"""Schema25 MLX model, trajectory memory, and strict release loading.""" from __future__ import annotations from collections import Counter, defaultdict from dataclasses import dataclass, replace import hashlib import json import math from pathlib import Path from types import SimpleNamespace from typing import Any import weakref import mlx.core as mx from mlx import nn from mlx.utils import tree_flatten from mlx_lm.tuner.lora import LoRALinear, LoRASwitchLinear from mlx_vlm.models.diffusion_gemma.config import ModelConfig from mlx_vlm.models.diffusion_gemma.diffusion_gemma import Model from mlx_vlm.models.diffusion_gemma.language import DecoderLayer, Router def _l2(value: mx.array) -> mx.array: return value * mx.rsqrt(mx.maximum(mx.sum(mx.square(value), axis=-1, keepdims=True), 1e-12)) class _GateProjection(nn.Module): def __init__(self, source: int, target: int, rank: int) -> None: super().__init__() self.down = nn.Linear(source, rank, bias=False) self.up = nn.Linear(rank, target, bias=False) def __call__(self, source: mx.array) -> mx.array: return self.up(self.down(source)) class GDN2Memory(nn.Module): def __init__(self, input_dim: int, heads: int, key_dim: int, value_dim: int, *, observation_dim: int | None = None) -> None: super().__init__() if min(input_dim, heads, key_dim, value_dim) <= 0: raise ValueError("GDN2 dimensions must be positive.") self.heads, self.key_dim, self.value_dim = heads, key_dim, value_dim observation_dim = input_dim if observation_dim is None else observation_dim if observation_dim <= 0: raise ValueError("GDN2 observation width must be positive.") self.q_proj = nn.Linear(input_dim, heads * key_dim, bias=False) self.k_proj = nn.Linear(observation_dim, heads * key_dim, bias=False) self.v_proj = nn.Linear(observation_dim, heads * value_dim, bias=False) self.f_proj = _GateProjection(observation_dim, heads * key_dim, min(observation_dim, key_dim)) self.b_proj = nn.Linear(observation_dim, heads * key_dim, bias=False) self.w_proj = nn.Linear(observation_dim, heads * value_dim, bias=False) self.g_proj = _GateProjection(input_dim, heads * value_dim, min(input_dim, value_dim)) self.o_proj = nn.Linear(heads * value_dim, input_dim, bias=False) self.a_log = mx.zeros((heads,), mx.float32) self.dt_bias = mx.full((heads, key_dim), -6.906255, mx.float32) def empty(self, *leading: int) -> mx.array: return mx.zeros((*leading, self.heads, self.key_dim, self.value_dim), mx.float32) def _projections(self, source: mx.array): normalized = source.astype(mx.float32) source = (normalized * mx.rsqrt(mx.mean(mx.square(normalized), axis=-1, keepdims=True) + 1e-6)).astype(source.dtype) shape = source.shape[:-1] key_shape = (*shape, self.heads, self.key_dim) value_shape = (*shape, self.heads, self.value_dim) k = _l2(nn.silu(self.k_proj(source).astype(mx.float32)).reshape(key_shape)) v = nn.silu(self.v_proj(source).astype(mx.float32)).reshape(value_shape) head_rate = mx.exp(self.a_log.astype(mx.float32)).reshape( *((1,) * (source.ndim - 1)), self.heads, 1) decay = mx.exp(-head_rate * nn.softplus( self.f_proj(source).astype(mx.float32).reshape(key_shape) + self.dt_bias.astype(mx.float32))) erase = mx.sigmoid(self.b_proj(source).astype(mx.float32).reshape(key_shape)) write = mx.sigmoid(self.w_proj(source).astype(mx.float32).reshape(value_shape)) return k, v, decay, erase, write def _query(self, source: mx.array) -> mx.array: return _l2(nn.silu(self.q_proj(source).astype(mx.float32)).reshape( *source.shape[:-1], self.heads, self.key_dim)) def _output(self, value: mx.array, source: mx.array) -> mx.array: gate = nn.silu(self.g_proj(source).astype(mx.float32).reshape(value.shape)) value = value * mx.rsqrt(mx.mean(mx.square(value), axis=-1, keepdims=True) + 1e-6) * gate return self.o_proj(value.reshape(*source.shape[:-1], -1).astype(source.dtype)) def read(self, state: mx.array, source: mx.array) -> mx.array: if state.shape != (*source.shape[:-1], self.heads, self.key_dim, self.value_dim): raise ValueError("GDN2 state and query leading dimensions differ.") value = mx.sum(self._query(source)[..., None] * state.astype(mx.float32), axis=-2) return self._output(value, source) def read_shared(self, state: mx.array, source: mx.array) -> mx.array: if source.ndim != 3 or state.shape != (source.shape[0], self.heads, self.key_dim, self.value_dim): raise ValueError("Shared GDN2 read requires [batch, canvas, width] queries.") query = self._query(source).transpose(0, 2, 1, 3) value = (query @ state.astype(mx.float32)).transpose(0, 2, 1, 3) return self._output(value, source) @staticmethod def _transition(state, k, v, decay, erase, write, valid): decayed = state.astype(mx.float32) * decay[..., None] old = mx.sum((erase * k)[..., None] * decayed, axis=-2) candidate = decayed + k[..., None] * (write * v - old)[..., None, :] return candidate if valid is None else mx.where( valid[..., None, None, None], candidate, state.astype(mx.float32)) def transition(self, state: mx.array, source: mx.array, valid: mx.array | None = None) -> mx.array: if state.shape != (*source.shape[:-1], self.heads, self.key_dim, self.value_dim): raise ValueError("GDN2 state and observation leading dimensions differ.") k, v, decay, erase, write = self._projections(source) if valid is not None: if valid.shape != source.shape[:-1]: raise ValueError("GDN2 valid mask must match observation rows.") return self._transition(state, k, v, decay, erase, write, valid) def write_sequence(self, state: mx.array, source: mx.array, valid: mx.array) -> mx.array: if source.ndim != 3 or valid.shape != source.shape[:2]: raise ValueError("GDN2 sequence and mask must share [batch, length].") if state.shape != (source.shape[0], self.heads, self.key_dim, self.value_dim): raise ValueError("GDN2 sequence state shape differs.") projected = self._projections(source) for index in range(source.shape[1]): state = self._transition(state, *(part[:, index] for part in projected), valid[:, index]) return state @dataclass(frozen=True) class GDN2TrajectoryState: cells: mx.array row: mx.array persistent: mx.array seen: mx.array def clear_refill(self, lengths: mx.array, head: mx.array) -> "GDN2TrajectoryState": batch, canvas = self.seen.shape if lengths.shape != (batch,) or head.shape != (batch,): raise ValueError("Commit lengths and canvas heads must be per row.") physical = mx.arange(canvas)[None, :] overwritten = ((physical - head[:, None]) % canvas) < lengths[:, None] return GDN2TrajectoryState( mx.where(overwritten[..., None, None, None], 0.0, self.cells), self.row, self.persistent, self.seen & ~overwritten, ) class GDN2TrajectoryMemory(nn.Module): def __init__(self, width: int, *, probes: int = 4, working_heads: int = 16, working_key: int = 64, working_value: int = 64, persistent_heads: int = 16, persistent_key: int = 128, persistent_value: int = 128, persistent_observation_dim: int | None = None) -> None: super().__init__() if probes <= 0: raise ValueError("Probe count must be positive.") self.probes = probes self.cell = GDN2Memory(width, working_heads, working_key, working_value) self.row = GDN2Memory(width, working_heads, working_key, working_value) self.persistent = GDN2Memory(width, persistent_heads, persistent_key, persistent_value, observation_dim=persistent_observation_dim) self.probe_embed = nn.Embedding(probes, width) self.probe_embed.weight = mx.random.normal((probes, width)) * 0.02 def read(self, state: GDN2TrajectoryState, query: mx.array) -> mx.array: result = (self.cell.read(state.cells, query) + self.row.read_shared(state.row, query) + self.persistent.read_shared(state.persistent, query)) return mx.where(state.seen[..., None], result, 0.0) def observe(self, state: GDN2TrajectoryState, observation: mx.array, live: mx.array, head: mx.array) -> GDN2TrajectoryState: batch, canvas, width = observation.shape if live.shape != (batch, canvas) or head.shape != (batch,): raise ValueError("Observation mask and head have incorrect shapes.") cells = self.cell.transition(state.cells, observation, live) seen = state.seen | live logical_idx = (head[:, None] + mx.arange(canvas)[None, :]) % canvas logical = mx.take_along_axis(observation, mx.broadcast_to(logical_idx[..., None], observation.shape), axis=1) logical_live = mx.take_along_axis(live, logical_idx, axis=1) row_state = state.row for probe in range(self.probes): lo = canvas * probe // self.probes hi = canvas * (probe + 1) // self.probes selected = logical_live[:, lo:hi] count = mx.sum(selected.astype(mx.float32), axis=1, keepdims=True) pooled = mx.sum(logical[:, lo:hi].astype(mx.float32) * selected[..., None], axis=1) pooled = (pooled / mx.maximum(count, 1.0)).astype(observation.dtype) pooled = pooled + self.probe_embed.weight[probe].astype(observation.dtype) row_state = self.row.transition(row_state, pooled, count[:, 0] > 0) return GDN2TrajectoryState(cells, row_state, state.persistent, seen) def fp32_attention( query: mx.array, keys: mx.array, values: mx.array, *, attn_mask: mx.array | None = None, ) -> mx.array: """Small memory-attention reductions in FP32 with model-dtype output.""" dtype = query.dtype if query.ndim == 4 and keys.ndim == 4 and keys.shape[-2] >= 64: # The working bus attends over the full 256-token canvas. Metal's # fused attention avoids materializing its FP32 score/probability grid # while retaining FP32 accumulation. mask = attn_mask if mask is not None: mask = (mx.where(mask, 0.0, -1.0e9) if mask.dtype == mx.bool_ else mask.astype(mx.float32)) return mx.fast.scaled_dot_product_attention( query.astype(mx.float32), keys.astype(mx.float32), values.astype(mx.float32), scale=1.0 / math.sqrt(max(query.shape[-1], 1)), mask=mask, ).astype(dtype) scores = mx.matmul(query.astype(mx.float32), mx.swapaxes(keys.astype(mx.float32), -1, -2)) scores = scores / math.sqrt(max(query.shape[-1], 1)) if attn_mask is not None: if attn_mask.dtype == mx.bool_: scores = mx.where(attn_mask, scores, -1.0e9) else: scores = scores + attn_mask.astype(mx.float32) probabilities = mx.nan_to_num(mx.softmax(scores, axis=-1), nan=0.0) return mx.matmul(probabilities, values.astype(mx.float32)).astype(dtype) class RMSNorm(nn.Module): def __init__(self, dim: int, eps: float = 1.0e-6) -> None: super().__init__() self.weight = mx.ones((dim,)) self.eps = eps def __call__(self, hidden: mx.array) -> mx.array: value = hidden.astype(mx.float32) scale = mx.rsqrt(mx.mean(mx.square(value), axis=-1, keepdims=True) + self.eps) return (value * scale * self.weight.astype(mx.float32)).astype(hidden.dtype) class SwiGLU(nn.Module): def __init__(self, dim: int, hidden: int) -> None: super().__init__() self.gate = nn.Linear(dim, hidden, bias=False) self.up = nn.Linear(dim, hidden, bias=False) self.down = nn.Linear(hidden, dim, bias=False) def __call__(self, hidden: mx.array) -> mx.array: return self.down(nn.silu(self.gate(hidden)) * self.up(hidden)) class RankAttention(nn.Module): def __init__(self, dim: int, num_heads: int, kv_rank: int) -> None: super().__init__() if kv_rank % num_heads: raise ValueError("Attention K/V rank must divide heads.") self.num_heads, self.kv_rank = num_heads, kv_rank self.head_dim = kv_rank // num_heads self.q_proj = nn.Linear(dim, kv_rank, bias=False) self.k_proj = nn.Linear(dim, kv_rank, bias=False) self.v_proj = nn.Linear(dim, kv_rank, bias=False) self.o_proj = nn.Linear(kv_rank, dim, bias=False) self.q_norm = RMSNorm(dim) self.k_norm = RMSNorm(dim) def __call__(self, query: mx.array, keys: mx.array, values: mx.array, attn_mask: mx.array | None = None) -> mx.array: batch, queries, _ = query.shape key_len = keys.shape[1] heads, head_dim = self.num_heads, self.head_dim reshape_q = lambda x: x.reshape(batch, queries, heads, head_dim).transpose(0, 2, 1, 3) reshape_k = lambda x: x.reshape(batch, key_len, heads, head_dim).transpose(0, 2, 1, 3) q = reshape_q(self.q_proj(self.q_norm(query))) k = reshape_k(self.k_proj(self.k_norm(keys))) v = reshape_k(self.v_proj(values)) mask = attn_mask if mask is not None and mask.ndim == 2: mask = mask[None, None, :, :] elif mask is not None and mask.ndim == 3: mask = mask[:, None] context = fp32_attention(q, k, v, attn_mask=mask) return self.o_proj(context.transpose(0, 2, 1, 3).reshape(batch, queries, self.kv_rank)) class _WorkingBus(nn.Module): def __init__(self, width: int, heads: int, readers: int, rank: int, canvas: int) -> None: super().__init__() self.num_readers, self.heads, self.rank = readers, heads, rank self.head_dim = rank // heads self.memory_norm = RMSNorm(width) # Retain the published address norm tensor in the weight topology. self.address_norm = RMSNorm(width) self.k_proj = nn.Linear(width, rank, bias=False) self.v_proj = nn.Linear(width, rank, bias=False) self.q_norm = RMSNorm(width) self.q_proj = [nn.Linear(width, rank, bias=False) for _ in range(readers)] self.o_proj = [nn.Linear(rank, width, bias=False) for _ in range(readers)] self.alpha = mx.zeros((max(readers, 1), heads)) self.rel_bias = mx.zeros((heads, 2 * canvas - 1)) def prepare_kv(self, working: mx.array, seen: mx.array): if not self.num_readers: return None normed = self.memory_norm(working) batch, canvas, _ = working.shape shape = (batch, canvas, self.heads, self.head_dim) keys = self.k_proj(normed).reshape(shape).transpose(0, 2, 1, 3) values = self.v_proj(normed).reshape(shape).transpose(0, 2, 1, 3) positions = mx.arange(canvas) relative = positions[:, None] - positions[None, :] + canvas - 1 bias = mx.take(self.rel_bias, relative, axis=1)[None].astype(mx.float32) bias = bias + mx.where(seen[:, None, None, :], 0.0, -1e9).astype(mx.float32) return keys, values, bias, seen def read(self, hidden, reader, keys, values, bias, seen): batch, canvas, _ = hidden.shape query = self.q_proj[reader](self.q_norm(hidden)) query = query.reshape(batch, canvas, self.heads, self.head_dim).transpose(0, 2, 1, 3) context = fp32_attention(query, keys, values, attn_mask=bias) context = context * mx.tanh(self.alpha[reader]).astype(hidden.dtype)[None, :, None, None] context = context.transpose(0, 2, 1, 3).reshape(batch, canvas, self.rank) return mx.where(seen[..., None], hidden + self.o_proj[reader](context), hidden) class _CanvasBlock(nn.Module): def __init__(self, width: int, heads: int, rank: int, ffn: int, window: int, global_attention: bool) -> None: super().__init__() self.norm = RMSNorm(width) self.attn = RankAttention(width, heads, rank) self.ff_norm = RMSNorm(width) self.ff = SwiGLU(width, ffn) self.window = window self.global_attention = global_attention def __call__(self, hidden: mx.array, seen: mx.array, offsets: mx.array) -> mx.array: allowed = mx.broadcast_to(seen[:, None, :], (hidden.shape[0], hidden.shape[1], hidden.shape[1])) if not self.global_attention: allowed = allowed & (mx.abs(offsets[:, :, None] - offsets[:, None, :]) < self.window) mask = mx.where(allowed, 0.0, -1e9).astype(mx.float32) normed = self.norm(hidden) hidden = hidden + self.attn(normed, normed, normed, mask) hidden = hidden + self.ff(self.ff_norm(hidden)) return mx.where(seen[..., None], hidden, mx.zeros_like(hidden)) class _PersistentBus(nn.Module): def __init__(self, memory: nn.Module, readers: int) -> None: super().__init__() object.__setattr__(self, "_memory_ref", weakref.ref(memory)) self.num_readers = readers self.alpha = mx.zeros((max(readers, 1), 1)) def prepare_kv(self, memory: mx.array, seen: mx.array | None = None): if self.num_readers <= 0: return None return memory, seen def read(self, hidden: mx.array, reader_index: int, memory: mx.array, seen: mx.array | None) -> mx.array: delta = self._memory_ref().read_shared(memory, hidden) if seen is not None: delta = delta * seen[..., None].astype(delta.dtype) return hidden + mx.tanh(self.alpha[reader_index]).astype(hidden.dtype) * delta class LatentDeliberationTransformer(nn.Module): def __init__(self, *, hidden_size: int, latent_dim: int, ffn_dim: int, num_layers: int, num_heads: int, local_attention_window: int, tape_probes: int, history_kv_rank: int, num_working_readers: int, num_persistent_readers: int, working_last_block_global: bool, commit_sequence_dim: int | None = None, max_canvas_length: int = 256) -> None: super().__init__() if hidden_size != latent_dim: raise ValueError("Schema25 GDN2 requires hidden_size == latent_dim.") packet_dim = int(commit_sequence_dim or history_kv_rank) self.trajectory = GDN2TrajectoryMemory( hidden_size, probes=tape_probes, persistent_observation_dim=packet_dim, ) self.blocks = [ _CanvasBlock(hidden_size, num_heads, history_kv_rank, ffn_dim, local_attention_window, bool(working_last_block_global and i == num_layers - 1)) for i in range(num_layers) ] self.output_norm = RMSNorm(hidden_size) self.working_memory_bus = _WorkingBus( hidden_size, num_heads, num_working_readers, history_kv_rank, max_canvas_length, ) self.persistent_memory_bus = _PersistentBus( self.trajectory.persistent, num_persistent_readers, ) self.experience_in = nn.Linear(hidden_size * 3, packet_dim, bias=False) self.reason_embed = nn.Embedding(6, packet_dim) def __call__(self, token_embeddings: mx.array, memory: GDN2TrajectoryState, canvas_head: mx.array) -> mx.array: hidden = self.trajectory.read(memory, token_embeddings) offsets = (mx.arange(memory.seen.shape[1])[None, :] - canvas_head[:, None]) % memory.seen.shape[1] for block in self.blocks: hidden = block(hidden, memory.seen, offsets) return self.output_norm(hidden) * memory.seen[..., None].astype(hidden.dtype) def commit_write(self, *, memory: mx.array, working_state: mx.array, heavy_hidden: mx.array, committed_token_embeddings: mx.array, canvas_head: mx.array) -> mx.array: batch, canvas, width = working_state.shape count = committed_token_embeddings.shape[1] index = (mx.arange(count)[None, :] + canvas_head[:, None]) % canvas gather = mx.broadcast_to(index[..., None], (batch, count, width)) selected_working = mx.take_along_axis(working_state, gather, axis=1) selected_heavy = mx.take_along_axis(heavy_hidden, gather, axis=1) def normalize(value): fp32 = value.astype(mx.float32) return (fp32 * mx.rsqrt(mx.mean(mx.square(fp32), axis=-1, keepdims=True) + 1.0)).astype(value.dtype) roles = tuple(normalize(value) for value in ( selected_heavy, selected_working, committed_token_embeddings, )) packet = self.experience_in(mx.concatenate(roles, axis=-1)) # Schema25 maps both continuing commit reasons to embedding row 5. packet = packet + self.reason_embed.weight[5] valid = mx.ones((batch, count), mx.bool_) return self.trajectory.persistent.write_sequence(memory, packet, valid) def create_mlx_latent(config: Any) -> LatentDeliberationTransformer: readers = sum(kind == "full_attention" for kind in config.text_config.layer_types) return LatentDeliberationTransformer( hidden_size=config.text_config.hidden_size, latent_dim=config.latent_dim, ffn_dim=config.latent_ffn_dim, num_layers=config.latent_num_layers, num_heads=config.latent_num_heads, local_attention_window=config.latent_local_attention_window, tape_probes=config.latent_tape_probes, history_kv_rank=config.latent_history_kv_rank, num_working_readers=readers if config.working_memory_bus else 0, num_persistent_readers=readers if config.persistent_memory_bus else 0, working_last_block_global=config.latent_working_last_block_global, commit_sequence_dim=config.commit_sequence_dim, max_canvas_length=config.canvas_length, ) @dataclass class MLXCanvasOutput: heavy_hidden: mx.array working_state: mx.array class MLXModilifyMk2(nn.Module): """Shared DiffusionGemma trunk plus native commit-only trajectory module.""" def __init__(self, backbone: Any, config: Any) -> None: super().__init__() if int(config.state_schema_version) != 25: raise RuntimeError("Native MLX ModilifyMk2 requires schema25.") self.model = backbone.model self.latent_deliberation = create_mlx_latent(config) self.config = config def _merge_context(self, token_embeddings: mx.array, context: mx.array) -> mx.array: """Frozen self-conditioning bridge with the schema25 residual cap.""" mapper = self.model.decoder.self_conditioning normed = mapper.pre_norm(context.astype(token_embeddings.dtype)) mapped = mapper.down_proj( nn.gelu_approx(mapper.gate_proj(normed)) * mapper.up_proj(normed) ) mapped_fp32 = mapped.astype(mx.float32) energy = mx.mean(mx.square(mapped_fp32), axis=-1, keepdims=True) token_rms = mx.sqrt(mx.mean(mx.square(token_embeddings.astype(mx.float32)), axis=-1, keepdims=True)) cap = 0.5 * token_rms scale = cap / mx.sqrt(energy + mx.square(cap) + 1.0e-12) combined = token_embeddings + (mapped_fp32 * scale).astype(mapped.dtype) return mapper.post_norm(combined) def forward_canvas(self, canvas: mx.array, *, prefix_cache: list[Any], state: GDN2TrajectoryState, canvas_head: mx.array, first_step: bool = False, has_commits: bool = True) -> MLXCanvasOutput: """Denoise physical ring state through logical-order trunk positions.""" batch, length = canvas.shape decoder = self.model.decoder token_embeddings = decoder.embed_tokens(canvas) * decoder.embed_scale working = (mx.zeros_like(token_embeddings) if first_step else self.latent_deliberation(token_embeddings, state, canvas_head)) physical = (canvas_head[:, None] + mx.arange(length)[None, :]) % length hidden_width = token_embeddings.shape[-1] gather = mx.broadcast_to(physical[:, :, None], (batch, length, hidden_width)) logical_tokens = mx.take_along_axis(token_embeddings, gather, axis=1) logical_working = mx.take_along_axis(working, gather, axis=1) hidden = self._merge_context(logical_tokens, logical_working) prefix = int(prefix_cache[0].offset) mask = mx.ones((batch, prefix + length), mx.bool_) masks = decoder._make_decoder_masks(hidden, prefix_cache, mask) offset = prefix_cache[0].offset working_bus = self.latent_deliberation.working_memory_bus persistent_bus = self.latent_deliberation.persistent_memory_bus logical_seen = mx.take_along_axis(state.seen, physical, axis=1) working_kv = None if first_step else working_bus.prepare_kv(logical_working, logical_seen) persistent_kv = persistent_bus.prepare_kv(state.persistent, logical_seen) if has_commits else None reader = 0 for layer, layer_cache in zip(decoder.layers, prefix_cache): hidden = layer(hidden, masks[layer.layer_type], layer_cache, decoder=True, offset=offset) if layer.layer_type == "full_attention": if working_kv is not None and reader < working_bus.num_readers: hidden = working_bus.read(hidden, reader, *working_kv) if persistent_kv is not None and reader < persistent_bus.num_readers: hidden = persistent_bus.read(hidden, reader, *persistent_kv) reader += 1 hidden = decoder.norm(hidden) inverse = (mx.arange(length)[None, :] - canvas_head[:, None]) % length heavy = mx.take_along_axis(hidden, mx.broadcast_to(inverse[:, :, None], hidden.shape), axis=1) return MLXCanvasOutput(heavy, working) def _install_sliding_encoder_masks(encoder: Any) -> None: """Align mlx-vlm's full-length masks with trimmed sliding KV chunks.""" if getattr(encoder, "_modilify_sliding_masks", False): return original = encoder._make_encoder_masks window = int(encoder.text_config.sliding_window) def make_masks(h, cache, attention_mask=None, mm_token_type_ids=None): masks = original(h, cache, attention_mask, mm_token_type_ids) if isinstance(masks, dict): return masks query_length = int(h.shape[1]) maximum = window - 1 + query_length return [ mask[..., -maximum:] if ( layer.layer_type == "sliding_attention" and isinstance(mask, mx.array) and mask.shape[-1] > maximum ) else mask for layer, mask in zip(encoder.decoder.layers, masks) ] encoder._make_encoder_masks = make_masks encoder._modilify_sliding_masks = True def _read_json(path: Path) -> dict: return json.loads(path.read_text(encoding="utf-8")) class _ReleaseRouter(Router): """Route with the published model's FP32 probabilities and weights.""" def __call__(self, x: mx.array) -> tuple[mx.array, mx.array]: normalized = mx.fast.rms_norm(x.astype(mx.float32), None, self.eps).astype(x.dtype) scores = self.proj(normalized * self.scale * self._root_size) probabilities = mx.softmax(scores.astype(mx.float32), axis=-1, precise=True) k = self.config.top_k_experts indices = mx.argpartition(-probabilities, kth=k - 1, axis=-1)[..., :k] weights = mx.take_along_axis(probabilities, indices, axis=-1) weights = weights / mx.maximum(weights.sum(axis=-1, keepdims=True), 1.17549435e-38) return indices, weights * self.per_expert_scale[indices] class _ReleaseLayer(DecoderLayer): """Keep expert weighting in FP32 and return the sum in activation dtype.""" def __call__(self, x, mask=None, cache=None, *, decoder=False, offset=None, layer_scalar=None): attention = self.self_attn(self.input_layernorm(x), mask, cache, decoder=decoder, offset=offset) residual = x + self.post_attention_layernorm(attention) dense = self.post_feedforward_layernorm_1(self.mlp(self.pre_feedforward_layernorm(residual))) flat = residual.reshape(-1, residual.shape[-1]) indices, weights = self.router(flat) routed = self.experts(self.pre_feedforward_layernorm_2(flat), indices) routed = (routed * weights[..., None]).sum(axis=-2).astype(x.dtype).reshape(x.shape) routed = self.post_feedforward_layernorm_2(routed) output = residual + self.post_feedforward_layernorm(dense + routed) return output * (self.layer_scalar if layer_scalar is None else layer_scalar) def _create_model(config: dict) -> MLXModilifyMk2: text_config = dict(config["text_config"], model_type="diffusion_gemma_text") backbone = Model(ModelConfig.from_dict({ "model_type": "diffusion_gemma", "text_config": text_config, "vision_config": None, })) lora = config["lora_config"] targets = set(lora["target_modules"]) for layer in backbone.model.decoder.layers: layer.__class__ = _ReleaseLayer layer.router = _ReleaseRouter(layer.router.config) for parent in (layer.self_attn, layer.mlp): for name, module in list(parent.named_modules()): if name in targets: setattr(parent, name, LoRALinear.from_base( module, r=lora["r"], scale=lora["alpha"] / lora["r"], )) for name in ("gate_up_proj", "down_proj"): module = getattr(layer.experts, name) setattr(layer.experts, name, LoRASwitchLinear.from_base( module, r=lora["expert_r"], scale=lora["expert_alpha"] / lora["expert_r"], )) settings = SimpleNamespace(**dict(config, text_config=SimpleNamespace(**text_config))) model = MLXModilifyMk2(backbone, settings) _install_sliding_encoder_masks(model.model.encoder) model.eval() model.freeze() return model def _published_name(native: str) -> tuple[str, bool]: """Return the release key and whether it requires dense LoRA transposition.""" if not native.startswith("model.decoder.layers."): return native, False if ".experts." in native: for projection, role in (("gate_up_proj", "gate_up"), ("down_proj", "down")): prefix = f".experts.{projection}." if prefix in native: if native.endswith(".linear.weight"): return native.replace(prefix + "linear.weight", f".experts.{projection}"), False return native.replace(prefix + "lora_", f".experts.lora_{role}_"), False return native.replace(".linear.weight", ".base.weight"), ".lora_" in native def _expected_dtype(name: str) -> mx.Dtype: if name.startswith("latent_deliberation.") and ( name.endswith((".dt_bias", ".a_log")) or (name.endswith(".weight") and "norm" in name.rsplit(".", 2)[-2]) ): return mx.float32 return mx.bfloat16 def _verify_shard(path: Path, info: dict) -> None: if path.stat().st_size != info["bytes"]: raise ValueError(f"Incorrect shard size: {path.name}") with path.open("rb") as source: if hashlib.file_digest(source, "sha256").hexdigest() != info["sha256"]: raise ValueError(f"Incorrect shard SHA-256: {path.name}") def load_model(root: Path) -> MLXModilifyMk2: config = _read_json(root / "config.json") manifest = _read_json(root / "export_manifest.json") protocol = {"state_schema_version": 25, "memory_architecture": "compact_gdn2_v2", "memory_scheme": "dual_timescale_gdn2_trajectory_memory"} if (manifest.get("format") != "modilify_mk2_mlx_inference_v1" or any(config.get(k) != v or manifest.get(k) != v for k, v in protocol.items()) or config.get("checkpoint_step") != 4400 or manifest.get("global_step") != 4400 or config.get("canvas_length") != 256 or config.get("dtype") != "bfloat16" or manifest.get("precision_policy") != "gdn2_small_fp32_v1"): raise ValueError("Expected the step 4400 schema25 compact_gdn2_v2 MLX release.") if not mx.metal.is_available(): raise RuntimeError("Modilify Mk2 MLX requires an Apple Silicon Metal device.") model = _create_model(config) weight_map = _read_json(root / "model.safetensors.index.json")["weight_map"] tensors = {} for native, value in tree_flatten(model.parameters()): name, transpose = _published_name(native) if name in tensors: raise ValueError(f"Duplicate release weight mapping: {name}") tensors[name] = (native, tuple(value.shape), _expected_dtype(name), transpose) if set(tensors) != set(weight_map) or len(tensors) != manifest.get("total_tensor_count") or len(tensors) != 1323: raise ValueError( f"Weight index mismatch. Missing: {sorted(set(tensors) - set(weight_map))[:8]}. " f"Unexpected: {sorted(set(weight_map) - set(tensors))[:8]}." ) by_shard = defaultdict(set) for name, shard in weight_map.items(): by_shard[shard].add(name) shards = {entry["file"]: entry for entry in manifest["shards"]} expected_files = {f"model-{i:05d}-of-00016.safetensors" for i in range(1, 17)} if set(shards) != expected_files or set(by_shard) != expected_files or len(manifest["shards"]) != 16: raise ValueError("Release manifest and index must describe all 16 shards.") counts = Counter() for shard in sorted(by_shard): path = root / shard _verify_shard(path, shards[shard]) raw = mx.load(str(path)) if set(raw) != by_shard[shard]: raise ValueError(f"Weight index and shard disagree: {shard}") updates = [] for name, value in raw.items(): native, shape, dtype, transpose = tensors[name] if transpose: value = value.T if value.shape != shape or value.dtype != dtype: raise ValueError(f"Invalid shape or dtype for {name}: {value.shape}, {value.dtype}") updates.append((native, value)) counts["float32" if dtype == mx.float32 else "bfloat16"] += 1 model.load_weights(updates, strict=False) mx.eval(*(value for _, value in updates)) del raw, updates if counts != {"bfloat16": 1297, "float32": 26} or counts != manifest.get("tensor_dtypes"): raise ValueError(f"Unexpected release precision counts: {dict(counts)}") return model