LlaVA in MLX (#461)

* add: llava mlx first draft

* add: weights comparision

* add forward pass skeleton

* update: now  imports weights correctly

* delete base

* latest

* adding config

* fix: use config

* add mlx config

* feat: add image processor for llava processor

* wip

* feat: llava working example

* chore: refactor generate script

* chore: clean up

* add: warning to user if no <image> token despite using one

* add: __call__ to LlavaModel

* add: call to LlavaModel

* update fp

* clean up var names

* update: native GeLU

* Cleanup

* update generate and readme

* remove todo comment

* rearrange tests

* fix example code

* nits in README

* update readme

* nit in readme

* nits in README

* chore(llava): refactor image embedding merging logic

* min mlx version

* nits in readmes

* fix cli prompt, some nits

* updates, slight simplify

---------

Co-authored-by: anchen <li.anchen.au@gmail.com>
Co-authored-by: Awni Hannun <awni@apple.com>

README.md

@@ -31,6 +31,7 @@ Some more useful examples are listed below.
 ### Multimodal models
 
 - Joint text and image embeddings with [CLIP](clip).
+- Text generation from image and text inputs with [LLaVA](llava).
 
 ### Other Models 
 

llava/.gitignore

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

llava/README.md

@@ -0,0 +1,61 @@
+# LLaVA
+
+An example of LLaVA: Large Language and Vision Assistant in MLX.[^1] LLlava is
+a multimodal model that can generate text given combined image and text inputs.
+
+## Setup
+
+Install the dependencies:
+
+```bash
+pip install -r requirements.txt
+```
+
+## Run
+
+You can use LLaVA to ask questions about images.
+
+For example, using the command line:
+
+```bash
+python generate.py \
+  --model llava-hf/llava-1.5-7b-hf \
+  --image "http://images.cocodataset.org/val2017/000000039769.jpg" \
+  --prompt "USER: <image>\nWhat are these?\nASSISTANT:" \
+  --max-tokens 128 \
+  --temp 0
+```
+
+This uses the following image:
+
+![alt text](http://images.cocodataset.org/val2017/000000039769.jpg)
+ 
+And generates the output:
+
+```
+These are two cats lying on a pink couch.
+```
+
+You can also use LLaVA in Python:
+
+```python
+from generate import load_model, prepare_inputs, generate_text
+
+processor, model = load_model("llava-hf/llava-1.5-7b-hf")
+
+max_tokens, temperature = 128, 0.0
+
+prompt = "USER: <image>\nWhat are these?\nASSISTANT:"
+image = "http://images.cocodataset.org/val2017/000000039769.jpg"
+input_ids, pixel_values = prepare_inputs(processor, image, prompt)
+
+reply = generate_text(
+    input_ids, pixel_values, model, processor, max_tokens, temperature
+)
+
+print(reply)
+```
+
+[^1]:
+    Refer to [LLaVA project webpage](https://llava-vl.github.io/) for more
+    information.

llava/generate.py

@@ -0,0 +1,130 @@
+# Copyright © 2024 Apple Inc.
+
+import argparse
+import codecs
+from pathlib import Path
+
+import mlx.core as mx
+import requests
+from PIL import Image
+from transformers import AutoProcessor
+
+from llava import LlavaModel
+
+
+def parse_arguments():
+    parser = argparse.ArgumentParser(
+        description="Generate text from an image using a model."
+    )
+    parser.add_argument(
+        "--model",
+        type=str,
+        default="llava-hf/llava-1.5-7b-hf",
+        help="The path to the local model directory or Hugging Face repo.",
+    )
+    parser.add_argument(
+        "--image",
+        type=str,
+        default="http://images.cocodataset.org/val2017/000000039769.jpg",
+        help="URL or path of the image to process.",
+    )
+    parser.add_argument(
+        "--prompt",
+        type=str,
+        default="USER: <image>\nWhat are these?\nASSISTANT:",
+        help="Message to be processed by the model.",
+    )
+    parser.add_argument(
+        "--max-tokens",
+        type=int,
+        default=100,
+        help="Maximum number of tokens to generate.",
+    )
+    parser.add_argument(
+        "--temp", type=float, default=0.3, help="Temperature for sampling."
+    )
+    return parser.parse_args()
+
+
+def load_image(image_source):
+    """
+    Helper function to load an image from either a URL or file.
+    """
+    if image_source.startswith(("http://", "https://")):
+        try:
+            response = requests.get(image_source, stream=True)
+            response.raise_for_status()
+            return Image.open(response.raw)
+        except Exception as e:
+            raise ValueError(
+                f"Failed to load image from URL: {image_source} with error {e}"
+            )
+    elif Path(image_source).is_file():
+        try:
+            return Image.open(image_source)
+        except IOError as e:
+            raise ValueError(f"Failed to load image {image_source} with error: {e}")
+    else:
+        raise ValueError(
+            f"The image {image_source} must be a valid URL or existing file."
+        )
+
+
+def prepare_inputs(processor, image, prompt):
+    if isinstance(image, str):
+        image = load_image(image)
+    inputs = processor(prompt, image, return_tensors="np")
+    pixel_values = mx.array(inputs["pixel_values"])
+    input_ids = mx.array(inputs["input_ids"])
+    return input_ids, pixel_values
+
+
+def load_model(model_path):
+    processor = AutoProcessor.from_pretrained(model_path)
+    model = LlavaModel.from_pretrained(model_path)
+    return processor, model
+
+
+def sample(logits, temperature=0.0):
+    if temperature == 0:
+        return mx.argmax(logits, axis=-1)
+    else:
+        return mx.random.categorical(logits * (1 / temperature))
+
+
+def generate_text(input_ids, pixel_values, model, processor, max_tokens, temperature):
+
+    logits, cache = model(input_ids, pixel_values)
+    logits = logits[:, -1, :]
+    y = sample(logits, temperature=temperature)
+    tokens = [y.item()]
+
+    for n in range(max_tokens - 1):
+        logits, cache = model.language_model(y[None], cache=cache)
+        logits = logits[:, -1, :]
+        y = sample(logits, temperature)
+        token = y.item()
+        if token == processor.tokenizer.eos_token_id:
+            break
+        tokens.append(token)
+
+    return processor.tokenizer.decode(tokens)
+
+
+def main():
+    args = parse_arguments()
+    processor, model = load_model(args.model)
+
+    prompt = codecs.decode(args.prompt, "unicode_escape")
+
+    input_ids, pixel_values = prepare_inputs(processor, args.image, prompt)
+
+    print(prompt)
+    generated_text = generate_text(
+        input_ids, pixel_values, model, processor, args.max_tokens, args.temp
+    )
+    print(generated_text)
+
+
+if __name__ == "__main__":
+    main()

llava/language.py

@@ -0,0 +1,231 @@
+# Copyright © 2024 Apple Inc.
+
+import inspect
+from dataclasses import dataclass
+from typing import Dict, Optional, Tuple, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+
+
+@dataclass
+class TextConfig:
+    model_type: str
+    hidden_size: int = 4096
+    num_hidden_layers: int = 32
+    intermediate_size: int = 11008
+    num_attention_heads: int = 32
+    rms_norm_eps: float = 1e-6
+    vocab_size: int = 32000
+    num_key_value_heads: int = None
+    rope_theta: float = 10000
+    rope_traditional: bool = False
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+
+    @classmethod
+    def from_dict(cls, params):
+        return cls(
+            **{
+                k: v
+                for k, v in params.items()
+                if k in inspect.signature(cls).parameters
+            }
+        )
+
+    def __post_init__(self):
+        if self.num_key_value_heads is None:
+            self.num_key_value_heads = self.num_attention_heads
+
+        if self.rope_scaling:
+            required_keys = {"factor", "type"}
+            if not all(key in self.rope_scaling for key in required_keys):
+                raise ValueError(f"rope_scaling must contain keys {required_keys}")
+
+            if self.rope_scaling["type"] != "linear":
+                raise ValueError("rope_scaling 'type' currently only supports 'linear'")
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, dims: int, eps: float = 1e-5):
+        super().__init__()
+        self.weight = mx.ones((dims,))
+        self.eps = eps
+
+    def _norm(self, x):
+        return x * mx.rsqrt(x.square().mean(-1, keepdims=True) + self.eps)
+
+    def __call__(self, x):
+        output = self._norm(x.astype(mx.float32)).astype(x.dtype)
+        return self.weight * output
+
+
+class Attention(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+
+        dim = config.hidden_size
+        self.n_heads = n_heads = config.num_attention_heads
+        self.n_kv_heads = n_kv_heads = config.num_key_value_heads
+
+        self.repeats = n_heads // n_kv_heads
+
+        head_dim = config.hidden_size // n_heads
+        self.scale = head_dim**-0.5
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=False)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=False)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=False)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=False)
+
+        rope_scale = (
+            1 / config.rope_scaling["factor"]
+            if config.rope_scaling is not None
+            and config.rope_scaling["type"] == "linear"
+            else 1
+        )
+        self.rope = nn.RoPE(
+            head_dim,
+            traditional=config.rope_traditional,
+            base=config.rope_theta,
+            scale=rope_scale,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Tuple[mx.array, mx.array]] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        # Prepare the queries, keys and values for the attention computation
+        queries = queries.reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        if self.repeats > 1:
+            keys = mx.repeat(keys, self.repeats, axis=1)
+            values = mx.repeat(values, self.repeats, axis=1)
+
+        if cache is not None:
+            key_cache, value_cache = cache
+            queries = self.rope(queries, offset=key_cache.shape[2])
+            keys = self.rope(keys, offset=key_cache.shape[2])
+            keys = mx.concatenate([key_cache, keys], axis=2)
+            values = mx.concatenate([value_cache, values], axis=2)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        scores = (queries * self.scale) @ keys.transpose(0, 1, 3, 2)
+        if mask is not None:
+            scores += mask
+        scores = mx.softmax(scores.astype(mx.float32), axis=-1).astype(scores.dtype)
+        output = (scores @ values).transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output), (keys, values)
+
+
+class MLP(nn.Module):
+    def __init__(self, dim, hidden_dim):
+        super().__init__()
+        self.gate_proj = nn.Linear(dim, hidden_dim, bias=False)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=False)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=False)
+
+    def __call__(self, x) -> mx.array:
+        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.num_attention_heads = config.num_attention_heads
+        self.hidden_size = config.hidden_size
+        self.self_attn = Attention(config)
+        self.mlp = MLP(config.hidden_size, config.intermediate_size)
+        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+        self.config = config
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Tuple[mx.array, mx.array]] = None,
+    ) -> mx.array:
+        r, cache = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        out = h + r
+        return out, cache
+
+
+class Llama(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.config = config
+        self.vocab_size = config.vocab_size
+        self.num_hidden_layers = config.num_hidden_layers
+        assert self.vocab_size > 0
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = [
+            TransformerBlock(config=config) for _ in range(config.num_hidden_layers)
+        ]
+        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+        inputs_embeds=None,
+    ):
+        # for passing merged input embeddings
+        if inputs_embeds is None:
+            h = self.embed_tokens(inputs)
+        else:
+            h = inputs_embeds
+
+        mask = None
+        if h.shape[1] > 1:
+            mask = nn.MultiHeadAttention.create_additive_causal_mask(h.shape[1])
+            mask = mask.astype(h.dtype)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        for e, layer in enumerate(self.layers):
+            h, cache[e] = layer(h, mask, cache[e])
+
+        return self.norm(h), cache
+
+
+class LanguageModel(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.model_type = config.model_type
+        if self.model_type != "llama":
+            raise ValueError(
+                f"Model type {self.model_type} not supported. Currently only 'llama' is supported"
+            )
+        self.model = Llama(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        cache=None,
+        inputs_embeds=None,
+    ):
+        out, cache = self.model(inputs, cache, inputs_embeds)
+        return self.lm_head(out), cache
+
+    @staticmethod
+    def sanitize(weights):
+        # Remove unused precomputed rotary freqs
+        return {
+            k: v for k, v in weights.items() if "self_attn.rotary_emb.inv_freq" not in k
+        }

llava/llava.py

@@ -0,0 +1,179 @@
+# Copyright © 2024 Apple Inc.
+
+import glob
+import inspect
+import json
+from dataclasses import dataclass
+from pathlib import Path
+from typing import Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+import numpy as np
+from huggingface_hub import snapshot_download
+from language import LanguageModel, TextConfig
+from vision import VisionConfig, VisionModel
+
+
+@dataclass
+class LlaVAConfig:
+    text_config: TextConfig
+    vision_config: VisionConfig
+    ignore_index: int = -100
+    image_token_index: int = 32000
+    vision_feature_select_strategy: str = "default"
+    vision_feature_layer: int = -2
+    vocab_size: int = 32000
+
+    @classmethod
+    def from_dict(cls, params):
+        return cls(
+            **{
+                k: v
+                for k, v in params.items()
+                if k in inspect.signature(cls).parameters
+            }
+        )
+
+
+class LlavaMultiModalProjector(nn.Module):
+    def __init__(self, config: LlaVAConfig):
+        super().__init__()
+        self.linear_1 = nn.Linear(
+            config.vision_config.hidden_size, config.text_config.hidden_size, bias=True
+        )
+        self.gelu = nn.GELU()
+        self.linear_2 = nn.Linear(
+            config.text_config.hidden_size, config.text_config.hidden_size, bias=True
+        )
+
+    def __call__(self, x: mx.array) -> mx.array:
+        x = self.linear_1(x)
+        x = self.gelu(x)
+        x = self.linear_2(x)
+        return x
+
+
+class LlavaModel(nn.Module):
+    def __init__(self, config: LlaVAConfig):
+        self.config = config
+        self.vision_tower = VisionModel(config.vision_config)
+        self.language_model = LanguageModel(config.text_config)
+        self.multi_modal_projector = LlavaMultiModalProjector(config)
+        self.vision_feature_layer = config.vision_feature_layer
+        self.vision_feature_select_strategy = config.vision_feature_select_strategy
+
+    def get_input_embeddings(
+        self,
+        input_ids: Optional[mx.array] = None,
+        pixel_values: Optional[mx.array] = None,
+    ):
+        if pixel_values is None:
+            return self.language_model(input_ids)
+
+        # Get the input embeddings from the language model
+        inputs_embeds = self.language_model.model.embed_tokens(input_ids)
+
+        # Get the ouptut hidden states from the vision model
+        *_, hidden_states = self.vision_tower(
+            pixel_values.transpose(0, 2, 3, 1), output_hidden_states=True
+        )
+
+        # Select the hidden states from the desired layer
+        selected_image_feature = hidden_states[self.vision_feature_layer]
+
+        if self.vision_feature_select_strategy == "default":
+            selected_image_feature = selected_image_feature[:, 1:]
+        elif self.vision_feature_select_strategy == "full":
+            selected_image_feature = selected_image_feature
+        else:
+            raise ValueError(
+                "Unexpected feature selection strategy: "
+                f"{self.vision_feature_select_strategy}"
+            )
+
+        # Pass image features through the multi-modal projector
+        image_features = self.multi_modal_projector(selected_image_feature)
+
+        # Insert special image tokens in the input_ids
+        final_inputs_embeds = self._merge_input_ids_with_image_features(
+            image_features, inputs_embeds, input_ids
+        )
+        return final_inputs_embeds
+
+    def _merge_input_ids_with_image_features(
+        self, image_features, inputs_embeds, input_ids
+    ):
+        image_token_index = self.config.image_token_index
+        num_images, num_image_patches, embed_dim = image_features.shape
+
+        # Positions of <image> tokens in input_ids, assuming batch size is 1
+        image_positions = np.where(input_ids[0] == image_token_index)[0].tolist()
+
+        if len(image_positions) != num_images:
+            raise ValueError(
+                f"The number of image tokens ({len(image_positions)}) does not "
+                f" match the number of image inputs ({num_images})."
+            )
+
+        text_segments = []
+        start_idx = 0
+
+        for position in image_positions:
+            text_segments.append(inputs_embeds[:, start_idx:position])
+            start_idx = position + 1
+
+        image_embeddings = mx.split(image_features, image_features.shape[0])
+        final_embeddings = [v for p in zip(text_segments, image_embeddings) for v in p]
+        final_embeddings += [inputs_embeds[:, start_idx:]]
+
+        # Create a final embedding of shape
+        # (1, num_image_patches*num_images + sequence_len, embed_dim)
+        return mx.concatenate(final_embeddings, axis=1)
+
+    def __call__(self, input_ids: mx.array, pixel_values: mx.array, cache=None):
+        input_embddings = self.get_input_embeddings(input_ids, pixel_values)
+        logits, cache = self.language_model(
+            input_ids, cache=cache, inputs_embeds=input_embddings
+        )
+        return logits, cache
+
+    @staticmethod
+    def from_pretrained(path_or_hf_repo: str):
+        path = Path(path_or_hf_repo)
+        if not path.exists():
+            path = Path(
+                snapshot_download(
+                    repo_id=path_or_hf_repo,
+                    allow_patterns=[
+                        "*.json",
+                        "*.safetensors",
+                        "*.py",
+                        "tokenizer.model",
+                        "*.tiktoken",
+                    ],
+                )
+            )
+
+        with open(path / "config.json", "r") as f:
+            model_config = json.load(f)
+
+        model_config = LlaVAConfig.from_dict(model_config)
+
+        model_config.vision_config = VisionConfig.from_dict(model_config.vision_config)
+        model_config.text_config = TextConfig.from_dict(model_config.text_config)
+
+        model = LlavaModel(model_config)
+        weight_files = glob.glob(str(path / "*.safetensors"))
+        if not weight_files:
+            raise FileNotFoundError(f"No safetensors found in {path}")
+
+        weights = {}
+        for wf in weight_files:
+            weights.update(mx.load(wf))
+
+        weights = VisionModel.sanitize(weights)
+        weights = LanguageModel.sanitize(weights)
+
+        model.load_weights(list(weights.items()))
+        return model

llava/requirements.txt

@@ -0,0 +1,6 @@
+mlx>=0.5.0
+numpy
+transformers
+torch
+huggingface_hub
+Pillow

llava/test.py

@@ -0,0 +1,162 @@
+# Copyright © 2024 Apple Inc.
+
+import unittest
+
+import mlx.core as mx
+import requests
+import torch
+from PIL import Image
+from transformers import AutoProcessor, LlavaForConditionalGeneration
+
+from llava import LlavaModel
+
+MODEL_PATH = "llava-hf/llava-1.5-7b-hf"
+PROMPT = "USER: <image>\nWhat are these?\nASSISTANT:"
+IMAGE_FILE = "http://images.cocodataset.org/val2017/000000039769.jpg"
+
+
+def load_mlx_models(path):
+    model = LlavaModel.from_pretrained(path)
+    model.eval()
+    return model
+
+
+def load_hf_models(path):
+    model = LlavaForConditionalGeneration.from_pretrained(path)
+    model.eval()
+    return model
+
+
+class TestVisionTower(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.mx_llava = load_mlx_models(MODEL_PATH)
+        cls.hf_llava = load_hf_models(MODEL_PATH)
+        cls.proc = AutoProcessor.from_pretrained(MODEL_PATH)
+
+    def test_image_features(self):
+        raw_image = Image.open(requests.get(IMAGE_FILE, stream=True).raw)
+        vision_feature_layer = -2
+        with torch.no_grad():
+            pixel_values = self.proc(PROMPT, raw_image, return_tensors="pt")[
+                "pixel_values"
+            ]
+
+            hf_pixel_values = pixel_values
+            mx_pixel_values = mx.array(pixel_values.numpy()).transpose(0, 2, 3, 1)
+
+            _, _, hidden_states = self.mx_llava.vision_tower(
+                mx_pixel_values,
+                output_hidden_states=True,
+            )
+
+            mx_elected_image_feature = hidden_states[vision_feature_layer]
+            mx_image_features = self.mx_llava.multi_modal_projector(
+                mx_elected_image_feature
+            )
+
+            hf_image_outputs = self.hf_llava.vision_tower(
+                hf_pixel_values, output_hidden_states=True
+            )
+            hf_elected_image_feature = hf_image_outputs.hidden_states[
+                vision_feature_layer
+            ]
+            hf_image_features = self.hf_llava.multi_modal_projector(
+                hf_elected_image_feature
+            )
+
+            self.assertTrue(
+                mx.allclose(
+                    mx_image_features,
+                    mx.array(hf_image_features.numpy()),
+                    atol=1e-2,
+                )
+            )
+
+
+class TestLlava(unittest.TestCase):
+    @classmethod
+    def setUpClass(cls):
+        cls.mx_llava = load_mlx_models(MODEL_PATH)
+        cls.hf_llava = load_hf_models(MODEL_PATH)
+        cls.proc = AutoProcessor.from_pretrained(MODEL_PATH)
+
+    def test_merge_input_ids_with_image_features(self):
+        raw_image = Image.open(requests.get(IMAGE_FILE, stream=True).raw)
+        vision_feature_layer = -2
+        with torch.no_grad():
+            values = self.proc(PROMPT, raw_image, return_tensors="pt")
+            pixel_values = values["pixel_values"]
+            input_ids = values["input_ids"]
+
+            hf_pixel_values = pixel_values
+            mx_pixel_values = mx.array(pixel_values.numpy()).transpose(0, 2, 3, 1)
+
+            _, _, hidden_states = self.mx_llava.vision_tower(
+                mx_pixel_values,
+                output_hidden_states=True,
+            )
+            mx_input_ids = mx.array(input_ids.numpy())
+            mx_elected_image_feature = hidden_states[vision_feature_layer]
+            mx_image_features = self.mx_llava.multi_modal_projector(
+                mx_elected_image_feature
+            )
+            mx_inputs_embeds = self.mx_llava.language_model.model.embed_tokens(
+                mx_input_ids
+            )
+            mx_final_embedding = self.mx_llava._merge_input_ids_with_image_features(
+                mx_image_features, mx_inputs_embeds, mx_input_ids
+            )
+
+            hf_image_outputs = self.hf_llava.vision_tower(
+                hf_pixel_values, output_hidden_states=True
+            )
+            hf_elected_image_feature = hf_image_outputs.hidden_states[
+                vision_feature_layer
+            ]
+            hf_image_features = self.hf_llava.multi_modal_projector(
+                hf_elected_image_feature
+            )
+            hf_inputs_embeds = self.hf_llava.get_input_embeddings()(input_ids)
+            hf_final_embedding, _, _, _ = (
+                self.hf_llava._merge_input_ids_with_image_features(
+                    hf_image_features,
+                    hf_inputs_embeds,
+                    input_ids,
+                    attention_mask=input_ids,
+                    labels=torch.ones_like(input_ids),
+                )
+            )
+
+            self.assertTrue(
+                mx.allclose(
+                    mx_final_embedding,
+                    mx.array(hf_final_embedding.numpy()),
+                    atol=1e-1,
+                )
+            )
+
+    def test_generated_tokens(self):
+        raw_image = Image.open(requests.get(IMAGE_FILE, stream=True).raw)
+        with torch.no_grad():
+            hf_inputs = self.proc(PROMPT, raw_image, return_tensors="pt")
+            hf_outputs = self.hf_llava(**hf_inputs)
+            hf_logits = hf_outputs.logits
+
+            mx_inputs = self.proc(PROMPT, raw_image, return_tensors="np")
+            pixel_values = mx.array(mx_inputs["pixel_values"])
+            input_ids = mx.array(mx_inputs["input_ids"])
+
+            mx_logits, _ = self.mx_llava(input_ids, pixel_values)
+
+            self.assertTrue(
+                mx.allclose(
+                    mx_logits[:, -1, :].argmax(axis=-1),
+                    mx.array(hf_logits.numpy())[:, -1, :].argmax(axis=-1),
+                    atol=1e-2,
+                )
+            )
+
+
+if __name__ == "__main__":
+    unittest.main()

llava/vision.py

@@ -0,0 +1,223 @@
+# Copyright © 2024 Apple Inc.
+
+import inspect
+import math
+from dataclasses import dataclass
+from typing import Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+
+@dataclass
+class VisionConfig:
+    model_type: str
+    num_hidden_layers: int = 24
+    hidden_size: int = 1024
+    intermediate_size: int = 4096
+    num_attention_heads: int = 16
+    image_size: int = 336
+    patch_size: int = 14
+    projection_dim: int = 768
+    vocab_size: int = 32000
+    num_channels: int = 3
+    layer_norm_eps: float = 1e-5
+
+    @classmethod
+    def from_dict(cls, params):
+        return cls(
+            **{
+                k: v
+                for k, v in params.items()
+                if k in inspect.signature(cls).parameters
+            }
+        )
+
+
+class Attention(nn.Module):
+    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,
+        bias: bool = False,
+    ):
+        super().__init__()
+
+        if (dims % num_heads) != 0:
+            raise ValueError(
+                "The input feature dimensions should be divisible by the "
+                f"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.q_proj = nn.Linear(query_input_dims, dims, bias=bias)
+        self.k_proj = nn.Linear(key_input_dims, dims, bias=bias)
+        self.v_proj = nn.Linear(value_input_dims, value_dims, bias=bias)
+        self.out_proj = nn.Linear(value_dims, value_output_dims, bias=bias)
+
+    def __call__(self, queries, keys, values, mask=None):
+        queries = self.q_proj(queries)
+        keys = self.k_proj(keys)
+        values = self.v_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)
+
+        scale = math.sqrt(1 / queries.shape[-1])
+        scores = (queries * scale) @ keys
+        if mask is not None:
+            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)
+
+
+class MLP(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.activation_fn = nn.GELU(approx="fast")
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        x = self.activation_fn(self.fc1(x))
+        x = self.fc2(x)
+        return x
+
+
+class EncoderLayer(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.self_attn = Attention(
+            config.hidden_size, config.num_attention_heads, bias=True
+        )
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = MLP(config)
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+
+    def __call__(self, x: mx.array, mask: Optional[mx.array] = None) -> mx.array:
+        y = self.layer_norm1(x)
+        y = self.self_attn(y, y, y, mask)
+        x = x + y
+        y = self.layer_norm2(x)
+        y = self.mlp(y)
+        return x + y
+
+
+class Encoder(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.layers = [EncoderLayer(config) for _ in range(config.num_hidden_layers)]
+
+
+class VisionEmbeddings(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.class_embedding = mx.zeros((config.hidden_size,))
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=config.num_channels,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+            bias=False,
+        )
+
+        self.num_patches = (self.image_size // self.patch_size) ** 2
+        self.num_positions = self.num_patches + 1
+        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        batch_size = x.shape[0]
+        patch_embeddings = self.patch_embedding(x)
+        patch_embeddings = mx.flatten(patch_embeddings, start_axis=1, end_axis=2)
+        embed_dim = patch_embeddings.shape[-1]
+        cls_embeddings = mx.broadcast_to(
+            self.class_embedding, (batch_size, 1, embed_dim)
+        )
+        embeddings = mx.concatenate((cls_embeddings, patch_embeddings), axis=1)
+        embeddings += self.position_embedding.weight
+        return embeddings
+
+
+class ClipVisionModel(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.embeddings = VisionEmbeddings(config)
+        self.pre_layrnorm = nn.LayerNorm(config.hidden_size)
+        self.encoder = Encoder(config)
+        self.post_layernorm = nn.LayerNorm(config.hidden_size)
+
+    def __call__(
+        self,
+        x: mx.array,
+        output_hidden_states: Optional[bool] = None,
+    ) -> mx.array:
+        x = self.embeddings(x)
+        x = self.pre_layrnorm(x)
+
+        encoder_states = (x,) if output_hidden_states else None
+
+        for l in self.encoder.layers:
+            x = l(x, mask=None)
+            if output_hidden_states:
+                encoder_states = encoder_states + (x,)
+
+        pooler_output = self.post_layernorm(x[:, 0, :])
+        return pooler_output, x, encoder_states
+
+
+class VisionModel(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+
+        self.model_type = config.model_type
+        if self.model_type != "clip_vision_model":
+            raise ValueError(f"Unsupported model type: {self.model_type}")
+
+        self.vision_model = ClipVisionModel(config)
+
+    def __call__(
+        self, x: mx.array, output_hidden_states: Optional[bool] = None
+    ) -> mx.array:
+        return self.vision_model(x, output_hidden_states)
+
+    @staticmethod
+    def sanitize(weights):
+        sanitized_weights = {}
+        for k, v in weights.items():
+            if "position_ids" in k:
+                # Remove unused position_ids
+                continue
+            elif "patch_embedding.weight" in k:
+                # PyTorch conv2d weight tensors have shape:
+                #   [out_channels, in_channels, kH, KW]
+                # MLX conv2d expects the weight be of shape:
+                #   [out_channels, kH, KW, in_channels]
+                sanitized_weights[k] = v.transpose(0, 2, 3, 1)
+            else:
+                sanitized_weights[k] = v
+
+        return sanitized_weights