Fix inclusion of non py files in package (#21546)

* Fix inclusion of non py files in package

* No need for the **

setup.py

@@ -413,7 +413,7 @@ install_requires = [
 
 setup(
     name="transformers",
-    version="4.26.0",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
+    version="4.26.1",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
     author="The Hugging Face team (past and future) with the help of all our contributors (https://github.com/huggingface/transformers/graphs/contributors)",
     author_email="transformers@huggingface.co",
     description="State-of-the-art Machine Learning for JAX, PyTorch and TensorFlow",
@@ -424,7 +424,8 @@ setup(
     url="https://github.com/huggingface/transformers",
     package_dir={"": "src"},
     packages=find_packages("src"),
-    package_data={"transformers": ["py.typed", "*.cu", "*.cpp", "*.cuh", "*.h"]},
+    include_package_data=True,
+    package_data={"transformers": ["*.cu", "*.cpp", "*.cuh", "*.h", "*.pyx"]},
     zip_safe=False,
     extras_require=extras,
     entry_points={"console_scripts": ["transformers-cli=transformers.commands.transformers_cli:main"]},

src/transformers/__init__.py

@@ -22,7 +22,7 @@
 # to defer the actual importing for when the objects are requested. This way `import transformers` provides the names
 # in the namespace without actually importing anything (and especially none of the backends).
 
-__version__ = "4.26.0"
+__version__ = "4.26.1"
 
 from typing import TYPE_CHECKING
 

src/transformers/models/longt5/modeling_longt5.py

@@ -276,6 +276,8 @@ class LongT5DenseActDense(nn.Module):
         hidden_states = self.wi(hidden_states)
         hidden_states = self.act(hidden_states)
         hidden_states = self.dropout(hidden_states)
+        if hidden_states.dtype != self.wo.weight.dtype and self.wo.weight.dtype != torch.int8:
+            hidden_states = hidden_states.to(self.wo.weight.dtype)
         hidden_states = self.wo(hidden_states)
         return hidden_states
 

src/transformers/models/mt5/modeling_mt5.py

@@ -146,6 +146,8 @@ class MT5DenseActDense(nn.Module):
         hidden_states = self.wi(hidden_states)
         hidden_states = self.act(hidden_states)
         hidden_states = self.dropout(hidden_states)
+        if hidden_states.dtype != self.wo.weight.dtype and self.wo.weight.dtype != torch.int8:
+            hidden_states = hidden_states.to(self.wo.weight.dtype)
         hidden_states = self.wo(hidden_states)
         return hidden_states
 
@@ -168,7 +170,8 @@ class MT5DenseGatedActDense(nn.Module):
 
         # To make 8bit quantization work for google/flan-t5-xxl, self.wo is kept in float32.
         # See https://github.com/huggingface/transformers/issues/20287
-        if hidden_states.dtype != self.wo.weight.dtype:
+        # we also make sure the weights are not in `int8` in case users will force `_keep_in_fp32_modules` to be `None``
+        if hidden_states.dtype != self.wo.weight.dtype and self.wo.weight.dtype != torch.int8:
             hidden_states = hidden_states.to(self.wo.weight.dtype)
 
         hidden_states = self.wo(hidden_states)

src/transformers/models/switch_transformers/modeling_switch_transformers.py

@@ -273,6 +273,8 @@ class SwitchTransformersDenseActDense(nn.Module):
         hidden_states = self.wi(hidden_states)
         hidden_states = self.act(hidden_states)
         hidden_states = self.dropout(hidden_states)
+        if hidden_states.dtype != self.wo.weight.dtype and self.wo.weight.dtype != torch.int8:
+            hidden_states = hidden_states.to(self.wo.weight.dtype)
         hidden_states = self.wo(hidden_states)
         return hidden_states
 

src/transformers/models/t5/modeling_t5.py

@@ -289,6 +289,8 @@ class T5DenseActDense(nn.Module):
         hidden_states = self.wi(hidden_states)
         hidden_states = self.act(hidden_states)
         hidden_states = self.dropout(hidden_states)
+        if hidden_states.dtype != self.wo.weight.dtype and self.wo.weight.dtype != torch.int8:
+            hidden_states = hidden_states.to(self.wo.weight.dtype)
         hidden_states = self.wo(hidden_states)
         return hidden_states
 
@@ -310,7 +312,8 @@ class T5DenseGatedActDense(nn.Module):
 
         # To make 8bit quantization work for google/flan-t5-xxl, self.wo is kept in float32.
         # See https://github.com/huggingface/transformers/issues/20287
-        if hidden_states.dtype != self.wo.weight.dtype:
+        # we also make sure the weights are not in `int8` in case users will force `_keep_in_fp32_modules` to be `None``
+        if hidden_states.dtype != self.wo.weight.dtype and self.wo.weight.dtype != torch.int8:
             hidden_states = hidden_states.to(self.wo.weight.dtype)
 
         hidden_states = self.wo(hidden_states)

src/transformers/trainer_utils.py

@@ -646,9 +646,9 @@ def find_executable_batch_size(
 
     if auto_find_batch_size:
         requires_backends(find_executable_batch_size, "accelerate")
-        import accelerate.memory_utils as mem_utils
+        from accelerate.utils import find_executable_batch_size as accelerate_find_executable_batch_size
 
-        return mem_utils.find_executable_batch_size(function=function, starting_batch_size=starting_batch_size)
+        return accelerate_find_executable_batch_size(function=function, starting_batch_size=starting_batch_size)
 
     return functools.partial(function, batch_size=starting_batch_size)
 

tests/mixed_int8/test_mixed_int8.py

@@ -163,6 +163,70 @@ class MixedInt8Test(BaseMixedInt8Test):
         self.assertTrue(model.decoder.block[0].layer[2].DenseReluDense.wo.weight.dtype == torch.float32)
 
 
+@require_bitsandbytes
+@require_accelerate
+@require_torch
+@require_torch_gpu
+@slow
+class MixedInt8T5Test(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.model_name = "t5-small"
+        cls.dense_act_model_name = "google/flan-t5-small"  # flan-t5 uses dense-act instead of dense-relu-dense
+        cls.tokenizer = AutoTokenizer.from_pretrained(cls.model_name)
+        cls.input_text = "Translate in German: Hello, my dog is cute"
+
+    def tearDown(self):
+        r"""
+        TearDown function needs to be called at the end of each test to free the GPU memory and cache, also to
+        avoid unexpected behaviors. Please see: https://discuss.pytorch.org/t/how-can-we-release-gpu-memory-cache/14530/27
+        """
+        gc.collect()
+        torch.cuda.empty_cache()
+
+    def test_inference_without_keep_in_fp32(self):
+        r"""
+        Test whether it is possible to mix both `int8` and `fp32` weights when using `keep_in_fp32_modules` correctly.
+        `flan-t5-small` uses `T5DenseGatedActDense` whereas `t5-small` uses `T5DenseReluDense`. We need to test
+        both cases.
+        """
+        from transformers import T5ForConditionalGeneration
+
+        T5ForConditionalGeneration._keep_in_fp32_modules = None
+
+        # test with `t5-small`
+        model = T5ForConditionalGeneration.from_pretrained(self.model_name, load_in_8bit=True, device_map="auto")
+        encoded_input = self.tokenizer(self.input_text, return_tensors="pt").to(0)
+        _ = model.generate(**encoded_input)
+
+        # test with `flan-t5-small`
+        model = T5ForConditionalGeneration.from_pretrained(
+            self.dense_act_model_name, load_in_8bit=True, device_map="auto"
+        )
+        encoded_input = self.tokenizer(self.input_text, return_tensors="pt").to(0)
+        _ = model.generate(**encoded_input)
+
+    def test_inference_with_keep_in_fp32(self):
+        r"""
+        Test whether it is possible to mix both `int8` and `fp32` weights when using `keep_in_fp32_modules` correctly.
+        `flan-t5-small` uses `T5DenseGatedActDense` whereas `t5-small` uses `T5DenseReluDense`. We need to test
+        both cases.
+        """
+        from transformers import T5ForConditionalGeneration
+
+        # test with `t5-small`
+        model = T5ForConditionalGeneration.from_pretrained(self.model_name, load_in_8bit=True, device_map="auto")
+        encoded_input = self.tokenizer(self.input_text, return_tensors="pt").to(0)
+        _ = model.generate(**encoded_input)
+
+        # test with `flan-t5-small`
+        model = T5ForConditionalGeneration.from_pretrained(
+            self.dense_act_model_name, load_in_8bit=True, device_map="auto"
+        )
+        encoded_input = self.tokenizer(self.input_text, return_tensors="pt").to(0)
+        _ = model.generate(**encoded_input)
+
+
 class MixedInt8ModelClassesTest(BaseMixedInt8Test):
     def setUp(self):
         super().setUp()