Python API debug symbols by CharlieL7 · Pull Request #4803 · ROCm/AMDMIGraphX

CharlieL7 · 2026-04-20T18:56:30Z

Motivation

Add debug symbols functionality to python API

Technical Details

Do not merge till Netron output update - use protobuff, API integration, debug symbols #4701 is merged.
Can now get/add/remove debug symbols to instructions through the python API.

Changelog Category

Add a CHANGELOG.md entry for any option other than Not Applicable

- Added: New functionality.
- Changed: Changes to existing functionality.
- Removed: Functionality or support that has been removed. (Compared to a previous release)
- Optimized: Component performance that has been optimized or improved.
- Resolved Issues: Known issues from a previous version that have been resolved.
- Not Applicable: This PR is not to be included in the changelog.

…symbols

Traverse over ins with empty debug symbols also just in case. Probably will not occur however

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

…ins_debug_symbols

…symbols

…put_update

…to netron_output_update

…_debug_symbols

…to python_api_debug_symbols

Copilot

Pull request overview

This PR extends MIGraphX’s Python bindings to expose instruction/module debug symbols and adds a Netron export path that emits an ONNX ModelProto binary (including debug symbols when present), updating the driver/docs/tests accordingly.

Changes:

Added Python API bindings for instruction_ref.get_debug_symbols() and module add/remove/has_debug_symbols.
Reworked Netron export from JSON to ONNX protobuf binary via write_netron_output(...), and added program APIs to write it to a file.
Added C++/C/Python tests covering Netron output and debug symbols.

Reviewed changes

Copilot reviewed 24 out of 24 changed files in this pull request and generated 6 comments.

Show a summary per file

File	Description
`src/py/migraphx_py.cpp`	Exposes debug symbols + `program.write_netron_output()` in pybind11 module.
`src/onnx/netron_output.cpp`	Implements ONNX `ModelProto` Netron export and includes debug symbols as node attributes.
`src/include/migraphx/netron_output.hpp`	Updates public Netron export API to a stream-based writer.
`src/api/api.cpp` / `tools/api/api.cpp`	Adds C API plumbing to write Netron output to a file.
`src/api/include/migraphx/migraphx.h` / `migraphx.hpp`	Adds `migraphx_program_write_netron_output` and C++ wrapper method.
`src/driver/main.cpp` + docs (`docs/*`)	Switches `--netron` output to protobuf binary and updates help/docs text.
`test/*`	Adds/updates C++ and Python tests for Netron export and debug symbols.
`src/program.cpp` / `src/include/migraphx/program.hpp`	Exposes program file version via getter and centralizes constant.
`src/CMakeLists.txt` / `codecov.yml`	Removes old `src/netron_output.cpp` from build/coverage ignore.
`CHANGELOG.md`	Adds release note about Netron export and debug symbols.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-04-20T19:05:43Z

+{
+    onnx::ModelProto model;
+    auto prog_value = prog.to_value();
+    model.set_ir_version(prog.get_program_file_version());


model.set_ir_version(prog.get_program_file_version()) uses MIGraphX's MXR serialization version as the ONNX IR version. ONNX consumers expect an ONNX IR version (e.g., onnx::IR_VERSION), and using an unrelated value can make the exported model invalid or rejected. Set ir_version to the ONNX constant and consider also populating opset_import to improve compatibility with ONNX tooling/Netron.

Suggested change

model.set_ir_version(prog.get_program_file_version());

model.set_ir_version(onnx::IR_VERSION);

Copilot · 2026-04-20T19:05:44Z

+#include <ostream>
 #include <migraphx/config.hpp>
 #include <migraphx/program.hpp>
+#include <migraphx/onnx/export.h>

 namespace migraphx {
 inline namespace MIGRAPHX_INLINE_NS {

-MIGRAPHX_EXPORT std::string make_netron_output(const program& prog);
+MIGRAPHX_ONNX_EXPORT void write_netron_output(const program& prog, std::ostream& os);


This header replaces the existing make_netron_output(const program&) -> std::string API with write_netron_output(program, ostream), which is a breaking change for C++ consumers. Consider keeping make_netron_output as a compatibility wrapper (e.g., implemented via std::ostringstream + write_netron_output) and/or deprecating it instead of removing it outright.

Copilot · 2026-04-20T19:05:44Z

 * Updated `argmin` and `argmax` ops to be implemented as reduction ops, so they now have JIT support and can fuse (#4620).
 * Replaced usages of `std::cout` and `std::cerr` with the logger (#4732)
 * Converted RNN variable sequence length operations (`rnn_var_sl_shift_sequence`, `rnn_var_sl_shift_output`, `rnn_var_sl_last_output`) from device implementation to JIT compilation (#4755).
+* Updated netron output to create an ONNX-like protobuff. Now also includes debug symbols if enabled. (#4701)


The changelog entry says "protobuff"; the correct term is "protobuf".

Suggested change

* Updated netron output to create an ONNX-like protobuff. Now also includes debug symbols if enabled. (#4701)

* Updated netron output to create an ONNX-like protobuf. Now also includes debug symbols if enabled. (#4701)

Copilot · 2026-04-20T19:05:45Z

+static void write_netron_output_file(const program& p, const char* filename)
+{
+    std::ofstream os(filename, std::ios::binary);
+    if(not os.is_open())
+        MIGRAPHX_THROW(migraphx_status_bad_param,
+                       "Failed to open file for writing: " + std::string(filename));
+    write_netron_output(p, os);


write_netron_output_file constructs std::ofstream and later builds std::string(filename) without validating filename. If the C API is called with a null pointer, this is undefined behavior and can crash before returning a status. Add an explicit filename == nullptr (and ideally empty-string) check and throw migraphx_status_bad_param before using it.

Copilot · 2026-04-20T19:05:45Z

+static void write_netron_output_file(const program& p, const char* filename)
+{
+    std::ofstream os(filename, std::ios::binary);
+    if(not os.is_open())
+        MIGRAPHX_THROW(migraphx_status_bad_param,
+                        "Failed to open file for writing: " + std::string(filename));
+    write_netron_output(p, os);


Template version has the same issue as the generated C API: write_netron_output_file uses filename without checking for null, and std::string(filename) will crash on nullptr. Add a filename == nullptr (and optionally empty-string) check before opening/formatting the filename.

pfultz2 · 2026-04-20T22:08:30Z

+            [](migraphx::module& mm, migraphx::instruction_ref ins) {
+                mm.remove_debug_symbols(ins);
+            },
+            py::arg("ins"))


Only get_debug_symbols should be exposed to the API.

What if a user wants to add a debug symbol to an instruction when manually adding instructions?

We should add a debug_symbols parameter to the add_instruction, insert_instruction, add_macro, insert_macro, etc methods. So they can write m.add_instruction(op("add"), [x, y], debug_symbols=['addOp']).

This will make it consistent and can handle the macros, which is a little more complicated.

Hold off on the C++ API as we need a better design to make it more extensible, and it doesnt even support macros right now.

pfultz2 · 2026-04-20T22:10:13Z


    void print() const { call(&migraphx_program_print, this->get_handle_ptr()); }

+    void write_netron_output(const char* filename) const


Why are we adding this to the C++ API?

This is from #4701. There wasn't a way to write out as ONNX-like protobuff from the C++ API before, only the driver.

Why is #4701 in this PR? The python API debug symbols are not related to this.

eddieliao · 2026-04-20T22:28:07Z

+    case shape::fp8e5m2_type: return onnx::TensorProto::FLOAT8E5M2;
+    case shape::fp8e5m2fnuz_type: return onnx::TensorProto::FLOAT8E5M2FNUZ;
+    case shape::tuple_type: return onnx::TensorProto::UNDEFINED;
+    case shape::fp4x2_type: return onnx::TensorProto::UINT4;


Can this be onnx::TensorProto::FLOAT4E2M1 or are we limited by ONNX opset version? Will using UINT4 result in display issues for fp4 types?

This change is also from #4701. The update should be doable. I made the netron output before ONNX had the type.

CharlieL7 and others added 30 commits January 13, 2026 16:41

progress

2e4d205

dd

bb622af

Merge branch 'develop' of github.com:ROCm/AMDMIGraphX into ins_debug_…

277ff30

…symbols

Add onnx_parser and module changes

432ed83

Change to use a set of debug_symbols when inserting ins

c2381cb

Incomplete first parse design

b00e7e2

Version 2 using scoped debug symbols in module

1517438

Cleanup

b8e1cdb

cleanup 2

ec7487e

Formatting

f499489

Add debug propagate with matcher context

8324d9b

formatting

c9f567f

tests and fixup

e66c848

Merge branch 'develop' of github.com:ROCm/AMDMIGraphX into ins_debug_…

42eaac8

…symbols

Add docs changes

a769c11

Have gather_replace_debug_symbols traverse empty

7321509

Traverse over ins with empty debug symbols also just in case. Probably will not occur however

Copilot suggestions

2794a66

more copilot

69c787a

Update CHANGELOG.md

d90382c

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Update src/module.cpp

107a7e7

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Revert matcher.hpp

f9bb2c8

Revert env_var doc

3d3bc42

Review progress

66b0a2f

Merge branch 'ins_debug_symbols' of github.com:ROCm/AMDMIGraphX into …

3cff180

…ins_debug_symbols

Module tracking number of instructions with debug_symbols

b3c8a1c

Cleanup

5830120

Merge branch 'develop' of github.com:ROCm/AMDMIGraphX into ins_debug_…

f3e9709

…symbols

Revert simplify_algebra

593fc75

Fix implementation and tests

fbdd1e8

splice fix and add gemm->add->add test

4e1cee9

CharlieL7 added 18 commits April 13, 2026 13:53

Merge branch 'develop' of github.com:ROCm/AMDMIGraphX into netron_out…

3f69f9b

…put_update

Fix merge conflicts leftover

fe15f32

Resolve conflict markers

57e649b

Make program ir version static constexpr

801bf36

netron output only main module

31fd8c1

api check output stream

40e8e99

Remove raw_data entry

8c9599a

Licensing

78445f8

Add ostream checking for python api

836cf4d

remove typos

ab03d6d

Update tests to have ONNX protobuff expected value

b6c35dc

license

930d480

formatting

6681273

add changelog

a99ab17

Merge branch 'netron_output_update' of github.com:ROCm/AMDMIGraphX in…

0a11c12

…to netron_output_update

Merge branch 'develop' of github.com:ROCm/AMDMIGraphX into python_api…

f4a960b

…_debug_symbols

Merge branch 'netron_output_update' of github.com:ROCm/AMDMIGraphX in…

9fae5c4

…to python_api_debug_symbols

Add tests and pybind funcs

7055913

CharlieL7 requested review from ahsan-ca, eddieliao and pfultz2 April 20, 2026 18:56

CharlieL7 self-assigned this Apr 20, 2026

Copilot AI review requested due to automatic review settings April 20, 2026 18:56

CharlieL7 requested review from a team and causten as code owners April 20, 2026 18:56

Copilot started reviewing on behalf of CharlieL7 April 20, 2026 18:57 View session

Copilot AI reviewed Apr 20, 2026

View reviewed changes

pfultz2 reviewed Apr 20, 2026

View reviewed changes

eddieliao reviewed Apr 20, 2026

View reviewed changes

	model.set_ir_version(prog.get_program_file_version());
	model.set_ir_version(onnx::IR_VERSION);

	* Updated netron output to create an ONNX-like protobuff. Now also includes debug symbols if enabled. (#4701)
	* Updated netron output to create an ONNX-like protobuf. Now also includes debug symbols if enabled. (#4701)


		void print() const { call(&migraphx_program_print, this->get_handle_ptr()); }

		void write_netron_output(const char* filename) const

Conversation

CharlieL7 commented Apr 20, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Motivation

Technical Details

Changelog Category

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Apr 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 20, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Copilot AI Apr 20, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 20, 2026

Choose a reason for hiding this comment

Uh oh!

pfultz2 Apr 20, 2026

Choose a reason for hiding this comment

Uh oh!

CharlieL7 Apr 21, 2026

Choose a reason for hiding this comment

Uh oh!

pfultz2 Apr 21, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

pfultz2 Apr 20, 2026

Choose a reason for hiding this comment

Uh oh!

CharlieL7 Apr 21, 2026

Choose a reason for hiding this comment

Uh oh!

pfultz2 Apr 21, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

eddieliao Apr 20, 2026

Choose a reason for hiding this comment

Uh oh!

CharlieL7 Apr 21, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

CharlieL7 commented Apr 20, 2026 •

edited

Loading

pfultz2 Apr 21, 2026 •

edited

Loading