Fixes for Adreno inference (Q8/Q4) and OUT_PROD #34
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
github-actions
bot
added
Nvidia GPU
Vulkan
examples
ggml
testing
build
devops
python
script
android
labels
infinitalo
force-pushed
the
italo/adreno_inference
branch
11 times, most recently
from
082fa25 to
0f376cd
Compare
olyasir
requested changes
Oct 16, 2025
infinitalo
force-pushed
the
italo/adreno_inference
branch
from
0f376cd to
0f6d6da
Compare
infinitalo
changed the base branch from
temp-latest
to
temp-latest-finetuning
Author
|
@olyasir I've addressed your comments and pushed some cleanups, please feel free to give it another look. I've also edited the PR to merge into |
infinitalo
force-pushed
the
italo/adreno_inference
branch
2 times, most recently
from
e40fed4 to
3f57c82
Compare
olyasir
approved these changes
Oct 22, 2025
olyasir
approved these changes
Oct 24, 2025
|
@infinitalo can you please also check the failed pipelines? |
added 11 commits
October 24, 2025 18:12
Shouldn't change any behavior since currently nb00 is always 1. Robustness is usually disabled for Q8/Q4 shaders since having it enabled impacts performance more significantly for those types than F16/F32.
Introduce a CMAKE option for disabling Adreno-specific shaders if needed, this improves build time, but should not be used when targeting Adreno devices.
Extend device detection to classify Qualcomm Adreno GPUs, enabling targeted workarounds and shader selection when those devices are present.
Avoid subgroup operations on Adreno by selecting safer paths to sidestep compiler/driver bugs while preserving behavior.
Similar to what we do for other vendors such as Intel.
This optimization broke inference on Adreno.
Add build-time generation of Adreno-targeted shader variants under a guard, so Adreno devices use safer code paths while other GPUs remain unaffected.
Increase OUT_PROD Q8 performance through improving memory locality.
This makes finetuning work without crashing on Adreno 830.
infinitalo
force-pushed
the
italo/adreno_inference
branch
from
3f57c82 to
2c9a0e8
Compare
gianni-cor
approved these changes
Oct 27, 2025
Signed-off-by: Marcus Edel <[email protected]>
Signed-off-by: Marcus Edel <[email protected]>
olyasir
merged commit f9e7293
into
tetherto:temp-latest-finetuning
39 of 47 checks passed
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.
Make sure to read the contributing guidelines before submitting a PR