Skip to content

vllm.models.muse_spark.common.audio

Waveform AutoEncoder for MuseSpark raw audio input.

Each patch of 1920 samples (80 ms @ 24 kHz) yields one latent token.

Classes:

Functions:

  • patchify –

    Group consecutive timesteps: [B, T, D] -> [B, T // p, p * D].

AutoEncoder

Bases: Module

Methods:

  • forward –

    Encode raw waveforms into [num_tokens, latent_channels] embeddings.

Source code in vllm/models/muse_spark/common/audio.py 
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
class AutoEncoder(nn.Module):
    def __init__(
        self,
        latent_channels: int,
        patch_size: int = 1920,
        kernel_size: int = 7,
        data_channels: int = 1,
        patchify_sizes: tuple[int, ...] = (16, 6, 5, 4),
        d_models: tuple[int, ...] = (256, 512, 1024, 2048),
        n_layers: tuple[int, ...] = (2, 2, 2, 2),
        prefix: str = "",
        disable_tp: bool = True,
        norm_eps: float = 1e-6,
        mlp_ratio: float = 4.0,
    ) -> None:
        super().__init__()
        self.latent_channels = latent_channels
        self.patch_size = patch_size
        self.kernel_size = kernel_size
        self.data_channels = data_channels

        d_model_list = [data_channels] + list(d_models)
        self.stages = nn.Sequential(
            *[
                self._make_stage(
                    d_input=d_model_list[stage_id],
                    d_model=d_model_list[stage_id + 1],
                    patch_size=patchify_sizes[stage_id],
                    n_layers_count=n_layers[stage_id],
                    s_kernel=kernel_size,
                    prefix=maybe_prefix(prefix, f"stages.{stage_id}"),
                    disable_tp=disable_tp,
                    norm_eps=norm_eps,
                    mlp_ratio=mlp_ratio,
                )
                for stage_id in range(len(n_layers))
            ]
        )

        self.norm = RMSNorm(d_model_list[-1], eps=norm_eps)
        self.o_proj = ColumnParallelLinear(
            input_size=d_model_list[-1],
            output_size=latent_channels,
            bias=True,
            gather_output=True,
            prefix=maybe_prefix(prefix, "o_proj"),
            disable_tp=disable_tp,
        )
        self.dual_norm = RMSNorm(latent_channels, eps=norm_eps)

    @staticmethod
    def _make_stage(
        d_input: int,
        d_model: int,
        patch_size: int,
        n_layers_count: int,
        s_kernel: int,
        prefix: str = "",
        disable_tp: bool = True,
        norm_eps: float = 1e-6,
        mlp_ratio: float = 4.0,
    ) -> nn.Sequential:
        layers: list[nn.Module] = [
            Patchifier(
                patch_size,
                d_input,
                d_model,
                prefix=maybe_prefix(prefix, "0"),
                disable_tp=disable_tp,
                norm_eps=norm_eps,
            )
        ]
        for i in range(n_layers_count):
            layers.append(
                Block(
                    d_model,
                    int(d_model * mlp_ratio),
                    kernel_size=s_kernel,
                    prefix=maybe_prefix(prefix, str(i + 1)),
                    disable_tp=disable_tp,
                    norm_eps=norm_eps,
                )
            )
        return nn.Sequential(*layers)

    def _encode_prefill(self, x: torch.Tensor) -> torch.Tensor:
        """Prefill encoding: [B, T, data_channels] -> [B, T', latent_channels]."""
        for stage in self.stages:
            for block in stage:
                if isinstance(block, Patchifier):
                    x = block(x)
                elif isinstance(block, Block):
                    x, _ = block(x, None)
        x = self.norm(x)
        x, _ = self.o_proj(x)
        x = self.dual_norm(x)
        return x

    def forward(self, wavs: list[torch.Tensor]) -> list[torch.Tensor]:
        """Encode raw waveforms into [num_tokens, latent_channels] embeddings."""
        if not wavs:
            return []

        def _pad(wav: torch.Tensor) -> torch.Tensor:
            wav = wav.reshape(-1, self.data_channels)
            pad_size = (self.patch_size - wav.shape[0]) % self.patch_size
            return F.pad(wav, (0, 0, 0, pad_size))

        embeds: list[torch.Tensor] = []
        for wav in wavs:
            wav_batch = _pad(wav).unsqueeze(0)  # [1, T, C]
            embed = self._encode_prefill(wav_batch)
            embeds.append(embed.squeeze(0))  # [num_tokens, latent_channels]
        return embeds

_encode_prefill(x)

Prefill encoding: [B, T, data_channels] -> [B, T', latent_channels].

Source code in vllm/models/muse_spark/common/audio.py 
240
241
242
243
244
245
246
247
248
249
250
251
def _encode_prefill(self, x: torch.Tensor) -> torch.Tensor:
    """Prefill encoding: [B, T, data_channels] -> [B, T', latent_channels]."""
    for stage in self.stages:
        for block in stage:
            if isinstance(block, Patchifier):
                x = block(x)
            elif isinstance(block, Block):
                x, _ = block(x, None)
    x = self.norm(x)
    x, _ = self.o_proj(x)
    x = self.dual_norm(x)
    return x

forward(wavs)

Encode raw waveforms into [num_tokens, latent_channels] embeddings.

Source code in vllm/models/muse_spark/common/audio.py 
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
def forward(self, wavs: list[torch.Tensor]) -> list[torch.Tensor]:
    """Encode raw waveforms into [num_tokens, latent_channels] embeddings."""
    if not wavs:
        return []

    def _pad(wav: torch.Tensor) -> torch.Tensor:
        wav = wav.reshape(-1, self.data_channels)
        pad_size = (self.patch_size - wav.shape[0]) % self.patch_size
        return F.pad(wav, (0, 0, 0, pad_size))

    embeds: list[torch.Tensor] = []
    for wav in wavs:
        wav_batch = _pad(wav).unsqueeze(0)  # [1, T, C]
        embed = self._encode_prefill(wav_batch)
        embeds.append(embed.squeeze(0))  # [num_tokens, latent_channels]
    return embeds

Block

Bases: Module

DepthwiseCausalConv -> RMSNorm -> FFN (GELU) with a residual.

Source code in vllm/models/muse_spark/common/audio.py 
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
class Block(nn.Module):
    """DepthwiseCausalConv -> RMSNorm -> FFN (GELU) with a residual."""

    def __init__(
        self,
        d_model: int,
        d_hidden: int,
        kernel_size: int,
        prefix: str = "",
        disable_tp: bool = True,
        norm_eps: float = 1e-6,
    ) -> None:
        super().__init__()
        self._d_model = d_model
        self._kernel_size = kernel_size
        self.conv = DepthwiseCausalConv(d_model, kernel_size)
        self.norm = RMSNorm(d_model, eps=norm_eps)
        self.linear1 = ColumnParallelLinear(
            input_size=d_model,
            output_size=d_hidden,
            bias=True,
            prefix=maybe_prefix(prefix, "linear1"),
            disable_tp=disable_tp,
        )
        self.linear2 = RowParallelLinear(
            input_size=d_hidden,
            output_size=d_model,
            bias=True,
            prefix=maybe_prefix(prefix, "linear2"),
            disable_tp=disable_tp,
        )
        self.activation = nn.GELU()

    @property
    def d_model(self) -> int:
        return self._d_model

    @property
    def kernel_size(self) -> int:
        return self._kernel_size

    def forward(
        self,
        inputs: torch.Tensor,
        cache: torch.Tensor | None = None,
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        x = inputs
        x, cache = self.conv(x, cache)
        x = self.norm(x)
        x, _ = self.linear1(x)
        x = self.activation(x)
        x, _ = self.linear2(x)
        return x + inputs, cache

DepthwiseCausalConv

Bases: Module

Depthwise causal 1-D convolution with optional streaming cache.

Methods:

  • forward –

    [B, T, D] -> [B, T, D]; cache is [B, D, kernel_size - 1] or None.

Source code in vllm/models/muse_spark/common/audio.py 
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
class DepthwiseCausalConv(nn.Module):
    """Depthwise causal 1-D convolution with optional streaming cache."""

    def __init__(self, d_model: int, kernel_size: int) -> None:
        super().__init__()
        self._kernel_size = kernel_size
        self.conv = nn.Conv1d(
            in_channels=d_model,
            out_channels=d_model,
            kernel_size=kernel_size,
            stride=1,
            padding=0,
            bias=True,
            groups=d_model,
        )

    def forward(
        self,
        x: torch.Tensor,
        cache: torch.Tensor | None = None,
    ) -> tuple[torch.Tensor, torch.Tensor | None]:
        """[B, T, D] -> [B, T, D]; ``cache`` is [B, D, kernel_size - 1] or None."""
        x = x.permute(0, 2, 1)  # [B, D, T]

        if cache is None:
            x = F.pad(x, (self._kernel_size - 1, 0))
            x = self.conv(x)
            new_cache = None
        else:
            x = torch.cat((cache, x), dim=-1)
            if torch.cuda.is_current_stream_capturing():
                # CUDA-graph safe: in-place update.
                cache.copy_(x[:, :, -self._kernel_size + 1 :])
                new_cache = cache
            else:
                new_cache = x[:, :, -self._kernel_size + 1 :]
            x = self.conv(x)

        # The downstream fused RMSNorm kernel requires a contiguous buffer.
        x = x.permute(0, 2, 1).contiguous()  # [B, T, D]
        return x, new_cache

forward(x, cache=None)

[B, T, D] -> [B, T, D]; cache is [B, D, kernel_size - 1] or None.

Source code in vllm/models/muse_spark/common/audio.py 
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
def forward(
    self,
    x: torch.Tensor,
    cache: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor | None]:
    """[B, T, D] -> [B, T, D]; ``cache`` is [B, D, kernel_size - 1] or None."""
    x = x.permute(0, 2, 1)  # [B, D, T]

    if cache is None:
        x = F.pad(x, (self._kernel_size - 1, 0))
        x = self.conv(x)
        new_cache = None
    else:
        x = torch.cat((cache, x), dim=-1)
        if torch.cuda.is_current_stream_capturing():
            # CUDA-graph safe: in-place update.
            cache.copy_(x[:, :, -self._kernel_size + 1 :])
            new_cache = cache
        else:
            new_cache = x[:, :, -self._kernel_size + 1 :]
        x = self.conv(x)

    # The downstream fused RMSNorm kernel requires a contiguous buffer.
    x = x.permute(0, 2, 1).contiguous()  # [B, T, D]
    return x, new_cache

Patchifier

Bases: Module

Patchify -> Linear -> RMSNorm.

Source code in vllm/models/muse_spark/common/audio.py 
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
class Patchifier(nn.Module):
    """Patchify -> Linear -> RMSNorm."""

    def __init__(
        self,
        patch_size: int,
        d_input: int,
        d_output: int,
        prefix: str = "",
        disable_tp: bool = True,
        norm_eps: float = 1e-6,
    ) -> None:
        super().__init__()
        self.patch_size = patch_size
        self.linear = ColumnParallelLinear(
            input_size=d_input * patch_size,
            output_size=d_output,
            bias=True,
            gather_output=True,
            prefix=maybe_prefix(prefix, "linear"),
            disable_tp=disable_tp,
        )
        self.norm = RMSNorm(d_output, eps=norm_eps)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        x = patchify(x, self.patch_size)
        x, _ = self.linear(x)
        return self.norm(x)

patchify(x, patch_size)

Group consecutive timesteps: [B, T, D] -> [B, T // p, p * D].

Source code in vllm/models/muse_spark/common/audio.py 
20
21
22
23
24
def patchify(x: torch.Tensor, patch_size: int) -> torch.Tensor:
    """Group consecutive timesteps: [B, T, D] -> [B, T // p, p * D]."""
    b, t, d = x.shape
    assert t % patch_size == 0, f"t={t} is not divisible by patch_size={patch_size}"
    return x.reshape(b, t // patch_size, patch_size * d)