generate and pass binary json before passing binary to python

tt_torch/csrc/bindings.cpp

@@ -168,16 +168,18 @@ std::string compile_stable_hlo_to_ttir(std::string_view code) {
   return ret;
 }
 
-std::tuple<py::bytes, std::string>
+std::tuple<py::bytes, std::string, std::string>
 compile_ttir_to_bytestream(std::string_view code) {
-  auto [binary, ttnn] = tt::torch::compileTTIRToTTNN(code);
+  auto [binary_ptr, ttnn] = tt::torch::compileTTIRToTTNN(code);
   auto size = ::flatbuffers::GetSizePrefixedBufferLength(
-      static_cast<const uint8_t *>(binary->get()));
+      static_cast<const uint8_t *>(binary_ptr->get()));
+  tt::runtime::Binary binary = tt::runtime::Binary(*binary_ptr);
 
-  std::string data_str(static_cast<const char *>(binary->get()), size);
-  delete binary;
+  std::string json = binary.asJson();
+  std::string data_str(static_cast<const char *>(binary_ptr->get()), size);
+  delete binary_ptr;
 
-  return std::make_tuple(py::bytes(data_str), ttnn);
+  return std::make_tuple(py::bytes(data_str), ttnn, json);
 }
 
 py::bytes compile_stablehlo_to_bytestream(std::string_view code) {

tt_torch/dynamo/backend.py

@@ -79,8 +79,8 @@ def compile_process(receiver, sender, ttir_event, ttnn_event):
     ttir = tt_mlir.compile_stable_hlo_to_ttir(asm)
     sender.put({"ttir": ttir})
     ttir_event.wait()
-    binary, ttnn = tt_mlir.compile_ttir_to_bytestream(ttir)
-    sender.put({"binary": binary, "ttnn": ttnn})
+    binary, ttnn, json = tt_mlir.compile_ttir_to_bytestream(ttir)
+    sender.put({"binary": binary, "ttnn": ttnn, "binary_json": json})
     ttnn_event.wait()
     sys.exit(0)
 
@@ -275,7 +275,7 @@ def compile_op(self, node, *inputs, **kwargs):
                 if "binary" in result:
                     binary = result["binary"]
                     op.binary = binary
-                    op.json = tt_mlir.bytestream_to_json(binary)
+                    op.json = result["binary_json"]
                     op.add_ttnn_graph(result["ttnn"])
                     ttnn_event.set()
                     op.compilation_status = OpCompilationStatus.CONVERTED_TO_TTNN
@@ -549,7 +549,7 @@ def _base_backend(gm: torch.fx.GraphModule, example_inputs, compiler_config):
     if compiler_config.enable_intermediate_verification:
         executor.register_intermediate_callback(verify_golden_callback)
 
-    binary, ttnn = tt_mlir.compile_ttir_to_bytestream(ttir)
+    binary, ttnn, _ = tt_mlir.compile_ttir_to_bytestream(ttir)
 
     if dump_intermediates:
         print("TTNN module", file=sys.stderr)