Dynamic concat gpu support#5032
Draft
turneram wants to merge 5 commits into
Draft
Conversation
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
| } | ||
|
|
||
| static std::vector<argument> ensure_gpu_kernel_args(const std::vector<argument>& args, | ||
| pmr::vector<argument>& temps) |
Contributor
There was a problem hiding this comment.
[format.py] reported by reviewdog 🐶
Suggested change
| pmr::vector<argument>& temps) | |
| pmr::vector<argument>& temps) |
Comment on lines
+91
to
+92
| const std::size_t num_concat = | ||
| v.get("num_concat_inputs", inputs.size()); |
Contributor
There was a problem hiding this comment.
[format.py] reported by reviewdog 🐶
Suggested change
| const std::size_t num_concat = | |
| v.get("num_concat_inputs", inputs.size()); | |
| const std::size_t num_concat = v.get("num_concat_inputs", inputs.size()); |
Comment on lines
+94
to
+97
| concat_shapes.assign(inputs.begin(), | ||
| inputs.begin() + std::min(num_concat, inputs.size())); | ||
| shape output_shape = v.contains("output_shape") ? from_value<shape>(v.at("output_shape")) | ||
| : inputs.back(); |
Contributor
There was a problem hiding this comment.
[format.py] reported by reviewdog 🐶
Suggested change
| concat_shapes.assign(inputs.begin(), | |
| inputs.begin() + std::min(num_concat, inputs.size())); | |
| shape output_shape = v.contains("output_shape") ? from_value<shape>(v.at("output_shape")) | |
| : inputs.back(); | |
| concat_shapes.assign(inputs.begin(), inputs.begin() + std::min(num_concat, inputs.size())); | |
| shape output_shape = | |
| v.contains("output_shape") ? from_value<shape>(v.at("output_shape")) : inputs.back(); |
Comment on lines
+99
to
+100
| options.inputs = inputs; | ||
| options.output = output_shape; |
Contributor
There was a problem hiding this comment.
[format.py] reported by reviewdog 🐶
Suggested change
| options.inputs = inputs; | |
| options.output = output_shape; | |
| options.inputs = inputs; | |
| options.output = output_shape; |
| auto args = v.at("args"); | ||
|
|
||
| // normalize() rewrites axis into reduced-dim space; kernel concat<Axis> uses full tensors. | ||
| std::size_t fast_axis = kernel_axis; |
Contributor
There was a problem hiding this comment.
[format.py] reported by reviewdog 🐶
Suggested change
| std::size_t fast_axis = kernel_axis; | |
| std::size_t fast_axis = kernel_axis; |
|
|
||
| const std::size_t nelem = | ||
| output_shape.dynamic() ? output_shape.element_space() : output_shape.elements(); | ||
| auto nelements_per_op = nelem / op_names.size(); |
Contributor
There was a problem hiding this comment.
[format.py] reported by reviewdog 🐶
Suggested change
| auto nelements_per_op = nelem / op_names.size(); | |
| auto nelements_per_op = nelem / op_names.size(); |
| auto psl = var("psl", {1, 64}); | ||
| using dd = migraphx::shape::dynamic_dimension; | ||
|
|
||
| migraphx::shape past_shape{migraphx::shape::half_type, {dd{1, 1}, dd{5, 5}, dd{psl}, dd{64, 64}}}; |
Contributor
There was a problem hiding this comment.
[format.py] reported by reviewdog 🐶
Suggested change
| migraphx::shape past_shape{migraphx::shape::half_type, {dd{1, 1}, dd{5, 5}, dd{psl}, dd{64, 64}}}; | |
| migraphx::shape past_shape{migraphx::shape::half_type, | |
| {dd{1, 1}, dd{5, 5}, dd{psl}, dd{64, 64}}}; |
Comment on lines
+69
to
+70
| auto* mm = p.get_main_module(); | ||
| auto past_key = mm->add_parameter("past_key_values.0.key", past_shape); |
Contributor
There was a problem hiding this comment.
[format.py] reported by reviewdog 🐶
Suggested change
| auto* mm = p.get_main_module(); | |
| auto past_key = mm->add_parameter("past_key_values.0.key", past_shape); | |
| auto* mm = p.get_main_module(); | |
| auto past_key = mm->add_parameter("past_key_values.0.key", past_shape); |
Comment on lines
+78
to
+79
| return {{"past_key_values.0.key", | ||
| migraphx::shape{migraphx::shape::half_type, {1, 5, 1, 64}}}}; |
Contributor
There was a problem hiding this comment.
[format.py] reported by reviewdog 🐶
Suggested change
| return {{"past_key_values.0.key", | |
| migraphx::shape{migraphx::shape::half_type, {1, 5, 1, 64}}}}; | |
| return { | |
| {"past_key_values.0.key", migraphx::shape{migraphx::shape::half_type, {1, 5, 1, 64}}}}; |
Comment on lines
+88
to
+91
| auto n = var("n", {2, 3}); | ||
| auto d0 = var("d0", {2, 4}); | ||
| auto d1 = var("d1", {3, 4}); | ||
| auto d2 = var("d2", {1, 5}); |
Contributor
There was a problem hiding this comment.
[format.py] reported by reviewdog 🐶
Suggested change
| auto n = var("n", {2, 3}); | |
| auto d0 = var("d0", {2, 4}); | |
| auto d1 = var("d1", {3, 4}); | |
| auto d2 = var("d2", {1, 5}); | |
| auto n = var("n", {2, 3}); | |
| auto d0 = var("d0", {2, 4}); | |
| auto d1 = var("d1", {3, 4}); | |
| auto d2 = var("d2", {1, 5}); |
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
1 participant
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