Enable more BERT models (#580)

* Update convert.py

* Update model.py

* Update test.py

* Update model.py

* Update convert.py

* Add files via upload

* Update convert.py

* format

* nit

* nit

---------

Co-authored-by: Awni Hannun <awni@apple.com>

bert/README.md

@@ -1,6 +1,6 @@
 # BERT
 
-An implementation of BERT [(Devlin, et al., 2019)](https://aclanthology.org/N19-1423/) within MLX.
+An implementation of BERT [(Devlin, et al., 2019)](https://aclanthology.org/N19-1423/) in MLX.
 
 ## Setup 
 
@@ -38,12 +38,12 @@ output, pooled = model(**tokens)
 ```
 
 The `output` contains a `Batch x Tokens x Dims` tensor, representing a vector
-for every input token. If you want to train anything at a **token-level**,
-you'll want to use this.
+for every input token. If you want to train anything at the **token-level**,
+use this.
 
 The `pooled` contains a `Batch x Dims` tensor, which is the pooled
 representation for each input. If you want to train a **classification**
-model, you'll want to use this.
+model, use this.
 
 
 ## Test

bert/convert.py

@@ -1,7 +1,7 @@
 import argparse
 
 import numpy
-from transformers import BertModel
+from transformers import AutoModel
 
 
 def replace_key(key: str) -> str:
@@ -20,7 +20,7 @@ def replace_key(key: str) -> str:
 
 
 def convert(bert_model: str, mlx_model: str) -> None:
-    model = BertModel.from_pretrained(bert_model)
+    model = AutoModel.from_pretrained(bert_model)
     # save the tensors
     tensors = {
         replace_key(key): tensor.numpy() for key, tensor in model.state_dict().items()
@@ -32,14 +32,9 @@ if __name__ == "__main__":
     parser = argparse.ArgumentParser(description="Convert BERT weights to MLX.")
     parser.add_argument(
         "--bert-model",
-        choices=[
-            "bert-base-uncased",
-            "bert-base-cased",
-            "bert-large-uncased",
-            "bert-large-cased",
-        ],
+        type=str,
         default="bert-base-uncased",
-        help="The huggingface name of the BERT model to save.",
+        help="The huggingface name of the BERT model to save. Any BERT-like model can be specified.",
     )
     parser.add_argument(
         "--mlx-model",

bert/model.py

@@ -8,31 +8,7 @@ import mlx.nn as nn
 import numpy
 import numpy as np
 from mlx.utils import tree_unflatten
-from transformers import BertTokenizer
-
-
-@dataclass
-class ModelArgs:
-    dim: int = 768
-    num_attention_heads: int = 12
-    num_hidden_layers: int = 12
-    vocab_size: int = 30522
-    attention_probs_dropout_prob: float = 0.1
-    hidden_dropout_prob: float = 0.1
-    layer_norm_eps: float = 1e-12
-    max_position_embeddings: int = 512
-
-
-model_configs = {
-    "bert-base-uncased": ModelArgs(),
-    "bert-base-cased": ModelArgs(),
-    "bert-large-uncased": ModelArgs(
-        dim=1024, num_attention_heads=16, num_hidden_layers=24
-    ),
-    "bert-large-cased": ModelArgs(
-        dim=1024, num_attention_heads=16, num_hidden_layers=24
-    ),
-}
+from transformers import AutoConfig, AutoTokenizer, PreTrainedTokenizerBase
 
 
 class TransformerEncoderLayer(nn.Module):
@@ -86,20 +62,29 @@ class TransformerEncoder(nn.Module):
 
 
 class BertEmbeddings(nn.Module):
-    def __init__(self, config: ModelArgs):
+    def __init__(self, config):
         super().__init__()
-        self.word_embeddings = nn.Embedding(config.vocab_size, config.dim)
-        self.token_type_embeddings = nn.Embedding(2, config.dim)
-        self.position_embeddings = nn.Embedding(
-            config.max_position_embeddings, config.dim
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.token_type_embeddings = nn.Embedding(
+            config.type_vocab_size, config.hidden_size
         )
-        self.norm = nn.LayerNorm(config.dim, eps=config.layer_norm_eps)
+        self.position_embeddings = nn.Embedding(
+            config.max_position_embeddings, config.hidden_size
+        )
+        self.norm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
 
-    def __call__(self, input_ids: mx.array, token_type_ids: mx.array) -> mx.array:
+    def __call__(
+        self, input_ids: mx.array, token_type_ids: mx.array = None
+    ) -> mx.array:
         words = self.word_embeddings(input_ids)
         position = self.position_embeddings(
             mx.broadcast_to(mx.arange(input_ids.shape[1]), input_ids.shape)
         )
+
+        if token_type_ids is None:
+            # If token_type_ids is not provided, default to zeros
+            token_type_ids = mx.zeros_like(input_ids)
+
         token_types = self.token_type_embeddings(token_type_ids)
 
         embeddings = position + words + token_types
@@ -107,20 +92,21 @@ class BertEmbeddings(nn.Module):
 
 
 class Bert(nn.Module):
-    def __init__(self, config: ModelArgs):
+    def __init__(self, config):
         super().__init__()
         self.embeddings = BertEmbeddings(config)
         self.encoder = TransformerEncoder(
             num_layers=config.num_hidden_layers,
-            dims=config.dim,
+            dims=config.hidden_size,
             num_heads=config.num_attention_heads,
+            mlp_dims=config.intermediate_size,
         )
-        self.pooler = nn.Linear(config.dim, config.dim)
+        self.pooler = nn.Linear(config.hidden_size, config.hidden_size)
 
     def __call__(
         self,
         input_ids: mx.array,
-        token_type_ids: mx.array,
+        token_type_ids: mx.array = None,
         attention_mask: mx.array = None,
     ) -> Tuple[mx.array, mx.array]:
         x = self.embeddings(input_ids, token_type_ids)
@@ -134,15 +120,19 @@ class Bert(nn.Module):
         return y, mx.tanh(self.pooler(y[:, 0]))
 
 
-def load_model(bert_model: str, weights_path: str) -> Tuple[Bert, BertTokenizer]:
+def load_model(
+    bert_model: str, weights_path: str
+) -> Tuple[Bert, PreTrainedTokenizerBase]:
     if not Path(weights_path).exists():
         raise ValueError(f"No model weights found in {weights_path}")
 
+    config = AutoConfig.from_pretrained(bert_model)
+
     # create and update the model
-    model = Bert(model_configs[bert_model])
+    model = Bert(config)
     model.load_weights(weights_path)
 
-    tokenizer = BertTokenizer.from_pretrained(bert_model)
+    tokenizer = AutoTokenizer.from_pretrained(bert_model)
 
     return model, tokenizer
 

bert/test.py

@@ -1,3 +1,4 @@
+import argparse
 from typing import List
 
 import model
@@ -16,19 +17,41 @@ def run_torch(bert_model: str, batch: List[str]):
 
 
 if __name__ == "__main__":
-    bert_model = "bert-base-uncased"
-    mlx_model = "weights/bert-base-uncased.npz"
-    batch = [
-        "This is an example of BERT working in MLX.",
-        "A second string",
-        "This is another string.",
-    ]
-    torch_output, torch_pooled = run_torch(bert_model, batch)
-    mlx_output, mlx_pooled = model.run(bert_model, mlx_model, batch)
-    assert np.allclose(
-        torch_output, mlx_output, rtol=1e-4, atol=1e-5
-    ), "Model output is different"
-    assert np.allclose(
-        torch_pooled, mlx_pooled, rtol=1e-4, atol=1e-5
-    ), "Model pooled output is different"
-    print("Tests pass :)")
+    parser = argparse.ArgumentParser(
+        description="Run a BERT-like model for a batch of text."
+    )
+    parser.add_argument(
+        "--bert-model",
+        type=str,
+        default="bert-base-uncased",
+        help="The model identifier for a BERT-like model from Hugging Face Transformers.",
+    )
+    parser.add_argument(
+        "--mlx-model",
+        type=str,
+        default="weights/bert-base-uncased.npz",
+        help="The path of the stored MLX BERT weights (npz file).",
+    )
+    parser.add_argument(
+        "--text",
+        nargs="+",
+        default=["This is an example of BERT working in MLX."],
+        help="A batch of texts to process. Multiple texts should be separated by spaces.",
+    )
+
+    args = parser.parse_args()
+
+    torch_output, torch_pooled = run_torch(args.bert_model, args.text)
+
+    mlx_output, mlx_pooled = model.run(args.bert_model, args.mlx_model, args.text)
+
+    if torch_pooled is not None and mlx_pooled is not None:
+        assert np.allclose(
+            torch_output, mlx_output, rtol=1e-4, atol=1e-5
+        ), "Model output is different"
+        assert np.allclose(
+            torch_pooled, mlx_pooled, rtol=1e-4, atol=1e-5
+        ), "Model pooled output is different"
+        print("Tests pass :)")
+    else:
+        print("Pooled outputs were not compared due to one or both being None.")