[In Progress] ONNX weight replacement#4957
Draft
kahmed10 wants to merge 16 commits into
Draft
Conversation
kahmed10
commented
Jun 12, 2026
Comment on lines
+114
to
+141
| Running a baked program in-process | ||
| ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ | ||
|
|
||
| ``create_program_with_weights`` deliberately does **not** finalize the program. | ||
| Finalizing uploads literal data to the device, which is wasted work if you only | ||
| intend to ``save`` the result (the bytes would be serialized after a redundant | ||
| host-to-device round trip). | ||
|
|
||
| The baked program therefore is not yet runnable on the device. The portable way | ||
| to make it runnable is to save it and load it back: loading a compiled MXR | ||
| finalizes it automatically, allocating device buffers and uploading the baked | ||
| literals. | ||
|
|
||
| .. code-block:: cpp | ||
|
|
||
| auto baked = migraphx::create_program_with_weights(prog, "weights_v1", t); | ||
| migraphx::save(baked, "model_v1.mxr"); // serialize the baked program | ||
|
|
||
| auto runnable = migraphx::load("model_v1.mxr"); // finalized on load | ||
| auto outputs = runnable.eval(params); | ||
|
|
||
| .. note:: | ||
|
|
||
| The underlying core library has a ``program::finalize(const target&)`` method | ||
| that finalizes a baked or loaded program in place, but it is **not** exposed | ||
| through the C or C++ API wrappers (``migraphx.h`` / ``migraphx.hpp``). From | ||
| C++, use the save/load round trip above. From Python the method *is* exposed | ||
| (see below) if you want to avoid touching disk. |
Collaborator
Author
There was a problem hiding this comment.
Looking for feedback on this
Regressions detected 🔴 * No develop baseline was found for this PR's branch point; compared against the latest available develop run instead. |
|
pfultz2
reviewed
Jun 15, 2026
| return gpu::allocate_gpu(s); | ||
| } | ||
|
|
||
| void target::lower_baked_literals(module& m) const |
Collaborator
There was a problem hiding this comment.
There should be a function that returns passes to lower the literals.
| } | ||
|
|
||
| if(result.is_compiled()) | ||
| t.lower_baked_literals(*mm); |
Collaborator
There was a problem hiding this comment.
You need to run the pass manager just in case there are weights in the submodules.
| std::unordered_map<std::string, module> modules; | ||
| std::vector<context> contexts; | ||
| std::vector<target> targets; | ||
| std::unordered_map<std::string, external_data_info> external_weight_map; |
Collaborator
There was a problem hiding this comment.
I dont think we should store this here. This can be stored in the IR directly by creating a onnx_externel_weights op.
| } | ||
|
|
||
| program | ||
| create_program_with_weights(const program& prog, const std::string& base_dir, const target& t) |
Collaborator
There was a problem hiding this comment.
This function is very onnx specific and should be moved to the onnx module. It should be name replace_onnx_external_weights.
| std::string external_data_path = ""; | ||
| /// When true, external-data initializers become parameters instead of literals, | ||
| /// enabling runtime weight swapping without re-parsing | ||
| bool external_weights_as_parameters = false; |
Collaborator
There was a problem hiding this comment.
This should be named keep_weights_external or defer_external_weights.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Motivation
The goal is to parameterize weights in onnx so that a user can quickly swap out weights without recompiling the program from scratch.
Technical Details
Included doc of rough outline of proposed changes.
Changelog Category
Add a
CHANGELOG.mdentry for any option other thanNot Applicable