Add delay between sending ttir and continuing compilation to allow pa…

…rent process to read results before potential crash

tt_torch/csrc/bindings.cpp

@@ -136,14 +136,15 @@ std::string compile_stable_hlo_to_ttir(std::string_view code) {
   return ret;
 }
 
-std::tuple<py::bytes, std::string_view>
+std::tuple<py::bytes, std::string>
 compile_ttir_to_bytestream(std::string_view code) {
   auto [binary, ttnn] = tt::torch::compileTTIRToTTNN(code);
   auto size = ::flatbuffers::GetSizePrefixedBufferLength(
       static_cast<const uint8_t *>(binary->get()));
 
   std::string data_str(static_cast<const char *>(binary->get()), size);
   delete binary;
+
   return std::make_tuple(py::bytes(data_str), ttnn);
 }
 

tt_torch/csrc/tt-mlir-interface.cpp

@@ -94,7 +94,7 @@ std::string compileStableHLOToTTIR(std::string_view code) {
   return buffer;
 }
 
-std::tuple<std::shared_ptr<void> *, std::string_view>
+std::tuple<std::shared_ptr<void> *, std::string>
 compileTTIRToTTNN(std::string_view code) {
 
   mlir::MLIRContext context;
@@ -141,10 +141,10 @@ compileTTIRToTTNN(std::string_view code) {
 
   std::string buffer;
   llvm::raw_string_ostream os(buffer);
-  mlir_module.get()->print(os);
+  mlir_module->print(os);
   os.flush();
 
-  return std::make_tuple(binary, std::string_view(buffer));
+  return std::make_tuple(binary, buffer);
 }
 
 std::shared_ptr<void> *Compile(std::string_view code) {

tt_torch/csrc/tt-mlir-interface.hpp

@@ -11,6 +11,6 @@
 namespace tt::torch {
 std::shared_ptr<void> *Compile(std::string_view code);
 std::string compileStableHLOToTTIR(std::string_view code);
-std::tuple<std::shared_ptr<void> *, std::string_view>
+std::tuple<std::shared_ptr<void> *, std::string>
 compileTTIRToTTNN(std::string_view code);
 } // namespace tt::torch

tt_torch/dynamo/backend.py

@@ -107,6 +107,7 @@ def compile_process(receiver, sender):
     asm = obj["asm"]
     ttir = tt_mlir.compile_stable_hlo_to_ttir(asm)
     sender.put({"ttir": ttir})
+    time.sleep(0.1)
     binary, ttnn = tt_mlir.compile_ttir_to_bytestream(ttir)
     sender.put({"binary": binary, "ttnn": ttnn})
     time.sleep(0.1)
@@ -246,15 +247,16 @@ def compile_op(self, node, *inputs, **kwargs):
         process.start()
         sender.put(obj)
         start = time.time()
-        result = {}
+        binary = None
         while True:
             try:
                 result = receiver.get_nowait()
                 if "ttir" in result:
                     op.compilation_status = OpCompilationStatus.CONVERTED_TO_TTIR
                     op.add_ttir_graph(result["ttir"])
                 if "binary" in result:
-                    op.binary = result["binary"]
+                    binary = result["binary"]
+                    op.binary = binary
                     op.add_ttnn_graph(result["ttnn"])
                     op.compilation_status = OpCompilationStatus.CONVERTED_TO_TTNN
                     break
@@ -267,7 +269,7 @@ def compile_op(self, node, *inputs, **kwargs):
                 break
             time.sleep(0.01)
         process.join()
-        return result["binary"], op
+        return binary, op
 
     def run_op(self, binary, *inputs):
         sender = mp.Queue()