Arm backend: Add DeepSeek-R1-Distill-Qwen layer tests#21010
Merged
bdemirb merged 6 commits intoJul 20, 2026
Merged
Conversation
This patch adds Arm backend layer tests for DeepSeek-R1-Distill-Qwen-1.5B. The tests use the checkpoint configuration from the Hugging Face model and the upstream Qwen2 layer implementations that back this distilled model. The test structure follows the existing Qwen3-VL layer-test approach. The config helper constructs the checkpoint-sized model configuration directly, the test file uses layer-level wrappers with dataclass-driven test cases, and coverage includes TOSA FP, TOSA BF16 reference-model, VGF no-quant, and VGF BF16 no-quant runtime paths. Token embedding is excluded because the full checkpoint embedding allocation is too large for regular CI. VGF quant coverage is also left out of this patch so CI resources stay focused on the requested BF16 path and export coverage. The covered layers include rotary embedding, rotary application, KV repetition, attention, RMSNorm, MLP, decoder layer, and final norm. Signed-off-by: Baris Demir <baris.demir@arm.com> Change-Id: Ia28581bbb4ffe070bc35af060fcceef2ac90084a
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/executorch/21010
Note: Links to docs will display an error until the docs builds have been completed. ❌ 2 New Failures, 1 Unrelated FailureAs of commit 4352396 with merge base 2b8cc80 ( NEW FAILURES - The following jobs have failed:
FLAKY - The following job failed but was likely due to flakiness present on trunk:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
meta-cla
Bot
added
the
CLA Signed
Collaborator
Author
|
@pytorchbot label "partner: arm" |
pytorch-bot
Bot
added
the
partner: arm
Collaborator
Author
|
@pytorchbot label "release notes: arm" |
Collaborator
Author
|
@pytorchbot label "partner: arm" |
Collaborator
Author
|
@pytorchbot label "release notes: arm" |
robell
approved these changes
Jul 20, 2026
Move DeepSeek checkpoint metadata fields that are not part of the typed Qwen2Config constructor into a post-construction config.update call. This preserves the checkpoint values while avoiding mypy call-arg errors for torch_dtype and use_mrope. Signed-off-by: Baris Demir <baris.demir@arm.com>
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
2 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.
This patch adds Arm backend layer tests for
DeepSeek-R1-Distill-Qwen-1.5B. The tests use the checkpoint configuration from the Hugging Face model and the upstream Qwen2 layer implementations that back this distilled model.
The test structure follows the existing Qwen3-VL layer-test approach. The config helper constructs the checkpoint-sized model configuration directly, the test file uses layer-level wrappers with dataclass-driven test cases, and coverage includes TOSA FP, TOSA BF16 reference-model, VGF no-quant, and VGF BF16 no-quant runtime paths.
Token embedding is excluded because the full checkpoint embedding allocation is too large for regular CI. VGF quant coverage is also left out of this patch so CI resources stay focused on the requested BF16 path and export coverage.
The covered layers include rotary embedding, rotary application, KV repetition, attention, RMSNorm, MLP, decoder layer, and final norm.
cc @digantdesai @freddan80 @per @zingo @oscarandersson8218 @mansnils @Sebastian-Larsson @robell @rascani