fix(llm): snapshot vision encoder output before caching

packages/react-native-executorch/common/runner/encoders/vision_encoder.cpp

@@ -112,7 +112,7 @@ Result<EValue> VisionEncoder::encode(const MultimodalInput &input) {
 
   auto it = embedding_cache_.find(path);
   if (it != embedding_cache_.end()) {
-    return it->second;
+    return EValue(*it->second.tensor);
   }
 
   auto shape = ET_UNWRAP(getInputShape());
@@ -128,9 +128,18 @@ Result<EValue> VisionEncoder::encode(const MultimodalInput &input) {
       chw.data(), sizes, ::executorch::aten::ScalarType::Float);
 
   auto result = ET_UNWRAP(module_->execute(kVisionEncoderMethod, image_tensor));
-  auto embedding = result[0];
-  embedding_cache_.emplace(path, embedding);
-  return embedding;
+  auto out_tensor = result[0].toTensor();
+
+  CachedEmbedding cached;
+  cached.bytes.resize(out_tensor.nbytes());
+  std::memcpy(cached.bytes.data(), out_tensor.const_data_ptr(),
+              out_tensor.nbytes());
+  cached.sizes.assign(out_tensor.sizes().begin(), out_tensor.sizes().end());
+  cached.dtype = out_tensor.scalar_type();
+  auto [entry, inserted] = embedding_cache_.emplace(path, std::move(cached));
+  entry->second.tensor = ::executorch::extension::from_blob(
+      entry->second.bytes.data(), entry->second.sizes, entry->second.dtype);
+  return EValue(*entry->second.tensor);
 }
 
 } // namespace executorch::extension::llm

packages/react-native-executorch/common/runner/encoders/vision_encoder.h

@@ -2,11 +2,14 @@
 #pragma once
 
 #include "iencoder.h"
+#include <cstdint>
 #include <executorch/extension/module/module.h>
+#include <executorch/extension/tensor/tensor.h>
 #include <executorch/runtime/core/evalue.h>
 #include <runner/multimodal_input.h>
 #include <string>
 #include <unordered_map>
+#include <vector>
 
 namespace executorch::extension::llm {
 
@@ -26,13 +29,23 @@ class VisionEncoder : public IEncoder {
     bool with_batch;
   };
 
+  // The method's output EValue aliases the runtime's reusable output buffer,
+  // which the NEXT vision_encoder.execute() overwrites — caching it directly
+  // silently turns earlier images into the most recently encoded one. Cache
+  // an owned byte snapshot instead and hand out a tensor over those bytes.
+  struct CachedEmbedding {
+    std::vector<uint8_t> bytes;
+    std::vector<::executorch::aten::SizesType> sizes;
+    ::executorch::aten::ScalarType dtype;
+    ::executorch::extension::TensorPtr tensor;
+  };
+
   ::executorch::runtime::Result<ImageShape> getInputShape() const;
   std::vector<float> preprocessImage(const std::string &path,
                                      const ImageShape &targetShape) const;
 
   ::executorch::extension::Module *module_;
-  std::unordered_map<std::string, ::executorch::runtime::EValue>
-      embedding_cache_;
+  std::unordered_map<std::string, CachedEmbedding> embedding_cache_;
 };
 
 } // namespace executorch::extension::llm