Dynamic concat gpu support#5032
Open
turneram wants to merge 18 commits into
Open
Conversation
Format Co-authored-by: github-actions[bot] <41898282+github-actions[bot]@users.noreply.github.com>
…ynamic-concat-pointwise
Regressions detected 🔴 |
|
turneram
marked this pull request as ready for review
July 21, 2026 18:10
pfultz2
reviewed
Jul 21, 2026
| const char* ptr = arg.data(); | ||
| if(ptr != nullptr and not is_device_ptr(ptr)) | ||
| { | ||
| temps.push_back(to_gpu(arg)); |
Collaborator
There was a problem hiding this comment.
The copies from host to device should happen in the IR not in the code object op. This can break hipgraph.
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
Comment on lines
+133
to
+135
| std::transform(input_shapes.begin(), input_shapes.end(), std::back_inserter(params), [&](const auto& s) { | ||
| return m.add_parameter("x" + std::to_string(params.size()), s); | ||
| }); |
Contributor
There was a problem hiding this comment.
[format.py] reported by reviewdog 🐶
Suggested change
| std::transform(input_shapes.begin(), input_shapes.end(), std::back_inserter(params), [&](const auto& s) { | |
| return m.add_parameter("x" + std::to_string(params.size()), s); | |
| }); | |
| std::transform( | |
| input_shapes.begin(), input_shapes.end(), std::back_inserter(params), [&](const auto& s) { | |
| return m.add_parameter("x" + std::to_string(params.size()), s); | |
| }); |
Contributor
There was a problem hiding this comment.
Pull request overview
This PR extends MIGraphX’s GPU pipeline to support concat with dynamic-shape inputs (including a KV-cache-style “past + present” concat), primarily by ensuring host-resident inputs are moved to GPU memory before JIT compilation/execution and by updating the concat JIT compilation metadata to handle dynamic/output-alias cases.
Changes:
- Insert
hip::copy_to_gpu(with per-input allocations) for host inputs feeding dynamic concat precompile/dynamic-code-object ops duringgpu::lower_device_ops. - Update GPU concat JIT compilation to carry
num_concat_inputsandoutput_shape, and to avoid vectorization in output-alias and dynamic cases. - Add/adjust tests covering dynamic concat GPU lowering behavior and update ONNX clip dynamic tests for the new multibroadcast/common-args behavior.
Reviewed changes
Copilot reviewed 10 out of 10 changed files in this pull request and generated 1 comment.
Show a summary per file
| File | Description |
|---|---|
| test/verify/test_dynamic_concat.cpp | Adds verify-program coverage for dynamic concat patterns (axis0, KV-cache axis2, symbolic). |
| test/onnx/parse/clip_dyn_min_only_test.cpp | Updates expected IR construction for dynamic clip min-only case using new multibroadcast wiring. |
| test/onnx/parse/clip_dyn_min_max_test.cpp | Updates expected IR construction for dynamic clip min/max case; adds missing include for to_value. |
| test/gpu/lower_device_ops.cpp | Adds unit tests asserting gpu::lower_device_ops inserts hip::copy_to_gpu only for dynamic concat cases. |
| src/targets/gpu/lower_device_ops.cpp | Adds dynamic-concat input preparation (inserting hip::copy_to_gpu) and preserves module_inputs() on replace. |
| src/targets/gpu/jit/concat.cpp | Extends concat JIT config to handle dynamic shapes/output aliasing; adjusts vectorization selection accordingly. |
| src/targets/gpu/include/migraphx/gpu/lower_device_ops.hpp | Documents the expanded responsibilities of gpu::lower_device_ops. |
| src/targets/gpu/include/migraphx/gpu/hip.hpp | Exposes is_device_ptr and simplifies hip_copy_to_gpu to use copy_to_gpu. |
| src/targets/gpu/hip.cpp | Makes is_device_ptr non-static so it matches the new header declaration. |
| src/common.cpp | Refactors dynamic multibroadcast insertion in insert_common_args to use an existing dynamic reference operand where possible. |
Comments suppressed due to low confidence (2)
test/verify/test_dynamic_concat.cpp:74
- The program does not add a return instruction, so the module has no outputs. Add an explicit mm->add_return({result}) after creating the concat instruction.
auto past_key = mm->add_parameter("past_key_values.0.key", past_shape);
auto current_key = mm->add_literal(migraphx::generate_literal(current_shape));
mm->add_instruction(migraphx::make_op("concat", {{"axis", 2}}), past_key, current_key);
return p;
test/verify/test_dynamic_concat.cpp:104
- The program does not add a return instruction, so the module has no outputs. Add an explicit return for the concat result.
auto x = mm->add_parameter("X", s0);
auto y = mm->add_parameter("Y", s1);
auto z = mm->add_parameter("Z", s2);
mm->add_instruction(migraphx::make_op("concat", {{"axis", 0}}), x, y, z);
return p;
Comment on lines
+41
to
+45
| auto x = mm->add_parameter("X", s0); | ||
| auto y = mm->add_parameter("Y", s1); | ||
| auto z = mm->add_parameter("Z", s2); | ||
| mm->add_instruction(migraphx::make_op("concat", {{"axis", 0}}), x, y, z); | ||
| return p; |
Collaborator
|
This seems confusing. It says its a dynamic concat input but it generates static JIT kernel. Is this trying to implement a dynamic concat_past_present? It seems like insert slice op is a better choice. |
Contributor
Author
Isn't that the current behavior for dynamic_code_object_op? Compiles static kernels for each value in {min,max}?
The pattern that's needed for at least the current model is this. Old: Dynamic past-seq-len: |
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
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
Dynamic concat is required to run dynamic kv-cache
Technical Details
Adds changes needed to run concat with dynamic shape inputs on gpu.
Changelog Category
Add a
CHANGELOG.mdentry for any option other thanNot Applicable