nexaai 1.0.29__cp310-cp310-macosx_14_0_universal2.whl

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

@@ -0,0 +1,465 @@
+import inspect
+from dataclasses import dataclass, field
+from typing import List, Optional, Tuple
+
+import mlx.core as mx
+import mlx.nn as nn
+
+
+@dataclass
+class VisionConfig:
+    model_type: str = "molmo"
+    num_channels: int = 3
+    image_default_input_size: Tuple[int, int] = (336, 336)
+    image_patch_size: int = 14
+    image_pos_patch_size: int = 14
+    hidden_size: int = 18944
+    image_emb_dim: int = 1024
+    image_num_heads: int = 16
+    image_num_key_value_heads: int = 16
+    image_num_layers: int = 23
+    image_head_dim: int = 64
+    image_mlp_dim: int = 4096
+    image_mlp_activations: str = "gelu"
+    image_dropout_rate: float = 0.0
+    image_num_pos: int = 577
+    image_norm_eps: float = 1e-5
+    attention_dropout: float = 0.0
+    residual_dropout: float = 0.0
+    initializer_range: float = 0.02
+    d_model: int = 3584
+    image_pooling_h: int = 2
+    image_pooling_w: int = 2
+    vit_layers: Optional[List[int]] = field(default_factory=lambda: [-2, -9])
+    image_pooling_2d: str = "attention-meanq"
+    image_padding_embed: str = "pad_and_partial_pad"
+    intermediate_size: Optional[int] = None
+
+    def __post_init__(self):
+        if self.intermediate_size is None:
+            self.intermediate_size = self.image_patch_size * self.image_patch_size * 3
+
+    @property
+    def image_num_patch(self):
+        h, w = self.image_default_input_size
+        return h // self.image_patch_size, w // self.image_patch_size
+
+    @property
+    def llm_patches_per_crop(self):
+        h, w = self.image_num_patch
+        # Round up in case we need to pad the image features for pooling
+        h = (h + self.image_pooling_h - 1) // self.image_pooling_h
+        w = (w + self.image_pooling_w - 1) // self.image_pooling_w
+        return h, w
+
+    @classmethod
+    def from_dict(cls, params):
+        return cls(
+            **{
+                k: v
+                for k, v in params.items()
+                if k in inspect.signature(cls).parameters
+            }
+        )
+
+
+class MLP(nn.Module):
+    def __init__(self, config: VisionConfig, input_dim: int):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size
+        self.w1 = nn.Linear(
+            input_dim,
+            self.hidden_size,
+            bias=False,
+        )
+        self.w2 = nn.Linear(
+            self.hidden_size,
+            config.d_model,
+            bias=False,
+        )
+        self.w3 = nn.Linear(
+            input_dim,
+            self.hidden_size,
+            bias=False,
+        )
+
+    def __call__(self, x: mx.array) -> mx.array:
+        x = self.w2(nn.silu(self.w1(x)) * self.w3(x))
+        return x
+
+
+class ViTMLP(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.config = config
+        self.w1 = nn.Linear(config.image_emb_dim, config.image_mlp_dim, bias=True)
+        self.w2 = nn.Linear(config.image_mlp_dim, config.image_emb_dim, bias=True)
+        self.act = nn.GELU(approx="fast")
+
+    def __call__(self, x: mx.array) -> mx.array:
+        x = self.w1(x)
+        x = self.act(x)
+        x = self.w2(x)
+        return x
+
+
+class MultiHeadDotProductAttention(nn.Module):
+    def __init__(self, config: VisionConfig, is_vit_layer: Optional[bool] = True):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.image_emb_dim
+        self.num_heads = config.image_num_heads
+        self.head_dim = config.image_head_dim
+        self.num_key_value_heads = config.image_num_key_value_heads
+        self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+        self.scale = self.head_dim**-0.5
+        self.is_vit_layer = is_vit_layer
+
+        n_layers = (
+            1 if (is_vit_layer or config.vit_layers is None) else len(config.vit_layers)
+        )
+
+        self.wq = nn.Linear(
+            n_layers * self.embed_dim, self.num_heads * self.head_dim, bias=True
+        )
+        self.wk = nn.Linear(
+            n_layers * self.embed_dim,
+            self.num_key_value_heads * self.head_dim,
+            bias=True,
+        )
+        self.wv = nn.Linear(
+            n_layers * self.embed_dim,
+            self.num_key_value_heads * self.head_dim,
+            bias=True,
+        )
+        self.wo = nn.Linear(self.num_heads * self.head_dim, self.embed_dim, bias=True)
+
+    def _split_heads(self, hidden_states, num_heads) -> mx.array:
+        return hidden_states.reshape(
+            hidden_states.shape[:2] + (num_heads, self.head_dim)
+        )
+
+    def _merge_heads(self, hidden_states) -> mx.array:
+        return hidden_states.reshape(hidden_states.shape[:2] + (self.embed_dim,))
+
+    def __call__(self, x: mx.array, kv: mx.array = None) -> mx.array:
+        batch_size, seq_len, _ = x.shape
+
+        if kv is None:
+            k = x
+            v = x
+        else:
+            k = kv
+            v = kv
+        q = self._split_heads(self.wq(x), self.num_heads).transpose(0, 2, 1, 3)
+
+        k = self._split_heads(self.wk(k), self.num_key_value_heads).transpose(
+            0, 2, 1, 3
+        )
+        v = self._split_heads(self.wv(v), self.num_key_value_heads).transpose(
+            0, 2, 1, 3
+        )
+
+        attn = mx.fast.scaled_dot_product_attention(q, k, v, scale=self.scale)
+        out = attn.transpose(0, 2, 1, 3)
+        out = self._merge_heads(out)
+        out = self.wo(out)
+        return out
+
+
+class ResidualAttentionBlock(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.config = config
+        self.attention = MultiHeadDotProductAttention(config)
+        self.feed_forward = ViTMLP(config)
+        self.attention_norm = nn.LayerNorm(
+            config.image_emb_dim, eps=config.image_norm_eps
+        )
+        self.ffn_norm = nn.LayerNorm(config.image_emb_dim, eps=config.image_norm_eps)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        x = x + self.attention(self.attention_norm(x))
+        x = x + self.feed_forward(self.ffn_norm(x))
+        return x
+
+
+class ResidualAttentionBlocks(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.resblocks = [
+            ResidualAttentionBlock(config) for _ in range(config.image_num_layers)
+        ]
+
+    def __call__(self, x: mx.array) -> mx.array:
+        h = []
+        for block in self.resblocks:
+            x = block(x)
+            h.append(x)
+        return h
+
+
+def _expand_token(token, batch_size: int):
+    return mx.broadcast_to(
+        mx.reshape(token, (1, 1, -1)), (batch_size, 1, token.shape[-1])
+    )
+
+
+def pad_to_multiple(x, target_size, pad_mode="edge", pad_value=0):
+    """
+    Pad the last dimension of input tensor to match target size.
+
+    Args:
+        x: Input tensor with shape [..., D]
+        target_size: Desired size for the last dimension
+        pad_mode: Padding mode ('constant', 'reflect', etc.)
+        pad_value: Value to use for constant padding
+
+    Returns:
+        Padded tensor with shape [..., target_size]
+    """
+    current_size = x.shape[-1]
+
+    # Return early if no padding needed
+    if current_size == target_size:
+        return x
+
+    # Ensure target size is larger
+    if current_size > target_size:
+        raise ValueError(
+            f"Current size {current_size} is larger than target size {target_size}"
+        )
+
+    # Calculate padding needed
+    pad_size = target_size - current_size
+
+    # Create padding configuration
+    # No padding for batch and channel dimensions (0,0), only pad the last dim
+    pad_config = [(0, 0)] * (len(x.shape) - 1) + [(0, pad_size)]
+
+    return mx.pad(x, pad_width=pad_config, mode=pad_mode, constant_values=pad_value)
+
+
+class VisionTransformer(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.config = config
+        self.class_embedding = mx.zeros((config.image_emb_dim,))
+        self.positional_embedding = mx.zeros(
+            (config.image_num_pos, config.image_emb_dim)
+        )
+        self.patch_embedding = nn.Linear(
+            config.intermediate_size,
+            config.image_emb_dim,
+            bias=False,
+        )
+        self.pre_ln = nn.LayerNorm(config.image_emb_dim, eps=config.image_norm_eps)
+        self.transformer = ResidualAttentionBlocks(config)
+
+    def add_pos_emb(self, x: mx.array, patch_num: int) -> mx.array:
+        cls_emb = self.positional_embedding[0:1]
+        pos_emb = self.positional_embedding[1:]
+
+        # Reshape into 2D grid
+        pos_emb_size = int(pos_emb.shape[0] ** 0.5)
+        pos_emb = mx.reshape(pos_emb, (pos_emb_size, pos_emb_size, pos_emb.shape[1]))
+
+        (patch_num_0, patch_num_1) = patch_num
+
+        if pos_emb.shape[0] != patch_num_0 or pos_emb.shape[1] != patch_num_1:
+            # Reshape for upsampling (add batch and channel dims)
+            pos_emb = mx.expand_dims(pos_emb, 0)
+            pos_emb = mx.transpose(pos_emb, (0, 3, 1, 2))
+
+            # Create and apply upsampler
+            upsampler = nn.Upsample(
+                scale_factor=(
+                    patch_num_0 / pos_emb.shape[2],
+                    patch_num_1 / pos_emb.shape[3],
+                ),
+                mode="linear",  # MLX doesn't have bicubic, using linear as closest alternative
+                align_corners=False,
+            )
+            pos_emb = upsampler(pos_emb)
+
+            # Restore original dimensions
+            pos_emb = mx.transpose(pos_emb, (0, 2, 3, 1))
+            pos_emb = mx.squeeze(pos_emb, 0)
+
+        pos_emb = mx.reshape(pos_emb, (-1, pos_emb.shape[-1]))
+
+        # Expand cls_emb and pos_emb
+        expanded_cls = cls_emb[None, :, :]
+        expanded_pos = pos_emb[None, :, :]
+
+        # Concatenate and add to x
+        pos_embedding = mx.concatenate([expanded_cls, expanded_pos], axis=1)
+        x = x + pos_embedding
+        return x
+
+    def __call__(self, x: mx.array, patch_num: int = None) -> List[mx.array]:
+        """
+        : param x: (batch_size, num_patch, n_pixels)
+        """
+        if patch_num is None:
+            patch_num = self.config.image_num_patch
+        B, N, D = x.shape
+
+        # (Optional) Due to quantization, pad around the image to match intermediate_size
+        x = pad_to_multiple(x, self.config.intermediate_size)
+
+        x = self.patch_embedding(x)
+
+        # class embeddings and positional embeddings
+        expanded_class_emb = _expand_token(self.class_embedding, x.shape[0])
+        expanded_class_emb = expanded_class_emb
+
+        x = mx.concatenate([expanded_class_emb, x], axis=1)
+        x = self.add_pos_emb(x, patch_num)
+
+        x = self.pre_ln(x)
+
+        hidden_states = self.transformer(x)
+        return hidden_states
+
+
+class VisionModel(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.model_type = config.model_type
+        if self.model_type != "molmo":
+            raise ValueError(
+                f"Model type {self.model_type} not supported. Currently only 'molmo' is supported"
+            )
+        self.image_vit = VisionTransformer(config)
+        self.num_prefix_tokens = 1
+
+        self.image_pooling_2d = MultiHeadDotProductAttention(config, is_vit_layer=False)
+        self.image_projector = MLP(config, config.image_emb_dim)
+        self.pad_embed = mx.zeros((2, config.image_emb_dim * 2))
+
+    def encode_image(self, images: mx.array) -> mx.array:
+        """
+        : param images: (batch_size, num_crops, num_patch, n_pixels)
+        """
+        cfg = self.config
+        B, T, N, D = images.shape
+
+        # Check for -1 values across dimensions 1 and 2
+        reshaped_images = mx.reshape(images, (B * T, N, D))
+        mask = ~mx.all(reshaped_images == -1, axis=(1, 2), keepdims=True)
+
+        # Output all hidden states
+        images = reshaped_images
+        image_features = self.image_vit(images)
+
+        if cfg.vit_layers is not None:
+            features = []
+            for layer in cfg.vit_layers:
+                features.append(image_features[layer])
+            image_features = mx.concatenate(features, axis=-1)
+        else:
+            image_features = image_features[-1]
+
+        cls_embed = None
+        if self.num_prefix_tokens > 0:
+            cls_embed = image_features[:, 0]
+            image_features = image_features[:, 1:]
+
+        image_features = image_features * mask
+        image_features = mx.reshape(image_features, (B, T, N, -1))
+
+        cls_embed = mx.reshape(cls_embed, (B, T, -1)) if cls_embed is not None else None
+
+        return image_features, cls_embed
+
+    def __call__(
+        self, images: mx.array, image_masks: mx.array
+    ) -> Tuple[mx.array, Optional[mx.array]]:
+        cfg = self.config
+
+        batch_size, num_image = images.shape[:2]
+        image_features, cls_embed = self.encode_image(images)
+
+        if cfg.image_padding_embed:
+            assert image_masks is not None
+            if cfg.image_padding_embed == "pad_embed":
+                all_pad = image_masks == 0
+                pad_embed = mx.reshape(self.pad_embed, (1, 1, 1, -1))
+                image_features = image_features + pad_embed * mx.expand_dims(
+                    all_pad, -1
+                )
+            elif cfg.image_padding_embed == "regress":
+                pad_embed = mx.reshape(self.pad_embed, (1, 1, 1, -1))
+                image_features = image_features + pad_embed * mx.expand_dims(
+                    mx.maximum(image_masks, mx.zeros_like(image_masks)), -1
+                )
+            elif cfg.image_padding_embed == "pad_and_partial_pad":
+                pad_embed = mx.reshape(self.pad_embed, (2, 1, 1, 1, -1))
+                all_pad = image_masks == 0
+                partial_pad = mx.logical_and(image_masks < 1, mx.logical_not(all_pad))
+                partial_pad = partial_pad
+                all_pad = all_pad
+                image_features = image_features + pad_embed[0] * mx.expand_dims(
+                    all_pad, -1
+                )
+                image_features = image_features + pad_embed[1] * mx.expand_dims(
+                    partial_pad, -1
+                )
+            else:
+                raise ValueError(cfg.image_padding_embed)
+
+        image_features = mx.reshape(
+            image_features, (batch_size, num_image) + cfg.image_num_patch + (-1,)
+        )
+
+        if cfg.image_num_patch[0] % cfg.image_pooling_h == 1:
+            # Pad so we can still pool 2x2 patches
+            image_features = mx.pad(
+                image_features, [(0, 0), (0, 0), (0, 1), (0, 1), (0, 0)]
+            )
+
+        # image pooling
+        # MLX equivalent of einops rearrange
+        h_blocks = image_features.shape[2] // cfg.image_pooling_h
+        w_blocks = image_features.shape[3] // cfg.image_pooling_w
+        image_features = mx.reshape(
+            mx.transpose(
+                mx.reshape(
+                    image_features,
+                    (
+                        batch_size,
+                        num_image,
+                        h_blocks,
+                        cfg.image_pooling_h,
+                        w_blocks,
+                        cfg.image_pooling_w,
+                        -1,
+                    ),
+                ),
+                (0, 1, 2, 4, 3, 5, 6),
+            ),
+            (
+                batch_size * num_image * h_blocks * w_blocks,
+                cfg.image_pooling_h * cfg.image_pooling_w,
+                -1,
+            ),
+        )
+
+        if cfg.image_pooling_2d == "attention-meanq":
+            query = mx.mean(image_features, axis=-2, keepdims=True)
+            image_features = self.image_pooling_2d(query, image_features)
+        elif cfg.image_pooling_2d not in {"none", "stack"}:
+            image_features = self.image_pooling_2d(
+                image_features[:, :1, :], image_features
+            )
+
+        h, w = cfg.llm_patches_per_crop
+        image_features = mx.reshape(image_features, (batch_size, num_image, h * w, -1))
+
+        # # MLP layer to map the feature
+        image_features = self.image_projector(image_features)
+
+        return image_features, cls_embed

nexaai/mlx_backend/vlm/modeling/models/multi_modality/__init__.py

@@ -0,0 +1,10 @@
+from .multi_modality import (
+    ImageProcessor,
+    LanguageModel,
+    Model,
+    ModelConfig,
+    ProjectorConfig,
+    TextConfig,
+    VisionConfig,
+    VisionModel,
+)

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

@@ -0,0 +1,230 @@
+import inspect
+from dataclasses import dataclass
+from typing import Dict, Optional, Tuple, 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 = 4096
+    num_hidden_layers: int = 32
+    intermediate_size: int = 11008
+    num_attention_heads: int = 32
+    rms_norm_eps: float = 1e-6
+    vocab_size: int = 102400
+    num_key_value_heads: int = None
+    rope_theta: float = 10000
+    rope_traditional: bool = False
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = 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
+            }
+        )
+
+    def __post_init__(self):
+        if self.num_key_value_heads is None:
+            self.num_key_value_heads = self.num_attention_heads
+
+        if self.rope_scaling:
+            required_keys = {"factor", "type"}
+            if not all(key in self.rope_scaling for key in required_keys):
+                raise ValueError(f"rope_scaling must contain keys {required_keys}")
+
+            if self.rope_scaling["type"] != "linear":
+                raise ValueError("rope_scaling 'type' currently only supports 'linear'")
+
+
+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.repeats = n_heads // n_kv_heads
+
+        head_dim = config.hidden_size // n_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)
+
+        rope_scale = (
+            1 / config.rope_scaling["factor"]
+            if config.rope_scaling is not None
+            and config.rope_scaling["type"] == "linear"
+            else 1
+        )
+        self.rope = nn.RoPE(
+            head_dim,
+            traditional=config.rope_traditional,
+            base=config.rope_theta,
+            scale=rope_scale,
+        )
+
+    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 Llama(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 = nn.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
+
+        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, config: TextConfig):
+        super().__init__()
+        self.config = config
+        self.model_type = config.model_type
+        if self.model_type != "llama":
+            raise ValueError(
+                f"Model type {self.model_type} not supported. Currently only 'llama' is supported"
+            )
+        self.model = Llama(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, mask=mask, cache=cache, inputs_embeds=inputs_embeds)
+        logits = self.lm_head(out)
+        return LanguageModelOutput(logits=logits)
+
+    @staticmethod
+    def sanitize(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 layers(self):
+        return self.model.layers
+
+    @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