"""Temporal/depth autoregression over two audio tracks and two text tracks. A frame holds `codebooks` audio codes for each of the two tracks, track 1 then track 2. In every track the acoustic codebooks run one frame behind codebook 0 inside the model. Callers pass and receive canonical undelayed frames. Text runs at a variable rate (training readiness owns the design). Sequence step s is one temporal position for the audio, followed by the text tokens of frame s of each present text track, track 1 first. A track's text ends with the advance token, which is only ever a target: each text event is predicted from the last input before it, and the step's last position conditions the depth transformer. """ from dataclasses import dataclass, replace import numpy as np import torch from torch import nn from torch.nn import functional as F from .kernels import depth_attention TRACKS = 2 # Moshi's SentencePiece ids are 0..31999; the advance token follows them. TEXT_IDS = 32000 ADVANCE = TEXT_IDS # Sampling forces the advance after this many tokens of one track in one frame. MAX_TEXT_PER_FRAME = 16 # Moshi's equation 7: codebook 0 of each track against 1 for each acoustic codebook, # and its padding weight for the advance token. SEMANTIC_WEIGHT = 100.0 ADVANCE_WEIGHT = 0.5 TEXT_OPTIONS = {"none": (), "one": (1,), "two": (0, 1)} # Text against audio per track, in 80 ms frames: a word starting at frame f is # written on track p at frame f - TEXT_SHIFT[p]. Positive puts the text ahead of the # audio, negative behind it, 0 at the word's start. Track 1 (listened to) is 480 ms # behind, a transcript; track 2 (spoken as) writes words where they start. The shift # follows the track after the speaker swap; windows apply it, generation follows it. # The default of `Config.text_shift`, which each checkpoint records. TEXT_SHIFT = (-6, 0) # Text output heads: one shared by both tracks, or one per track. TEXT_HEADS = ("shared", "per-track") @dataclass(frozen=True) class Config: layers: int = 12 width: int = 768 heads: int = 6 depth_layers: int = 2 depth_width: int = 256 depth_heads: int = 4 # Temporal positions per window: frames plus text tokens, plus one flush step. context: int = 4096 codebooks: int = 8 cardinality: int = 2048 fps: float = 12.5 # The training option; the parameters are the same for all three. text: str = "none" # Frames each track's text sits from its word start (see TEXT_SHIFT). text_shift: tuple = TEXT_SHIFT text_heads: str = "shared" @property def total_codebooks(self): return self.codebooks * TRACKS @property def text_tracks(self): return TEXT_OPTIONS[self.text] def __post_init__(self): for width, heads in ( (self.width, self.heads), (self.depth_width, self.depth_heads), ): if heads < 1 or width % heads or (width // heads) % 2: raise ValueError("Head dimensions must be even") if ( min( self.layers, self.depth_layers, self.context, self.codebooks, self.cardinality, ) < 1 ): raise ValueError("Dimensions must be positive") if self.fps <= 0: raise ValueError("fps must be positive") if self.text not in TEXT_OPTIONS: raise ValueError(f"text must be one of {sorted(TEXT_OPTIONS)}") # A config read from JSON holds a list. object.__setattr__(self, "text_shift", tuple(int(x) for x in self.text_shift)) if len(self.text_shift) != TRACKS: raise ValueError("text_shift has one value per track") if self.text_heads not in TEXT_HEADS: raise ValueError(f"text_heads must be one of {TEXT_HEADS}") # Sequence order of the codebooks and text. Checkpoints carry it; a model trained # under one factorization is meaningless under another. def factorization(config): return "two-tracks-delay-0-1-1-1-1-1-1-1-text-advance-shift" + "".join( f"{shift:+d}" for shift in config.text_shift ) FACTORIZATION = factorization(Config()) def tiny_config(text="none"): # Real codec vocabulary, small transformers. No fake-codec checkpoint confusion. return Config( layers=2, width=32, heads=2, depth_layers=1, depth_width=32, depth_heads=2, context=64, text=text, ) def norm(x): return F.rms_norm(x, (x.shape[-1],)) def rotary(positions, head_width): """Cos and sin tables [positions, head_width / 2], built once per model.""" freq = torch.arange(0, head_width, 2, dtype=torch.float32) / head_width angles = torch.outer(torch.arange(positions), 100000.0**-freq) return torch.stack((angles.cos(), angles.sin())) def rotate(y, cos, sin): a, b = y.chunk(2, dim=-1) return torch.cat((a * cos + b * sin, -a * sin + b * cos), dim=-1) class Attention(nn.Module): def __init__(self, width, heads): super().__init__() self.heads = heads self.c_q = nn.Linear(width, width, bias=False) self.c_k = nn.Linear(width, width, bias=False) self.c_v = nn.Linear(width, width, bias=False) self.c_proj = nn.Linear(width, width, bias=False) def forward(self, x, rope): b, t, w = x.shape d = w // self.heads q, k, v = [ layer(x).reshape(b, t, self.heads, d).transpose(1, 2) for layer in (self.c_q, self.c_k, self.c_v) ] cos, sin = rope[0, :t].to(x.dtype), rope[1, :t].to(x.dtype) q, k = norm(rotate(q, cos, sin)) * 1.2, norm(rotate(k, cos, sin)) * 1.2 y = F.scaled_dot_product_attention(q, k, v, is_causal=True) return self.c_proj(y.transpose(1, 2).reshape(x.shape)) class MLP(nn.Module): def __init__(self, width): super().__init__() self.c_fc = nn.Linear(width, 4 * width, bias=False) self.c_proj = nn.Linear(4 * width, width, bias=False) def forward(self, x): return self.c_proj(F.relu(self.c_fc(x)).square()) class Block(nn.Module): def __init__(self, width, heads): super().__init__() self.attn = Attention(width, heads) self.mlp = MLP(width) def forward(self, x, rope): x = x + self.attn(norm(x), rope) return x + self.mlp(norm(x)) def per_position(x, weight): """x [rows, positions, in] through weight [positions, out, in]: one matrix per position.""" y = torch.bmm(x.transpose(0, 1), weight.transpose(1, 2).to(x.dtype)) return y.transpose(0, 1) class DepthBlock(nn.Module): """A depth block with its own weights at every codebook position. Same widths and arithmetic as `Block`; each [positions, out, in] matrix holds one nn.Linear weight per position, initialised as nn.Linear initialises. """ def __init__(self, positions, width, heads): super().__init__() self.heads = heads def weight(out, inner): bound = inner**-0.5 return nn.Parameter( torch.empty(positions, out, inner).uniform_(-bound, bound) ) self.c_q, self.c_k, self.c_v, self.c_proj = ( weight(width, width) for _ in range(4) ) self.c_fc = weight(4 * width, width) self.c_out = weight(width, 4 * width) def forward(self, x, rope): n, p, w = x.shape d = w // self.heads h = norm(x) q, k, v = [ per_position(h, m).reshape(n, p, self.heads, d) for m in (self.c_q, self.c_k, self.c_v) ] cos = rope[0, :p].to(x.dtype).unsqueeze(1) sin = rope[1, :p].to(x.dtype).unsqueeze(1) q, k = norm(rotate(q, cos, sin)) * 1.2, norm(rotate(k, cos, sin)) * 1.2 # Rotation and QK norm run at the residual precision; the kernel takes # the projection precision. y = depth_attention(q.to(v.dtype), k.to(v.dtype), v) x = x + per_position(y.reshape(n, p, w), self.c_proj) h = F.relu(per_position(norm(x), self.c_fc)).square() return x + per_position(h, self.c_out) @dataclass class TextLayout: """Where text sits in a padded batch, built by `text_layout` and `collate_text`. source [batch, positions] indexes the concatenation of the audio steps [frames + 1], the text tokens [tokens] and one zero row for padding. token and track [batch, tokens] are the text tokens' ids and tracks. last [batch, frames + 1] is the position whose output conditions each step's depth. event_* [batch, events] are the text predictions: the position predicting it, its track, its target id and its loss weight (0 for padding and the flush step). pattern [batch] is the bitmask of present text tracks. """ source: torch.Tensor token: torch.Tensor track: torch.Tensor last: torch.Tensor event_position: torch.Tensor event_track: torch.Tensor event_target: torch.Tensor event_weight: torch.Tensor pattern: torch.Tensor def text_layout(frames, text): """One window's text layout as numpy arrays; text is per track None (absent) or an int array [n, 2] of (frame, id) rows in word order, frames in [0, frames).""" present = [j for j in range(TRACKS) if text[j] is not None] counts = np.zeros((frames + 1, TRACKS), dtype=np.int64) parts = [] for j in present: rows = np.asarray(text[j], dtype=np.int64).reshape(-1, 2) if len(rows) and (rows[:, 0].min() < 0 or rows[:, 0].max() >= frames): raise ValueError("Text frame outside the window") if len(rows) and (rows[:, 1].min() < 0 or rows[:, 1].max() >= TEXT_IDS): raise ValueError("Text id outside the vocabulary") np.add.at(counts[:, j], rows[:, 0], 1) parts.append( np.column_stack( (rows[:, 0], np.full(len(rows), j), np.arange(len(rows)), rows[:, 1]) ) ) tokens = np.concatenate(parts) if parts else np.zeros((0, 4), dtype=np.int64) # Frame, then track, then word order. tokens = tokens[np.lexsort((tokens[:, 2], tokens[:, 1], tokens[:, 0]))] per_step = counts.sum(1) start = np.concatenate(([0], np.cumsum(1 + per_step)[:-1])) before = np.concatenate(([0], np.cumsum(per_step))) positions = int(frames + 1 + per_step.sum()) rank = np.arange(len(tokens)) - before[tokens[:, 0]] token_position = start[tokens[:, 0]] + 1 + rank source = np.arange(frames + 1 + len(tokens)) order = np.empty(positions, dtype=np.int64) order[start] = np.arange(frames + 1) order[token_position] = source[frames + 1 :] # Events: each token from the input before it; each advance from the last input # of its track in the step, or where the track's first token would come from. position = [token_position - 1] track = [tokens[:, 1]] target = [tokens[:, 3]] weight = [np.ones(len(tokens))] written = np.zeros(frames + 1, dtype=np.int64) for j in present: written = written + counts[:, j] position.append(start + written) track.append(np.full(frames + 1, j)) target.append(np.full(frames + 1, ADVANCE)) weight.append(np.r_[np.full(frames, ADVANCE_WEIGHT), 0.0]) position, track, target, weight = ( np.concatenate(x) for x in (position, track, target, weight) ) # Sampling order: by position, then track. events = np.lexsort((track, position)) return { "source": order, "token": tokens[:, 3], "track": tokens[:, 1], "last": start + per_step, "event_position": position[events], "event_track": track[events], "event_target": target[events], "event_weight": weight[events].astype(np.float32), "pattern": sum(1 << j for j in present), } def collate_text(layouts, frames, positions, device): """Pads per-window layouts to [batch, positions] and [batch, frames + 1].""" b = len(layouts) tokens = max(1, max(len(x["token"]) for x in layouts)) events = max(1, max(len(x["event_target"]) for x in layouts)) pad = frames + 1 + tokens source = torch.full((b, positions), pad, dtype=torch.long) token = torch.zeros((b, tokens), dtype=torch.long) track = torch.zeros((b, tokens), dtype=torch.long) last = torch.zeros((b, frames + 1), dtype=torch.long) event = [torch.zeros((b, events), dtype=torch.long) for _ in range(3)] event_weight = torch.zeros((b, events)) for i, x in enumerate(layouts): steps, n, m = len(x["last"]), len(x["token"]), len(x["event_target"]) if len(x["source"]) > positions or steps > frames + 1: raise ValueError("Window exceeds the padded batch") s = torch.as_tensor(x["source"]) # Token sources are rebased past this batch's padded audio steps. source[i, : len(s)] = torch.where(s < steps, s, s - steps + frames + 1) token[i, :n] = torch.as_tensor(x["token"]) track[i, :n] = torch.as_tensor(x["track"]) last[i, :steps] = torch.as_tensor(x["last"]) for tensor, key in zip( event, ("event_position", "event_track", "event_target"), strict=True ): tensor[i, :m] = torch.as_tensor(x[key]) event_weight[i, :m] = torch.as_tensor(x["event_weight"]) pattern = torch.tensor([x["pattern"] for x in layouts]) return TextLayout( source.to(device), token.to(device), track.to(device), last.to(device), *(t.to(device) for t in event), event_weight.to(device), pattern.to(device), ) class DuplexModel(nn.Module): def __init__(self, config: Config): super().__init__() self.config = config c = config q = c.total_codebooks # One table per role with codebook offsets: a single gather instead of # one per codebook. Each track has its own rows. self.frame_embeddings = nn.EmbeddingBag(q * c.cardinality, c.width, mode="sum") # A zero BOS stays zero through bias-free blocks, amplifying RMSNorm's # derivative at zero until deep models overflow on the first backward. self.bos = nn.Parameter(torch.randn(c.width)) # Text: one token table for both tracks, a vector per track, an embedding of # which text tracks are present added to every audio position (patterns 1 to # 3; none adds nothing), and a query per track before the shared head. self.text_embeddings = nn.Embedding(TEXT_IDS, c.width) self.text_tracks = nn.Parameter(torch.zeros(TRACKS, c.width)) self.text_presence = nn.Parameter(torch.zeros(2**TRACKS - 1, c.width)) self.text_queries = nn.Parameter(torch.zeros(TRACKS, c.width)) if c.text_heads == "shared": self.text_head = nn.Linear(c.width, TEXT_IDS + 1, bias=False) else: # [track, events, width], initialised as nn.Linear initialises one head. bound = c.width**-0.5 self.text_head = nn.Parameter( torch.empty(TRACKS, TEXT_IDS + 1, c.width).uniform_(-bound, bound) ) self.temporal = nn.ModuleList( [Block(c.width, c.heads) for _ in range(c.layers)] ) self.to_depth = nn.Linear(c.width, c.depth_width, bias=False) self.depth_embeddings = nn.Embedding((q - 1) * c.cardinality, c.depth_width) self.depth_positions = nn.Parameter(torch.zeros(q, c.depth_width)) self.depth = nn.ModuleList( [DepthBlock(q, c.depth_width, c.depth_heads) for _ in range(c.depth_layers)] ) # Per-codebook output heads [codebook, cardinality, depth_width]. bound = c.depth_width**-0.5 self.outputs = nn.Parameter( torch.empty(q, c.cardinality, c.depth_width).uniform_(-bound, bound) ) self.register_buffer( "offsets", torch.arange(q) * c.cardinality, persistent=False ) # Codebook 0 of each track has delay zero; its acoustic codebooks one. self.delays = [int(k % c.codebooks != 0) for k in range(q)] self.register_buffer( "semantic", torch.tensor([d == 0 for d in self.delays]), persistent=False, ) self.register_buffer( "track_of", torch.arange(q) // c.codebooks, persistent=False ) self.register_buffer( "temporal_rope", # One position more than the context flushes the last acoustic frame. rotary(c.context + 1, c.width // c.heads), persistent=False, ) self.register_buffer( "depth_rope", rotary(q, c.depth_width // c.depth_heads), persistent=False, ) def delayed(self, tokens, valid=None): """Token ids and presence for sequence steps 0..frames, [batch, frames + 1, codebooks]. Step s holds codebook 0 of frame s and the acoustic codebooks of frame s - 1 in every track. valid is [batch, frames, tracks]: a token is absent before the first frame, after the last, on padded frames and on absent tracks. """ b, t, _ = tokens.shape if valid is None: valid = tokens.new_ones((b, t, TRACKS), dtype=torch.bool) valid = valid[..., self.track_of] ids = torch.where( self.semantic, F.pad(tokens, (0, 0, 0, 1)), F.pad(tokens, (0, 0, 1, 0)) ) present = torch.where( self.semantic, F.pad(valid, (0, 0, 0, 1)), F.pad(valid, (0, 0, 1, 0)), ) return ids, present def audio_steps(self, tokens, valid=None): """Temporal inputs of the audio positions, [batch, frames + 1, width]: BOS, then the sum of every present token of the step before.""" b, t, q = tokens.shape if q != self.config.total_codebooks: raise ValueError("Expected [batch, frames, codebooks x tracks]") x = self.bos.expand(b, 1, -1) if t: ids, present = self.delayed(tokens, valid) weights = self.frame_embeddings.weight embedded = self.frame_embeddings( (ids[:, :t] + self.offsets).reshape(-1, q), per_sample_weights=present[:, :t].reshape(-1, q).to(weights.dtype), ) x = torch.cat((x, embedded.reshape(b, t, -1)), dim=1) return x def temporal_outputs(self, tokens, valid=None, text=None): """Normalised temporal outputs [batch, positions, width].""" x = self.audio_steps(tokens, valid) if text is not None: b, _, w = x.shape presence = torch.cat((x.new_zeros(1, w), self.text_presence)) x = x + presence[text.pattern][:, None] words = self.text_embeddings(text.token) + self.text_tracks[text.track] table = torch.cat((x, words.to(x.dtype), x.new_zeros(b, 1, w)), dim=1) x = table.gather(1, text.source[..., None].expand(-1, -1, w)) if x.shape[1] > self.config.context + 1: raise ValueError("Window exceeds the temporal context") x = norm(x) for block in self.temporal: x = block(x, self.temporal_rope) return norm(x) def text_logits(self, hidden, track): """Logits [..., TEXT_IDS + 1] of text events from normalised temporal outputs.""" x = norm(hidden + self.text_queries[track].to(hidden.dtype)) if self.config.text_heads == "shared": return self.text_head(x) weights = self.text_head.to(x.dtype) track = torch.as_tensor(track, device=x.device) if track.dim() == 0: return x @ weights[track].T # Both heads for every event, each event keeping its own track's: fixed # shapes, so the compiled step stays fast (mask selection has # data-dependent shapes and trained at about half the step rate). return torch.where(track[..., None] == 0, x @ weights[0].T, x @ weights[1].T) def depth_logits(self, context, inputs): """Logits [batch, steps, codebooks, cardinality] in delayed step order. inputs is [batch, steps, codebooks - 1, depth_width]: the embedding of the token before each depth position, zero where there is none. """ b, s, _ = context.shape q = self.config.total_codebooks x = context.unsqueeze(2) + self.depth_positions x = torch.cat((x[:, :, :1], x[:, :, 1:] + inputs), dim=2) x = x.reshape(b * s, q, -1) for block in self.depth: x = block(x, self.depth_rope) # [codebook, steps, width] x [codebook, width, cardinality] x = norm(x).transpose(0, 1) logits = torch.bmm(x, self.outputs.transpose(1, 2).to(x.dtype)) return logits.transpose(0, 1).reshape(b, s, q, -1) def forward(self, tokens, valid=None, text=None): """Canonical frames in; audio logits aligned to the same [batch, frames, codebooks], and text logits [batch, events, TEXT_IDS + 1] or None. valid is the [batch, frames, tracks] mask of real frames of present tracks. text is a `TextLayout` or None for audio only. Fillers never appear as targets: codebook 0 of frame t is read from step t and its acoustic codebooks from step t + 1, so the final step flushes the last frame and every source token is scored once. """ hidden = self.temporal_outputs(tokens, valid, text) steps = hidden if text is not None: w = hidden.shape[-1] steps = hidden.gather(1, text.last[..., None].expand(-1, -1, w)) context = self.to_depth(steps) ids, present = self.delayed(tokens, valid) # Each depth position after the first takes the token before it in # step order. The step that flushes a short row's last frame sees no # codebook 0, as in generation, instead of batch padding. inputs = self.depth_embeddings(ids[:, :, :-1] + self.offsets[:-1]) inputs = inputs * present[:, :, :-1, None].to(inputs.dtype) logits = self.depth_logits(context, inputs) audio = torch.where(self.semantic[:, None], logits[:, :-1], logits[:, 1:]) if text is None: return audio, None w = hidden.shape[-1] events = hidden.gather(1, text.event_position[..., None].expand(-1, -1, w)) return audio, self.text_logits(events, text.event_track) @torch.no_grad() def generate(self, prefix, frames, temperature=1.0, top_k=250): """Reference audio sampler without KV cache or text; returns canonical tokens.""" if frames < 0 or temperature < 0 or top_k < 0: raise ValueError("frames, temperature and top_k must be nonnegative") c = self.config b, forced, q = prefix.shape result = torch.cat((prefix, prefix.new_zeros((b, frames, q))), dim=1) if not frames: return result # One step past the last frame flushes its acoustic codebooks. for step in range(forced, forced + frames + 1): window = result[:, max(0, step - c.context) : step] context = self.to_depth(self.temporal_outputs(window)[:, -1:]) inputs = context.new_zeros((b, 1, q - 1, c.depth_width)) for position, delay in enumerate(self.delays): frame = step - delay if not 0 <= frame < forced + frames: continue # The prefix keeps its own tokens, including the acoustic # codebooks of its last frame; later frames are sampled. if frame >= forced: logits = self.depth_logits(context, inputs)[:, 0, position].float() result[:, frame, position] = sample( logits, temperature, top_k ).squeeze(-1) if position < q - 1: inputs[:, 0, position] = self.depth_embeddings( result[:, frame, position] + self.offsets[position] ) return result def sample(logits, temperature, top_k=0, generator=None): """One draw per row of [..., classes] logits; top_k 0 keeps every class.""" if temperature == 0: return logits.argmax(-1, keepdim=True) if top_k and top_k < logits.shape[-1]: floor = logits.topk(top_k).values[..., -1:] logits = logits.masked_fill(logits < floor, float("-inf")) probabilities = (logits / temperature).softmax(-1) return torch.multinomial( probabilities.reshape(-1, logits.shape[-1]), 1, generator=generator ).reshape(*logits.shape[:-1], 1) def from_legacy(config, state): """A model in this format from a 14 November checkpoint (one or two streams, shared depth weights, no text). A two-stream checkpoint converts exactly: the shared depth block weights are copied to every position and the new text weights stay unused while no text is present. A one-stream checkpoint is widened as two-stream training began: each track gets a copy of the stream's embeddings, positions and heads, and track 1's codebook 7 feeding track 2's codebook 0 starts at zero. """ config = dict(config) config.pop("factorization", None) streams = config.pop("streams", 1) if streams not in (1, 2): raise ValueError("Expected a one- or two-stream checkpoint") model = DuplexModel(Config(**config)) c = model.config new = model.state_dict() state = dict(state) if streams == 1: for name in ("frame_embeddings.weight", "depth_positions", "outputs"): value = state[name] state[name] = value.repeat(TRACKS, *[1] * (value.dim() - 1)) value = state["depth_embeddings.weight"] joint = value.new_zeros((c.cardinality, value.shape[1])) state["depth_embeddings.weight"] = torch.cat([value, joint, value]) names = { "attn.c_q": "c_q", "attn.c_k": "c_k", "attn.c_v": "c_v", "attn.c_proj": "c_proj", "mlp.c_fc": "c_fc", "mlp.c_proj": "c_out", } for layer in range(c.depth_layers): for old, name in names.items(): weight = state.pop(f"depth.{layer}.{old}.weight") state[f"depth.{layer}.{name}"] = weight.repeat(c.total_codebooks, 1, 1) for name, value in new.items(): if name.startswith("text_"): state[name] = value model.load_state_dict(state, strict=True) return model def with_text(model, text): """The same weights under another training option, such as a later stage's.""" widened = DuplexModel(replace(model.config, text=text)).to(model.bos.device) widened.load_state_dict(model.state_dict(), strict=True) return widened