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

mlx_vlm/models/gemma3/gemma3.py

@@ -0,0 +1,194 @@
+from typing import Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+import numpy as np
+
+from ..base import InputEmbeddingsFeatures
+from .config import ModelConfig
+from .language import LanguageModel, RMSNorm
+from .vision import VisionModel
+
+
+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,
+        **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).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 InputEmbeddingsFeatures(
+            inputs_embeds=final_inputs_embeds, attention_mask_4d=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_features = self.get_input_embeddings(
+            input_ids, pixel_values, mask
+        )
+        inputs_embeds = input_embeddings_features.inputs_embeds
+        attention_mask = input_embeddings_features.attention_mask_4d
+
+        logits = self.language_model(
+            inputs=input_ids,
+            cache=cache,
+            inputs_embeds=inputs_embeds,
+            mask=attention_mask,
+        )
+        return logits

mlx_vlm/models/gemma3/language.py

@@ -0,0 +1,293 @@
+from functools import partial
+from typing import Any, 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, RotatingKVCache
+from .config import TextConfig
+
+
+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.window_size = config.sliding_window
+        self.sliding_window_pattern = config.sliding_window_pattern
+        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:
+            global_mask = create_attention_mask(
+                h, cache[self.sliding_window_pattern - 1]
+            )
+
+            if self.sliding_window_pattern > 1:
+                sliding_window_mask = create_attention_mask(
+                    h,
+                    cache[0],
+                    window_size=self.window_size,
+                )
+            else:
+                sliding_window_mask = None
+
+        for i, (layer, c) in enumerate(zip(self.layers, cache)):
+            is_global = (
+                i % self.sliding_window_pattern == self.sliding_window_pattern - 1
+            )
+
+            local_mask = mask
+            if mask is None and is_global:
+                local_mask = global_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,
+        **kwargs,
+    ):
+        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

mlx_vlm/models/gemma3/vision.py

@@ -0,0 +1,215 @@
+from typing import Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+import numpy as np
+
+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 = (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 not in ["siglip_vision_model", "gemma3", "gemma3_vision"]:
+            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 "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/gemma3n/__init__.py

@@ -0,0 +1,2 @@
+from .config import AudioConfig, ModelConfig, TextConfig, VisionConfig
+from .gemma3n import AudioModel, LanguageModel, Model, VisionModel