Skip to content

vllm.models.muse_spark.common.layers

Platform-neutral MuseSpark1x building blocks shared by the NVIDIA and AMD stacks.

Classes:

Functions:

MulTanhSoftCap

Bases: Module

Computes soft_cap_temp * tanh(x * multiplier / soft_cap_temp) in fp32.

Source code in vllm/models/muse_spark/common/layers.py 
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
class MulTanhSoftCap(nn.Module):
    """Computes ``soft_cap_temp * tanh(x * multiplier / soft_cap_temp)`` in fp32."""

    def __init__(self, multiplier: float, soft_cap_temp: float) -> None:
        super().__init__()
        self.multiplier = multiplier
        self.soft_cap_temp = soft_cap_temp

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        assert x.dtype in (torch.float32, torch.float16, torch.bfloat16)
        if x.is_cuda and x.is_contiguous():
            if x.dtype != torch.float32:
                x = x.float()
            return ops.muse_spark_mul_tanh_soft_cap(
                x, self.multiplier, self.soft_cap_temp
            )
        x = x.float() * self.multiplier
        return self.soft_cap_temp * torch.tanh(x / self.soft_cap_temp)

MuseSpark1xNormPreservingResidualConnection

Bases: MuseSpark1xResidualConnection

Gated residual sum with norm-preserving weights: alpha*a + beta*b.

With a the stream and b the branch, and gate parameter w:

beta = sigmoid(w / temperature)
alpha = sqrt(max(eps, sigmoid(-w) * (1 + beta)))   # = sqrt(1 - beta^2)

The sum is computed in fp32 and cast back to the stream dtype.

Methods:

Source code in vllm/models/muse_spark/common/layers.py 
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
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
class MuseSpark1xNormPreservingResidualConnection(MuseSpark1xResidualConnection):
    """Gated residual sum with norm-preserving weights: ``alpha*a + beta*b``.

    With ``a`` the stream and ``b`` the branch, and gate parameter ``w``:

        beta = sigmoid(w / temperature)
        alpha = sqrt(max(eps, sigmoid(-w) * (1 + beta)))   # = sqrt(1 - beta^2)

    The sum is computed in fp32 and cast back to the stream dtype.
    """

    def __init__(
        self,
        dim: int,
        activation: str = "sigmoid",
        temperature: float = 0.3,
        use_stream_gate: bool = True,
        channel_wise: bool = True,
        eps: float = 1e-3,
    ):
        super().__init__()
        self.eps = eps
        self.activation_type = activation
        self.temperature = temperature
        self.use_stream_gate = use_stream_gate
        self.gate = nn.Parameter(torch.empty(dim if channel_wise else 1))

    def coefficients(
        self, w: torch.Tensor
    ) -> tuple[torch.Tensor | float, torch.Tensor]:
        """Compute post-activation coefficients for the stream and branch."""
        if self.activation_type != "sigmoid":
            raise ValueError(f"Unsupported activation: {self.activation_type}")
        beta = torch.sigmoid(w)
        if self.use_stream_gate:
            alpha_sq = torch.sigmoid(-w) * (1.0 + beta)  # = 1 - beta^2 (stable)
            alpha = torch.sqrt(torch.clamp(alpha_sq, min=self.eps))
        else:
            alpha = 1.0
        return alpha, beta

    def materialize_coefficients(self) -> None:
        """Cache alpha/beta as fp32 buffers. Call once, after weight loading."""
        w = self.gate.detach().float() / self.temperature
        alpha, beta = self.coefficients(w)
        self.register_buffer("beta", beta.contiguous(), persistent=False)
        if isinstance(alpha, torch.Tensor):
            self.register_buffer("alpha", alpha.contiguous(), persistent=False)
        else:
            self.alpha = alpha
        self._coefficients_ready = True

    def forward(self, stream: torch.Tensor, branch: torch.Tensor) -> torch.Tensor:
        dtype = stream.dtype
        branch = branch.to(dtype)

        if getattr(self, "_coefficients_ready", False):
            alpha, beta = self.alpha, self.beta
        else:
            w = self.gate.view(*([1] * (stream.ndim - 1)), -1).float()
            alpha, beta = self.coefficients(w / self.temperature)
        out = alpha * stream + beta * branch
        return out.to(dtype)

coefficients(w)

Compute post-activation coefficients for the stream and branch.

Source code in vllm/models/muse_spark/common/layers.py 
251
252
253
254
255
256
257
258
259
260
261
262
263
def coefficients(
    self, w: torch.Tensor
) -> tuple[torch.Tensor | float, torch.Tensor]:
    """Compute post-activation coefficients for the stream and branch."""
    if self.activation_type != "sigmoid":
        raise ValueError(f"Unsupported activation: {self.activation_type}")
    beta = torch.sigmoid(w)
    if self.use_stream_gate:
        alpha_sq = torch.sigmoid(-w) * (1.0 + beta)  # = 1 - beta^2 (stable)
        alpha = torch.sqrt(torch.clamp(alpha_sq, min=self.eps))
    else:
        alpha = 1.0
    return alpha, beta

materialize_coefficients()

Cache alpha/beta as fp32 buffers. Call once, after weight loading.

Source code in vllm/models/muse_spark/common/layers.py 
265
266
267
268
269
270
271
272
273
274
def materialize_coefficients(self) -> None:
    """Cache alpha/beta as fp32 buffers. Call once, after weight loading."""
    w = self.gate.detach().float() / self.temperature
    alpha, beta = self.coefficients(w)
    self.register_buffer("beta", beta.contiguous(), persistent=False)
    if isinstance(alpha, torch.Tensor):
        self.register_buffer("alpha", alpha.contiguous(), persistent=False)
    else:
        self.alpha = alpha
    self._coefficients_ready = True

MuseSpark1xQKVOGateParallelLinear

Bases: ColumnParallelLinear

Fused [Q, K, V, O-gate] projection with head-aware TP sharding.

Per-rank rows are [q, k, v, o_gate], so checkpoint tensors are sliced by head group. KV-head padding is not supported.

Source code in vllm/models/muse_spark/common/layers.py 
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
class MuseSpark1xQKVOGateParallelLinear(ColumnParallelLinear):
    """Fused [Q, K, V, O-gate] projection with head-aware TP sharding.

    Per-rank rows are [q, k, v, o_gate], so checkpoint tensors are sliced by
    head group. KV-head padding is not supported.
    """

    def __init__(
        self,
        hidden_size: int,
        head_size: int,
        total_num_heads: int,
        total_num_kv_heads: int,
        quant_config: QuantizationConfig | None = None,
        prefix: str = "",
        *,
        tp_rank: int | None = None,
        tp_size: int | None = None,
    ) -> None:
        self.head_size = head_size
        super().__init__(
            input_size=hidden_size,
            output_size=2 * (total_num_heads + total_num_kv_heads) * head_size,
            bias=False,
            quant_config=quant_config,
            prefix=prefix,
            tp_rank=tp_rank,
            tp_size=tp_size,
        )
        assert total_num_heads % self.tp_size == 0
        assert total_num_kv_heads % self.tp_size == 0, (
            "KV-head padding is not supported for the fused QKV+O-gate "
            f"projection: {total_num_kv_heads=}, tp_size={self.tp_size}."
        )
        self.num_heads = total_num_heads // self.tp_size
        self.num_kv_heads = total_num_kv_heads // self.tp_size

    def weight_loader(
        self,
        param: nn.Parameter,
        loaded_weight: torch.Tensor,
        loaded_shard_id: str | None = None,
    ) -> None:
        if loaded_shard_id is None:
            assert param.shape == loaded_weight.shape
            param.data.copy_(loaded_weight)
            return
        q_size = self.num_heads * self.head_size
        kv_size = self.num_kv_heads * self.head_size
        if loaded_shard_id == "q":
            shard_offset, shard_size = 0, q_size
        elif loaded_shard_id == "k":
            shard_offset, shard_size = q_size, kv_size
        elif loaded_shard_id == "v":
            shard_offset, shard_size = q_size + kv_size, kv_size
        elif loaded_shard_id == "o_gate":
            shard_offset, shard_size = q_size + 2 * kv_size, q_size
        else:
            raise ValueError(
                f"{self.prefix}: unknown shard id {loaded_shard_id!r} for the "
                "fused QKV+O-gate projection."
            )
        start_idx = self.tp_rank * shard_size
        loaded_weight = loaded_weight.narrow(0, start_idx, shard_size)
        param_shard = param.data.narrow(0, shard_offset, shard_size)
        assert param_shard.shape == loaded_weight.shape
        param_shard.copy_(loaded_weight)

    def load_weights(
        self, weights: Iterable[tuple[str, torch.Tensor]]
    ) -> Iterable[str]:
        for name, loaded_weight in weights:
            shard_id = getattr(loaded_weight, "shard_id", None)
            if name == "bias" and self.bias is None:
                continue
            param = getattr(self, name)
            self.weight_loader(param, loaded_weight, shard_id)
            yield name

MuseSpark1xRMSNorm

Bases: Module

RMSNorm in fp32, rounded once at the store to out_dtype.

Deliberately not F.rms_norm: that path can flip greedy token selection.

Methods:

Attributes:

Source code in vllm/models/muse_spark/common/layers.py 
 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
class MuseSpark1xRMSNorm(nn.Module):
    """RMSNorm in fp32, rounded once at the store to ``out_dtype``.

    Deliberately not ``F.rms_norm``: that path can flip greedy token selection.
    """

    # Which side of the normalization the weight lands on; read by fused kernels.
    applies_weight_before_norm = False

    def __init__(
        self,
        hidden_size: int,
        eps: float = 1e-6,
        var_hidden_size: int | None = None,
        has_weight: bool = True,
        is_post_norm: bool = False,
        zero_centered_gamma: bool = False,
        post_norm_gain_center_type: str = "one",
        residual_id: int | None = None,
        output_norm_gain_center_type: str = "zero",
    ) -> None:
        super().__init__()

        self.hidden_size = hidden_size
        self.variance_epsilon = eps
        self.variance_size_override = (
            None if var_hidden_size == hidden_size else var_hidden_size
        )
        self.has_weight = has_weight
        if self.has_weight:
            self.weight = nn.Parameter(torch.empty(hidden_size), requires_grad=False)
        else:
            self.weight = None

        self.zero_centered_gamma = zero_centered_gamma
        self.gain_center = 1.0
        if residual_id is not None and is_post_norm:
            self.gain_center = calculate_gain_center(
                post_norm_gain_center_type, residual_id
            )
        elif residual_id == -1:
            self.gain_center = calculate_gain_center(
                output_norm_gain_center_type, residual_id
            )

    @property
    def fused_gain_center(self) -> float:
        """``gain_center`` as the fused kernels take it: 0.0 when unused."""
        return self.gain_center if self.has_weight and self.zero_centered_gamma else 0.0

    def forward_native(
        self,
        x: torch.Tensor,
        residual: torch.Tensor | None = None,
        out_dtype: torch.dtype | None = None,
    ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
        orig_dtype = x.dtype
        out_dtype = out_dtype or orig_dtype
        x = x.to(torch.float32)
        if residual is not None:
            x = x + residual.to(torch.float32)
            residual = x.to(orig_dtype)

        hidden_size = x.shape[-1]
        if hidden_size != self.hidden_size:
            raise ValueError(
                "Expected hidden_size to be "
                f"{self.hidden_size}, but found: {hidden_size}"
            )

        if self.variance_size_override is None:
            x_var = x
        else:
            if hidden_size < self.variance_size_override:
                raise ValueError(
                    "Expected hidden_size to be at least "
                    f"{self.variance_size_override}, but found: {hidden_size}"
                )
            x_var = x[..., : self.variance_size_override]

        variance = x_var.pow(2).mean(dim=-1, keepdim=True)

        x = x * torch.rsqrt(variance + self.variance_epsilon)
        if self.has_weight:
            weight = self.weight.float()
            if self.zero_centered_gamma:
                weight = weight + self.gain_center
            x = x * weight
        x = x.to(out_dtype)
        if residual is None:
            return x
        return x, residual

    def forward(
        self,
        x: torch.Tensor,
        residual: torch.Tensor | None = None,
        out_dtype: torch.dtype | None = None,
    ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
        if residual is None and self.supports_fused_kernel(x):
            return fused_rmsnorm_fwd(
                x,
                self.weight if self.has_weight else None,
                self.variance_epsilon,
                out_dtype=out_dtype,
                gain_center=self.fused_gain_center,
                input_scale=self.applies_weight_before_norm and self.has_weight,
                # Only matters past one 8192-element tile; a device query here
                # would break the compiled graph.
                use_kahan_summation=True,
            )
        return self.forward_native(x, residual, out_dtype)

    def supports_fused_kernel(self, x: torch.Tensor) -> bool:
        """Whether the Triton kernel can serve this norm for input ``x``."""
        return (
            self.variance_size_override is None
            and x.is_contiguous()
            and x.shape[-1] == self.hidden_size
        )

fused_gain_center property

gain_center as the fused kernels take it: 0.0 when unused.

supports_fused_kernel(x)

Whether the Triton kernel can serve this norm for input x.

Source code in vllm/models/muse_spark/common/layers.py 
157
158
159
160
161
162
163
def supports_fused_kernel(self, x: torch.Tensor) -> bool:
    """Whether the Triton kernel can serve this norm for input ``x``."""
    return (
        self.variance_size_override is None
        and x.is_contiguous()
        and x.shape[-1] == self.hidden_size
    )

MuseSpark1xRMSNormWithInputScale

Bases: MuseSpark1xRMSNorm

Weight-first RMSNorm variant: rms_norm(x * weight).

Source code in vllm/models/muse_spark/common/layers.py 
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
class MuseSpark1xRMSNormWithInputScale(MuseSpark1xRMSNorm):
    """Weight-first RMSNorm variant: ``rms_norm(x * weight)``."""

    applies_weight_before_norm = True

    def forward_native(
        self,
        x: torch.Tensor,
        residual: torch.Tensor | None = None,
        out_dtype: torch.dtype | None = None,
    ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
        orig_dtype = x.dtype
        out_dtype = out_dtype or orig_dtype
        x = x.to(torch.float32)
        if residual is not None:
            x = x + residual.to(torch.float32)
            residual = x.to(orig_dtype)

        if self.weight is not None:
            weight = self.weight.float()
            if self.zero_centered_gamma:
                weight = weight + self.gain_center
            x = x * weight
        x = F.rms_norm(x, (self.hidden_size,), None, self.variance_epsilon)
        x = x.to(out_dtype)

        if residual is None:
            return x
        return x, residual

MuseSpark1xResidualConnection

Bases: Module

Vanilla residual sum: stream + branch.

Source code in vllm/models/muse_spark/common/layers.py 
217
218
219
220
221
class MuseSpark1xResidualConnection(nn.Module):
    """Vanilla residual sum: ``stream + branch``."""

    def forward(self, stream: torch.Tensor, branch: torch.Tensor) -> torch.Tensor:
        return stream + branch.to(stream.dtype)

MuseSpark1xResidualStream dataclass

A residual stream whose last blend is deferred to fuse with the next norm.

Source code in vllm/models/muse_spark/common/layers.py 
289
290
291
292
293
294
295
296
@dataclass(slots=True)
class MuseSpark1xResidualStream:
    """A residual stream whose last blend is deferred to fuse with the next norm."""

    stream: torch.Tensor
    branch: torch.Tensor
    post_norm: MuseSpark1xRMSNorm
    connection: MuseSpark1xResidualConnection

calculate_gain_center(gain_center_type, residual_id, n_layers=None)

Gain center for the given type and residual ID (1-based indices).

Source code in vllm/models/muse_spark/common/layers.py 
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
def calculate_gain_center(
    gain_center_type: str,
    residual_id: int,
    n_layers: int | None = None,
) -> float:
    """Gain center for the given type and residual ID (1-based indices)."""
    if gain_center_type == "one":
        return 1.0
    if gain_center_type == "zero":
        return 0.0
    if gain_center_type == "inverse_residual_index":
        return 1.0 / (2 * math.ceil(residual_id / 2))
    if gain_center_type in (
        "inverse_residual_index_v2",
        "inverse_residual_index_p1",
    ):
        return 1.0 / (residual_id + 1)
    if gain_center_type == "inverse_sqrt_residual_index":
        return 1.0 / math.sqrt(2 * math.ceil(residual_id / 2))
    if gain_center_type == "inverse_sqrt_layers":
        assert n_layers is not None, "n_layers must be provided for this option."
        return 1.0 / math.sqrt(n_layers)
    raise ValueError(f"Unsupported gain_center_type: {gain_center_type}")

materialize_residual_path_constants(model)

Precompute parameter-only constants after weight loading; idempotent.

Source code in vllm/models/muse_spark/common/layers.py 
325
326
327
328
329
330
331
332
333
334
335
336
def materialize_residual_path_constants(model: nn.Module) -> int:
    """Precompute parameter-only constants after weight loading; idempotent."""
    n_gates = 0
    for module in model.modules():
        if isinstance(module, MuseSpark1xNormPreservingResidualConnection):
            module.materialize_coefficients()
            n_gates += 1
        else:
            fold = getattr(module, "fold_router_output_multiplier", None)
            if callable(fold):
                fold()
    return n_gates

norm_then_gated_residual(x)

alpha * stream + beta * norm(branch), in one kernel where possible.

Source code in vllm/models/muse_spark/common/layers.py 
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
def norm_then_gated_residual(x: MuseSpark1xResidualStream) -> torch.Tensor:
    """``alpha * stream + beta * norm(branch)``, in one kernel where possible."""
    norm, connection = x.post_norm, x.connection
    branch, stream = x.branch, x.stream
    if (
        isinstance(connection, MuseSpark1xNormPreservingResidualConnection)
        and getattr(connection, "_coefficients_ready", False)
        and branch.is_cuda
        and stream.shape == branch.shape
        and stream.is_contiguous()
        and norm.supports_fused_kernel(branch)
    ):
        return fused_rmsnorm_fwd(
            branch,
            norm.weight if norm.has_weight else None,
            norm.variance_epsilon,
            gain_center=norm.fused_gain_center,
            input_scale=norm.applies_weight_before_norm and norm.has_weight,
            use_kahan_summation=True,
            residual_stream=stream,
            alpha=connection.alpha,
            beta=connection.beta,
        )
    return connection(branch=norm(branch), stream=stream)