from typing import Any, Optional import mlx.core as mx import mlx.nn as nn import numpy as np from ..base import ( LanguageModelOutput, create_attention_mask, scaled_dot_product_attention, ) from ..rope_utils import ( apply_rotary_pos_emb_even_odd, compute_mrope_frequencies, mrope_position_selector, ) 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 partial_rotary_factor = config.partial_rotary_factor head_dim = config.head_dim dim = int(head_dim * partial_rotary_factor) attention_factor = 1.0 inv_freq = 1.0 / ( base ** (mx.arange(0, dim, 2, dtype=mx.int64).astype(mx.float32) / dim) ) return inv_freq, attention_factor class GlmOcrRotaryEmbedding(nn.Module): def __init__(self, config: TextConfig): super().__init__() self.rope_type = config.rope_parameters.get("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.mrope_section = config.rope_parameters.get("mrope_section", [16, 24, 24]) 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) self._position_selector = mrope_position_selector( "split_select", self.mrope_section, self._inv_freq.shape[0], ) def __call__(self, x, position_ids): freqs = compute_mrope_frequencies( position_ids, self._inv_freq, self.mrope_section, style="split_select", position_selector=self._position_selector, ) 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 apply_rotary_pos_emb(q, k, cos, sin): """ Applies Rotary Position Embedding to the query and key tensors. Matches PyTorch's GLM-OCR implementation exactly. Args: q: Query tensor of shape (batch, n_heads, seq_len, head_dim) k: Key tensor of shape (batch, n_kv_heads, seq_len, head_dim) cos: Cosine tensor of shape (batch, seq_len, head_dim) sin: Sine tensor of shape (batch, seq_len, head_dim) """ return apply_rotary_pos_emb_even_odd(q, k, cos, sin) class GlmOcrAttention(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 self.head_dim = args.head_dim self.scale = self.head_dim**-0.5 self.q_proj = nn.Linear(dim, n_heads * self.head_dim, bias=args.attention_bias) self.k_proj = nn.Linear( dim, n_kv_heads * self.head_dim, bias=args.attention_bias ) self.v_proj = nn.Linear( dim, n_kv_heads * self.head_dim, bias=args.attention_bias ) self.o_proj = nn.Linear(n_heads * self.head_dim, dim, bias=False) self.rope_parameters = args.rope_parameters 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, _ = x.shape queries = self.q_proj(x) keys = self.k_proj(x) values = 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_rotary_pos_emb(queries, keys, cos, sin) 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 GlmOcrMLP(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_up_proj = nn.Linear( self.hidden_size, self.intermediate_size * 2, bias=False ) self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False) def __call__(self, x): x = self.gate_up_proj(x) gate, x = mx.split(x, 2, axis=-1) return self.down_proj(nn.silu(gate) * x) class GlmOcrDecoderLayer(nn.Module): def __init__(self, config: TextConfig): super().__init__() self.self_attn = GlmOcrAttention(config) self.mlp = GlmOcrMLP(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 ) self.post_self_attn_layernorm = nn.RMSNorm( config.hidden_size, eps=config.rms_norm_eps ) self.post_mlp_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 = x x = self.self_attn(self.input_layernorm(x), mask, cache, position_embeddings) x = self.post_self_attn_layernorm(x) x = r + x r = x x = self.post_attention_layernorm(x) x = self.mlp(x) x = self.post_mlp_layernorm(x) x = r + x return x class GlmOcrTextModel(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 = [ GlmOcrDecoderLayer(config) for _ 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 = GlmOcrRotaryEmbedding(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[0] if cache and cache[0] is not None else 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 = GlmOcrTextModel(args) self.lm_head = nn.Linear(args.hidden_size, args.vocab_size, bias=False) self._rope_deltas = None self._position_ids = None def get_vision_position_ids( self, start_position: int, grid_thw: list[int], temp_merge_size: int = 1, ): spatial_merge_size = self.config.vision_config.spatial_merge_size llm_grid_t = int(grid_thw[0]) // temp_merge_size llm_grid_h = int(grid_thw[1]) // spatial_merge_size llm_grid_w = int(grid_thw[2]) // spatial_merge_size image_seq_length = llm_grid_t * llm_grid_h * llm_grid_w position_temporal = np.full((image_seq_length,), start_position, dtype=np.int64) position_height = np.repeat( np.arange( start_position, start_position + llm_grid_h, dtype=np.int64, ), llm_grid_w * llm_grid_t, ) position_width = np.tile( np.arange( start_position, start_position + llm_grid_w, dtype=np.int64, ), llm_grid_h * llm_grid_t, ) return np.stack([position_temporal, position_height, position_width], axis=0) 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, ): batch_size, seq_length = input_ids.shape position_ids_np = np.zeros((3, batch_size, seq_length), dtype=np.int64) 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 mrope_position_deltas = [] if input_ids is not None and ( image_grid_thw is not None or video_grid_thw is not None ): input_ids_list = input_ids.tolist() attention_mask_list = ( attention_mask.tolist() if attention_mask is not None and attention_mask.shape[-1] == input_ids.shape[-1] else None ) image_grids = image_grid_thw.tolist() if image_grid_thw is not None else [] video_grids = video_grid_thw.tolist() if video_grid_thw is not None else [] image_index = 0 video_index = 0 for batch_idx, current_input_ids in enumerate(input_ids_list): if attention_mask_list is not None: valid_indices = [ i for i, m in enumerate(attention_mask_list[batch_idx]) if m == 1 ] current_input_ids = [current_input_ids[i] for i in valid_indices] else: valid_indices = list(range(len(current_input_ids))) token_types = [] for token_id in current_input_ids: if token_id == image_token_id: token_types.append(1) elif token_id == video_token_id: token_types.append(2) else: token_types.append(0) groups = [] if token_types: start = 0 current_type = token_types[0] for idx, token_type in enumerate(token_types[1:], start=1): if token_type != current_type: groups.append((current_type, start, idx)) start = idx current_type = token_type groups.append((current_type, start, len(token_types))) current_pos = 0 video_group_index = 0 llm_pos_ids_list: list = [] for modality_type, start_idx, end_idx in groups: if modality_type == 0: text_len = end_idx - start_idx text_positions = np.arange(text_len, dtype=np.int64)[None, :] llm_pos_ids_list.append( np.broadcast_to(text_positions, (3, text_len)) + current_pos ) current_pos += text_len continue if modality_type == 2: grid_thw = video_grids[video_index] video_group_index += 1 if video_group_index >= int(grid_thw[0]): video_index += 1 video_group_index = 0 else: grid_thw = image_grids[image_index] image_index += 1 temp_merge_size = int(grid_thw[0]) llm_pos_ids_list.append( self.get_vision_position_ids( current_pos, grid_thw, temp_merge_size=temp_merge_size, ) ) current_pos += max(int(grid_thw[1]), int(grid_thw[2])) // ( spatial_merge_size ) llm_positions = np.concatenate(llm_pos_ids_list, axis=1).reshape(3, -1) position_ids_np[:, batch_idx, valid_indices] = llm_positions mrope_position_deltas.append( int(llm_positions.max()) + 1 - len(current_input_ids) ) return ( mx.array(position_ids_np, dtype=input_ids.dtype), mx.array(mrope_position_deltas, dtype=input_ids.dtype).reshape(-1, 1), ) 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.zeros_like(position_ids), position_ids ) max_position_ids = position_ids.max(axis=-1, keepdims=True) position_ids = mx.broadcast_to( position_ids[None, :, :], (3, *position_ids.shape) ) 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) rope_deltas_kw = kwargs.pop("rope_deltas", None) if pixel_values is not None: self._rope_deltas = None # Use ``cache._idx`` — the Python-int token counter — instead of # syncing on ``cache[0].offset``. See Qwen2.5-VL for details. cache_offset = 0 cache_offsets = None if cache and cache[0] is not None: c0 = cache[0] cache_offset = c0._idx if hasattr(c0, "_idx") else c0.offset if ( isinstance(c0.offset, mx.array) and c0.offset.ndim > 0 and c0.offset.size > 1 ): cache_offsets = c0.offset # Check if mask shape matches input shape (for chunked prefill compatibility) rope_mask = mask if mask is not None and mask.shape[-1] != inputs.shape[-1]: rope_mask = None if position_ids is None and (rope_mask is None or rope_mask.ndim == 2): # Calculate RoPE index once per generation in the pre-fill stage only is_prefill = ( cache is None or cache[0] is None or (cache_offsets is None and cache_offset == 0) ) seq_length = inputs.shape[1] if ( self._position_ids is not None and cache_offsets is None and cache_offset + seq_length <= self._position_ids.shape[-1] ): position_ids = self._position_ids[ :, :, cache_offset : cache_offset + seq_length ] elif is_prefill or self._rope_deltas is None: if self._position_ids is None: position_ids, rope_deltas = self.get_rope_index( inputs, image_grid_thw, video_grid_thw, rope_mask ) self._rope_deltas = rope_deltas self._position_ids = position_ids else: position_ids = self._position_ids[:, :, :seq_length] else: # Use the prev pre-calculated rope-deltas to get the correct position ids batch_size, seq_length = inputs.shape rope_deltas_src = ( rope_deltas_kw if rope_deltas_kw is not None else self._rope_deltas ) if cache_offsets is not None: offsets = cache_offsets[:batch_size] rope_deltas = rope_deltas_src if rope_deltas.shape[0] > batch_size: rope_deltas = rope_deltas[:batch_size] delta = (offsets + rope_deltas.squeeze(-1))[:, None] else: delta = mx.array( cache_offset + rope_deltas_src if cache is not None else 0 ) 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: delta = delta[:batch_size] position_ids = mx.arange(seq_length).reshape(1, -1) position_ids = mx.broadcast_to(position_ids, (batch_size, seq_length)) 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, mask=mask, ) out = self.lm_head(out) return LanguageModelOutput(logits=out) def sanitize(self, weights): return weights @property def layers(self): return self.model.layers @property def n_kv_heads(self): return self.args.num_key_value_heads