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

mlx_vlm/models/glm4v_moe/language.py

@@ -0,0 +1,674 @@
+from typing import Any, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+import numpy as np
+from mlx_lm.models.switch_layers import SwitchGLU
+
+from ..base import (
+    LanguageModelOutput,
+    create_attention_mask,
+    scaled_dot_product_attention,
+)
+from .config import ModelConfig, TextConfig
+
+
+def _compute_default_rope_parameters(
+    config: Optional[TextConfig] = None,
+    **rope_kwargs,
+) -> tuple[mx.array, float]:
+
+    if config is not None and len(rope_kwargs) > 0:
+        raise ValueError(
+            "Unexpected arguments: `**rope_kwargs` and `config` are mutually exclusive in "
+            f"`_compute_default_rope_parameters`, got `rope_kwargs`={rope_kwargs} and `config`={config}"
+        )
+    if len(rope_kwargs) > 0:
+        base = rope_kwargs["base"]
+        dim = rope_kwargs["dim"]
+    elif config is not None:
+        base = config.rope_theta or config.rope_parameters["rope_theta"]
+        partial_rotary_factor = (
+            config.partial_rotary_factor
+            if hasattr(config, "partial_rotary_factor")
+            else 1.0
+        )
+        head_dim = (
+            getattr(config, "head_dim", None)
+            or config.hidden_size // config.num_attention_heads
+        )
+        dim = int(head_dim * partial_rotary_factor)
+
+    attention_factor = 1.0  # Unused in this type of RoPE
+
+    # Compute the inverse frequencies
+    inv_freq = 1.0 / (
+        base ** (mx.arange(0, dim, 2, dtype=mx.int64).astype(mx.float32) / dim)
+    )
+    return inv_freq, attention_factor
+
+
+class GLM4VRotaryEmbedding(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+
+        if hasattr(config, "rope_scaling") and config.rope_scaling is not None:
+            self.rope_type = config.rope_scaling.get(
+                "rope_type", config.rope_scaling.get("type")
+            )
+        else:
+            self.rope_type = "default"
+        self.max_seq_len_cached = config.max_position_embeddings
+        self.original_max_seq_len = config.max_position_embeddings
+
+        self.config = config
+
+        self.rope_init_fn = _compute_default_rope_parameters
+
+        inv_freq, self.attention_scaling = self.rope_init_fn(self.config)
+        self._inv_freq = mx.array(inv_freq, dtype=mx.float32)
+        self._original_inv_freq = mx.array(inv_freq, dtype=mx.float32)
+
+    def __call__(self, x, position_ids):
+        inv_freq_expanded = self._inv_freq[None, None, :, None].astype(mx.float32)
+        inv_freq_expanded = mx.broadcast_to(
+            inv_freq_expanded, (3, position_ids.shape[1], self._inv_freq.shape[0], 1)
+        )
+        position_ids_expanded = position_ids[:, :, None, :].astype(mx.float32)
+
+        freqs = (
+            inv_freq_expanded.astype(mx.float32)
+            @ position_ids_expanded.astype(mx.float32)
+        ).transpose(0, 1, 3, 2)
+        emb = mx.concatenate((freqs, freqs), axis=-1)
+        cos = mx.cos(emb) * self.attention_scaling
+        sin = mx.sin(emb) * self.attention_scaling
+
+        return cos.astype(x.dtype), sin.astype(x.dtype)
+
+
+def rotate_half_llm(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return mx.concatenate([-x2, x1], axis=-1)
+
+
+def apply_multimodal_rotary_pos_emb(q, k, cos, sin, mrope_section):
+    """
+    Applies Rotary Position Embedding with Multimodal Sections to the query and key tensors.
+    Args:
+        q (mx.array): The query tensor.
+        k (mx.array): The key tensor.
+        cos (mx.array): The cosine part of the rotary embedding.
+        sin (mx.array): The sine part of the rotary embedding.
+        mrope_section (List[int]): Multimodal rope section for channel dimension of temporal, height and width.
+        unsqueeze_dim (int, optional): Dimension to unsqueeze. Defaults to 1.
+    Returns:
+        tuple(mx.array): The rotated query and key tensors.
+    """
+
+    mrope_section = np.cumsum(mrope_section * 2)[:-1].tolist()
+
+    cos = mx.concatenate(
+        [m[i % 3] for i, m in enumerate(mx.split(cos, mrope_section, axis=-1))], axis=-1
+    )[
+        :, None, :, :
+    ]  # unsqueeze dim 1
+    sin = mx.concatenate(
+        [m[i % 3] for i, m in enumerate(mx.split(sin, mrope_section, axis=-1))], axis=-1
+    )[:, None, :, :]
+
+    rotary_dim = cos.shape[-1]
+    q_rot = q[..., :rotary_dim]
+    q_pass = q[..., rotary_dim:]
+
+    k_rot = k[..., :rotary_dim]
+    k_pass = k[..., rotary_dim:]
+
+    q_embed = (q_rot * cos) + (rotate_half_llm(q_rot) * sin)
+    k_embed = (k_rot * cos) + (rotate_half_llm(k_rot) * sin)
+
+    q_embed = mx.concatenate([q_embed, q_pass], axis=-1)
+    k_embed = mx.concatenate([k_embed, k_pass], axis=-1)
+
+    return q_embed, k_embed
+
+
+class Glm4vMoeAttention(nn.Module):
+    def __init__(self, args: TextConfig):
+        super().__init__()
+
+        dim = args.hidden_size
+        self.n_heads = n_heads = args.num_attention_heads
+        self.n_kv_heads = n_kv_heads = args.num_key_value_heads
+
+        head_dim = args.head_dim
+        self.scale = head_dim**-0.5
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=args.attention_bias)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=args.attention_bias)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=args.attention_bias)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=False)
+
+        self.rope_parameters = args.rope_parameters or args.rope_scaling
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+        position_embeddings: Optional[mx.array] = None,
+    ) -> mx.array:
+        B, L, D = 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)
+        keys = keys.reshape(B, L, self.n_kv_heads, -1)
+
+        queries = queries.transpose(0, 2, 1, 3)
+        keys = keys.transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        cos, sin = position_embeddings
+
+        queries, keys = apply_multimodal_rotary_pos_emb(
+            queries, keys, cos, sin, self.rope_parameters["mrope_section"]
+        )
+
+        if cache is not None:
+            keys, values = cache.update_and_fetch(keys, values)
+
+        if mask is not None and isinstance(mask, mx.array):
+            mask = mask[..., : keys.shape[-2]]
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache=cache, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class Glm4vMoeMLP(nn.Module):
+    def __init__(
+        self, config: TextConfig, hidden_size: int = None, intermediate_size: int = None
+    ):
+        super().__init__()
+        self.config = config
+        self.hidden_size = config.hidden_size if hidden_size is None else hidden_size
+        self.intermediate_size = (
+            config.intermediate_size if intermediate_size is None else intermediate_size
+        )
+
+        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+
+    def __call__(self, x):
+        down_proj = self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+        return down_proj
+
+
+@mx.compile
+def group_expert_select(
+    gates,
+    e_score_correction_bias,
+    top_k,
+    n_group,
+    topk_group,
+    routed_scaling_factor,
+    norm_topk_prob,
+):
+
+    scores = mx.sigmoid(gates.astype(mx.float32))
+    orig_scores = scores
+    scores = scores + e_score_correction_bias
+    if n_group > 1:
+        k = top_k
+        scores = mx.unflatten(scores, axis=-1, shape=(n_group, -1))
+        group_scores = mx.topk(scores, 2, axis=-1).sum(axis=-1, keepdims=True)
+        k = n_group - topk_group
+        group_idx = mx.argpartition(group_scores, kth=k - 1, axis=-2)[..., :k, :]
+        scores = mx.put_along_axis(
+            scores, mx.stop_gradient(group_idx), mx.array(0.0), axis=-2
+        )
+        scores = mx.flatten(scores, -2, -1)
+
+    k = top_k
+    inds = mx.argpartition(-scores, kth=k - 1, axis=-1)[..., :k]
+    scores = mx.take_along_axis(orig_scores, inds, axis=-1)
+    if top_k > 1 and norm_topk_prob:
+        denominator = scores.sum(axis=-1, keepdims=True)
+        scores = scores / denominator
+    scores = scores * routed_scaling_factor
+
+    return inds, scores
+
+
+class MoEGate(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.config = config
+        self.top_k = config.num_experts_per_tok
+        self.norm_topk_prob = config.norm_topk_prob
+        self.n_routed_experts = config.n_routed_experts
+        self.routed_scaling_factor = config.routed_scaling_factor
+        self.n_group = config.n_group
+        self.topk_group = config.topk_group
+        self.weight = mx.zeros((self.n_routed_experts, config.hidden_size))
+        self.e_score_correction_bias = mx.zeros((self.n_routed_experts,))
+        assert config.topk_method == "noaux_tc", "Unsupported topk method."
+
+    def __call__(self, x):
+        return group_expert_select(
+            x @ self.weight.T,
+            self.e_score_correction_bias,
+            self.top_k,
+            self.n_group,
+            self.topk_group,
+            self.routed_scaling_factor,
+            self.norm_topk_prob,
+        )
+
+
+class MoE(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.config = config
+        self.num_experts_per_tok = config.num_experts_per_tok
+        self.switch_mlp = SwitchGLU(
+            config.hidden_size,
+            config.moe_intermediate_size,
+            config.n_routed_experts,
+        )
+
+        self.gate = MoEGate(config)
+        if config.n_shared_experts is not None:
+            intermediate_size = config.moe_intermediate_size * config.n_shared_experts
+            self.shared_experts = Glm4vMoeMLP(
+                config=config, intermediate_size=intermediate_size
+            )
+
+    def __call__(self, x):
+        inds, scores = self.gate(x)
+        y = self.switch_mlp(x, inds)
+        y = (y * scores[..., None]).sum(axis=-2).astype(y.dtype)
+        if self.config.n_shared_experts is not None:
+            y = y + self.shared_experts(x)
+
+        return y
+
+
+class Glm4vMoeDecoderLayer(nn.Module):
+    def __init__(self, config: TextConfig, layer_idx: int):
+        super().__init__()
+        self.self_attn = Glm4vMoeAttention(config)
+        self.mlp = (
+            MoE(config)
+            if (
+                config.n_routed_experts is not None
+                and layer_idx >= config.first_k_dense_replace
+            )
+            else Glm4vMoeMLP(config)
+        )
+        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
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+        position_embeddings: Optional[mx.array] = None,
+    ) -> mx.array:
+
+        r = self.self_attn(self.input_layernorm(x), mask, cache, position_embeddings)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        return h + r
+
+
+class GLM4VModel(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.vocab_size = config.vocab_size
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = [
+            Glm4vMoeDecoderLayer(config, idx) for idx in range(config.num_hidden_layers)
+        ]
+        self.start_idx = 0
+        self.end_idx = len(self.layers)
+        self.num_layers = self.end_idx
+
+        self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+        self.rotary_emb = GLM4VRotaryEmbedding(config)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        inputs_embeds: Optional[mx.array] = None,
+        cache: Optional[Any] = None,
+        mask: Optional[mx.array] = None,
+        position_ids: Optional[mx.array] = None,
+    ) -> mx.array:
+
+        if inputs_embeds is None:
+            h = self.embed_tokens(inputs)
+        else:
+            h = inputs_embeds.astype(self.norm.weight.dtype)
+
+        if position_ids is None:
+            position_ids = mx.arange(cache[0].offset, cache[0].offset + h.shape[-2])
+            position_ids = mx.expand_dims(position_ids, axis=0)
+            position_ids = mx.tile(position_ids, (3, 1, 1))
+
+        position_embeddings = self.rotary_emb(h, position_ids)
+
+        if mask is None:
+            mask = create_attention_mask(h, cache)
+
+        if cache is None:
+            cache = [None] * self.num_layers
+
+        for i in range(self.num_layers):
+            h = self.layers[self.start_idx + i](h, mask, cache[i], position_embeddings)
+
+        return self.norm(h)
+
+
+class LanguageModel(nn.Module):
+    def __init__(self, args: TextConfig, config: ModelConfig = None):
+        super().__init__()
+        self.args = args
+        self.config = config
+        self.model_type = args.model_type
+        self.model = GLM4VModel(args)
+        self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False)
+        self._rope_deltas = None
+
+    def get_rope_index(
+        self,
+        input_ids: mx.array,
+        image_grid_thw: Optional[mx.array] = None,
+        video_grid_thw: Optional[mx.array] = None,
+        attention_mask: Optional[mx.array] = None,
+    ):
+        # Calculate RoPE index for image/video tokens
+        batch_size, seq_length = input_ids.shape
+        position_ids = mx.arange(seq_length, dtype=mx.int32)
+        position_ids = mx.broadcast_to(position_ids[None, :], (batch_size, seq_length))
+        spatial_merge_size = self.config.vision_config.spatial_merge_size
+        image_token_id = self.config.image_token_id
+        video_token_id = self.config.video_token_id
+        vision_start_token_id = self.config.vision_start_token_id
+        mrope_position_deltas = []
+        if input_ids is not None and (
+            image_grid_thw is not None or video_grid_thw is not None
+        ):
+            total_input_ids = input_ids
+            if attention_mask is None:
+                attention_mask = mx.ones_like(input_ids)
+            position_ids = mx.ones(
+                (3, input_ids.shape[0], input_ids.shape[1]), dtype=input_ids.dtype
+            )
+            image_index, video_index = 0, 0
+            for i, input_ids in enumerate(total_input_ids):
+                input_ids = mx.where(
+                    attention_mask[i] == 1, input_ids, mx.zeros_like(input_ids)
+                )
+                image_nums, video_nums = 0, 0
+                vision_start_indices = mx.sum(
+                    mx.where(
+                        input_ids == vision_start_token_id,
+                        mx.arange(input_ids.shape[0]),
+                        mx.zeros_like(input_ids),
+                    )
+                )
+                vision_tokens = input_ids[vision_start_indices + 1]
+                image_nums = (vision_tokens == image_token_id).sum().item()
+                video_nums = (vision_tokens == video_token_id).sum().item()
+                input_tokens = input_ids.tolist()
+                llm_pos_ids_list: list = []
+                st = 0
+                remain_images, remain_videos = image_nums, video_nums
+                for _ in range(image_nums + video_nums):
+                    if image_token_id in input_tokens and remain_images > 0:
+                        ed_image = input_tokens.index(image_token_id, st)
+                    else:
+                        ed_image = len(input_tokens) + 1
+                    if video_token_id in input_tokens and remain_videos > 0:
+                        ed_video = input_tokens.index(video_token_id, st)
+                    else:
+                        ed_video = len(input_tokens) + 1
+                    if ed_image < ed_video:
+                        t, h, w = (
+                            image_grid_thw[image_index][0],
+                            image_grid_thw[image_index][1],
+                            image_grid_thw[image_index][2],
+                        )
+                        image_index += 1
+                        remain_images -= 1
+                        ed = ed_image
+                    else:
+                        t, h, w = (
+                            video_grid_thw[video_index][0],
+                            video_grid_thw[video_index][1],
+                            video_grid_thw[video_index][2],
+                        )
+                        video_index += 1
+                        remain_videos -= 1
+                        ed = ed_video
+                    llm_grid_t, llm_grid_h, llm_grid_w = (
+                        t.item(),
+                        h.item() // spatial_merge_size,
+                        w.item() // spatial_merge_size,
+                    )
+                    text_len = ed - st
+                    st_idx = (
+                        llm_pos_ids_list[-1].max() + 1
+                        if len(llm_pos_ids_list) > 0
+                        else 0
+                    )
+                    index = mx.arange(text_len).reshape(1, text_len)
+                    index = mx.broadcast_to(index, (3, text_len))
+                    index = index + st_idx
+                    llm_pos_ids_list.append(index)
+                    t_index = mx.arange(llm_grid_t).reshape(
+                        llm_grid_t, 1
+                    )  # Equivalent to .view(-1, 1)
+                    t_index = mx.broadcast_to(
+                        t_index, (llm_grid_t, llm_grid_h * llm_grid_w)
+                    )  # Equivalent to expand()
+                    t_index = t_index.flatten()  # Flattens to 1D
+
+                    h_index = mx.arange(llm_grid_h).reshape(
+                        1, llm_grid_h, 1
+                    )  # Equivalent to .view(1, -1)
+                    h_index = mx.broadcast_to(
+                        h_index, (llm_grid_t, llm_grid_h, llm_grid_w)
+                    )  # Equivalent to expand()
+                    h_index = h_index.flatten()  # Flattens to 1D
+
+                    w_index = mx.arange(llm_grid_w).reshape(
+                        1, 1, llm_grid_w
+                    )  # Equivalent to .view(1, -1)
+                    w_index = mx.broadcast_to(
+                        w_index, (llm_grid_t, llm_grid_h, llm_grid_w)
+                    )  # Equivalent to expand()
+                    w_index = w_index.flatten()  # Flattens to 1D
+
+                    llm_pos_ids_list.append(
+                        mx.stack([t_index, h_index, w_index]) + text_len + st_idx
+                    )
+                    st = ed + llm_grid_t * llm_grid_h * llm_grid_w
+                if st < len(input_tokens):
+                    st_idx = (
+                        llm_pos_ids_list[-1].max() + 1
+                        if len(llm_pos_ids_list) > 0
+                        else 0
+                    )
+                    text_len = len(input_tokens) - st
+
+                    t_index = mx.arange(text_len).reshape(
+                        1, text_len
+                    )  # Equivalent to .view(-1, 1)
+                    t_index = mx.broadcast_to(
+                        t_index, (3, text_len)
+                    )  # Equivalent to expand(3, -1)
+
+                    llm_pos_ids_list.append(t_index + st_idx)
+
+                llm_positions = mx.concatenate(llm_pos_ids_list, axis=1).reshape(3, -1)
+                mask = mx.array(attention_mask[i] == 1)
+                expanded_mask = mx.expand_dims(mask, axis=0)
+                expanded_mask = mx.broadcast_to(expanded_mask, (3, 1, mask.shape[0]))
+                expanded_positions = mx.expand_dims(llm_positions, axis=1)
+                new_positions = mx.where(
+                    expanded_mask, expanded_positions, position_ids[:, i : i + 1, :]
+                )
+                updated_position_ids = mx.concatenate(
+                    [
+                        position_ids[:, :i, :],
+                        new_positions,
+                        position_ids[:, i + 1 :, :],
+                    ],
+                    axis=1,
+                )
+                position_ids = updated_position_ids
+                mrope_position_deltas.append(
+                    llm_positions.max() + 1 - len(total_input_ids[i])
+                )
+            mrope_position_deltas = mx.array(mrope_position_deltas)[0]
+            return position_ids, mrope_position_deltas
+        else:
+            if attention_mask is not None:
+                position_ids = mx.cumsum(attention_mask.astype(mx.int64), axis=-1) - 1
+                position_ids = mx.where(
+                    attention_mask == 0, mx.ones_like(position_ids), position_ids
+                )
+                position_ids = mx.expand_dims(position_ids[0], axis=0)
+                position_ids = mx.tile(position_ids, (3, 1, 1))
+                max_position_ids = position_ids.max(0, keepdims=False)[0].max(
+                    -1, keepdims=True
+                )[0]
+                mrope_position_deltas = max_position_ids + 1 - attention_mask.shape[-1]
+            else:
+                position_ids = mx.arange(input_ids.shape[1]).reshape(1, -1)
+                position_ids = mx.broadcast_to(
+                    position_ids, (3, input_ids.shape[0], input_ids.shape[1])
+                )
+                mrope_position_deltas = mx.zeros(
+                    [input_ids.shape[0], 1],
+                    dtype=input_ids.dtype,
+                )
+            return position_ids, mrope_position_deltas
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        inputs_embeds: Optional[mx.array] = None,
+        mask: Optional[mx.array] = None,
+        cache=None,
+        **kwargs,
+    ):
+
+        position_ids = kwargs.pop("position_ids", None)
+        pixel_values = kwargs.pop("pixel_values", None)
+        image_grid_thw = kwargs.pop("image_grid_thw", None)
+        video_grid_thw = kwargs.pop("video_grid_thw", None)
+        # reset rope_deltas when processing a new image/video
+        if pixel_values is not None:
+            self._rope_deltas = None
+
+        cache_offset = 0
+        if cache and cache[0] is not None:
+            offset = cache[0].offset
+            if isinstance(offset, int):
+                cache_offset = offset
+            elif isinstance(offset, mx.array):
+                cache_offset = (offset if offset.ndim == 0 else offset[0]).item()
+            else:
+                raise ValueError(f"Unexpected cache offset type: {type(offset)}")
+
+        if position_ids is None and (mask is None or mask.ndim == 2):
+            # Calculate RoPE index once per generation in the pre-fill stage only
+            if (
+                (cache is not None and cache[0] is not None and (cache_offset == 0))
+                or self._rope_deltas is None
+                or cache is None
+            ):
+                position_ids, rope_deltas = self.get_rope_index(
+                    inputs, image_grid_thw, video_grid_thw, mask
+                )
+                self._rope_deltas = rope_deltas
+            else:
+                # Use the prev pre-calculated rope-deltas to get the correct position ids
+                batch_size, seq_length = inputs.shape
+                delta = mx.array(
+                    cache_offset + self._rope_deltas if cache is not None else 0
+                )
+                position_ids = mx.arange(seq_length).reshape(1, -1)
+                position_ids = mx.broadcast_to(position_ids, (batch_size, seq_length))
+
+                if cache_offset is not None:
+                    if delta.ndim == 0:
+                        delta = mx.expand_dims(delta, axis=0)
+
+                    if delta.shape[0] < batch_size:
+                        delta = mx.tile(delta, (batch_size, 1))
+                    else:
+                        # Slice delta to match batch
+                        delta = delta[:batch_size]
+
+                position_ids = mx.add(position_ids, delta)[None, ...]
+                position_ids = mx.broadcast_to(
+                    position_ids, (3, batch_size, seq_length)
+                )
+
+        out = self.model(
+            inputs, cache=cache, inputs_embeds=inputs_embeds, position_ids=position_ids
+        )
+
+        out = self.lm_head(out)
+        return LanguageModelOutput(logits=out)
+
+    def sanitize(self, weights):
+        mpt_layer = self.args.num_hidden_layers
+
+        # Stack experts
+        for l in range(self.args.num_hidden_layers):
+            prefix = f"language_model.model.layers.{l}"
+            for n, m in [("w1", "gate_proj"), ("w2", "down_proj"), ("w3", "up_proj")]:
+                for k in ["weight", "scales", "biases"]:
+                    if f"{prefix}.mlp.experts.0.{m}.{k}" in weights:
+                        to_join = [
+                            weights.pop(f"{prefix}.mlp.experts.{e}.{m}.{k}")
+                            for e in range(self.args.n_routed_experts)
+                        ]
+                        weights[f"{prefix}.mlp.switch_mlp.{m}.{k}"] = mx.stack(to_join)
+
+        # Remove multi-token prediction layer
+        return {
+            k: v
+            for k, v in weights.items()
+            if not k.startswith(f"language_model.model.layers.{mpt_layer}")
+        }
+
+    @property
+    def layers(self):
+        return self.model.layers
+
+    @property
+    def cast_predicate(self):
+        def predicate(k):
+            return "e_score_correction_bias" not in k
+
+        return predicate
+
+    @property
+    def n_kv_heads(self):
+        return self.args.num_key_value_heads