nexaai 1.0.29__cp310-cp310-macosx_14_0_universal2.whl

nexaai/mlx_backend/vlm/modeling/models/gemma3/gemma3.py

@@ -0,0 +1,213 @@
+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 .language import LanguageModel, RMSNorm, 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 = 262144
+    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 Gemma3MultiModalProjector(nn.Module):
+    def __init__(self, config: ModelConfig):
+        super().__init__()
+        self.mm_input_projection_weight = mx.ones(
+            (config.vision_config.hidden_size, config.text_config.hidden_size)
+        )
+
+        self.mm_soft_emb_norm = RMSNorm(
+            config.vision_config.hidden_size, eps=config.vision_config.layer_norm_eps
+        )
+        self.patches_per_image = int(
+            config.vision_config.image_size // config.vision_config.patch_size
+        )
+        self.tokens_per_side = int(config.text_config.mm_tokens_per_image**0.5)
+        self.kernel_size = self.patches_per_image // self.tokens_per_side
+        self.avg_pool = nn.AvgPool2d(
+            kernel_size=self.kernel_size, stride=self.kernel_size
+        )
+
+    def __call__(self, x: mx.array) -> mx.array:
+        b, _, l = x.shape
+
+        reshaped_vision_outputs = x.transpose(0, 2, 1)
+        reshaped_vision_outputs = reshaped_vision_outputs.reshape(
+            b, l, self.patches_per_image, self.patches_per_image
+        )
+
+        # Transpose to place h, w in indices 1, 2
+        reshaped_vision_outputs = reshaped_vision_outputs.transpose(0, 2, 3, 1)
+        pooled_vision_outputs = self.avg_pool(reshaped_vision_outputs)
+        pooled_vision_outputs = pooled_vision_outputs.transpose(0, 3, 1, 2).flatten(2)
+        pooled_vision_outputs = pooled_vision_outputs.transpose(0, 2, 1)
+
+        normed_vision_outputs = self.mm_soft_emb_norm(pooled_vision_outputs)
+
+        projected_vision_outputs = mx.einsum(
+            "btm,md->btd", normed_vision_outputs, self.mm_input_projection_weight
+        )
+        return projected_vision_outputs.astype(x.dtype)
+
+
+def masked_scatter(
+    final_embedding: mx.array,
+    image_mask_expanded: mx.array,
+    scaled_image_features: mx.array,
+):
+    # Reshape the tensors to 1D
+    final_embedding_shape = final_embedding.shape
+    scaled_image_features_flattened = mx.flatten(scaled_image_features)
+    final_embedding_flattened = mx.flatten(final_embedding)
+    image_mask_expanded_flattened = mx.flatten(image_mask_expanded)
+
+    # Scatter the scaled image features into the special image token positions
+    image_positions = mx.array(np.where(image_mask_expanded_flattened)[0], mx.uint32)
+    final_embedding_flattened[image_positions] = scaled_image_features_flattened
+
+    # Reshape back to the original shape
+    final_embedding = mx.reshape(final_embedding_flattened, final_embedding_shape)
+
+    return final_embedding
+
+
+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 = Gemma3MultiModalProjector(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 = self.multi_modal_projector(hidden_state)
+
+        final_inputs_embeds, final_attention_mask_4d = (
+            self.prepare_inputs_for_multimodal(
+                self.config.hidden_size,
+                self.config.pad_token_id,
+                self.config.image_token_index,
+                image_features,
+                inputs_embeds,
+                input_ids,
+                mask,
+            )
+        )
+        return final_inputs_embeds, final_attention_mask_4d
+
+    @staticmethod
+    def prepare_inputs_for_multimodal(
+        hidden_size,
+        pad_token_id,
+        image_token_index,
+        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 / (hidden_size**0.5)
+        final_embedding = mx.zeros((batch_size, sequence_length, embed_dim))
+
+        pad_token_id = pad_token_id
+        pad_token_id = pad_token_id if pad_token_id is not None else 0
+        text_mask = (input_ids != image_token_index) & (input_ids != pad_token_id)
+        image_mask = input_ids == image_token_index
+        pad_mask = input_ids == 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)
+        image_mask_expanded = mx.expand_dims(image_mask, -1)
+        image_mask_expanded = mx.repeat(image_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
+        )
+
+        # insert image token embeddings
+        final_embedding = masked_scatter(
+            final_embedding, image_mask_expanded, scaled_image_features
+        )
+
+        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.astype(inputs_embeds.dtype), 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,
+        )
+        return logits

nexaai/mlx_backend/vlm/modeling/models/gemma3/language.py

@@ -0,0 +1,315 @@
+import inspect
+from dataclasses import dataclass
+from functools import partial
+from typing import Any, Dict, List, 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, RotatingKVCache
+
+
+@dataclass
+class TextConfig:
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int = 8
+    head_dim: int = 256
+    rms_norm_eps: float = 1.0e-6
+    vocab_size: int = 262208
+    num_key_value_heads: int = 4
+    rope_global_base_freq: float = 1_000_000.0
+    rope_local_base_freq: float = 10_000.0
+    rope_traditional: bool = False
+    query_pre_attn_scalar: float = 256
+    sliding_window: int = 1024
+    rope_scaling: Optional[Dict[str, Union[float, List[float]]]] = None
+    mm_tokens_per_image: int = 256
+    sliding_window_pattern: int = 6
+    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-5):
+        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, layer_idx: int):
+        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.repeats = n_heads // n_kv_heads
+        self.head_dim = head_dim = config.head_dim
+        self.layer_idx = layer_idx
+
+        self.scale = 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.q_norm = RMSNorm(dims=head_dim, eps=config.rms_norm_eps)
+        self.k_norm = RMSNorm(dims=head_dim, eps=config.rms_norm_eps)
+        self.is_sliding = (layer_idx + 1) % config.sliding_window_pattern != 0
+
+        self.rope = nn.RoPE(
+            head_dim,
+            traditional=config.rope_traditional,
+            base=(
+                config.rope_local_base_freq
+                if self.is_sliding
+                else config.rope_global_base_freq
+            ),
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+    ) -> mx.array:
+        B, L, _ = 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)
+
+        queries = self.q_norm(queries)
+        keys = self.k_norm(keys)
+
+        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)
+
+        # Sliding window
+        if mask is not None and isinstance(mask, mx.array):
+            if mask.shape[-1] != keys.shape[-2]:
+                mask = mask[..., -keys.shape[-2] :]
+
+        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:
+        # This should not be GELU approx, jax.nn.gelu
+        return self.down_proj(nn.gelu_approx(self.gate_proj(x)) * self.up_proj(x))
+
+
+@partial(mx.compile, shapeless=True)
+def clip_residual(x, y=None):
+    bound = mx.finfo(mx.float16).max
+    if y is None:
+        if x.dtype == mx.float16:
+            return mx.clip(x.astype(mx.float32), -bound, bound).astype(mx.float16)
+        else:
+            return x
+
+    if x.dtype != mx.float16:
+        return x + y
+
+    return mx.clip(x.astype(mx.float32) + y.astype(mx.float32), -bound, bound).astype(
+        mx.float16
+    )
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, config: TextConfig, layer_idx: int):
+        super().__init__()
+        self.num_attention_heads = config.num_attention_heads
+        self.hidden_size = config.hidden_size
+        self.self_attn = Attention(config, layer_idx)
+        self.mlp = MLP(config.hidden_size, config.intermediate_size)
+        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.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[Any] = None,
+    ) -> mx.array:
+
+        # Clip the input to avoid overflow in float16
+        # Float16 has a max value of 65504. When values exceed this limit, they become inf.
+        # Example: If x contains 70000.0 in float16, it becomes inf, causing gradient issues.
+        # We upcast to float32 for operations that might exceed the limit, then clip and
+        # convert back to float16 to maintain numerical stability.
+
+        # Clip input to avoid overflow in float16
+        x = clip_residual(x)
+
+        # Self-attention block
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = self.post_attention_layernorm(r)
+
+        # Add residual connection with overflow protection for float16
+        h = clip_residual(x + h)
+
+        # MLP block
+        r = self.mlp(self.pre_feedforward_layernorm(h))
+        out = self.post_feedforward_layernorm(r)
+
+        # Add residual connection with overflow protection for float16
+        out = clip_residual(h + out)
+
+        return out
+
+
+class Gemma3Model(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, layer_idx=layer_idx)
+            for layer_idx 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: mx.array = None,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        if inputs_embeds is None:
+            h = self.embed_tokens(inputs)
+        else:
+            h = inputs_embeds
+
+        h *= mx.array(self.config.hidden_size**0.5, mx.bfloat16).astype(h.dtype)
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        if mask is None:
+            j = self.config.sliding_window_pattern
+            full_mask = create_attention_mask(h, cache[j - 1 : j])
+            sliding_window_mask = create_attention_mask(h, cache)
+
+        for i, (layer, c) in enumerate(zip(self.layers, cache)):
+            is_global = (
+                i % self.config.sliding_window_pattern
+                == self.config.sliding_window_pattern - 1
+            )
+
+            local_mask = mask
+            if mask is None and is_global:
+                local_mask = full_mask
+            elif mask is None:
+                local_mask = sliding_window_mask
+
+            h = layer(h, local_mask, c)
+
+        return 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 = Gemma3Model(config)
+        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,
+    ):
+        out = self.model(inputs, inputs_embeds=inputs_embeds, mask=mask, cache=cache)
+        out = self.lm_head(out)
+        return LanguageModelOutput(logits=out)
+
+    def sanitize(self, weights):
+        if "lm_head.weight" not in weights:
+            weights["language_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.head_dim
+
+    @property
+    def n_kv_heads(self):
+        return self.config.num_key_value_heads
+
+    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