Update

[ghstack-poisoned]

Author: JCNTH

backends/webgpu/runtime/WebGPUBackend.cpp

@@ -104,19 +104,18 @@ Error WebGPUBackend::execute(
   // Copy inputs from EValue tensors to GPU buffers
   std::vector<InputData> inputs;
   inputs.reserve(num_inputs);
-  for (size_t i = 0; i < num_inputs; i++) {
-    const auto& tensor = args[i]->toTensor();
-    const bool host_is_int64 =
-        tensor.scalar_type() == executorch::aten::ScalarType::Long;
-    inputs.push_back({tensor.const_data_ptr(), tensor.nbytes(), host_is_int64});
-  }
   // Fail loud as a runtime Error so a throw never crosses the backend boundary.
   try {
-    // Dynamic shapes: shrink each input to its live sizes before upload
-    // (mirrors Vulkan maybe_resize_input). No-op when unchanged, so a static
-    // graph is byte-identical.
+    // Build the input list and, for dynamic shapes, shrink each input to its
+    // live sizes before upload (mirrors Vulkan maybe_resize_input). No-op when
+    // unchanged, so a static graph is byte-identical.
     for (size_t i = 0; i < num_inputs; i++) {
-      const auto sizes = args[i]->toTensor().sizes();
+      const auto& tensor = args[i]->toTensor();
+      const bool host_is_int64 =
+          tensor.scalar_type() == executorch::aten::ScalarType::Long;
+      inputs.push_back(
+          {tensor.const_data_ptr(), tensor.nbytes(), host_is_int64});
+      const auto sizes = tensor.sizes();
       std::vector<int64_t> new_dims(sizes.begin(), sizes.end());
       graph->resize_input(graph->input_ids()[i], new_dims);
     }

backends/webgpu/runtime/WebGPUGraph.cpp

@@ -142,16 +142,14 @@ void WebGPUGraph::update_symints_from_inputs(
     // Reads the [0,..,index,..,0] element; symint sources are scalar-ish.
     const int64_t offset = static_cast<int64_t>(index) * stride;
     const void* host = inputs[pos].data;
-    // Stored elem_size (live nbytes/numel mis-derives for a dynamic source).
-    const size_t elem_size = tensors_[src.input_tensor_id].elem_size;
+    // Interpret the HOST buffer by its scalar type, not the tensor's serialized
+    // elem_size: copy_inputs narrows an int64 host input to an int32 buffer, so
+    // elem_size (buffer-derived) would misread int64 host data as int32.
     int32_t val;
-    if (elem_size == sizeof(int64_t)) {
+    if (inputs[pos].host_is_int64) {
       val = static_cast<int32_t>(static_cast<const int64_t*>(host)[offset]);
-    } else if (elem_size == sizeof(int32_t)) {
-      val = static_cast<const int32_t*>(host)[offset];
     } else {
-      throw std::runtime_error(
-          "select_as_symint: unsupported input element size");
+      val = static_cast<const int32_t*>(host)[offset];
     }
     set_symint(src.symint_id, val);
   }

backends/webgpu/runtime/ops/select_as_symint/SelectAsSymint.cpp

@@ -91,7 +91,9 @@ void register_sym_binary(
   if (graph.get_value_type(a) == WebGPUGraph::ValueType::SymInt) {
     graph.add_resize_hook(a, recompute);
   }
-  if (graph.get_value_type(b) == WebGPUGraph::ValueType::SymInt) {
+  // b != a: for a self-op (e.g. x + x) both operands share one id; register the
+  // recompute hook once, not twice.
+  if (b != a && graph.get_value_type(b) == WebGPUGraph::ValueType::SymInt) {
     graph.add_resize_hook(b, recompute);
   }
 }