Repository navigation
Vulkan: convert quantized linear weight scales to the output dtype - #24023
Merged
Merged
Conversation
The quantized linear shaders read the weight scales in the output dtype, but quantized_linear_impl prepacked them in whatever dtype they were serialized in. An fp16 model quantized from a bf16 or fp32 checkpoint can keep fp32 scales (e.g. torchao's weight_scale_dtype=torch.float32), which the partitioner accepts, and the shaders then read the fp32 bytes as fp16. Prepack the scales in the output dtype, as Q4gswLinear already does.
pytorch-bot
Bot
added
the
module: vulkan
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/24023
Note: Links to docs will display an error until the docs builds have been completed. ❌ 1 New Failure, 2 Unrelated FailuresAs of commit 7947020 with merge base a39d597 ( NEW FAILURE - The following job has failed:
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
executorch-triage
Bot
added
the
community: contribution
nil-is-all
added
the
release notes: vulkan
SS-JIA
approved these changes
Oct 26, 2026
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Summary
The quantized linear shaders read the weight scales in the output dtype, but
quantized_linear_implprepacked them in their serialized dtype. An fp16 model with fp32 scales (e.g. torchaoweight_scale_dtype=torch.float32, or a bf16 checkpoint cast to fp16 around its scales) is accepted by the partitioner and silently produces garbage. This prepacks the scales in the output dtype, asQ4gswLinearalready does.Test plan
Galaxy S26 Ultra (Adreno 840), fp16 8da4w
Linear(1536, 2048)g32, vs the same model in fp32:No unit test: the 8da4w delegate test is disabled on SwiftShader and the
q4gsw_linearop test skips fp16dq8careferences.cc @SS-JIA @manuelcandales @digantdesai @cbilgin @nil-is-all