fount-vlm-nell-02 0.3.11__py3-none-any.whl

mlx_vlm/lora.py

@@ -0,0 +1,207 @@
+import argparse
+import json
+import logging
+import os
+
+import mlx.optimizers as optim
+from datasets import load_dataset
+from tqdm import tqdm
+
+from .prompt_utils import apply_chat_template
+from .trainer import Dataset, Trainer, save_adapter
+from .trainer.utils import apply_lora_layers, find_all_linear_names, get_peft_model
+from .utils import load, load_image_processor
+
+logging.basicConfig(level=logging.INFO)
+logger = logging.getLogger(__name__)
+
+
+def custom_print(*args, **kwargs):
+    tqdm.write(" ".join(map(str, args)), **kwargs)
+
+
+def main(args):
+    logger.info(f"\033[32mLoading model from {args.model_path}\033[0m")
+    model, processor = load(
+        args.model_path, processor_config={"trust_remote_code": True}
+    )
+    config = model.config.__dict__
+    image_processor = load_image_processor(args.model_path)
+
+    logger.info(f"\033[32mLoading dataset from {args.dataset}\033[0m")
+    dataset = load_dataset(args.dataset, split=args.split)
+
+    if "messages" not in dataset.column_names:
+        raise ValueError("Dataset must have a 'messages' column")
+    if "images" not in dataset.column_names:
+        raise ValueError("Dataset must have an 'images' column")
+
+    if args.apply_chat_template:
+        logger.info(f"\033[32mApplying chat template to the dataset\033[0m")
+
+        def process_data(examples):
+            if config["model_type"] == "pixtral":
+                conversations = apply_chat_template(
+                    config=config,
+                    processor=processor,
+                    prompt=examples["messages"],
+                    return_messages=True,
+                )
+                examples["messages"] = [
+                    json.dumps(item, ensure_ascii=False) for item in conversations
+                ]
+            else:
+                examples["messages"] = apply_chat_template(
+                    config=config,
+                    processor=processor,
+                    prompt=examples["messages"],
+                    return_messages=True,
+                )
+            return examples
+
+        dataset = dataset.map(process_data)
+
+    dataset = Dataset(
+        dataset,
+        config,
+        processor,
+        image_processor=image_processor,
+        image_resize_shape=args.image_resize_shape,
+    )
+
+    adapter_path = args.adapter_path
+    if adapter_path:
+        logger.info(f"\033[32mResuming from adapter path {adapter_path}\033[0m")
+        logger.info(
+            f"\033[32mLora rank, alpha, and dropout will be loaded from adapter_config.json file\033[0m"
+        )
+
+        model = apply_lora_layers(model, adapter_path)
+
+    else:
+        logger.info(f"\033[32mSetting up LoRA\033[0m")
+
+        list_of_modules = find_all_linear_names(model.language_model)
+        model = get_peft_model(
+            model,
+            list_of_modules,
+            rank=args.lora_rank,
+            alpha=args.lora_alpha,
+            dropout=args.lora_dropout,
+        )
+
+    logger.info(f"\033[32mSetting up optimizer\033[0m")
+    optimizer = optim.Adam(learning_rate=args.learning_rate)
+
+    logger.info(f"\033[32mSetting up trainer\033[0m")
+    trainer = Trainer(model, optimizer)
+
+    model.train()
+
+    # Training loop
+    logger.info(f"\033[32mTraining model\033[0m")
+    for epoch in range(args.epochs):
+        if args.steps == 0:
+            args.steps = len(dataset) // args.batch_size
+
+        progress_bar = tqdm(range(args.steps), position=0, leave=True)
+        for i in progress_bar:
+            loss = trainer.train_step(
+                dataset[i * args.batch_size : (i + 1) * args.batch_size]
+            )
+            # Update progress bar
+            progress_bar.update(1)
+            progress_bar.set_postfix(
+                {"Epoch": epoch, "Step": i, "Loss": f"{loss.item():.4f}"}
+            )
+
+            if i % args.print_every == 0:
+                # Log additional information
+                custom_print(
+                    {
+                        "Epoch": epoch,
+                        "Step": i,
+                        "Loss": f"{loss.item():.4f}",
+                    }
+                )
+        # Save the interim adapter after each epoch except the last.
+        if args.save_after_epoch and (epoch < (args.epochs - 1)):
+            head, tail = os.path.split(args.output_path)
+            save_adapter(model, head + os.sep + "epoch_" + str(epoch) + "_" + tail)
+
+    # Save the adapter
+    save_adapter(model, args.output_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="Train NanoLLaVA model")
+    parser.add_argument(
+        "--model-path",
+        type=str,
+        default="mlx-community/Qwen2-VL-2B-Instruct-bf16",
+        help="Path to the pre-trained model",
+    )
+    parser.add_argument(
+        "--dataset", type=str, required=True, help="Path to the dataset"
+    )
+    parser.add_argument(
+        "--split", type=str, default="train", help="Split to use for training"
+    )
+    parser.add_argument(
+        "--image-resize-shape",
+        type=int,
+        nargs=2,
+        default=None,
+        help="Resize images to this shape",
+    )
+    parser.add_argument(
+        "--apply-chat-template",
+        action="store_false",
+        help="Apply chat template to the dataset",
+    )
+    parser.add_argument(
+        "--learning-rate",
+        type=float,
+        default=1e-4,
+        help="Learning rate for the optimizer",
+    )
+    parser.add_argument(
+        "--batch-size", type=int, default=1, help="Batch size for training"
+    )
+    parser.add_argument(
+        "--epochs", type=int, default=1, help="Number of epochs to train"
+    )
+    parser.add_argument(
+        "--steps", type=int, default=0, help="Number of steps per epoch"
+    )
+    parser.add_argument(
+        "--print-every", type=int, default=10, help="Print loss every n steps"
+    )
+    parser.add_argument(
+        "--lora-alpha",
+        type=float,
+        default=0.1,
+        help="LoRA scaling factor (alpha / rank)",
+    )
+    parser.add_argument("--lora-rank", type=int, default=10, help="LoRA rank")
+    parser.add_argument("--lora-dropout", type=float, default=0.1, help="LoRA dropout")
+    parser.add_argument(
+        "--output-path",
+        type=str,
+        default="adapters",
+        help="Path to save the trained adapter",
+    )
+    parser.add_argument(
+        "--adapter-path",
+        type=str,
+        default=None,
+        help="Load path to resume training from a previously saved adapter",
+    )
+    parser.add_argument(
+        "--save-after-epoch",
+        action="store_true",
+        help="Save interim versions of adapter files after each epoch",
+    )
+
+    args = parser.parse_args()
+    main(args)

mlx_vlm/models/aya_vision/__init__.py

@@ -0,0 +1,2 @@
+from .aya_vision import LanguageModel, Model, VisionModel
+from .config import ModelConfig, TextConfig, VisionConfig

mlx_vlm/models/aya_vision/aya_vision.py

@@ -0,0 +1,188 @@
+from typing import Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+import numpy as np
+
+from ..base import InputEmbeddingsFeatures
+from .config import ModelConfig
+from .language import LanguageModel
+from .vision import VisionModel
+
+
+class AyaVisionMultiModalProjector(nn.Module):
+    def __init__(self, config: ModelConfig):
+        super().__init__()
+        self.config = config
+        self.downsample_factor = config.downsample_factor
+        self.alignment_intermediate_size = getattr(
+            config, "alignment_intermediate_size", config.text_config.hidden_size
+        )
+        if config.model_type == "aya_vision":
+            self.layernorm = nn.LayerNorm(
+                config.vision_config.hidden_size * (config.downsample_factor**2),
+                eps=config.adapter_layer_norm_eps,
+            )
+
+        self.linear_1 = nn.Linear(
+            config.vision_config.hidden_size * (config.downsample_factor**2),
+            self.alignment_intermediate_size,
+            bias=True,
+        )
+
+        self.act = nn.SiLU()  # SwiGLU uses SiLU activation
+
+        # For SwiGLU, project down to half size since we split intermediate dim
+        self.linear_2 = nn.Linear(
+            self.alignment_intermediate_size // 2,
+            config.text_config.hidden_size,
+            bias=True,
+        )
+
+    def __call__(self, image_features):
+        image_features = self.pixel_shuffle(image_features)
+        if self.config.model_type == "aya_vision":
+            image_features = self.layernorm(image_features)
+        hidden_states = self.linear_1(image_features)
+
+        # Split along last dimension and apply SwiGLU
+        x, gate = mx.split(hidden_states, 2, axis=-1)
+        hidden_states = self.act(gate) * x
+
+        hidden_states = self.linear_2(hidden_states)
+        return hidden_states
+
+    def pixel_shuffle(self, image_features):  # B, S, D
+        batch_size, seq_length, feature_dim = image_features.shape
+        height = width = int(seq_length**0.5)
+        image_features = image_features.reshape(
+            image_features.shape[0], width, height, -1
+        )
+        channels = image_features.shape[-1]
+        image_features = image_features.reshape(
+            batch_size,
+            width,
+            int(height / self.downsample_factor),
+            int(channels * self.downsample_factor),
+        )
+        image_features = image_features.transpose(0, 2, 1, 3)
+        image_features = image_features.reshape(
+            batch_size,
+            int(height / self.downsample_factor),
+            int(width / self.downsample_factor),
+            -1,
+        )
+        image_features = image_features.transpose(0, 2, 1, 3)
+        return image_features
+
+
+class Model(nn.Module):
+    def __init__(self, config: ModelConfig):
+        super().__init__()
+        self.config = config
+        self.vision_tower = VisionModel(config.vision_config)
+        self.language_model = LanguageModel(config.text_config)
+        self.multi_modal_projector = AyaVisionMultiModalProjector(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,
+        **kwargs,
+    ):
+        if pixel_values is None:
+            return InputEmbeddingsFeatures(
+                inputs_embeds=self.language_model.model.embed_tokens(input_ids)
+            )
+
+        # Get the input embeddings from the language model
+        inputs_embeds = self.language_model.model.embed_tokens(input_ids)
+
+        spatial_shapes = kwargs.get("spatial_shapes", None)
+        # Get the ouptut hidden states from the vision model
+        *_, hidden_states = self.vision_tower(
+            pixel_values.transpose(0, 2, 3, 1),
+            spatial_shapes=spatial_shapes,
+            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 InputEmbeddingsFeatures(inputs_embeds=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
+
+        # Positions of <image> tokens in input_ids, assuming batch size is 1
+        image_positions = np.where(input_ids[0] == image_token_index)[0].tolist()
+        num_images, _, _, vision_hidden_size = image_features.shape
+
+        reshaped_image_hidden_states = image_features.reshape(-1, vision_hidden_size)
+
+        # cast to the dtype of the input_embeds to support quantized models
+        reshaped_image_hidden_states = reshaped_image_hidden_states.astype(
+            inputs_embeds.dtype
+        )
+        inputs_embeds[:, image_positions, :] = reshaped_image_hidden_states
+        return inputs_embeds
+
+    @property
+    def layers(self):
+        return self.language_model.model.layers
+
+    def __call__(
+        self,
+        input_ids: mx.array,
+        pixel_values: mx.array,
+        mask: mx.array,
+        cache=None,
+        **kwargs,
+    ):
+
+        input_embeddings_features = self.get_input_embeddings(
+            input_ids, pixel_values, **kwargs
+        )
+        logits = self.language_model(
+            input_ids,
+            cache=cache,
+            inputs_embeds=input_embeddings_features.inputs_embeds,
+        )
+        return logits
+
+    def sanitize(self, weights):
+        def transform_key(key):
+            if "model.vision_tower" in key:
+                key = key.replace("model.vision_tower", "vision_tower")
+            if "model.multi_modal_projector" in key:
+                key = key.replace(
+                    "model.multi_modal_projector", "multi_modal_projector"
+                )
+            if "model.language_model" in key:
+                key = key.replace("model.language_model", "language_model.model")
+            if "lm_head" in key and not key.startswith("language_model"):
+                key = key.replace("lm_head", "language_model.lm_head")
+            return key
+
+        return {transform_key(k): v for k, v in weights.items()}

mlx_vlm/models/aya_vision/config.py

@@ -0,0 +1,52 @@
+from dataclasses import dataclass
+from typing import List, Optional
+
+from ..base import BaseModelConfig
+
+
+@dataclass
+class TextConfig(BaseModelConfig):
+    model_type: str
+    hidden_size: int = 8192
+    head_dim: int = 128
+    num_hidden_layers: int = 40
+    intermediate_size: int = 14336
+    num_attention_heads: int = 64
+    num_key_value_heads: int = 8
+    rope_theta: float = 50000.0
+    vocab_size: int = 256000
+    layer_norm_eps: float = 1e-05
+    logit_scale: float = 0.0625
+    attention_bias: bool = False
+    layer_norm_bias: bool = False
+    sliding_window: int = 4096
+    sliding_window_pattern: int = 4
+    max_position_embeddings: int = 4096
+
+
+@dataclass
+class VisionConfig(BaseModelConfig):
+    model_type: str
+    hidden_size: int
+    num_attention_heads: int
+    patch_size: int
+    num_hidden_layers: int = 12
+    intermediate_size: int = 3072
+    image_size: int = 224
+    num_channels: int = 3
+    layer_norm_eps: float = 1e-6
+
+
+@dataclass
+class ModelConfig(BaseModelConfig):
+    text_config: TextConfig
+    vision_config: VisionConfig
+    model_type: str
+    image_token_index: int = 255036
+    max_splits_per_img: int = 12
+    downsample_factor: int = 2
+    alignment_intermediate_size: int = 28672
+    adapter_layer_norm_eps: float = 1e-06
+    vision_feature_layer: int = -1
+    vision_feature_select_strategy: str = "full"
+    eos_token_id: Optional[List[int]] = None

mlx_vlm/models/aya_vision/language.py

@@ -0,0 +1,202 @@
+from typing import Optional, Tuple
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from ..base import (
+    LanguageModelOutput,
+    create_attention_mask,
+    scaled_dot_product_attention,
+)
+from ..cache import KVCache, RotatingKVCache
+from .config import TextConfig
+
+
+class Attention(nn.Module):
+    def __init__(self, config: TextConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+
+        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.head_dim = head_dim = config.head_dim
+        if (head_dim * n_heads) != dim:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {dim}"
+                f" and `num_heads`: {n_heads})."
+            )
+        self.scale = head_dim**-0.5
+
+        attetion_bias = config.attention_bias
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=attetion_bias)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=attetion_bias)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=attetion_bias)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=attetion_bias)
+
+        self.rope = nn.RoPE(head_dim, traditional=True, base=config.rope_theta)
+
+        self.use_sliding_window = (layer_idx + 1) % config.sliding_window_pattern != 0
+
+    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)
+
+        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)
+
+        # Apply RoPE only if sliding window is enabled
+        if self.use_sliding_window:
+            if cache is None:
+                queries = self.rope(queries)
+                keys = self.rope(keys)
+            else:
+                queries = self.rope(queries, offset=cache.offset)
+                keys = self.rope(keys, offset=cache.offset)
+
+        if cache is not None:
+            keys, values = cache.update_and_fetch(keys, values)
+
+        if self.use_sliding_window and mask is not None and isinstance(mask, mx.array):
+            key_len = keys.shape[-2]
+            if mask.shape[-1] != key_len:
+                mask = mask[..., -key_len:]
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache, scale=self.scale, mask=mask
+        )
+
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class MLP(nn.Module):
+    def __init__(self, dim, hidden_dim):
+        super().__init__()
+        self.gate_proj = nn.Linear(dim, hidden_dim, bias=False)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=False)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=False)
+
+    def __call__(self, x):
+        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, config: TextConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.n_heads = config.num_attention_heads
+
+        self.self_attn = Attention(config, layer_idx)
+        self.mlp = MLP(config.hidden_size, config.intermediate_size)
+        self.input_layernorm = nn.LayerNorm(
+            config.hidden_size, eps=config.layer_norm_eps, bias=config.layer_norm_bias
+        )
+        self.config = config
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Tuple[mx.array, mx.array]] = None,
+    ) -> mx.array:
+        h = self.input_layernorm(x)
+        attn_h = self.self_attn(h, mask, cache)
+        ff_h = self.mlp(h)
+        return attn_h + ff_h + x
+
+
+class CohereModel(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, layer_idx=i)
+            for i in range(config.num_hidden_layers)
+        ]
+        self.norm = nn.LayerNorm(
+            config.hidden_size, eps=config.layer_norm_eps, bias=config.layer_norm_bias
+        )
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        inputs_embeds: mx.array = None,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        if inputs_embeds is None:
+            h = self.embed_tokens(inputs)
+        else:
+            h = inputs_embeds
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        if mask is None:
+            j = self.config.sliding_window_pattern
+            mask = create_attention_mask(h, cache[j - 1 : j])
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, c)
+
+        return self.norm(h)
+
+
+class LanguageModel(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.model_type = config.model_type
+        self.model = CohereModel(config)
+        self.config = config
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        inputs_embeds: mx.array = None,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        out = self.model(inputs, inputs_embeds, mask, cache)
+        out = self.model.embed_tokens.as_linear(out)
+        out = out * self.model.config.logit_scale
+        return LanguageModelOutput(logits=out)
+
+    def make_cache(self):
+        caches = []
+        for i in range(self.config.num_hidden_layers):
+            if (
+                i % self.config.sliding_window_pattern
+                == self.config.sliding_window_pattern - 1
+            ):
+                caches.append(KVCache())
+            else:
+                caches.append(
+                    RotatingKVCache(max_size=self.config.sliding_window, keep=0)
+                )
+        return caches
+
+    @property
+    def layers(self):
+        return self.model.layers
+
+    @property
+    def head_dim(self):
+        return self.model.config.head_dim
+
+    @property
+    def n_kv_heads(self):
+        return self.model.config.num_key_value_heads