fix: embedding context deadlock (#402)

Author: giladgd

.github/workflows/build.yml

@@ -23,7 +23,8 @@ jobs:
       - name: Download latest llama.cpp release
         env:
           CI: true
-        run: node ./dist/cli/cli.js source download --release latest --skipBuild --noBundle --noUsageExample --updateBinariesReleaseMetadataAndSaveGitBundle
+        # pinned to `b4291` temporarily until the Windows on Arm64 build is fixed
+        run: node ./dist/cli/cli.js source download --release b4291 --skipBuild --noBundle --noUsageExample --updateBinariesReleaseMetadataAndSaveGitBundle
       - name: Upload build artifact
         uses: actions/upload-artifact@v4
         with:

src/evaluator/LlamaModel/LlamaModel.ts

@@ -517,14 +517,7 @@ export class LlamaModel {
         if (this._vocabOnly)
             throw new Error("Model is loaded in vocabOnly mode, so no context can be created");
 
-        return await withLock(this._llama._memoryLock, LlamaLocks.loadToMemory, options.createSignal, async () => {
-            const preventDisposalHandle = this._backendModelDisposeGuard.createPreventDisposalHandle();
-            try {
-                return await LlamaEmbeddingContext._create({_model: this}, options);
-            } finally {
-                preventDisposalHandle.dispose();
-            }
-        });
+        return await LlamaEmbeddingContext._create({_model: this}, options);
     }
 
     /**