"""Incremental sampling for one DuplexModel sequence, for serving. `DuplexModel.generate` recomputes the whole temporal context for every sequence step. `Sampler` keeps a key/value cache instead, so a step costs its temporal positions and the depth positions it needs. It computes the same distribution as training: step s runs the audio position, then each present text track's tokens of frame s until its advance token, then predicts codebook 0 of frame s and the acoustic codebooks of frame s - 1 in every track from the step's last position. The text written on track p at step s is the words starting at frame s + the checkpoint's `Config.text_shift[p]`, as in training windows: track 2, spoken as, at their start, and track 1, listened to, a few frames after it (6 by default). Each audio position of a step is forced to a given token, sampled, left absent, or supplied later. A later position must follow every sampled position of its step, because no prediction in the step may depend on it; it must be supplied with `complete` before the next step. Live serving forces the caller's track this way while it arrives, frame by frame. Text never depends on the step's audio. """ import numpy as np import torch from torch.nn import functional as F from .network import ADVANCE, MAX_TEXT_PER_FRAME, TRACKS, norm, rotate, sample SAMPLE, LATER, ABSENT = -1, -2, -3 # The sampling standard: audio at temperature 1 from the 250 most likely codes; # text at temperature 1 from every token. TEMPERATURE, TOP_K = 1.0, 250 def _project(attention, x, rope, position): """Rotated, normalised queries and keys and the values for one position per row. x is [rows, width] and position an int or a [rows] tensor; each result is [rows, heads, 1, head_width]. """ head_width = x.shape[-1] // attention.heads q, k, v = [ layer(x).reshape(x.shape[0], attention.heads, 1, head_width) for layer in (attention.c_q, attention.c_k, attention.c_v) ] cos = rope[0, position].to(x.dtype) sin = rope[1, position].to(x.dtype) if cos.dim() == 2: cos, sin = cos[:, None, None], sin[:, None, None] return norm(rotate(q, cos, sin)) * 1.2, norm(rotate(k, cos, sin)) * 1.2, v class _Cache: """Keys and values of one transformer stack, [layers, rows, heads, positions, head_width].""" def __init__(self, layers, heads, width, positions, reference, rows=1): shape = (layers, rows, heads, positions, width // heads) self.keys = reference.new_zeros(shape) self.values = reference.new_zeros(shape) def attend(self, i, q, k, v, position): self.keys[i, :, :, position] = k[:, :, 0].to(self.keys.dtype) self.values[i, :, :, position] = v[:, :, 0].to(self.values.dtype) keys = self.keys[i, :, :, : position + 1].to(q.dtype) values = self.values[i, :, :, : position + 1].to(q.dtype) return F.scaled_dot_product_attention(q, keys, values) def run(self, blocks, x, rope, position): """One temporal position through every block, attending to it and those before.""" for i, block in enumerate(blocks): q, k, v = _project(block.attn, norm(x), rope, position) y = self.attend(i, q, k, v, position) x = x + block.attn.c_proj(y.reshape(x.shape)) x = x + block.mlp(norm(x)) return x def run_rows(self, blocks, x, rope, positions, length): """One temporal position per row, each at its own position [rows], attending to its own earlier positions; length bounds every row's position + 1.""" rows = torch.arange(x.shape[0], device=x.device) mask = torch.arange(length, device=x.device) <= positions[:, None] mask = mask[:, None, None] for i, block in enumerate(blocks): q, k, v = _project(block.attn, norm(x), rope, positions) self.keys[i, rows, :, positions] = k[:, :, 0].to(self.keys.dtype) self.values[i, rows, :, positions] = v[:, :, 0].to(self.values.dtype) keys = self.keys[i, :, :, :length].to(q.dtype) values = self.values[i, :, :, :length].to(q.dtype) y = F.scaled_dot_product_attention(q, keys, values, attn_mask=mask) x = x + block.attn.c_proj(y.reshape(x.shape)) x = x + block.mlp(norm(x)) return x def run_depth(self, blocks, x, rope, position): """One depth position through every block, with that position's own weights.""" for i, block in enumerate(blocks): heads = block.heads d = x.shape[-1] // heads h = norm(x) q, k, v = [ (h @ m[position].T.to(h.dtype)).reshape(x.shape[0], heads, 1, d) for m in (block.c_q, block.c_k, block.c_v) ] cos = rope[0, position].to(x.dtype) sin = rope[1, position].to(x.dtype) q, k = norm(rotate(q, cos, sin)) * 1.2, norm(rotate(k, cos, sin)) * 1.2 y = self.attend(i, q, k, v, position).reshape(x.shape) x = x + y @ block.c_proj[position].T.to(y.dtype) h = F.relu(norm(x) @ block.c_fc[position].T.to(x.dtype)).square() x = x + h @ block.c_out[position].T.to(h.dtype) return x class Sampler: """One sequence, one step at a time. Up to `context` + 1 positions; no rollover. text names the present text tracks (0 for track 1, 1 for track 2); the default is the checkpoint's training option. With `record`, `log` collects the log probability of every text event and every audio token the steps produce or are given, in sampling order: ("text", track, id, logp) and ("audio", position, id, logp). """ def __init__( self, model, temperature=TEMPERATURE, generator=None, top_k=TOP_K, text=None, record=False, text_sampling=None, ): if temperature < 0 or top_k < 0: raise ValueError("temperature and top_k must be nonnegative") self.model = model self.temperature = temperature # 0 samples from every code; k samples from the k most likely at each position. self.top_k = top_k # Text's (temperature, top_k); default the audio temperature from every token. self.text_sampling = text_sampling or (temperature, 0) self.generator = generator c = model.config self.text = tuple(model.config.text_tracks if text is None else text) if any(j not in range(TRACKS) for j in self.text): raise ValueError("Text tracks are 0 and 1") self.pattern = sum(1 << j for j in self.text) self.q = c.total_codebooks weights = model.bos.detach() self.temporal = _Cache(c.layers, c.heads, c.width, c.context + 1, weights) self.steps = 0 self.positions = 0 # Tokens of the previous step in delayed order; ABSENT where none. self.previous = None self.log = [] if record else None def room(self): """Whether one more step fits in the context at the per-frame text limit.""" need = 1 + len(self.text) * MAX_TEXT_PER_FRAME return self.positions + need <= self.model.config.context + 1 def _position(self, x): if self.positions > self.model.config.context: raise ValueError("Sequence exceeds the model context") m = self.model h = self.temporal.run(m.temporal, norm(x), m.temporal_rope, self.positions) self.positions += 1 return norm(h) @torch.no_grad() def step(self, forced, text=None): """Run the next step. forced has one entry per codebook position: a token, SAMPLE, LATER or ABSENT. Acoustic positions of step 0 are absent whatever is passed. text has one entry per present text track: SAMPLE, or the list of that frame's given token ids (the advance is implied); default all SAMPLE. Returns the step's audio tokens, with LATER where still to be supplied, and the text tokens written on each present track.""" m = self.model if self.previous is not None and LATER in self.previous: raise ValueError("Complete the previous step first") tokens = [int(t) for t in forced] if len(tokens) != self.q: raise ValueError(f"Expected {self.q} positions") if self.steps == 0: tokens = [t if d == 0 else ABSENT for t, d in zip(tokens, m.delays)] text = [SAMPLE] * len(self.text) if text is None else list(text) if len(text) != len(self.text): raise ValueError("One text entry per present text track") sampled = [k for k, t in enumerate(tokens) if t == SAMPLE] later = [k for k, t in enumerate(tokens) if t == LATER] if later and sampled and min(later) < max(sampled): raise ValueError("A later position precedes a sampled one") if self.previous is None: x = m.bos[None] else: present = [k for k, t in enumerate(self.previous) if t >= 0] ids = torch.tensor([self.previous[k] for k in present], device=m.bos.device) rows = m.frame_embeddings.weight[ids + m.offsets[present]] x = rows.sum(0, keepdim=True).to(m.bos.dtype) if self.pattern: x = x + m.text_presence[self.pattern - 1].to(x.dtype) h = self._position(x) written = [] for track, given in zip(self.text, text, strict=True): out = [] while True: logits = m.text_logits(h, torch.tensor(track, device=h.device))[0] logits = logits.float() if given == SAMPLE: if len(out) == MAX_TEXT_PER_FRAME: token = ADVANCE else: token = int(sample(logits, *self.text_sampling, self.generator)) else: token = given[len(out)] if len(out) < len(given) else ADVANCE if self.log is not None: self.log.append( ("text", track, token, float(logits.log_softmax(-1)[token])) ) if token == ADVANCE: break out.append(token) e = m.text_embeddings.weight[token] + m.text_tracks[track] h = self._position(e[None].to(h.dtype)) written.append(out) context = m.to_depth(h) scored = [k for k, t in enumerate(tokens) if t >= 0 or t == SAMPLE] last = max(scored if self.log is not None else sampled, default=-1) if last >= 0: c = m.config depth = _Cache( c.depth_layers, c.depth_heads, c.depth_width, self.q, context ) for position in range(last + 1): x = context + m.depth_positions[position] if position and tokens[position - 1] >= 0: x = x + m.depth_embeddings.weight[ tokens[position - 1] + m.offsets[position - 1] ].to(x.dtype) x = depth.run_depth(m.depth, x, m.depth_rope, position) if tokens[position] != SAMPLE and ( self.log is None or tokens[position] < 0 ): continue logits = (norm(x) @ m.outputs[position].T.to(x.dtype))[0].float() if tokens[position] == SAMPLE: tokens[position] = int( sample(logits, self.temperature, self.top_k, self.generator) ) if self.log is not None: self.log.append( ( "audio", position, tokens[position], float(logits.log_softmax(-1)[tokens[position]]), ) ) self.previous = tokens self.steps += 1 return list(tokens), written def complete(self, supplied): """Supply the LATER positions of the last step: {position: token or ABSENT}. Positions the step made absent, the acoustics of step 0, are skipped.""" for position, token in supplied.items(): if self.previous[position] == ABSENT: continue if self.previous[position] != LATER: raise ValueError(f"Position {position} was not left for later") self.previous[position] = int(token) class Duplex: """Live two-track sampling: one caller frame in, one model frame out per call. With `model_stream` 2 the caller is track 1 and the model hears each caller frame in the step that samples its own. Its frame s is complete one step later, so the first call returns None. With `model_stream` 1 the model samples each step before the caller's frame arrives and returns the frame of the same index. The model writes the text of its own track when its training option has it. Under the `two` option the caller's text track is written by the model too, as its reading of the caller. `text` holds the latest text tokens by track. """ def __init__( self, model, model_stream, temperature=TEMPERATURE, generator=None, top_k=TOP_K ): if model_stream not in (1, 2): raise ValueError("model_stream is 1 or 2") own = model_stream - 1 tracks = model.config.text_tracks # Track 2 is the one trained to speak with its own text at its start frames. if tracks and own != 1: raise ValueError("A checkpoint with text speaks only as track 2") self.model_stream = model_stream self.sampler = Sampler(model, temperature, generator, top_k) self.previous = None # the caller's last frame self.semantic = None # the model's codebook 0 awaiting its acoustics self.text = {} if model_stream == 1: tokens, written = self.sampler.step([SAMPLE] * 8 + [LATER] * 8) self.semantic = tokens[0] self.text = dict(zip(tracks, written, strict=True)) def full(self): return not self.sampler.room() def push(self, caller): """The caller's next frame, 8 codes; returns the model's next frame or None.""" caller = [int(c) for c in caller] given = [caller[0]] + (self.previous[1:] if self.previous else [ABSENT] * 7) self.previous = caller tracks = self.sampler.text if self.model_stream == 1: self.sampler.complete(dict(enumerate(given, start=8))) tokens, written = self.sampler.step([SAMPLE] * 8 + [LATER] * 8) frame, self.semantic = [self.semantic] + tokens[1:8], tokens[0] self.text = dict(zip(tracks, written, strict=True)) return frame tokens, written = self.sampler.step(given + [SAMPLE] * 8) self.text = dict(zip(tracks, written, strict=True)) frame = None if self.semantic is None else [self.semantic] + tokens[9:] self.semantic = tokens[8] return frame def generate( model, frames, temperature=TEMPERATURE, generator=None, top_k=TOP_K, text=None, given=None, text_sampling=None, ): """Fill the SAMPLE entries of canonical frames [frames, codebooks] in delayed order. A prefix of given frames followed by SAMPLE rows matches `DuplexModel.generate` when no text track is present. Any other mix forces some positions while sampling the rest, such as one track given and the other generated; ABSENT marks an empty track. text names the present text tracks (default: the checkpoint's option), each sampled unless `given` maps it to the token ids to write in each frame (an empty list writes only the advance; SAMPLE samples that frame). text_sampling is text's (temperature, top_k), by default the audio temperature from every token. Returns the tokens and, per present text track, the token ids written in each frame. """ given = given or {} result = frames.clone() total = len(result) sampler = Sampler( model, temperature, generator, top_k, text, text_sampling=text_sampling ) written = {track: [] for track in sampler.text} for step in range(total + 1): rows = [step - d for d in model.delays] forced = [ int(result[f, k]) if 0 <= f < total else ABSENT for k, f in enumerate(rows) ] # The flush step has no frame, so its text tracks write nothing. tokens, out = sampler.step( forced, [given[t][step] if t in given else SAMPLE for t in sampler.text] if step < total else [[] for _ in sampler.text], ) for track, ids in zip(sampler.text, out, strict=True): if step < total: written[track].append(ids) for k, f in enumerate(rows): if 0 <= f < total and int(result[f, k]) != ABSENT: result[f, k] = tokens[k] return result, written # Uniform draws per clip and step: each text track's events, then each codebook. _TEXT_SLOTS = MAX_TEXT_PER_FRAME + 1 def _draw(logits, temperature, top_k, u): """One inverse-CDF draw per row of logits [rows, classes] from uniforms u [rows], over the top_k most likely classes (0 keeps every class); argmax at temperature 0.""" if temperature == 0: return logits.argmax(-1) k = top_k if top_k and top_k < logits.shape[-1] else logits.shape[-1] values, indices = logits.topk(k) cdf = (values / temperature).softmax(-1).cumsum(-1) target = (u.to(cdf.dtype) * cdf[:, -1]).unsqueeze(-1) pick = torch.searchsorted(cdf, target).clamp(max=k - 1) return indices.gather(-1, pick).squeeze(-1) @torch.no_grad() def generate_batch( model, frames, seeds, temperature=TEMPERATURE, top_k=TOP_K, text_sampling=None, given=None, ): """`generate` for a batch of clips [rows, frames, codebooks]. Text is sampled unless `given[row]` maps a track to its token ids per frame, as in `generate` (SAMPLE samples that frame); forced rows ignore their draws. Each row keeps its own temporal position, because rows write different numbers of text tokens, and draws from its own uniforms, seeded by `seeds[row]`, so a row's output does not depend on the rest of the batch, up to float rounding. The distribution is the batch-one sampler's; at temperature 0 the tokens are its tokens. Returns the tokens and, per present text track, each row's token ids written in each frame. """ m, c = model, model.config device = m.bos.device result = frames.to(device).clone() rows, total, q = result.shape if len(seeds) != rows: raise ValueError("One seed per row") text_temperature, text_top_k = text_sampling or (temperature, 0) tracks = tuple(c.text_tracks) uniforms = torch.from_numpy( np.stack( [ np.random.default_rng(seed).random( (total + 1, len(tracks) * _TEXT_SLOTS + q) ) for seed in seeds ] ) ).to(device) # One spare position: rows that already advanced write a placeholder there. cache = _Cache(c.layers, c.heads, c.width, c.context + 2, m.bos.detach(), rows) positions = torch.zeros(rows, dtype=torch.long, device=device) furthest = 0 written = {track: [[] for _ in range(rows)] for track in tracks} delays = torch.tensor(m.delays, device=device) codebooks = torch.arange(q, device=device) previous = None def advance(x, moving): nonlocal furthest h = cache.run_rows( m.temporal, norm(x), m.temporal_rope, positions, furthest + 1 ) positions.add_(moving.long()) furthest = int(positions.max()) if furthest > c.context + 1: raise ValueError("Sequence exceeds the model context") return norm(h) given = given or [{} for _ in range(rows)] if len(given) != rows: raise ValueError("One given map per row") everyone = torch.ones(rows, dtype=torch.bool, device=device) for step in range(total + 1): frame = step - delays inside = (frame >= 0) & (frame < total) forced = result[:, frame.clamp(0, total - 1), codebooks] forced = torch.where(inside, forced, torch.full_like(forced, ABSENT)) if previous is None: x = m.bos.expand(rows, -1) else: present = previous >= 0 embedded = m.frame_embeddings.weight[previous.clamp(min=0) + m.offsets] x = (embedded * present[..., None].to(embedded.dtype)).sum(1) x = x.to(m.bos.dtype) if tracks: x = x + m.text_presence[sum(1 << j for j in tracks) - 1].to(x.dtype) h = advance(x, everyone) # The flush step has no frame, so its text tracks write nothing. for slot, track in enumerate(tracks if step < total else ()): active = everyone forced_text = [ g[track][step] if track in g and g[track][step] != SAMPLE else None for g in given ] longest = max((len(t) for t in forced_text if t is not None), default=0) # A sampled track advances after MAX_TEXT_PER_FRAME tokens without a draw; # a forced one after its given tokens. for count in range(max(MAX_TEXT_PER_FRAME, longest + 1)): logits = m.text_logits(h, torch.tensor(track, device=device)).float() u = uniforms[:, step, slot * _TEXT_SLOTS + min(count, _TEXT_SLOTS - 1)] token = _draw(logits, text_temperature, text_top_k, u) if count >= MAX_TEXT_PER_FRAME: token = torch.full_like(token, ADVANCE) for row, ids in enumerate(forced_text): if ids is not None: token[row] = ids[count] if count < len(ids) else ADVANCE moving = active & (token != ADVANCE) if not bool(moving.any()): break for row in moving.nonzero()[:, 0].tolist(): if count == 0: written[track][row].append([]) written[track][row][-1].append(int(token[row])) e = m.text_embeddings.weight[token.clamp(max=ADVANCE - 1)] e = e + m.text_tracks[track] h = torch.where(moving[:, None], advance(e.to(h.dtype), moving), h) active = moving for row in range(rows): if len(written[track][row]) == step: written[track][row].append([]) tokens = forced.clone() sampled = forced == SAMPLE wanted = sampled.any(0).nonzero() if len(wanted): context = m.to_depth(h) depth = _Cache( c.depth_layers, c.depth_heads, c.depth_width, q, context, rows ) for position in range(int(wanted.max()) + 1): x = context + m.depth_positions[position] if position: before = tokens[:, position - 1] e = m.depth_embeddings.weight[ before.clamp(min=0) + m.offsets[position - 1] ] x = x + (e * (before >= 0)[:, None].to(e.dtype)).to(x.dtype) x = depth.run_depth(m.depth, x, m.depth_rope, position) if not bool(sampled[:, position].any()): continue logits = (norm(x) @ m.outputs[position].T.to(x.dtype)).float() u = uniforms[:, step, len(tracks) * _TEXT_SLOTS + position] draw = _draw(logits, temperature, top_k, u) tokens[:, position] = torch.where( sampled[:, position], draw, tokens[:, position] ) previous = tokens keep = inside & (result[:, frame.clamp(0, total - 1), codebooks] != ABSENT) update = torch.where( keep, tokens, result[:, frame.clamp(0, total - 1), codebooks] ) result[:, frame.clamp(0, total - 1)[inside], codebooks[inside]] = update[ :, inside ] return result, written