…ding (#585) ### Summary Replace the custom ucbdrive-based DLA PyTorch implementation and its pretrained weight loading with [timm (PyTorch Image Models)](https://github.com/huggingface/pytorch-image-models/blob/main/timm/models/dla.py), which mirrors the same architecture and ImageNet checkpoints via HuggingFace Hub. This resolves a recurring download failure that has blocked DLA model testing in tt-forge-onnx CI pipelines - https://github.com/tenstorrent/tt-forge-onnx/actions/runs/24430244784/job/71374191041. ### Problem The DLA model loader relied on a custom implementation adapted from [ucbdrive/dla](https://github.com/ucbdrive/dla), which fetched pretrained ImageNet weights from `http://dl.yf.io/dla/models/imagenet/` at runtime: ``` Downloading: "http://dl.yf.io/dla/models/imagenet/dla60x-d15cacda.pth" [Errno 113] No route to host ``` This failure has been reported repeatedly by the community across [multiple open issues](https://github.com/ucbdrive/dla/issues) in the upstream repository (issues [#26](ucbdrive/dla#26), [#27](ucbdrive/dla#27), [#28](ucbdrive/dla#28), [#29](ucbdrive/dla#29)) We have encountered the same issue ourselves in [tenstorrent/tt-forge-onnx#2077](tenstorrent/tt-forge-onnx#2077), where the workaround at the time was to set `pretrained=None` — skipping weight loading entirely and testing only the model architecture. While that unblocked CI, it leaves the DLA models running without pretrained weights, which is not representative of real-world usage. Rather than continuing to work around a broken upstream dependency, it is cleaner and more reliable to migrate to timm, which provides the same models with working, maintained weight hosting. ### Solution Migrate all DLA model variants to load via `timm.create_model()`. This is the correct fix because timm explicitly states its DLA implementation was *"adapted from Official PyTorch impl at: https://github.com/ucbdrive/dla"* and is computationally equivalent. timm re-hosts all 10 DLA ImageNet checkpoints on HuggingFace Hub (`timm/dla34.in1k`, `timm/dla60x.in1k`, etc.) at a stable, well-maintained URL, making pretrained weight loading reliable. --- ### Changes **`third_party/tt_forge_models/dla/pytorch/loader.py`** - All 10 DLA variants now use `ModelSource.TIMM` with the corresponding `<name>.in1k` timm model identifier. - `load_model` simplified to a single `timm.create_model(model_name, pretrained=True)` call — source branching removed. - `load_inputs` simplified to always derive preprocessing from `resolve_data_config` + `create_transform` on the loaded model — hardcoded torchvision transform branch removed. - Removed imports: `torchvision.transforms`, `PIL.Image`, `from .src import dla_model`. - Removed redundant `DLA34_IN1K` variant (was identical to `DLA34`). **`third_party/tt_forge_models/dla/pytorch/src/dla_model.py`** (+ `src/` directory) - Deleted entirely — the custom ucbdrive DLA implementation is no longer needed.

link down #29

Description

Metadata

Metadata

Assignees

Labels

Type

Fields

Projects

Milestone

Relationships

Development

Issue actions