fix: temporarily pin to a working llama.cpp release

Author: giladgd

.github/workflows/build.yml

@@ -23,8 +23,8 @@ jobs:
       - name: Download latest llama.cpp release
         env:
           CI: true
-        # pinned to `b4357` temporarily until the Windows on Arm64 build is fixed
-        run: node ./dist/cli/cli.js source download --release b4357 --skipBuild --noBundle --noUsageExample --updateBinariesReleaseMetadataAndSaveGitBundle
+        # pinned to `b4333` temporarily until the Windows on Arm64 build is fixed
+        run: node ./dist/cli/cli.js source download --release b4333 --skipBuild --noBundle --noUsageExample --updateBinariesReleaseMetadataAndSaveGitBundle
       - name: Upload build artifact
         uses: actions/upload-artifact@v4
         with:

llama/addon/AddonSampler.cpp

@@ -350,10 +350,15 @@ Napi::Value AddonSampler::ApplyConfig(const Napi::CallbackInfo& info) {
 
         if (shouldCreateSampler) {
             repeatPenaltySampler = llama_sampler_init_penalties(
+                llama_n_vocab(model->model),
+                llama_token_eos(model->model),
+                llama_token_nl(model->model),
                 repeatPenaltyMaxTokens,
                 repeatPenalty,
                 repeatPenaltyFrequencyPenalty,
-                repeatPenaltyPresencePenalty
+                repeatPenaltyPresencePenalty,
+                true,
+                false
             );
             repeatPenalty_lastTokens = RingBuffer<llama_token>(repeatPenaltyMaxTokens);