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

mlx_vlm/models/llava_bunny/config.py

@@ -0,0 +1,85 @@
+import inspect
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Union
+
+from ..base import BaseModelConfig
+
+
+@dataclass
+class TextConfig(BaseModelConfig):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    rms_norm_eps: float
+    vocab_size: int
+    attention_bias: bool = True
+    num_key_value_heads: int = None
+    rope_theta: float = 1000000
+    rope_traditional: bool = False
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+    max_position_embeddings: int = 4096
+    tie_word_embeddings: bool = True
+
+    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'")
+
+
+@dataclass
+class VisionConfig(BaseModelConfig):
+    model_type: str
+    num_hidden_layers: int = 27
+    hidden_size: int = 1152
+    intermediate_size: int = 4304
+    num_attention_heads: int = 16
+    image_size: int = 384
+    patch_size: int = 14
+    projection_dim: int = 768
+    vocab_size: int = 32000
+    num_channels: int = 3
+    layer_norm_eps: float = 1e-6
+
+
+@dataclass
+class ModelConfig(BaseModelConfig):
+    text_config: TextConfig
+    vision_config: VisionConfig
+    model_type: str
+    auto_map: dict
+    hidden_size: int
+    mm_hidden_size: int
+    mm_projector_type: str = "mlp2x_gelu"
+    ignore_index: int = -100
+    image_token_index: int = -200
+    vocab_size: int = 151936
+    eos_token_id: Optional[List[int]] = None
+
+    @classmethod
+    def from_dict(cls, params):
+        if not params.get("text_config", {}):
+            # Copy text config parameters from root level
+            excluded_keys = {"vision_config"}
+            params["text_config"] = dict(
+                filter(lambda x: x[0] not in excluded_keys, params.items())
+            )
+        if not params.get("vision_config", {}).get("model_type", {}):
+            # Set default model type
+            params["vision_config"]["model_type"] = "siglip_vision_model"
+
+        return cls(
+            **{
+                k: v
+                for k, v in params.items()
+                if k in inspect.signature(cls).parameters
+            }
+        )

mlx_vlm/models/llava_bunny/language.py

@@ -0,0 +1,194 @@
+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 .config import TextConfig
+
+
+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
+
+        head_dim = config.hidden_size // n_heads
+        self.scale = head_dim**-0.5
+
+        if hasattr(config, "attention_bias"):
+            attention_bias = config.attention_bias
+        else:
+            attention_bias = False
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=attention_bias)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=attention_bias)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=attention_bias)
+        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:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        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.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=None,
+    ) -> mx.array:
+        r = 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
+
+
+class Qwen2Model(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)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        inputs_embeds: Optional[mx.array] = None,
+        mask: Optional[mx.array] = None,
+        cache=None,
+    ):
+        # for passing merged input embeddings
+        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:
+            mask = create_attention_mask(h, cache)
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, c)
+
+        return self.lm_head(self.norm(h))
+
+
+class LanguageModel(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.config = config
+        self.model_type = config.model_type
+        self.model = Qwen2Model(config)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        inputs_embeds: Optional[mx.array] = None,
+        mask: Optional[mx.array] = None,
+        cache=None,
+    ):
+        out = self.model(inputs, mask=mask, cache=cache, inputs_embeds=inputs_embeds)
+        return LanguageModelOutput(logits=out)
+
+    def sanitize(self, weights):
+        if (
+            self.config.tie_word_embeddings
+            and "language_model.model.lm_head.weight" not in weights
+        ):
+            weights["language_model.model.lm_head.weight"] = weights[
+                "language_model.model.embed_tokens.weight"
+            ]
+
+        return {
+            k: v for k, v in weights.items() if "self_attn.rotary_emb.inv_freq" not in k
+        }
+
+    @property
+    def layers(self):
+        return self.model.layers
+
+    @property
+    def head_dim(self):
+        return self.config.hidden_size // self.config.num_attention_heads
+
+    @property
+    def n_kv_heads(self):
+        return self.config.num_key_value_heads

mlx_vlm/models/llava_bunny/llava_bunny.py

@@ -0,0 +1,217 @@
+import re
+from functools import partial, reduce
+from typing import Optional, Tuple
+
+import mlx.core as mx
+import mlx.nn as nn
+from PIL import Image
+from transformers.image_transforms import (
+    convert_to_rgb,
+    normalize,
+    rescale,
+    resize,
+    to_channel_dimension_format,
+)
+from transformers.image_utils import to_numpy_array
+
+from ..base import BaseImageProcessor, InputEmbeddingsFeatures
+from .config import ModelConfig, VisionConfig
+from .language import LanguageModel
+from .vision import VisionModel
+
+
+class ImageProcessor(BaseImageProcessor):
+    def preprocess(self, images):
+        if isinstance(images, Image.Image):
+            images = [images]
+        else:
+            assert isinstance(images, list)
+
+        transforms = [
+            convert_to_rgb,
+            to_numpy_array,
+            partial(
+                resize,
+                size=self.size,
+                resample=self.resample,
+                data_format=self.data_format,
+            ),
+            partial(rescale, scale=self.rescale_factor, data_format=self.data_format),
+            partial(
+                normalize,
+                mean=self.image_mean,
+                std=self.image_std,
+                data_format=self.data_format,
+            ),
+            partial(
+                to_channel_dimension_format,
+                channel_dim=self.data_format,
+                input_channel_dim=self.data_format,
+            ),
+        ]
+
+        images = reduce(lambda x, f: [*map(f, x)], transforms, images)
+
+        return images
+
+
+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 SigLipVisionTower(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.vision_tower = VisionModel(config)
+
+    def __call__(
+        self, x: mx.array, output_hidden_states: Optional[bool] = None
+    ) -> mx.array:
+        return self.vision_tower(x, output_hidden_states)
+
+
+class Model(nn.Module):
+    def __init__(self, config: ModelConfig):
+        super().__init__()
+        self.model_type = config.model_type
+        self.config = config
+
+        self.vision_tower = SigLipVisionTower(config.vision_config)
+        self.language_model = LanguageModel(config.text_config)
+        self.mm_projector = LlavaMultiModalProjector(config)
+
+    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)
+            )
+
+        inputs_embeds = self.language_model.model.embed_tokens(input_ids)
+
+        *_, hidden_state = self.vision_tower(
+            pixel_values.transpose(0, 2, 3, 1), output_hidden_states=True
+        )
+
+        image_features = hidden_state[-1].astype(pixel_values.dtype)
+        assert image_features.shape[-2] == 729
+
+        image_features = self.mm_projector(image_features)
+
+        final_inputs_embeds = self._prepare_inputs_for_multimodal(
+            image_features, inputs_embeds, input_ids
+        )
+        return InputEmbeddingsFeatures(inputs_embeds=final_inputs_embeds)
+
+    def _prepare_inputs_for_multimodal(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
+
+        batch_size, seq_length, embed_dim = inputs_embeds.shape
+        num_images, num_image_patches, _ = image_features.shape
+
+        # Positions of <image> tokens in input_ids for each batch
+        image_positions = mx.argmax(input_ids == image_token_index, axis=1)
+
+        final_embeddings = []
+        for b in range(batch_size):
+            text_segments = []
+            start_idx = 0
+            position = int(image_positions[b].item())
+
+            text_segments.append(inputs_embeds[b : b + 1, start_idx:position])
+            text_segments.append(image_features[b : b + 1])
+            text_segments.append(inputs_embeds[b : b + 1, position + 1 :])
+
+            batch_embeddings = mx.concatenate(text_segments, axis=1)
+            final_embeddings.append(batch_embeddings)
+
+        # Create a final embedding of shape
+        # (batch_size, num_image_patches + sequence_len, embed_dim)
+        return mx.concatenate(final_embeddings, axis=0)
+
+    @property
+    def layers(self):
+        return self.language_model.model.layers
+
+    def __call__(
+        self,
+        input_ids: mx.array,
+        pixel_values: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Tuple[mx.array, mx.array]] = None,
+        **kwargs,
+    ):
+        input_embeddings_features = self.get_input_embeddings(input_ids, pixel_values)
+        logits = self.language_model(
+            inputs=input_ids,
+            cache=cache,
+            inputs_embeds=input_embeddings_features.inputs_embeds,
+            mask=None,  # TODO: add mask
+        )
+        return logits
+
+    def sanitize(self, weights):
+        weights = {
+            (
+                f"{k.split('.', 1)[1]}"
+                if re.match(r"^model\.vision_tower", k)
+                else (
+                    f"mm_projector.linear_1.{k.split('.')[-1]}"
+                    if re.match(r"^model\.mm_projector\.0", k)
+                    else (
+                        f"mm_projector.linear_2.{k.split('.')[-1]}"
+                        if re.match(r"^model\.mm_projector\.2", k)
+                        else (
+                            f"language_model.model.{k}"
+                            if re.match(r"^lm_head", k)
+                            else (
+                                f"language_model.{k}"
+                                if re.match(r"^model\.(embed_tokens|norm|layers)", k)
+                                else k
+                            )
+                        )
+                    )
+                )
+            ): v
+            for k, v in weights.items()
+        }
+
+        weights = {
+            (
+                f"vision_tower.vision_tower.vision_model.head.attention.in_proj.bias"
+                if re.match(
+                    r"^vision_tower\.vision_tower\.vision_model\.head\.attention\.in_proj_bias",
+                    k,
+                )
+                else (
+                    f"vision_tower.vision_tower.vision_model.head.attention.in_proj.weight"
+                    if re.match(
+                        r"^vision_tower\.vision_tower\.vision_model\.head\.attention\.in_proj_weight",
+                        k,
+                    )
+                    else k
+                )
+            ): v
+            for k, v in weights.items()
+        }
+
+        return weights