nexaai 1.0.19rc7__cp310-cp310-macosx_14_0_universal2.whl → 1.0.19rc8__cp310-cp310-macosx_14_0_universal2.whl

nexaai/binds/nexa_mlx/py-lib/vlm/modeling/models/idefics3/vision.py

@@ -1,233 +0,0 @@
-import inspect
-from dataclasses import dataclass
-from typing import Optional
-
-import mlx.core as mx
-import mlx.nn as nn
-import numpy as np
-
-
-@dataclass
-class VisionConfig:
-    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
-
-    @classmethod
-    def from_dict(cls, params):
-        return cls(
-            **{
-                k: v
-                for k, v in params.items()
-                if k in inspect.signature(cls).parameters
-            }
-        )
-
-
-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 = (self.image_size // self.patch_size) ** 2
-        self.num_positions = self.num_patches
-        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
-
-    def __call__(self, x: mx.array) -> mx.array:
-        patch_embeddings = self.patch_embedding(x)
-        patch_embeddings = mx.flatten(patch_embeddings, start_axis=1, end_axis=2)
-        position_ids = mx.array(mx.arange(self.num_positions)[None, :])
-        embeddings = patch_embeddings
-        embeddings += self.position_embedding(position_ids)
-        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,
-        output_hidden_states: Optional[bool] = None,
-    ) -> mx.array:
-        x = self.embeddings(x)
-        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

nexaai/binds/nexa_mlx/py-lib/vlm/modeling/models/internvl_chat/__init__.py

@@ -1,9 +0,0 @@
-from .internvl_chat import (
-    LanguageModel,
-    Model,
-    ModelConfig,
-    TextConfig,
-    VisionConfig,
-    VisionModel,
-)
-from .processor import InternVLChatProcessor, InternVLImageProcessor

nexaai/binds/nexa_mlx/py-lib/vlm/modeling/models/internvl_chat/internvl_chat.py

@@ -1,140 +0,0 @@
-import glob
-import inspect
-import json
-from dataclasses import dataclass
-from pathlib import Path
-from typing import List, Optional
-
-import mlx.core as mx
-import mlx.nn as nn
-import numpy as np
-from huggingface_hub import snapshot_download
-
-from ..base import pixel_shuffle
-from .language import LanguageModel, TextConfig
-from .vision import VisionConfig, VisionModel
-
-
-@dataclass
-class ModelConfig:
-    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
-
-    @classmethod
-    def from_dict(cls, params):
-        return cls(
-            **{
-                k: v
-                for k, v in params.items()
-                if k in inspect.signature(cls).parameters
-            }
-        )
-
-
-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,
-    ):
-
-        if pixel_values is None:
-            return 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 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_embddings = self.get_input_embeddings(input_ids, pixel_values)
-        logits = self.language_model(None, cache=cache, inputs_embeds=input_embddings)
-        return logits

nexaai/binds/nexa_mlx/py-lib/vlm/modeling/models/internvl_chat/language.py

@@ -1,220 +0,0 @@
-import inspect
-from dataclasses import dataclass
-from typing import Dict, Optional, Union
-
-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
-
-
-@dataclass
-class TextConfig:
-    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
-    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
-
-    @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, args: TextConfig):
-        super().__init__()
-
-        dim = args.hidden_size
-        self.n_heads = n_heads = args.num_attention_heads
-        assert args.num_key_value_heads is not None
-        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
-
-        self.head_dim = head_dim = args.hidden_size // n_heads
-        self.scale = head_dim**-0.5
-
-        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=True)
-        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=True)
-        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=True)
-        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=False)
-
-        self.rotary_emb = nn.RoPE(
-            head_dim,
-            base=args.rope_theta,
-            traditional=args.rope_traditional,
-        )
-
-    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, self.head_dim).transpose(
-            0, 2, 1, 3
-        )
-        keys = keys.reshape(B, L, self.n_kv_heads, self.head_dim).transpose(0, 2, 1, 3)
-        values = values.reshape(B, L, self.n_kv_heads, self.head_dim).transpose(
-            0, 2, 1, 3
-        )
-
-        offset = cache.offset if cache else 0
-
-        if mask is not None and isinstance(mask, mx.array):
-            mask = mask[..., : keys.shape[-2]]
-
-        queries = self.rotary_emb(queries, offset=offset)
-        keys = self.rotary_emb(keys, offset=offset)
-
-        if cache is not None:
-            keys, values = cache.update_and_fetch(keys, values)
-
-        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 Qwen2VLDecoderLayer(nn.Module):
-    def __init__(self, args: TextConfig):
-        super().__init__()
-        self.num_attention_heads = args.num_attention_heads
-        self.hidden_size = args.hidden_size
-        self.self_attn = Attention(args)
-        self.mlp = MLP(args.hidden_size, args.intermediate_size)
-        self.input_layernorm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
-        self.post_attention_layernorm = nn.RMSNorm(
-            args.hidden_size, eps=args.rms_norm_eps
-        )
-        self.args = args
-
-    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 Qwen2Model(nn.Module):
-    def __init__(self, args: TextConfig):
-        super().__init__()
-        self.args = args
-        self.vocab_size = args.vocab_size
-        self.num_hidden_layers = args.num_hidden_layers
-        assert self.vocab_size > 0
-        self.embed_tokens = nn.Embedding(args.vocab_size, args.hidden_size)
-        self.layers = [
-            Qwen2VLDecoderLayer(args=args) for _ in range(args.num_hidden_layers)
-        ]
-        self.norm = nn.RMSNorm(args.hidden_size, eps=args.rms_norm_eps)
-
-    def __call__(
-        self,
-        inputs: mx.array,
-        inputs_embeds: Optional[mx.array] = None,
-        mask: Optional[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:
-            mask = create_attention_mask(h, cache)
-
-        for layer, c in zip(self.layers, cache):
-            h = layer(h, mask, c)
-
-        return self.norm(h)
-
-
-class LanguageModel(nn.Module):
-    def __init__(self, args: TextConfig):
-        super().__init__()
-        self.args = args
-        self.model_type = args.model_type
-        self.model = Qwen2Model(args)
-
-        if not args.tie_word_embeddings:
-            self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
-
-    def __call__(
-        self,
-        inputs: mx.array,
-        inputs_embeds: Optional[mx.array] = None,
-        mask: Optional[mx.array] = None,
-        cache=None,
-    ):
-        out = self.model(inputs, cache=cache, inputs_embeds=inputs_embeds)
-        if self.args.tie_word_embeddings:
-            out = self.model.embed_tokens.as_linear(out)
-        else:
-            out = self.lm_head(out)
-        return LanguageModelOutput(logits=out)
-
-    @property
-    def layers(self):
-        return self.model.layers
-
-    @property
-    def head_dim(self):
-        return self.args.hidden_size // self.args.num_attention_heads
-
-    @property
-    def n_kv_heads(self):
-        return self.args.num_key_value_heads