Update torch minimum version to 2.3

Signed-off-by: Yuanyuan Chen <[email protected]>

Author: cyyever

src/transformers/integrations/executorch.py

@@ -32,7 +32,6 @@
 from ..modeling_utils import ALL_ATTENTION_FUNCTIONS, PreTrainedModel
 from ..pytorch_utils import (
     is_torch_greater_or_equal,
-    is_torch_greater_or_equal_than_2_3,
     is_torch_greater_or_equal_than_2_6,
 )
 
@@ -764,8 +763,6 @@ def convert_and_export_with_cache(
     Returns:
         Exported program (`torch.export.ExportedProgram`): The exported program generated via `torch.export`.
     """
-    if not is_torch_greater_or_equal_than_2_3:
-        raise ImportError("torch >= 2.3 is required.")
 
     import torch.export._trace
 
@@ -1034,8 +1031,6 @@ def export_with_dynamic_cache(
     Returns:
         Exported program (`torch.export.ExportedProgram`): The exported program generated via `torch.export`.
     """
-    if not is_torch_greater_or_equal_than_2_3:
-        raise ImportError("torch >= 2.3 is required.")
 
     # This is the same as sdpa, but mask creation does not use `vmap` which is not exportable
     ALL_MASK_ATTENTION_FUNCTIONS.register("sdpa_without_vmap", sdpa_mask_without_vmap)

src/transformers/modeling_utils.py

@@ -450,23 +450,20 @@ def load_sharded_checkpoint(model, folder, strict=True, prefer_safe=True):
     "U8": torch.uint8,
     "I8": torch.int8,
     "I16": torch.int16,
+    "U16": torch.uint16,
     "F16": torch.float16,
     "BF16": torch.bfloat16,
     "I32": torch.int32,
+    "U32": torch.uint32,
     "F32": torch.float32,
     "F64": torch.float64,
     "I64": torch.int64,
+    "U64": torch.uint64,
     "F8_E4M3": torch.float8_e4m3fn,
     "F8_E5M2": torch.float8_e5m2,
 }
 
 
-if is_torch_greater_or_equal("2.3.0"):
-    str_to_torch_dtype["U16"] = torch.uint16
-    str_to_torch_dtype["U32"] = torch.uint32
-    str_to_torch_dtype["U64"] = torch.uint64
-
-
 def load_state_dict(
     checkpoint_file: Union[str, os.PathLike],
     is_quantized: bool = False,

src/transformers/pytorch_utils.py

@@ -37,9 +37,9 @@
 is_torch_greater_or_equal_than_2_8 = is_torch_greater_or_equal("2.8", accept_dev=True)
 is_torch_greater_or_equal_than_2_6 = is_torch_greater_or_equal("2.6", accept_dev=True)
 is_torch_greater_or_equal_than_2_4 = is_torch_greater_or_equal("2.4", accept_dev=True)
-is_torch_greater_or_equal_than_2_3 = is_torch_greater_or_equal("2.3", accept_dev=True)
 
 # For backwards compatibility (e.g. some remote codes on Hub using those variables).
+is_torch_greater_or_equal_than_2_3 = is_torch_greater_or_equal("2.3", accept_dev=True)
 is_torch_greater_or_equal_than_2_2 = is_torch_greater_or_equal("2.2", accept_dev=True)
 is_torch_greater_or_equal_than_2_1 = is_torch_greater_or_equal("2.1", accept_dev=True)
 is_torch_greater_or_equal_than_2_0 = is_torch_greater_or_equal("2.0", accept_dev=True)

src/transformers/utils/import_utils.py

@@ -236,9 +236,9 @@ def _is_package_available(pkg_name: str, return_version: bool = False) -> Union[
 
 _torch_available, _torch_version = _is_package_available("torch", return_version=True)
 if _torch_available:
-    _torch_available = version.parse(_torch_version) >= version.parse("2.2.0")
+    _torch_available = version.parse(_torch_version) >= version.parse("2.3.0")
     if not _torch_available:
-        logger.warning(f"Disabling PyTorch because PyTorch >= 2.2 is required but found {_torch_version}")
+        logger.warning(f"Disabling PyTorch because PyTorch >= 2.3 is required but found {_torch_version}")
 
 
 _essentia_available = importlib.util.find_spec("essentia") is not None

tests/test_executorch.py

@@ -23,16 +23,12 @@
     TorchExportableModuleWithHybridCache,
     TorchExportableModuleWithStaticCache,
 )
-from transformers.pytorch_utils import is_torch_greater_or_equal_than_2_3
 from transformers.testing_utils import require_torch
 
 
 @require_torch
 class ExecutorchTest(unittest.TestCase):
     def setUp(self):
-        if not is_torch_greater_or_equal_than_2_3:
-            self.skipTest("torch >= 2.3 is required")
-
         set_seed(0)
         self.model = AutoModelForCausalLM.from_pretrained("hf-internal-testing/tiny-random-LlamaForCausalLM")
         self.model.eval()