Set torch upper bound to <2.1.0 (#363)

* Set torch upper bound to <2.1.0

Some changes in PyTorch 2.1.0 and later are incompatible with
Curated Transformers 1.x. Fixing these issues would require
API changes. So we set the upper bound on supported PyTorch
versions. We will soon release Curated Transformers 2.0.0,
which is compatible with the lastest PyTorch versions.

* black

curated_transformers/models/falcon/layer.py

@@ -70,9 +70,11 @@ def __init__(
                 n_key_value_heads=attention_config.n_key_value_heads,
             ),
             rotary_embeds=rotary_embeds,
-            qkv_mode=QkvMode.MERGED_SPLIT_AFTER
-            if attention_config.n_key_value_heads == 1
-            else QkvMode.MERGED_SPLIT_BEFORE,
+            qkv_mode=(
+                QkvMode.MERGED_SPLIT_AFTER
+                if attention_config.n_key_value_heads == 1
+                else QkvMode.MERGED_SPLIT_BEFORE
+            ),
             use_bias=attention_config.use_bias,
             device=device,
         )

curated_transformers/tests/models/util.py

@@ -58,9 +58,7 @@ class JITMethod(Enum):
     TorchCompile = 1
     TorchScriptTrace = 2
 
-    def convert(
-        self, model: Module, with_torch_sdp: bool, *args
-    ) -> Tuple[
+    def convert(self, model: Module, with_torch_sdp: bool, *args) -> Tuple[
         Union[Module, torch.ScriptModule],
         Callable[[Union[ModelOutput, Dict[str, torch.Tensor]]], Tensor],
     ]:

curated_transformers/tokenizers/legacy/legacy_tokenizer.py

@@ -155,12 +155,10 @@ def _convert_strings(
     @abstractmethod
     def _decode(
         self, input: Iterable[Iterable[int]], skip_special_pieces: bool
-    ) -> List[str]:
-        ...
+    ) -> List[str]: ...
 
     @abstractmethod
-    def _encode(self, input: Iterable[MergedInputChunks]) -> PiecesWithIds:
-        ...
+    def _encode(self, input: Iterable[MergedInputChunks]) -> PiecesWithIds: ...
 
 
 class AddBosEosPreEncoder(PreEncoder):

requirements.txt

@@ -1,7 +1,7 @@
 curated-tokenizers>=0.9.1,<1.0.0
 huggingface-hub>=0.14
 tokenizers>=0.13.3
-torch>=1.12.0
+torch>=1.12.0,<2.1.0
 
 # Development dependencies
 mypy>=0.990,<1.1.0; platform_machine != "aarch64"

setup.cfg

@@ -17,7 +17,7 @@ install_requires =
     curated-tokenizers>=0.9.1,<1.0.0
     huggingface-hub>=0.14
     tokenizers>=0.13.3
-    torch>=1.12.0
+    torch>=1.12.0,<2.1.0
 
 [options.extras_require]
 quantization =