[ET-VK] Enable automatic dtype conversion when copying to/from staging buffer #14222
Conversation
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
…g buffer ## Context During export, Vulkan sometimes converts certain tensor dtypes. The most common case of this is that int64 and float64 are internally represented as int32 and float32 tensors. The primary reason for this is to reduce the number of dtype variants that need to be generated for each shader, and also due to the fact that 64-bit types are not guaranteed to be supported. However, this raises an issue if an int64 or float64 tensor is marked as an input/output tensor of the model. The source/destination ETensor will have a different dtype than the internal representation, meaning that the input/output bytes will be interpreted incorrectly. ## Changes This diff fixes this behaviour by introducing the concept of a "staging dtype". This allows the staging buffer of a tensor to have a different dtype than the underlying GPU buffer or texture. When copying to/from the GPU resource, the dtype can then be converted to the correct dtype expected by the client code. As a bonus, also add an optional setting to force fp16 to be used internally for fp32 tensors. This allows models to access half precision inference without needing to incur the cost of dtype conversion ops being inserted into the graph, or needing to manually convert inputs/outputs to half type. Differential Revision: [D82234180](https://our.internmc.facebook.com/intern/diff/D82234180/) [ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/14222
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New Failure, 3 Pending, 6 Unrelated FailuresAs of commit fe8a683 with merge base b2ae2b4 ( BROKEN TRUNK - The following jobs failed but were present on the merge base:👉 Rebase onto the `viable/strict` branch to avoid these failures
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
bot
added
the
CLA Signed
|
This pull request was exported from Phabricator. Differential Revision: D82234180 |
This PR needs a
|
…from staging buffer" ## Context During export, Vulkan sometimes converts certain tensor dtypes. The most common case of this is that int64 and float64 are internally represented as int32 and float32 tensors. The primary reason for this is to reduce the number of dtype variants that need to be generated for each shader, and also due to the fact that 64-bit types are not guaranteed to be supported. However, this raises an issue if an int64 or float64 tensor is marked as an input/output tensor of the model. The source/destination ETensor will have a different dtype than the internal representation, meaning that the input/output bytes will be interpreted incorrectly. ## Changes This diff fixes this behaviour by introducing the concept of a "staging dtype". This allows the staging buffer of a tensor to have a different dtype than the underlying GPU buffer or texture. When copying to/from the GPU resource, the dtype can then be converted to the correct dtype expected by the client code. As a bonus, also add an optional setting to force fp16 to be used internally for fp32 tensors. This allows models to access half precision inference without needing to incur the cost of dtype conversion ops being inserted into the graph, or needing to manually convert inputs/outputs to half type. Differential Revision: [D82234180](https://our.internmc.facebook.com/intern/diff/D82234180/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D82234180 |
manuelcandales
approved these changes
Sep 11, 2025
…from staging buffer" ## Context During export, Vulkan sometimes converts certain tensor dtypes. The most common case of this is that int64 and float64 are internally represented as int32 and float32 tensors. The primary reason for this is to reduce the number of dtype variants that need to be generated for each shader, and also due to the fact that 64-bit types are not guaranteed to be supported. However, this raises an issue if an int64 or float64 tensor is marked as an input/output tensor of the model. The source/destination ETensor will have a different dtype than the internal representation, meaning that the input/output bytes will be interpreted incorrectly. ## Changes This diff fixes this behaviour by introducing the concept of a "staging dtype". This allows the staging buffer of a tensor to have a different dtype than the underlying GPU buffer or texture. When copying to/from the GPU resource, the dtype can then be converted to the correct dtype expected by the client code. As a bonus, also add an optional setting to force fp16 to be used internally for fp32 tensors. This allows models to access half precision inference without needing to incur the cost of dtype conversion ops being inserted into the graph, or needing to manually convert inputs/outputs to half type. Differential Revision: [D82234180](https://our.internmc.facebook.com/intern/diff/D82234180/) [ghstack-poisoned]
|
This pull request was exported from Phabricator. Differential Revision: D82234180 |
facebook-github-bot
merged commit 2ca7304
into
gh/SS-JIA/327/base
115 of 125 checks passed
facebook-github-bot
temporarily deployed
to
cherry-pick-bot
GitHub Actions
Inactive
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 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.
Stack from ghstack (oldest at bottom):
Context
During export, Vulkan sometimes converts certain tensor dtypes. The most common case of this is that int64 and float64 are internally represented as int32 and float32 tensors. The primary reason for this is to reduce the number of dtype variants that need to be generated for each shader, and also due to the fact that 64-bit types are not guaranteed to be supported.
However, this raises an issue if an int64 or float64 tensor is marked as an input/output tensor of the model. The source/destination ETensor will have a different dtype than the internal representation, meaning that the input/output bytes will be interpreted incorrectly.
Changes
This diff fixes this behaviour by introducing the concept of a "staging dtype". This allows the staging buffer of a tensor to have a different dtype than the underlying GPU buffer or texture. When copying to/from the GPU resource, the dtype can then be converted to the correct dtype expected by the client code.
As a bonus, also add an optional setting to force fp16 to be used internally for fp32 tensors. This allows models to access half precision inference without needing to incur the cost of dtype conversion ops being inserted into the graph, or needing to manually convert inputs/outputs to half type.
Differential Revision: D82234180