Bring back TIMM model inductor CI test (pytorch#87730)

Summary: pytorch#87588 has solved the
inductor compilation speed regression, so we can try to run TIMM models
with fewer shards and also enable pretained model downloading which
should resolve the flakyness we have seen previously.

cc @jansel @mlazos @soumith @voznesenskym @yanboliang @penguinwu @anijain2305
Pull Request resolved: pytorch#87730
Approved by: https://github.com/anijain2305

.github/workflows/inductor.yml

@@ -22,8 +22,13 @@ jobs:
       cuda-arch-list: 8.6
       test-matrix: |
         { include: [
-          { config: "inductor", shard: 1, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
-          { config: "inductor", shard: 2, num_shards: 2, runner: "linux.g5.4xlarge.nvidia.gpu" },
+          { config: "inductor", shard: 1, num_shards: 7, runner: "linux.g5.4xlarge.nvidia.gpu" },
+          { config: "inductor", shard: 2, num_shards: 7, runner: "linux.g5.4xlarge.nvidia.gpu" },
+          { config: "inductor", shard: 3, num_shards: 7, runner: "linux.g5.4xlarge.nvidia.gpu" },
+          { config: "inductor", shard: 4, num_shards: 7, runner: "linux.g5.4xlarge.nvidia.gpu" },
+          { config: "inductor", shard: 5, num_shards: 7, runner: "linux.g5.4xlarge.nvidia.gpu" },
+          { config: "inductor", shard: 6, num_shards: 7, runner: "linux.g5.4xlarge.nvidia.gpu" },
+          { config: "inductor", shard: 7, num_shards: 7, runner: "linux.g5.4xlarge.nvidia.gpu" },
         ]}
 
   linux-bionic-cuda11_6-py3_10-gcc7-inductor-test:

.jenkins/pytorch/test.sh

@@ -281,7 +281,7 @@ test_inductor_timm_shard() {
   TEST_REPORTS_DIR=/tmp/test-reports
   mkdir -p "$TEST_REPORTS_DIR"
   python benchmarks/dynamo/timm_models.py --ci --training --accuracy \
-    --device cuda --inductor --float32 --total-partitions 8 --partition-id "$1" \
+    --device cuda --inductor --float32 --total-partitions 5 --partition-id "$1" \
     --output "$TEST_REPORTS_DIR"/inductor_timm_"$1".csv
   python benchmarks/dynamo/check_csv.py -f "$TEST_REPORTS_DIR"/inductor_timm_"$1".csv
 }
@@ -749,6 +749,13 @@ elif [[ "${TEST_CONFIG}" == *inductor* && "${SHARD_NUMBER}" == 2 && $NUM_TEST_SH
   install_triton
   install_huggingface
   test_inductor_huggingface_shard 0
+elif [[ "${TEST_CONFIG}" == *inductor* && $SHARD_NUMBER -lt 8 && $NUM_TEST_SHARDS -gt 1 ]]; then
+  install_torchvision
+  install_filelock
+  install_triton
+  install_timm
+  id=$((SHARD_NUMBER-3))
+  test_inductor_timm_shard $id
 elif [[ "${SHARD_NUMBER}" == 1 && $NUM_TEST_SHARDS -gt 1 ]]; then
   test_without_numpy
   install_torchvision

benchmarks/dynamo/timm_models.py

@@ -205,8 +205,7 @@ def load_model(
                     drop_rate=0.0,
                     drop_path_rate=None,
                     drop_block_rate=None,
-                    # Skip downloading pretrained models for speedy CI
-                    pretrained=not self.args.ci,
+                    pretrained=True,
                     # global_pool=kwargs.pop('gp', 'fast'),
                     # num_classes=kwargs.pop('num_classes', None),
                     # drop_rate=kwargs.pop('drop', 0.),