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

mlx_vlm/models/idefics3/language.py

@@ -0,0 +1,157 @@
+from typing import Optional
+
+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
+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
+        head_dim = config.hidden_size // n_heads
+        self.n_kv_heads = n_kv_heads = config.num_key_value_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)
+
+        self.rope = nn.RoPE(
+            head_dim,
+            traditional=config.rope_traditional,
+            base=config.rope_theta,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[KVCache] = 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: Optional[KVCache] = 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 LanguageModel(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.config = config
+        self.model_type = config.model_type
+        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,
+        **kwargs,
+    ):
+        # for passing merged input embeddings
+        if inputs_embeds is None:
+            h = self.embed_tokens(inputs)
+        else:
+            h = inputs_embeds.astype(self.norm.weight.dtype)
+
+        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)
+
+        logits = self.lm_head(self.norm(h))
+        return LanguageModelOutput(logits=logits)
+
+    def sanitize(self, 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
+        }
+
+    @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/idefics3/vision.py

@@ -0,0 +1,265 @@
+from typing import Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .config import VisionConfig
+
+
+def check_array_shape(arr):
+    shape = arr.shape
+
+    # Check if the shape has 4 dimensions
+    if len(shape) != 4:
+        return False
+
+    out_channels, kH, KW, _ = shape
+
+    # Check if out_channels is the largest, and kH and KW are the same
+    if (out_channels >= kH) and (out_channels >= KW) and (kH == KW):
+        return True
+    else:
+        return False
+
+
+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 = True,
+    ):
+        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
+        head_dim = dims // num_heads
+        self.scale = head_dim**-0.5
+
+        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, x, mask=None):
+        queries = self.q_proj(x)
+        keys = self.k_proj(x)
+        values = self.v_proj(x)
+
+        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, 1, 3)
+        values = values.reshape(B, S, num_heads, -1).transpose(0, 2, 1, 3)
+
+        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.out_proj(output)
+
+
+class MLP(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.activation_fn = nn.GELU(approx="precise")
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size, bias=True)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size, bias=True)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        x = self.fc1(x)
+        x = self.activation_fn(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:
+        r = self.self_attn(self.layer_norm1(x), mask)
+        h = x + r
+        r = self.mlp(self.layer_norm2(h))
+        return h + r
+
+
+class Encoder(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.layers = [EncoderLayer(config) for _ in range(config.num_hidden_layers)]
+
+    def __call__(
+        self,
+        x: mx.array,
+        output_hidden_states: Optional[bool] = None,
+        mask: Optional[mx.array] = None,
+    ) -> mx.array:
+        encoder_states = (x,) if output_hidden_states else None
+        h = x
+        for l in self.layers:
+            x = l(x, mask=mask)
+            if output_hidden_states:
+                encoder_states = encoder_states + (x,)
+
+            h = x
+
+        return (h, encoder_states)
+
+
+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.patch_embedding = nn.Conv2d(
+            in_channels=config.num_channels,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+        )
+
+        self.num_patches_per_side = self.image_size // self.patch_size
+        self.num_patches = self.num_patches_per_side**2
+        self.num_positions = self.num_patches
+        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
+
+    def __call__(self, x: mx.array, patch_attention_mask: mx.array = None) -> mx.array:
+        batch_size, max_im_h, max_im_w, _ = x.shape
+        patch_embeds = self.patch_embedding(x)
+        embeddings = mx.flatten(patch_embeds, start_axis=1, end_axis=2)
+
+        seq_len = embeddings.shape[1]
+
+        if patch_attention_mask is None:
+            position_ids = mx.tile(mx.arange(seq_len), (batch_size, 1))
+        else:
+            boundaries = mx.arange(
+                1 / self.num_patches_per_side, 1.0, 1 / self.num_patches_per_side
+            )
+
+            # Flatten mask to match sequence length (handles both (B,H,W) and (B,H,W,1))
+            if patch_attention_mask.ndim == 4:
+                flat_mask = patch_attention_mask.squeeze(-1).reshape(batch_size, -1)[
+                    :, :seq_len
+                ]
+            else:
+                flat_mask = patch_attention_mask.reshape(batch_size, -1)[:, :seq_len]
+
+            # Compute valid patches per image (channels-last indexing)
+            nb_patches_h = mx.maximum(patch_attention_mask[:, :, 0].sum(axis=1), 1)
+            nb_patches_w = mx.maximum(patch_attention_mask[:, 0, :].sum(axis=1), 1)
+
+            position_ids = mx.zeros((batch_size, seq_len), dtype=mx.int32)
+
+            for batch_idx in range(batch_size):
+                nb_h = int(nb_patches_h[batch_idx])
+                nb_w = int(nb_patches_w[batch_idx])
+
+                # Compute fractional coordinates
+                fractional_h = mx.arange(nb_h, dtype=mx.float32) / nb_h
+                fractional_w = mx.arange(nb_w, dtype=mx.float32) / nb_w
+                fractional_h = mx.clip(fractional_h, a_min=0.0, a_max=1.0 - 1e-6)
+                fractional_w = mx.clip(fractional_w, a_min=0.0, a_max=1.0 - 1e-6)
+
+                # Bucket into position IDs
+                bucket_h = mx.sum(fractional_h[:, None] >= boundaries[None, :], axis=1)
+                bucket_w = mx.sum(fractional_w[:, None] >= boundaries[None, :], axis=1)
+
+                # Create 2D grid: iterate over height, then width (row-major)
+                pos_ids = (
+                    bucket_h[:, None] * self.num_patches_per_side + bucket_w[None, :]
+                ).reshape(-1)
+
+                valid_len = min(pos_ids.shape[0], seq_len)
+                position_ids[batch_idx, :valid_len] = pos_ids[:valid_len]
+
+            # Zero out position embeddings for padding
+            mask_expanded = flat_mask[:, :, None]  # (batch, seq_len, 1)
+
+        pos_embeddings = self.position_embedding(position_ids)
+
+        # Apply mask to zero out padding position embeddings
+        if patch_attention_mask is not None:
+            pos_embeddings = pos_embeddings * mask_expanded
+
+        embeddings = embeddings + pos_embeddings
+        return embeddings
+
+
+class VisionModel(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.model_type = config.model_type
+        if self.model_type not in [
+            "siglip_vision_model",
+            "idefics3",
+            "idefics3_vision",
+            "smolvlm_vision",
+        ]:
+            raise ValueError(f"Unsupported model type: {self.model_type}")
+
+        self.embeddings = VisionEmbeddings(config)
+        self.encoder = Encoder(config)
+        self.post_layernorm = nn.LayerNorm(config.hidden_size)
+
+    def __call__(
+        self,
+        x: mx.array,
+        patch_attention_mask: Optional[mx.array] = None,
+        output_hidden_states: Optional[bool] = None,
+    ) -> mx.array:
+        x = self.embeddings(x, patch_attention_mask)
+        x = x.astype(self.embeddings.patch_embedding.weight.dtype)
+        encoder_outputs = self.encoder(
+            x=x, output_hidden_states=output_hidden_states, mask=None
+        )
+        pooler_output = self.post_layernorm(encoder_outputs[0])
+        return pooler_output, x, encoder_outputs[-1]
+
+    def sanitize(self, 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]
+                if check_array_shape(v):
+                    sanitized_weights[k] = v
+                else:
+                    sanitized_weights[k] = v.transpose(0, 2, 3, 1)
+            else:
+                sanitized_weights[k] = v
+
+        return sanitized_weights

mlx_vlm/models/internvl_chat/__init__.py

@@ -0,0 +1,3 @@
+from .config import ModelConfig, TextConfig, VisionConfig
+from .internvl_chat import LanguageModel, Model, VisionModel
+from .processor import InternVLChatProcessor, InternVLImageProcessor

mlx_vlm/models/internvl_chat/config.py

@@ -0,0 +1,89 @@
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Union
+
+from ..base import BaseModelConfig
+
+
+@dataclass
+class VisionConfig(BaseModelConfig):
+    model_type: str
+    hidden_size: int = 1024
+    num_attention_heads: int = 16
+    patch_size: int = 14
+    num_hidden_layers: int = 24
+    intermediate_size: int = 4096
+    image_size: int = 448
+    num_channels: int = 3
+    layer_norm_eps: float = 1e-6
+    drop_path_rate: float = 0.1
+    qkv_bias: bool = True
+    qk_normalization: bool = False
+    norm_type: str = "layer_norm"
+
+    @classmethod
+    def from_dict(cls, params):
+        def normalize_dim(v):
+            if isinstance(v, (list, tuple)):
+                if len(v) > 0:
+                    try:
+                        return int(v[0])
+                    except Exception:
+                        return v[0]
+            try:
+                return int(v)
+            except Exception:
+                return v
+
+        p = dict(params)
+        if "image_size" in p:
+            p["image_size"] = normalize_dim(p["image_size"])
+        if "patch_size" in p:
+            p["patch_size"] = normalize_dim(p["patch_size"])
+
+        import inspect as _inspect
+
+        return cls(
+            **{k: v for k, v in p.items() if k in _inspect.signature(cls).parameters}
+        )
+
+
+@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
+    max_window_layers: int
+    hidden_act: str
+    num_key_value_heads: Optional[int] = 8
+    head_dim: Optional[int] = None
+    max_position_embeddings: Optional[int] = 40960
+    rope_theta: float = 1000000.0
+    rope_traditional: bool = False
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+    tie_word_embeddings: bool = False
+    sliding_window: int = 32768
+    use_sliding_window: bool = False
+    use_cache: bool = True
+
+    def __post_init__(self):
+        if self.num_key_value_heads is None:
+            self.num_key_value_heads = self.num_attention_heads
+
+
+@dataclass
+class ModelConfig(BaseModelConfig):
+    text_config: TextConfig
+    vision_config: VisionConfig
+    model_type: str
+    ignore_index: int = -100
+    image_token_index: int = 151667
+    video_token_index: int = 151656
+    vision_feature_select_strategy: str = "default"
+    vision_feature_layer: int = -1
+    vocab_size: int = 32000
+    downsample_ratio: float = 0.5
+    eos_token_id: Optional[List[int]] = None

mlx_vlm/models/internvl_chat/internvl_chat.py

@@ -0,0 +1,115 @@
+from typing import Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+import numpy as np
+
+from ..base import InputEmbeddingsFeatures, pixel_shuffle
+from .config import ModelConfig
+from .language import LanguageModel
+from .vision import VisionModel
+
+
+class Model(nn.Module):
+    def __init__(self, config: ModelConfig):
+        super().__init__()
+        self.config = config
+        self.vision_model = VisionModel(config.vision_config)
+        self.language_model = LanguageModel(config.text_config)
+
+        self.downsample_ratio = config.downsample_ratio
+
+        vit_hidden_size = self.config.vision_config.hidden_size
+        llm_hidden_size = self.config.text_config.hidden_size
+
+        self.mlp1 = [
+            nn.LayerNorm(vit_hidden_size * int(1 / self.downsample_ratio) ** 2),
+            nn.Linear(
+                vit_hidden_size * int(1 / self.downsample_ratio) ** 2, llm_hidden_size
+            ),
+            nn.GELU(),
+            nn.Linear(llm_hidden_size, llm_hidden_size),
+        ]
+
+    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)
+            )
+
+        dtype = self.vision_model.embeddings.patch_embedding.weight.dtype
+        pixel_values = pixel_values.astype(dtype)
+
+        # TODO: Remove this after transformers implementation is merged
+        if pixel_values.ndim == 5:
+            pixel_values = pixel_values[0]
+
+        # 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_model(
+            pixel_values.transpose(0, 2, 3, 1), output_hidden_states=True
+        )
+
+        # Extract vision embeddings, removing the class token (first token)
+        hidden_states = hidden_states[:, 1:, :]
+
+        # Apply pixel shuffle with downsampling
+        hidden_states = pixel_shuffle(
+            hidden_states, shuffle_ratio=self.downsample_ratio
+        )
+
+        # Apply MLP transformation
+        for layer in self.mlp1:
+            hidden_states = layer(hidden_states)
+
+        # Insert special image tokens in the input_ids
+        final_inputs_embeds = self._merge_input_ids_with_image_features(
+            hidden_states, 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
+    ):
+        B, N, C = inputs_embeds.shape
+        image_token_index = self.config.image_token_index
+        video_token_index = self.config.video_token_index
+
+        # Positions of <image> tokens in input_ids, assuming batch size is 1
+        image_positions = input_ids == image_token_index
+        if mx.sum(image_positions) == 0:
+            image_positions = input_ids == video_token_index
+
+        image_indices = np.where(image_positions)[1].tolist()
+
+        image_features = image_features.reshape(-1, image_features.shape[-1])
+
+        inputs_embeds[:, image_indices, :] = image_features
+
+        return inputs_embeds.reshape(B, N, C)
+
+    @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)
+        logits = self.language_model(
+            None, cache=cache, inputs_embeds=input_embeddings_features.inputs_embeds
+        )
+        return logits