41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167 | class FastDepth:
"""Cache depth attention and defer host reads to the end of each frame."""
def __init__(self, model, mode="cached"):
self.model, self.mode = model, mode
self.functions = {}
self.original = None
def initial(self, first, hidden):
import mlx.core as mx
m = self.model.depth_decoder.model
if m.backbone_hidden_state_projector is not None:
hidden = m.backbone_hidden_state_projector(hidden)
ids = mx.broadcast_to(first.reshape(1, 1), (hidden.shape[0], 1))
embeds = mx.concatenate((hidden[:, None, :], m.embed_tokens(ids)), axis=1)
x = m.inputs_embeds_projector(embeds)
caches = [FrameKVCache() for _ in m.layers]
for layer, cache in zip(m.layers, caches):
x = layer(x, "causal", cache)
return x, caches
def advance(self, token, codebook_index, caches, batch):
import mlx.core as mx
m = self.model.depth_decoder.model
ids = mx.broadcast_to(token.reshape(1, 1), (batch, 1))
# At absolute depth position p>=1, upstream uses offset (p-1)*vocab.
embeds = m.embed_tokens(ids + codebook_index * m.vocab_size)
x = m.inputs_embeds_projector(embeds)
for layer, cache in zip(m.layers, caches):
x = layer(x, None, cache)
return x
def logits(self, x, head):
d = self.model.depth_decoder
return d.model.norm(x[:, -1, :]) @ d.codebooks_head.weight[head]
def _function(self, temperature, top_p, top_k, cfg_scale, use_cfg):
import mlx.core as mx
import mlx.nn as nn
from mlx_audio.lm.sample_utils import make_sampler
key = (temperature, top_p, top_k, cfg_scale, use_cfg)
if key in self.functions:
return self.functions[key]
valid = self.model.vocab_size
effective_k = min(top_k, valid) if top_k else 0
if effective_k == valid:
effective_k = 0
sample = make_sampler(temp=temperature, top_p=top_p, top_k=effective_k)
def frame(first, hidden):
x, caches = self.initial(first, hidden)
tokens = [first.reshape(1)]
for head in range(self.model.num_codebooks - 1):
scores = self.logits(x, head)
if use_cfg:
scores = scores[1:2] + cfg_scale * (scores[:1] - scores[1:2])
scores = self.model._mask_reserved_codec_logits(scores)[..., :valid]
token = (
sample(nn.log_softmax(scores, axis=-1)).astype(mx.int32).reshape(1)
)
tokens.append(token)
if head + 1 < self.model.num_codebooks - 1:
x = self.advance(token, head + 1, caches, hidden.shape[0])
return mx.concatenate(tokens)
fn = (
mx.compile(frame, inputs=mx.random.state, outputs=mx.random.state)
if self.mode == "compiled"
else frame
)
self.functions[key] = fn
return fn
def generate_frame(
self,
first_codebook,
conditional_hidden,
*,
unconditional_hidden,
cfg_scale,
temperature,
top_p,
top_k,
):
import mlx.core as mx
if conditional_hidden.shape[0] != 1:
raise ValueError("FastDepth currently supports one utterance at a time.")
if self.model.num_codebooks != self.model.depth_decoder.model.num_codebooks:
raise ValueError("Wrapper/depth codebook counts do not match.")
hidden = conditional_hidden
use_cfg = unconditional_hidden is not None
if use_cfg:
hidden = mx.concatenate((hidden, unconditional_hidden), axis=0)
fn = self._function(temperature, top_p, top_k, cfg_scale, use_cfg)
codes = fn(mx.array([first_codebook], dtype=mx.int32), hidden)
# One host read for all remaining codebooks instead of .item() per code.
return codes.tolist()
def install(self):
if self.mode == "native":
return
target = self.model
cls = type(target)
if getattr(cls._depth_tokens, "_fast_depth_owner", None) is not None:
return
original = cls._depth_tokens
self.original = original
owner = self
def optimized(obj, *args, **kwargs):
if obj is target:
return owner.generate_frame(*args, **kwargs)
return original(obj, *args, **kwargs)
optimized._fast_depth_owner = self
cls._depth_tokens = optimized
def close(self):
if self.original is not None:
cls = type(self.model)
if getattr(cls._depth_tokens, "_fast_depth_owner", None) is self:
cls._depth_tokens = self.original
self.original = None
|