nexaai 1.0.29__cp310-cp310-macosx_14_0_universal2.whl

nexaai/mlx_backend/vlm/modeling/models/paligemma/language.py

@@ -0,0 +1,282 @@
+import inspect
+from dataclasses import dataclass
+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
+
+
+@dataclass
+class TextConfig:
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    num_key_value_heads: int
+    vocab_size: int
+    head_dim: int = 256
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 10000
+    rope_traditional: bool = False
+    attn_logit_softcapping: Optional[float] = None
+    final_logit_softcapping: Optional[float] = None
+    query_pre_attn_scalar: Optional[float] = None
+    max_position_embeddings: int = 4096
+
+    @classmethod
+    def from_dict(cls, params):
+        return cls(
+            **{
+                k: v
+                for k, v in params.items()
+                if k in inspect.signature(cls).parameters
+            }
+        )
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, dims: int, eps: float = 1e-6):
+        super().__init__()
+        self.weight = mx.ones((dims,))
+        self.eps = eps
+
+    def __call__(self, x):
+        return mx.fast.rms_norm(x, 1.0 + self.weight, self.eps)
+
+
+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.model_type = config.model_type
+        self.attn_logit_softcapping = config.attn_logit_softcapping
+        self.repeats = n_heads // n_kv_heads
+        self.head_dim = head_dim = (
+            config.hidden_size // n_heads
+            if self.model_type == "gemma"
+            else config.head_dim
+        )
+        self.scale = (
+            head_dim**-0.5
+            if self.model_type == "gemma"
+            else 1.0 / (config.query_pre_attn_scalar**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)
+
+        if self.model_type == "gemma":
+            output = scaled_dot_product_attention(
+                queries, keys, values, cache, scale=self.scale, mask=mask
+            )
+        else:
+            queries = queries * self.scale
+
+            if self.repeats > 1:
+                queries = queries.reshape(
+                    B, self.n_kv_heads, self.repeats, L, self.head_dim
+                )
+                keys = mx.expand_dims(keys, 2)
+                values = mx.expand_dims(values, 2)
+
+            scores = queries @ keys.swapaxes(-1, -2)
+            scores = mx.tanh(scores / self.attn_logit_softcapping)
+            scores *= self.attn_logit_softcapping
+
+            if mask is not None and isinstance(mask, mx.array):
+                scores = scores + mask
+            scores = mx.softmax(scores, precise=True, axis=-1)
+            output = scores @ values
+            if self.repeats > 1:
+                output = output.reshape(B, self.n_heads, L, self.head_dim)
+
+        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, model_type):
+        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)
+        self.gelu = nn.GELU() if model_type == "gemma" else nn.GELU(approx="precise")
+
+    def __call__(self, x) -> mx.array:
+        return self.down_proj(self.gelu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.model_type = config.model_type
+        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, config.model_type)
+        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+        self.config = config
+
+        if config.model_type == "gemma2":
+            self.pre_feedforward_layernorm = RMSNorm(
+                config.hidden_size, eps=config.rms_norm_eps
+            )
+            self.post_feedforward_layernorm = RMSNorm(
+                config.hidden_size, eps=config.rms_norm_eps
+            )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[KVCache] = None,
+    ) -> mx.array:
+        # Self attention block
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+
+        if self.model_type == "gemma":
+            # Gemma: Post-attention residual connection then MLP
+            h = x + r
+            r = self.mlp(self.post_attention_layernorm(h))
+            out = h + r
+        else:
+            # Gemma2: Normalized residual connections with pre/post norms
+            h = x + self.post_attention_layernorm(r)
+            r = self.mlp(self.pre_feedforward_layernorm(h))
+            out = h + self.post_feedforward_layernorm(r)
+        return out
+
+
+class GemmaModel(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 = RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    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
+
+        h *= self.config.hidden_size**0.5
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        if mask is None or cache[0].offset > 0:
+            mask = create_attention_mask(h, cache, return_array=True)
+
+        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.config = config
+        self.final_logit_softcapping = config.final_logit_softcapping
+        self.model_type = config.model_type
+        self.model = GemmaModel(config)
+
+        if self.model_type not in ["gemma", "gemma2"]:
+            raise ValueError(
+                f"Model type {self.model_type} not supported. Currently only 'gemma' is supported"
+            )
+
+    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)
+        out = self.model.embed_tokens.as_linear(out)
+
+        if self.model_type == "gemma2":
+            out = mx.tanh(out / self.final_logit_softcapping)
+            out = out * self.final_logit_softcapping
+        return LanguageModelOutput(logits=out)
+
+    def sanitize(self, weights):
+        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
+            if self.model_type == "gemma"
+            else self.config.head_dim
+        )
+
+    @property
+    def n_kv_heads(self):
+        return self.config.num_key_value_heads

nexaai/mlx_backend/vlm/modeling/models/paligemma/paligemma.py

@@ -0,0 +1,160 @@
+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
+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
+    model_type: str
+    vocab_size: int = 257152
+    ignore_index: int = -100
+    image_token_index: int = 257152
+    hidden_size: int = 2048
+    pad_token_id: int = 0
+    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 PaliGemmaMultiModalProjector(nn.Module):
+    def __init__(self, config: ModelConfig):
+        super().__init__()
+        self.linear = nn.Linear(
+            config.vision_config.hidden_size,
+            config.vision_config.projection_dim,
+            bias=True,
+        )
+
+    def __call__(self, x: mx.array) -> mx.array:
+        output = self.linear(x)
+        return output
+
+
+class Model(nn.Module):
+    def __init__(self, config: ModelConfig):
+        super().__init__()
+        self.model_type = config.model_type
+        self.config = config
+
+        self.vision_tower = VisionModel(config.vision_config)
+        self.language_model = LanguageModel(config.text_config)
+        self.multi_modal_projector = PaliGemmaMultiModalProjector(config)
+
+    def get_input_embeddings(
+        self,
+        input_ids: Optional[mx.array] = None,
+        pixel_values: Optional[mx.array] = None,
+        mask: Optional[mx.array] = None,
+    ):
+        if pixel_values is None:
+            return self.language_model.model.embed_tokens(input_ids), None
+
+        inputs_embeds = self.language_model.model.embed_tokens(input_ids)
+
+        hidden_state, _, _ = self.vision_tower(
+            pixel_values.transpose(0, 2, 3, 1).astype(inputs_embeds.dtype),
+            output_hidden_states=True,
+        )
+
+        image_features = hidden_state[None, :].astype(pixel_values.dtype)
+        image_features = self.multi_modal_projector(image_features)
+
+        final_inputs_embeds, final_attention_mask_4d = (
+            self._prepare_inputs_for_multimodal(
+                image_features, inputs_embeds, input_ids, mask
+            )
+        )
+        return final_inputs_embeds, final_attention_mask_4d
+
+    def _prepare_inputs_for_multimodal(
+        self, image_features, inputs_embeds, input_ids, attention_mask
+    ):
+        _, _, embed_dim = image_features.shape
+
+        batch_size, sequence_length = input_ids.shape
+        scaled_image_features = image_features / (self.config.hidden_size**0.5)
+        final_embedding = mx.zeros((batch_size, sequence_length, embed_dim))
+
+        text_mask = (input_ids != self.config.image_token_index) & (
+            input_ids != self.config.pad_token_id
+        )
+        image_mask = input_ids == self.config.image_token_index
+        pad_mask = input_ids == self.config.pad_token_id
+
+        # expand masks to match embedding dimension
+        text_mask_expanded = mx.expand_dims(text_mask, -1)
+        text_mask_expanded = mx.repeat(text_mask_expanded, embed_dim, axis=-1)
+        pad_mask_expanded = mx.expand_dims(pad_mask, -1)
+        pad_mask_expanded = mx.repeat(pad_mask_expanded, embed_dim, axis=-1)
+
+        # insert padding and text token embeddings
+        final_embedding = mx.where(text_mask_expanded, inputs_embeds, final_embedding)
+        final_embedding = mx.where(
+            pad_mask_expanded, mx.zeros_like(final_embedding), final_embedding
+        )
+        pad_size = final_embedding.shape[1] - scaled_image_features.shape[1]
+        scaled_image_features = mx.pad(
+            scaled_image_features, ((0, 0), (0, pad_size), (0, 0))
+        )
+        # insert image embeddings - the image mask is always less or equal to the sentence in length
+        image_mask_expanded = mx.expand_dims(image_mask, -1)
+        image_mask_expanded = mx.repeat(image_mask_expanded, embed_dim, axis=-1)
+        final_embedding = mx.where(
+            image_mask_expanded, scaled_image_features, final_embedding
+        )
+
+        final_embedding = mx.where(
+            pad_mask_expanded, mx.zeros_like(final_embedding), final_embedding
+        )
+
+        attention_mask_expanded_1 = mx.expand_dims(attention_mask, 1)
+        attention_mask_expanded_2 = mx.expand_dims(attention_mask, 2)
+        final_attention_mask_4d = attention_mask_expanded_1 * attention_mask_expanded_2
+        final_attention_mask_4d = final_attention_mask_4d
+        final_attention_mask_4d = mx.expand_dims(final_attention_mask_4d, 1)
+        final_embedding = mx.array(final_embedding)
+        return final_embedding, final_attention_mask_4d
+
+    @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[mx.array] = None,
+        **kwargs,
+    ):
+        input_embeddings, final_attention_mask_4d = self.get_input_embeddings(
+            input_ids, pixel_values, mask
+        )
+
+        logits = self.language_model(
+            inputs=input_ids,
+            cache=cache,
+            inputs_embeds=input_embeddings,
+            mask=final_attention_mask_4d,
+        )
+        return logits

nexaai/mlx_backend/vlm/modeling/models/paligemma/vision.py

@@ -0,0 +1,242 @@
+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
+    num_hidden_layers: int
+    hidden_size: int
+    intermediate_size: int
+    num_attention_heads: int
+    patch_size: int
+    projection_dim: int
+    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[0]
+
+        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(np.arange(self.num_positions)[None, :])
+        embeddings = patch_embeddings
+        embeddings += self.position_embedding(position_ids)
+        return embeddings
+
+
+class SigLipVisionModel(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        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)
+
+        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]
+
+
+class VisionModel(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.model_type = config.model_type
+        if self.model_type != "siglip_vision_model":
+            raise ValueError(f"Unsupported model type: {self.model_type}")
+
+        self.vision_model = SigLipVisionModel(config)
+
+    def __call__(
+        self, x: mx.array, output_hidden_states: Optional[bool] = None
+    ) -> mx.array:
+        return self.vision_model(x, output_hidden_states)
+
+    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/mlx_backend/vlm/modeling/models/phi3_v/__init__.py

@@ -0,0 +1,8 @@
+from .phi3_v import (
+    LanguageModel,
+    Model,
+    ModelConfig,
+    TextConfig,
+    VisionConfig,
+    VisionModel,
+)