Patch release: v4.27.8

Revert "Error (also in original) model, scaling only q matrix not qk.T dot product (qk.T/sqrt(dim_per_head)) (#21627)"

This reverts commit bad8300837.

setup.py

@@ -418,7 +418,7 @@ install_requires = [
 
 setup(
     name="transformers",
-    version="4.27.1",  # expected format is one of x.y.z.dev0, or x.y.z.rc1 or x.y.z (no to dashes, yes to dots)
+    version="4.27.4",  # 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",

src/transformers/__init__.py

@@ -18,7 +18,7 @@
 # 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__ = "4.27.1"
+__version__ = "4.27.4"
 
 from typing import TYPE_CHECKING
 

src/transformers/modeling_utils.py

@@ -2563,7 +2563,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
             elif device_map in ["balanced", "balanced_low_0"] and get_balanced_memory is None:
                 raise ValueError(f"`device_map={device_map}` requires a source install of Accelerate.")
 
-            kwargs = {"no_split_module_classes": no_split_modules, "max_memory": max_memory}
+            kwargs = {"no_split_module_classes": no_split_modules}
             if "special_dtypes" in inspect.signature(infer_auto_device_map).parameters:
                 kwargs["special_dtypes"] = special_dtypes
             elif len(special_dtypes) > 0:
@@ -2576,8 +2576,10 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                     model,
                     dtype=torch_dtype,
                     low_zero=(device_map == "balanced_low_0"),
+                    max_memory=max_memory,
                     **kwargs,
                 )
+            kwargs["max_memory"] = max_memory
             # Make sure tied weights are tied before creating the device map.
             model.tie_weights()
             device_map = infer_auto_device_map(model, dtype=torch_dtype if not load_in_8bit else torch.int8, **kwargs)

src/transformers/models/flaubert/modeling_flaubert.py

@@ -172,7 +172,8 @@ class MultiHeadAttention(nn.Module):
                     k, v = cache[self.layer_id]
             cache[self.layer_id] = (k, v)
 
-        scores = torch.matmul(q, k.transpose(2, 3)) / math.sqrt(dim_per_head)  # (bs, n_heads, qlen, klen)
+        q = q / math.sqrt(dim_per_head)  # (bs, n_heads, qlen, dim_per_head)
+        scores = torch.matmul(q, k.transpose(2, 3))  # (bs, n_heads, qlen, klen)
         mask = (mask == 0).view(mask_reshape).expand_as(scores)  # (bs, n_heads, qlen, klen)
         scores.masked_fill_(mask, torch.finfo(scores.dtype).min)  # (bs, n_heads, qlen, klen)
 

src/transformers/models/xlm/modeling_xlm.py

@@ -176,7 +176,8 @@ class MultiHeadAttention(nn.Module):
                     k, v = cache[self.layer_id]
             cache[self.layer_id] = (k, v)
 
-        scores = torch.matmul(q, k.transpose(2, 3)) / math.sqrt(dim_per_head)  # (bs, n_heads, qlen, klen)
+        q = q / math.sqrt(dim_per_head)  # (bs, n_heads, qlen, dim_per_head)
+        scores = torch.matmul(q, k.transpose(2, 3))  # (bs, n_heads, qlen, klen)
         mask = (mask == 0).view(mask_reshape).expand_as(scores)  # (bs, n_heads, qlen, klen)
         scores.masked_fill_(mask, torch.finfo(scores.dtype).min)  # (bs, n_heads, qlen, klen)