Merge pull request #43 from jbarrow/main

BERT implementation

bert/README.md

@@ -0,0 +1,78 @@
+# BERT
+
+An implementation of BERT [(Devlin, et al., 2019)](https://aclanthology.org/N19-1423/) within MLX.
+
+## Downloading and Converting Weights
+
+The `convert.py` script relies on `transformers` to download the weights, and exports them as a single `.npz` file.
+
+```
+python convert.py \
+    --bert-model bert-base-uncased
+    --mlx-model weights/bert-base-uncased.npz
+```
+
+## Usage
+
+To use the `Bert` model in your own code, you can load it with:
+
+```python
+from model import Bert, load_model
+
+model, tokenizer = load_model(
+    "bert-base-uncased",
+    "weights/bert-base-uncased.npz")
+
+batch = ["This is an example of BERT working on MLX."]
+tokens = tokenizer(batch, return_tensors="np", padding=True)
+tokens = {key: mx.array(v) for key, v in tokens.items()}
+
+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.
+
+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.
+
+## Comparison with 🤗 `transformers` Implementation
+
+In order to run the model, and have it forward inference on a batch of examples:
+
+```sh
+python model.py \
+  --bert-model bert-base-uncased \
+  --mlx-model weights/bert-base-uncased.npz
+```
+
+Which will show the following outputs:
+```
+MLX BERT:
+[[[-0.52508914 -0.1993871  -0.28210318 ... -0.61125606  0.19114694
+    0.8227601 ]
+  [-0.8783862  -0.37107834 -0.52238125 ... -0.5067165   1.0847603
+    0.31066895]
+  [-0.70010054 -0.5424497  -0.26593682 ... -0.2688697   0.38338926
+    0.6557663 ]
+  ...
+```
+
+They can be compared against the 🤗 implementation with:
+
+```sh
+python hf_model.py \
+  --bert-model bert-base-uncased
+```
+
+Which will show:
+```
+ HF BERT:
+[[[-0.52508944 -0.1993877  -0.28210333 ... -0.6112575   0.19114678
+    0.8227603 ]
+  [-0.878387   -0.371079   -0.522381   ... -0.50671494  1.0847601
+    0.31066933]
+  [-0.7001008  -0.5424504  -0.26593733 ... -0.26887015  0.38339025
+    0.65576553]
+  ...
+```

bert/convert.py

@@ -0,0 +1,47 @@
+from transformers import BertModel
+
+import argparse
+import numpy
+
+
+def replace_key(key: str) -> str:
+    key = key.replace(".layer.", ".layers.")
+    key = key.replace(".self.key.", ".key_proj.")
+    key = key.replace(".self.query.", ".query_proj.")
+    key = key.replace(".self.value.", ".value_proj.")
+    key = key.replace(".attention.output.dense.", ".attention.out_proj.")
+    key = key.replace(".attention.output.LayerNorm.", ".ln1.")
+    key = key.replace(".output.LayerNorm.", ".ln2.")
+    key = key.replace(".intermediate.dense.", ".linear1.")
+    key = key.replace(".output.dense.", ".linear2.")
+    key = key.replace(".LayerNorm.", ".norm.")
+    key = key.replace("pooler.dense.", "pooler.")
+    return key
+
+
+def convert(bert_model: str, mlx_model: str) -> None:
+    model = BertModel.from_pretrained(bert_model)
+    # save the tensors
+    tensors = {
+        replace_key(key): tensor.numpy() for key, tensor in model.state_dict().items()
+    }
+    numpy.savez(mlx_model, **tensors)
+
+
+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"],
+        default="bert-base-uncased",
+        help="The huggingface name of the BERT model to save.",
+    )
+    parser.add_argument(
+        "--mlx-model",
+        type=str,
+        default="weights/bert-base-uncased.npz",
+        help="The output path for the MLX BERT weights.",
+    )
+    args = parser.parse_args()
+
+    convert(args.bert_model, args.mlx_model)

bert/hf_model.py

@@ -0,0 +1,36 @@
+from transformers import AutoModel, AutoTokenizer
+
+import argparse
+
+
+def run(bert_model: str):
+    batch = [
+        "This is an example of BERT working on MLX.",
+        "A second string",
+        "This is another string.",
+    ]
+
+    tokenizer = AutoTokenizer.from_pretrained(bert_model)
+    torch_model = AutoModel.from_pretrained(bert_model)
+    torch_tokens = tokenizer(batch, return_tensors="pt", padding=True)
+    torch_forward = torch_model(**torch_tokens)
+    torch_output = torch_forward.last_hidden_state.detach().numpy()
+    torch_pooled = torch_forward.pooler_output.detach().numpy()
+
+    print("\n HF BERT:")
+    print(torch_output)
+    print("\n\n HF Pooled:")
+    print(torch_pooled[0, :20])
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Run the BERT model using HuggingFace Transformers.")
+    parser.add_argument(
+        "--bert-model",
+        choices=["bert-base-uncased", "bert-base-cased", "bert-large-uncased", "bert-large-cased"],
+        default="bert-base-uncased",
+        help="The huggingface name of the BERT model to save.",
+    )
+    args = parser.parse_args()
+
+    run(args.bert_model)

bert/model.py

@@ -0,0 +1,248 @@
+from typing import Optional
+from dataclasses import dataclass
+from transformers import BertTokenizer
+from mlx.utils import tree_unflatten
+
+import mlx.core as mx
+import mlx.nn as nn
+import argparse
+import numpy
+import math
+
+
+@dataclass
+class ModelArgs:
+    intermediate_size: 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(
+        intermediate_size=1024, num_attention_heads=16, num_hidden_layers=24
+    ),
+    "bert-large-cased": ModelArgs(
+        intermediate_size=1024, num_attention_heads=16, num_hidden_layers=24
+    ),
+}
+
+
+class MultiHeadAttention(nn.Module):
+    """
+    Minor update to the MultiHeadAttention module to ensure that the
+    projections use bias.
+    """
+
+    def __init__(
+        self,
+        dims: int,
+        num_heads: int,
+        query_input_dims: Optional[int] = None,
+        key_input_dims: Optional[int] = None,
+        value_input_dims: Optional[int] = None,
+        value_dims: Optional[int] = None,
+        value_output_dims: Optional[int] = None,
+    ):
+        super().__init__()
+
+        if (dims % num_heads) != 0:
+            raise ValueError(
+                f"The input feature dimensions should be divisible by the number of heads ({dims} % {num_heads}) != 0"
+            )
+
+        query_input_dims = query_input_dims or dims
+        key_input_dims = key_input_dims or dims
+        value_input_dims = value_input_dims or key_input_dims
+        value_dims = value_dims or dims
+        value_output_dims = value_output_dims or dims
+
+        self.num_heads = num_heads
+        self.query_proj = nn.Linear(query_input_dims, dims, True)
+        self.key_proj = nn.Linear(key_input_dims, dims, True)
+        self.value_proj = nn.Linear(value_input_dims, value_dims, True)
+        self.out_proj = nn.Linear(value_dims, value_output_dims, True)
+
+    def __call__(self, queries, keys, values, mask=None):
+        queries = self.query_proj(queries)
+        keys = self.key_proj(keys)
+        values = self.value_proj(values)
+
+        num_heads = self.num_heads
+        B, L, D = queries.shape
+        _, S, _ = keys.shape
+        queries = queries.reshape(B, L, num_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, S, num_heads, -1).transpose(0, 2, 3, 1)
+        values = values.reshape(B, S, num_heads, -1).transpose(0, 2, 1, 3)
+
+        # Dimensions are [batch x num heads x sequence x hidden dim]
+        scale = math.sqrt(1 / queries.shape[-1])
+        scores = (queries * scale) @ keys
+        if mask is not None:
+            mask = self.convert_mask_to_additive_causal_mask(mask)
+            mask = mx.expand_dims(mask, (1, 2))
+            mask = mx.broadcast_to(mask, scores.shape)
+            scores = scores + mask.astype(scores.dtype)
+        scores = mx.softmax(scores, axis=-1)
+        values_hat = (scores @ values).transpose(0, 2, 1, 3).reshape(B, L, -1)
+
+        return self.out_proj(values_hat)
+
+    def convert_mask_to_additive_causal_mask(
+        self, mask: mx.array, dtype: mx.Dtype = mx.float32
+    ) -> mx.array:
+        mask = mask == 0
+        mask = mask.astype(dtype) * -1e9
+        return mask
+
+
+class TransformerEncoderLayer(nn.Module):
+    """
+    A transformer encoder layer with (the original BERT) post-normalization.
+    """
+
+    def __init__(
+        self,
+        dims: int,
+        num_heads: int,
+        mlp_dims: Optional[int] = None,
+        layer_norm_eps: float = 1e-12,
+    ):
+        super().__init__()
+        mlp_dims = mlp_dims or dims * 4
+        self.attention = MultiHeadAttention(dims, num_heads)
+        self.ln1 = nn.LayerNorm(dims, eps=layer_norm_eps)
+        self.ln2 = nn.LayerNorm(dims, eps=layer_norm_eps)
+        self.linear1 = nn.Linear(dims, mlp_dims)
+        self.linear2 = nn.Linear(mlp_dims, dims)
+        self.gelu = nn.GELU()
+
+    def __call__(self, x, mask):
+        attention_out = self.attention(x, x, x, mask)
+        add_and_norm = self.ln1(x + attention_out)
+
+        ff = self.linear1(add_and_norm)
+        ff_gelu = self.gelu(ff)
+        ff_out = self.linear2(ff_gelu)
+        x = self.ln2(ff_out + add_and_norm)
+
+        return x
+
+
+class TransformerEncoder(nn.Module):
+    def __init__(
+        self, num_layers: int, dims: int, num_heads: int, mlp_dims: Optional[int] = None
+    ):
+        super().__init__()
+        self.layers = [
+            TransformerEncoderLayer(dims, num_heads, mlp_dims)
+            for i in range(num_layers)
+        ]
+
+    def __call__(self, x, mask):
+        for layer in self.layers:
+            x = layer(x, mask)
+
+        return x
+
+
+class BertEmbeddings(nn.Module):
+    def __init__(self, config: ModelArgs):
+        self.word_embeddings = nn.Embedding(config.vocab_size, config.intermediate_size)
+        self.token_type_embeddings = nn.Embedding(2, config.intermediate_size)
+        self.position_embeddings = nn.Embedding(
+            config.max_position_embeddings, config.intermediate_size
+        )
+        self.norm = nn.LayerNorm(config.intermediate_size, eps=config.layer_norm_eps)
+
+    def __call__(self, input_ids: mx.array, token_type_ids: mx.array) -> mx.array:
+        words = self.word_embeddings(input_ids)
+        position = self.position_embeddings(
+            mx.broadcast_to(mx.arange(input_ids.shape[1]), input_ids.shape)
+        )
+        token_types = self.token_type_embeddings(token_type_ids)
+
+        embeddings = position + words + token_types
+        return self.norm(embeddings)
+
+
+class Bert(nn.Module):
+    def __init__(self, config: ModelArgs):
+        self.embeddings = BertEmbeddings(config)
+        self.encoder = TransformerEncoder(
+            num_layers=config.num_hidden_layers,
+            dims=config.intermediate_size,
+            num_heads=config.num_attention_heads,
+        )
+        self.pooler = nn.Linear(config.intermediate_size, config.vocab_size)
+
+    def __call__(
+        self,
+        input_ids: mx.array,
+        token_type_ids: mx.array,
+        attention_mask: Optional[mx.array] = None,
+    ) -> tuple[mx.array, mx.array]:
+        x = self.embeddings(input_ids, token_type_ids)
+        y = self.encoder(x, attention_mask)
+        return y, mx.tanh(self.pooler(y[:, 0]))
+
+
+def load_model(bert_model: str, weights_path: str) -> tuple[Bert, BertTokenizer]:
+    # load the weights npz
+    weights = mx.load(weights_path)
+    weights = tree_unflatten(list(weights.items()))
+    # create and update the model
+    model = Bert(model_configs[bert_model])
+    model.update(weights)
+
+    tokenizer = BertTokenizer.from_pretrained(bert_model)
+
+    return model, tokenizer
+
+
+def run(bert_model: str, mlx_model: str):
+    model, tokenizer = load_model(bert_model, mlx_model)
+
+    batch = [
+        "This is an example of BERT working on MLX.",
+        "A second string",
+        "This is another string.",
+    ]
+    
+    tokens = tokenizer(batch, return_tensors="np", padding=True)
+    tokens = {key: mx.array(v) for key, v in tokens.items()}
+
+    mlx_output, mlx_pooled = model(**tokens)
+    mlx_output = numpy.array(mlx_output)
+    mlx_pooled = numpy.array(mlx_pooled)
+
+    print("MLX BERT:")
+    print(mlx_output)
+
+    print("\n\nMLX Pooled:")
+    print(mlx_pooled[0, :20])
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Run the BERT model using MLX.")
+    parser.add_argument(
+        "--bert-model",
+        type=str,
+        default="bert-base-uncased",
+        help="The huggingface name of the BERT model to save.",
+    )
+    parser.add_argument(
+        "--mlx-model",
+        type=str,
+        default="weights/bert-base-uncased.npz",
+        help="The path of the stored MLX BERT weights (npz file).",
+    )
+    args = parser.parse_args()
+
+    run(args.bert_model, args.mlx_model)

bert/requirements.txt

@@ -0,0 +1,3 @@
+mlx
+transformers
+numpy

bert/weights/.gitignore

@@ -0,0 +1 @@
+*.npz