mlx-vlm 0.0.1__tar.gz

mlx_vlm-0.0.1/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright © 2023 Apple Inc.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

mlx_vlm-0.0.1/MANIFEST.in

@@ -0,0 +1,2 @@
+include ./requirements.txt
+recursive-include mlx_vlm/ *.py

mlx_vlm-0.0.1/PKG-INFO

@@ -0,0 +1,49 @@
+Metadata-Version: 2.1
+Name: mlx-vlm
+Version: 0.0.1
+Summary: Vision LLMs on Apple silicon with MLX and the Hugging Face Hub
+Home-page: https://github.com/Blaizzy/mlx-vlm
+Author: MLX Contributors
+Author-email: prince.gdt@gmail.com
+License: MIT
+Requires-Python: >=3.8
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Requires-Dist: mlx>=0.8.0
+Requires-Dist: mlx-lm>=0.4.0
+Requires-Dist: numpy
+Requires-Dist: transformers
+Requires-Dist: torch
+Requires-Dist: huggingface_hub
+Requires-Dist: Pillow
+Requires-Dist: requests
+
+# MLX-VLM
+
+MLX-VLM a package for running Vision LLMs on your Mac using MLX.
+
+
+## Get started
+
+The easiest way to get started is to install the `mlx-vlm` package:
+
+**With `pip`**:
+
+```sh
+pip install mlx-vlm
+```
+
+**Inference**:
+
+```sh
+python -m mlx_vlm.generate --model mlx-community/llava-1.5-7b-hf-4bit --max-tokens 10 --temp 0.0
+```
+
+**Convert models**:
+
+```sh
+pip -m mlx_vlm.convert
+    --hf-path  llava-hf/llava-1.5-7b-hf\
+    -q \
+    --upload-repo mlx-community/llava-1.5-7b-hf-4bit
+```

mlx_vlm-0.0.1/README.md

@@ -0,0 +1,29 @@
+# MLX-VLM
+
+MLX-VLM a package for running Vision LLMs on your Mac using MLX.
+
+
+## Get started
+
+The easiest way to get started is to install the `mlx-vlm` package:
+
+**With `pip`**:
+
+```sh
+pip install mlx-vlm
+```
+
+**Inference**:
+
+```sh
+python -m mlx_vlm.generate --model mlx-community/llava-1.5-7b-hf-4bit --max-tokens 10 --temp 0.0
+```
+
+**Convert models**:
+
+```sh
+pip -m mlx_vlm.convert
+    --hf-path  llava-hf/llava-1.5-7b-hf\
+    -q \
+    --upload-repo mlx-community/llava-1.5-7b-hf-4bit
+```

mlx_vlm-0.0.1/mlx_vlm/__init__.py

@@ -0,0 +1,2 @@
+from .utils import convert, generate, load
+from .version import __version__

mlx_vlm-0.0.1/mlx_vlm/convert.py

@@ -0,0 +1,62 @@
+# Copyright © 2023-2024 Apple Inc.
+
+import argparse
+
+from .utils import convert
+
+
+def configure_parser() -> argparse.ArgumentParser:
+    """
+    Configures and returns the argument parser for the script.
+
+    Returns:
+        argparse.ArgumentParser: Configured argument parser.
+    """
+    parser = argparse.ArgumentParser(
+        description="Convert Hugging Face model to MLX format"
+    )
+
+    parser.add_argument("--hf-path", type=str, help="Path to the Hugging Face model.")
+    parser.add_argument(
+        "--mlx-path", type=str, default="mlx_model", help="Path to save the MLX model."
+    )
+    parser.add_argument(
+        "-q", "--quantize", help="Generate a quantized model.", action="store_true"
+    )
+    parser.add_argument(
+        "--q-group-size", help="Group size for quantization.", type=int, default=64
+    )
+    parser.add_argument(
+        "--q-bits", help="Bits per weight for quantization.", type=int, default=4
+    )
+    parser.add_argument(
+        "--dtype",
+        help="Type to save the parameters, ignored if -q is given.",
+        type=str,
+        choices=["float16", "bfloat16", "float32"],
+        default="float16",
+    )
+    parser.add_argument(
+        "--upload-repo",
+        help="The Hugging Face repo to upload the model to.",
+        type=str,
+        default=None,
+    )
+    parser.add_argument(
+        "-d",
+        "--dequantize",
+        help="Dequantize a quantized model.",
+        action="store_true",
+        default=False,
+    )
+    return parser
+
+
+def main():
+    parser = configure_parser()
+    args = parser.parse_args()
+    convert(**vars(args))
+
+
+if __name__ == "__main__":
+    main()

mlx_vlm-0.0.1/mlx_vlm/generate.py

@@ -0,0 +1,93 @@
+import argparse
+import codecs
+
+import mlx.core as mx
+from .utils import load, load_config, load_image_processor, generate, get_model_path
+
+
+MODEL_TYPE = ""
+def parse_arguments():
+    parser = argparse.ArgumentParser(
+        description="Generate text from an image using a model."
+    )
+    parser.add_argument(
+        "--model",
+        type=str,
+        default="qnguyen3/nanoLLaVA",
+        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="What are these?",
+        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 get_model_and_processors(model_path):
+    model_path = get_model_path(model_path)
+    model, processor = load(model_path, {"trust_remote_code": True})
+    config = load_config(model_path)
+    image_processor = load_image_processor(config)
+    return model, processor, image_processor
+
+
+def sample(logits, temperature=0.0):
+    if temperature == 0:
+        return mx.argmax(logits, axis=-1)
+    else:
+        return mx.random.categorical(logits * (1 / temperature))
+
+
+
+def main():
+    args = parse_arguments()
+    model, processor, image_processor = get_model_and_processors(args.model)
+
+    prompt = codecs.decode(args.prompt, "unicode_escape")
+
+    if "chat_template" in processor.__dict__.keys():
+        prompt = processor.apply_chat_template(
+            [{"role": "user", "content": f"<image>\n{prompt}"}],
+            tokenize=False,
+            add_generation_prompt=True,
+        )
+    elif "tokenizer" in processor.__dict__.keys():
+        prompt = processor.tokenizer.apply_chat_template(
+            [{"role": "user", "content": f"<image>\n{prompt}"}],
+            tokenize=False,
+            add_generation_prompt=True,
+        )
+    else:
+        ValueError("Error: processor does not have 'chat_template' or 'tokenizer' attribute.")
+
+    generate(
+        model,
+        processor,
+        args.image,
+        prompt,
+        image_processor,
+        args.temp,
+        args.max_tokens,
+        True
+    )
+
+
+if __name__ == "__main__":
+    main()

mlx_vlm-0.0.1/mlx_vlm/models/base.py

@@ -0,0 +1,32 @@
+
+from abc import ABC, abstractmethod
+from typing import Dict
+
+from transformers.image_processing_utils import get_size_dict
+from transformers.image_utils import (ChannelDimension, PILImageResampling )
+
+
+class BaseImageProcessor(ABC):
+    def __init__(self,
+                 image_mean=(0.5, 0.5, 0.5),
+                 image_std=(0.5, 0.5, 0.5),
+                 size=(384, 384),
+                 crop_size: Dict[str, int] = None,
+                 resample=PILImageResampling.BICUBIC,
+                 rescale_factor=1 / 255,
+                 data_format=ChannelDimension.FIRST):
+        crop_size = crop_size if crop_size is not None else {"height": 384, "width": 384}
+        crop_size = get_size_dict(crop_size, default_to_square=True, param_name="crop_size")
+
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.size = size
+        self.resample = resample
+        self.rescale_factor = rescale_factor
+        self.data_format = data_format
+        self.crop_size = crop_size
+
+    @abstractmethod
+    def preprocess(self, images):
+        pass
+

mlx_vlm-0.0.1/mlx_vlm/models/llava/__init__.py

@@ -0,0 +1 @@
+from .llava import Model, ModelConfig, VisionConfig, TextConfig, VisionModel, LanguageModel

mlx_vlm-0.0.1/mlx_vlm/models/llava/language.py

@@ -0,0 +1,211 @@
+
+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 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 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)
+
+        output = mx.fast.scaled_dot_product_attention(
+            queries, keys, values, scale=self.scale, mask=mask
+        )
+        output = output.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 = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = nn.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 = nn.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
+        }

mlx_vlm-0.0.1/mlx_vlm/models/llava/llava.py

@@ -0,0 +1,179 @@
+
+
+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 ModelConfig:
+    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: ModelConfig):
+        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 Model(nn.Module):
+    def __init__(self, config: ModelConfig):
+        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 = ModelConfig.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 = Model(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