +228,7 @@ function Test-LitertlmVision {
       -TimeoutSec 180 -UseBasicParsing
     return ($resp.StatusCode -eq 200)
   } catch {
-    Log "vision probe did not answer: $_"
+    Log "vision probe for '$InferenceModel' did not answer: $_"
     return $false
   }
 }
@@ -367,23 +368,32 @@ function Get-LitertLmVersion {
   }
 }
 
-# A HEALTHY SERVE IS NOT ENOUGH TO SKIP PROVISIONING.
+# A HEALTHY CONTROL ROUTE IS NOT ENOUGH TO SKIP PROVISIONING.
 #
 # `Test-ServeHealthy` substring-matches the bare model id, and every litert-lm
 # version answers that identically - including the ones that reject the `,gpu`
-# model id the GPU delegate is selected with. So this fast path, which exists to
-# make a repeat run cheap, was also the path that let an already-installed
-# 0.14.0 declare itself ready forever: the hosts that most need the pinned
-# runtime are exactly the ones with a serve already answering. The version is
-# therefore part of the short-circuit condition, not a check further down that a
-# healthy serve never reaches.
+# model id the GPU delegate is selected with, and without building an inference
+# engine. So this fast path may run only after BOTH the pinned version check and
+# a real image inference. Otherwise a failed first install can leave /v1/models
+# answering, and a reinstall falsely declares success without re-testing the
+# operation that just failed.
 $installedLitertLmVersion = Get-LitertLmVersion
-if ((Test-ServeHealthy) -and ($installedLitertLmVersion -eq $LitertLmVersion)) {
-  Complete-Setup
-  Stop-SetupLock
-  exit 0
+$serveHealthy = Test-ServeHealthy
+if ($serveHealthy -and ($installedLitertLmVersion -eq $LitertLmVersion)) {
+  if (Test-LitertlmVision -InferenceModel "$ModelId,gpu") {
+    Complete-Setup
+    Stop-SetupLock
+    exit 0
+  }
+  if (Test-LitertlmVision -InferenceModel $ModelId) {
+    Log "local AI healthy on the previously configured full CPU fallback (slower, but serving)"
+    Complete-Setup
+    Stop-SetupLock
+    exit 0
+  }
+  Log "WARN: the pinned serve answers /v1/models but not a real GPU or CPU image inference; reprovisioning local AI"
 }
-if ((Test-ServeHealthy) -and ($installedLitertLmVersion -ne $LitertLmVersion)) {
+if ($serveHealthy -and ($installedLitertLmVersion -ne $LitertLmVersion)) {
   Log "serve is healthy but litert-lm is $(if ($installedLitertLmVersion) { $installedLitertLmVersion } else { 'unknown' }) not pinned $LitertLmVersion; reinstalling and restarting the serve"
 }
 
@@ -686,7 +696,7 @@ if (-not $healthy) {
 # never builds an engine -- so it cannot see a GPU vision encoder that fails on
 # first USE. If the probe inference does not answer we fall back to the CPU
 # encoder rather than leave this host unable to sense at all.
-if (-not (Test-LitertlmVision)) {
+if (-not (Test-LitertlmVision -InferenceModel "$ModelId,gpu")) {
   Log "WARN: GPU vision encoder did not answer a probe inference; falling back to the CPU encoder"
   Write-LitertlmConfig | Out-Null
   & powershell.exe -NoProfile -ExecutionPolicy Bypass -File $registration -LitertLmPath $litertlmPath -Port $Port *>> $LogPath
@@ -700,7 +710,9 @@ if (-not (Test-LitertlmVision)) {
     Stop-SetupLock
     exit 1
   }
-  # RE-PROBE, with the same probe that sent us here.
+  # RE-PROBE the CPU-vision tier, first keeping GPU inference for hosts where
+  # only the image encoder failed, then using the same bare-model CPU fallback
+  # that LiteRtVlmProvider uses at runtime when the main GPU delegate fails.
   #
   # Test-ServeHealthy above only proves /v1/models answers, and this file's own
   # comment six lines up says that route never builds an engine. So before this
@@ -710,15 +722,18 @@ if (-not (Test-LitertlmVision)) {
   # install: the cascade detected the failure, downgraded a tier, and then
   # verified the downgrade with a check that cannot fail for this reason.
   #
-  # A CPU fallback that cannot infer is not a slower tier. It is a broken
+  # A full CPU fallback that cannot infer is not a slower tier. It is a broken
   # install, and the operator has to hear so at setup time rather than discover
   # it as silence from a camera weeks later.
-  if (-not (Test-LitertlmVision)) {
-    Fail-LocalAi "Gemma is installed and the local AI service is answering, but it could not analyze a test image on either the GPU or the CPU vision encoder. See $LogPath."
+  if (Test-LitertlmVision -InferenceModel "$ModelId,gpu") {
+    Log "local AI healthy on the CPU vision encoder with GPU inference (slower, but serving)"
+  } elseif (Test-LitertlmVision -InferenceModel $ModelId) {
+    Log "local AI healthy on the full CPU fallback (slower, but serving)"
+  } else {
+    Fail-LocalAi "Gemma is installed and the local AI service is answering, but it could not analyze a test image with either GPU inference or the full CPU fallback. See $LogPath."
     Stop-SetupLock
     exit 1
   }
-  Log "local AI healthy on the CPU vision encoder (slower, but serving)"
 } else {
   Log "local AI healthy with the GPU vision encoder"
 }

package/dist/EdgeAgent.js

@@ -44,7 +44,7 @@ const edge_ads_1 = require("@trillboards/edge-ads");
 const CommandRouter_1 = require("./CommandRouter");
 const config_1 = require("./config");
 const dataDir_1 = require("./dataDir");
-const SDK_VERSION = '0.2.109';
+const SDK_VERSION = '0.2.110';
 class EdgeAgent extends events_1.EventEmitter {
     constructor(config, deps) {
         super();

package/dist/diagnose.d.ts

@@ -31,7 +31,7 @@ export interface DiagnosticReport {
         skipped: number;
     };
 }
-declare const SDK_VERSION = "0.2.109";
+declare const SDK_VERSION = "0.2.110";
 declare function checkNodeVersion(): CheckResult;
 declare function checkPlatform(): CheckResult;
 declare function checkMemory(): CheckResult;

package/dist/diagnose.js

@@ -63,7 +63,7 @@ const localNetworkAccess_1 = require("./localNetworkAccess");
 // ---------------------------------------------------------------------------
 // SDK version — mirrors package.json; kept in sync by CI
 // ---------------------------------------------------------------------------
-const SDK_VERSION = '0.2.109';
+const SDK_VERSION = '0.2.110';
 exports._SDK_VERSION = SDK_VERSION;
 // ---------------------------------------------------------------------------
 // ANSI helpers (no external deps)

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@trillboards/edge-sdk",
-  "version": "0.2.109",
+  "version": "0.2.110",
   "description": "Composable cross-platform Edge AI SDK for audience sensing on DOOH/CTV devices",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",