[Inductor][MPS] Fix half-precision type mismatches in Metal shader codegen (#176436)#177193
Merged
malfet merged 1 commit intorelease/2.11from Mar 11, 2026
Merged
[Inductor][MPS] Fix half-precision type mismatches in Metal shader codegen (#176436)#177193malfet merged 1 commit intorelease/2.11from
malfet merged 1 commit intorelease/2.11from
Conversation
…degen (#176436) Metal Shading Language rejects implicit float-to-bfloat conversions, so bare float literals like `0.0` in generated shaders cause compilation failures when the target variable is `bfloat` (or `half`). Three codegen methods were affected: - `constant()` ignored its `dtype` parameter and returned raw literals. - `masked()` assigned a bare literal in the else-branch (`} else tmp = 0.0;`). - `where()` passed a bare literal through the ternary without casting. All three now emit `static_cast<bfloat>(...)` / `static_cast<half>(...)` where needed. Tests added for half-precision constants, reductions, and conditionals. Pull Request resolved: #176436 Approved by: https://github.com/malfet Test plan: Run `python -c "import torch;F=torch.nn.functional;print(torch.compile(lambda x: F.pad(F.gelu(x), [1, 0]))(torch.randn(4, device='mps', dtype=torch.bfloat16)))"` (cherry picked from commit 3b161e7)
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/177193
Note: Links to docs will display an error until the docs builds have been completed. ❌ 2 New Failures, 1 Cancelled Job, 67 PendingAs of commit 57929f7 with merge base 0fd766e ( NEW FAILURES - The following jobs have failed:
UNSTABLE - The following jobs are marked as unstable, possibly due to flakiness on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
This PR needs a
|
Contributor
Verify the bug is fixed in release 2.11.0.Repro Summary: Metal Shading Language bfloat/half implicit float conversionTest: Ran the MRE from the bug report using Result: Fixed
if not resolved: Local ReproduceSetupuv venv mps-repro
source mps-repro/bin/activate
uv pip install torch==2.11.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/test/cpu
python -c "import torch; print(torch.__version__)"
# output: 2.11.0
python -c "import torch;F=torch.nn.functional;print(torch.compile(lambda x: F.pad(F.gelu(x), [1, 0]))(torch.randn(4, device='mps', dtype=torch.bfloat16)))"OutputVersions |
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
4 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.
Metal Shading Language rejects implicit float-to-bfloat conversions, so
bare float literals like
0.0in generated shaders cause compilationfailures when the target variable is
bfloat(orhalf). Three codegenmethods were affected:
constant()ignored itsdtypeparameter and returned raw literals.masked()assigned a bare literal in the else-branch (} else tmp = 0.0;).where()passed a bare literal through the ternary without casting.All three now emit
static_cast<bfloat>(...)/static_cast<half>(...)where needed. Tests added for half-precision constants, reductions, and
conditionals.
Pull Request resolved: #176436
Approved by: https://github.com/malfet
Test plan: Run
python -c "import torch;F=torch.nn.functional;print(torch.compile(lambda x: F.pad(F.gelu(x), [1, 0]))(torch.randn(4, device='mps', dtype=torch.bfloat16)))"(cherry picked from commit 3b161e7)