Fix test-samsung-models-linux: accelerate dep and the datasets/torchvision collate

.github/workflows/trunk.yml

@@ -1133,5 +1133,7 @@ jobs:
         # Test models
         # TODO: Switch to run_tests.py once CI device provisioning is stable
         # python -m executorch.backends.samsung.test.utils.run_tests --chipset E9955
-        pip install evaluate
+        # accelerate is for test_mobilebert_qat, whose TrainingArguments needs it;
+        # transformers reports the same error when it is absent and when it is too old.
+        pip install evaluate 'accelerate>=1.1.0'
         python -m unittest discover -s backends/samsung/test/models -p "test_*.py"

examples/samsung/scripts/mobilebert_finetune_QAT.py

@@ -9,14 +9,15 @@
 from pathlib import Path
 from typing import Optional
 
+import datasets.config
+
 import evaluate
 import numpy as np
 import requests
 
 import torch
 import torch.nn as nn
 import torchao
-
 from datasets import ClassLabel, DatasetDict, load_dataset
 
 from executorch.backends.samsung.quantizer import EnnQuantizer, Precision
@@ -140,6 +141,14 @@ def preprocess_function(examples):
 
         print("Preprocessing data...")
         tokenized_datasets = raw_datasets.map(preprocess_function, batched=True)
+        # datasets' torch formatter does `from torchvision.io import VideoReader`
+        # whenever torchvision merely imports, but OSS torchvision ships that symbol
+        # only in Meta-internal builds -- its io/__init__ swallows the import with
+        # `except ImportError: pass`. timm puts torchvision in sys.modules, so
+        # collating the tensors below raises ImportError. This dataset is text only
+        # and never decodes an image or a video, so the torchvision path is dead
+        # weight here. Drop the two lines once datasets guards that import.
+        datasets.config.TORCHVISION_AVAILABLE = False
         tokenized_datasets.set_format(
             type="torch", columns=["input_ids", "attention_mask", "label"]
         )