Skip to content

vllm.model_executor.layers.layernorm

Custom normalization layers.

GemmaRMSNorm

Bases: CustomOp

RMS normalization for Gemma.

Two differences from the above RMSNorm
  1. x * (1 + w) instead of x * w.
  2. (x * w).to(orig_dtype) instead of x.to(orig_dtype) * w.
Source code in vllm/model_executor/layers/layernorm.py 
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
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
@CustomOp.register("gemma_rms_norm")
class GemmaRMSNorm(CustomOp):
    """RMS normalization for Gemma.

    Two differences from the above RMSNorm:
        1. x * (1 + w) instead of x * w.
        2. (x * w).to(orig_dtype) instead of x.to(orig_dtype) * w.
    """

    def __init__(
        self,
        hidden_size: int,
        eps: float = 1e-6,
    ) -> None:
        super().__init__()
        self.weight = nn.Parameter(torch.zeros(hidden_size))
        self.variance_epsilon = eps

    @staticmethod
    def forward_static(
        weight: torch.Tensor,
        variance_epsilon: float,
        x: torch.Tensor,
        residual: Optional[torch.Tensor],
    ) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
        """PyTorch-native implementation equivalent to forward()."""
        orig_dtype = x.dtype
        if residual is not None:
            if orig_dtype == torch.float16:
                x = x + residual.float()
            else:
                x = x + residual
            residual = x

        x = x.float()
        variance = x.pow(2).mean(dim=-1, keepdim=True)
        x = x * torch.rsqrt(variance + variance_epsilon)
        # Llama does x.to(float16) * w whilst Gemma is (x * w).to(float16)
        # See https://github.com/huggingface/transformers/pull/29402
        x = x * (1.0 + weight.float())
        x = x.to(orig_dtype)
        return x if residual is None else (x, residual)

    def forward_native(
        self,
        x: torch.Tensor,
        residual: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
        """PyTorch-native implementation equivalent to forward()."""
        return self.forward_static(self.weight.data, self.variance_epsilon, x,
                                   residual)

    def forward_cuda(
        self,
        x: torch.Tensor,
        residual: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
        if torch.compiler.is_compiling():
            return self.forward_native(x, residual)

        if not getattr(self, "_is_compiled", False):
            self.forward_static = torch.compile(  # type: ignore
                self.forward_static)
            self._is_compiled = True
        return self.forward_native(x, residual)

variance_epsilon instance-attribute 

variance_epsilon = eps

weight instance-attribute 

weight = Parameter(zeros(hidden_size))

__init__ 

__init__(hidden_size: int, eps: float = 1e-06) -> None
Source code in vllm/model_executor/layers/layernorm.py 
232
233
234
235
236
237
238
239
def __init__(
    self,
    hidden_size: int,
    eps: float = 1e-6,
) -> None:
    super().__init__()
    self.weight = nn.Parameter(torch.zeros(hidden_size))
    self.variance_epsilon = eps

forward_cuda 

forward_cuda(
    x: Tensor, residual: Optional[Tensor] = None
) -> Union[Tensor, tuple[Tensor, Tensor]]
Source code in vllm/model_executor/layers/layernorm.py 
275
276
277
278
279
280
281
282
283
284
285
286
287
def forward_cuda(
    self,
    x: torch.Tensor,
    residual: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
    if torch.compiler.is_compiling():
        return self.forward_native(x, residual)

    if not getattr(self, "_is_compiled", False):
        self.forward_static = torch.compile(  # type: ignore
            self.forward_static)
        self._is_compiled = True
    return self.forward_native(x, residual)

forward_native 

forward_native(
    x: Tensor, residual: Optional[Tensor] = None
) -> Union[Tensor, tuple[Tensor, Tensor]]

PyTorch-native implementation equivalent to forward().

Source code in vllm/model_executor/layers/layernorm.py 
266
267
268
269
270
271
272
273
def forward_native(
    self,
    x: torch.Tensor,
    residual: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
    """PyTorch-native implementation equivalent to forward()."""
    return self.forward_static(self.weight.data, self.variance_epsilon, x,
                               residual)

forward_static staticmethod 

forward_static(
    weight: Tensor,
    variance_epsilon: float,
    x: Tensor,
    residual: Optional[Tensor],
) -> Union[Tensor, tuple[Tensor, Tensor]]

PyTorch-native implementation equivalent to forward().

Source code in vllm/model_executor/layers/layernorm.py 
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
@staticmethod
def forward_static(
    weight: torch.Tensor,
    variance_epsilon: float,
    x: torch.Tensor,
    residual: Optional[torch.Tensor],
) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
    """PyTorch-native implementation equivalent to forward()."""
    orig_dtype = x.dtype
    if residual is not None:
        if orig_dtype == torch.float16:
            x = x + residual.float()
        else:
            x = x + residual
        residual = x

    x = x.float()
    variance = x.pow(2).mean(dim=-1, keepdim=True)
    x = x * torch.rsqrt(variance + variance_epsilon)
    # Llama does x.to(float16) * w whilst Gemma is (x * w).to(float16)
    # See https://github.com/huggingface/transformers/pull/29402
    x = x * (1.0 + weight.float())
    x = x.to(orig_dtype)
    return x if residual is None else (x, residual)

RMSNorm

Bases: CustomOp

Root mean square normalization.

Computes x -> w * x / sqrt(E[x^2] + eps) where w is the learned weight. Refer to https://arxiv.org/abs/1910.07467

Source code in vllm/model_executor/layers/layernorm.py 
 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
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
@CustomOp.register("rms_norm")
class RMSNorm(CustomOp):
    """Root mean square normalization.

    Computes x -> w * x / sqrt(E[x^2] + eps) where w is the learned weight.
    Refer to https://arxiv.org/abs/1910.07467
    """

    def __init__(
        self,
        hidden_size: int,
        eps: float = 1e-6,
        var_hidden_size: Optional[int] = None,
        has_weight: bool = True,
        dtype: Optional[torch.dtype] = None,
    ) -> 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 dtype is not None:
            self.weight = torch.ones(hidden_size, dtype=dtype)
        else:
            self.weight = torch.ones(hidden_size)
        if self.has_weight:
            self.weight = nn.Parameter(self.weight)

    def forward_native(
        self,
        x: torch.Tensor,
        residual: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
        """PyTorch-native implementation equivalent to forward()."""
        orig_dtype = x.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)
        x = x.to(orig_dtype)
        if self.has_weight:
            x = x * self.weight
        if residual is None:
            return x
        else:
            return x, residual

    def forward_cuda(
        self,
        x: torch.Tensor,
        residual: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
        if self.variance_size_override is not None:
            return self.forward_native(x, residual)

        add_residual = residual is not None
        norm_func = dispatch_cuda_rmsnorm_func(add_residual)

        if add_residual:
            return norm_func(x, residual, self.weight.data,
                             self.variance_epsilon)
        else:
            return norm_func(x, self.weight.data, self.variance_epsilon)

    def forward_hpu(
        self,
        x: torch.Tensor,
        residual: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
        from vllm_hpu_extension.kernels import rms_norm
        HPUFusedRMSNorm = rms_norm()
        if HPUFusedRMSNorm is None:
            return self.forward_native(x, residual)
        if residual is not None:
            orig_shape = x.shape
            residual += x.view(residual.shape)
            # Note: HPUFusedRMSNorm requires 3D tensors as inputs
            x = HPUFusedRMSNorm.apply(residual, self.weight,
                                      self.variance_epsilon)
            return x.view(orig_shape), residual

        x = HPUFusedRMSNorm.apply(x, self.weight, self.variance_epsilon)
        return x

    def forward_xpu(
        self,
        x: torch.Tensor,
        residual: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
        if self.variance_size_override is not None:
            return self.forward_native(x, residual)

        from vllm._ipex_ops import ipex_ops as ops

        if residual is not None:
            ops.fused_add_rms_norm(
                x,
                residual,
                self.weight.data,
                self.variance_epsilon,
            )
            return x, residual
        return ops.rms_norm(
            x,
            self.weight.data,
            self.variance_epsilon,
        )

    def extra_repr(self) -> str:
        s = f"hidden_size={self.weight.data.size(0)}"
        s += f", eps={self.variance_epsilon}"
        return s

has_weight instance-attribute 

has_weight = has_weight

hidden_size instance-attribute 

hidden_size = hidden_size

variance_epsilon instance-attribute 

variance_epsilon = eps

variance_size_override instance-attribute 

variance_size_override = (
    None
    if var_hidden_size == hidden_size
    else var_hidden_size
)

weight instance-attribute 

weight = ones(hidden_size, dtype=dtype)

__init__ 

__init__(
    hidden_size: int,
    eps: float = 1e-06,
    var_hidden_size: Optional[int] = None,
    has_weight: bool = True,
    dtype: Optional[dtype] = None,
) -> None
Source code in vllm/model_executor/layers/layernorm.py 
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
def __init__(
    self,
    hidden_size: int,
    eps: float = 1e-6,
    var_hidden_size: Optional[int] = None,
    has_weight: bool = True,
    dtype: Optional[torch.dtype] = None,
) -> 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 dtype is not None:
        self.weight = torch.ones(hidden_size, dtype=dtype)
    else:
        self.weight = torch.ones(hidden_size)
    if self.has_weight:
        self.weight = nn.Parameter(self.weight)

extra_repr 

extra_repr() -> str
Source code in vllm/model_executor/layers/layernorm.py 
217
218
219
220
def extra_repr(self) -> str:
    s = f"hidden_size={self.weight.data.size(0)}"
    s += f", eps={self.variance_epsilon}"
    return s

forward_cuda 

forward_cuda(
    x: Tensor, residual: Optional[Tensor] = None
) -> Union[Tensor, tuple[Tensor, Tensor]]
Source code in vllm/model_executor/layers/layernorm.py 
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
def forward_cuda(
    self,
    x: torch.Tensor,
    residual: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
    if self.variance_size_override is not None:
        return self.forward_native(x, residual)

    add_residual = residual is not None
    norm_func = dispatch_cuda_rmsnorm_func(add_residual)

    if add_residual:
        return norm_func(x, residual, self.weight.data,
                         self.variance_epsilon)
    else:
        return norm_func(x, self.weight.data, self.variance_epsilon)

forward_hpu 

forward_hpu(
    x: Tensor, residual: Optional[Tensor] = None
) -> Union[Tensor, tuple[Tensor, Tensor]]
Source code in vllm/model_executor/layers/layernorm.py 
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
def forward_hpu(
    self,
    x: torch.Tensor,
    residual: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
    from vllm_hpu_extension.kernels import rms_norm
    HPUFusedRMSNorm = rms_norm()
    if HPUFusedRMSNorm is None:
        return self.forward_native(x, residual)
    if residual is not None:
        orig_shape = x.shape
        residual += x.view(residual.shape)
        # Note: HPUFusedRMSNorm requires 3D tensors as inputs
        x = HPUFusedRMSNorm.apply(residual, self.weight,
                                  self.variance_epsilon)
        return x.view(orig_shape), residual

    x = HPUFusedRMSNorm.apply(x, self.weight, self.variance_epsilon)
    return x

forward_native 

forward_native(
    x: Tensor, residual: Optional[Tensor] = None
) -> Union[Tensor, tuple[Tensor, Tensor]]

PyTorch-native implementation equivalent to forward().

Source code in vllm/model_executor/layers/layernorm.py 
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
def forward_native(
    self,
    x: torch.Tensor,
    residual: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
    """PyTorch-native implementation equivalent to forward()."""
    orig_dtype = x.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)
    x = x.to(orig_dtype)
    if self.has_weight:
        x = x * self.weight
    if residual is None:
        return x
    else:
        return x, residual

forward_xpu 

forward_xpu(
    x: Tensor, residual: Optional[Tensor] = None
) -> Union[Tensor, tuple[Tensor, Tensor]]
Source code in vllm/model_executor/layers/layernorm.py 
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
def forward_xpu(
    self,
    x: torch.Tensor,
    residual: Optional[torch.Tensor] = None,
) -> Union[torch.Tensor, tuple[torch.Tensor, torch.Tensor]]:
    if self.variance_size_override is not None:
        return self.forward_native(x, residual)

    from vllm._ipex_ops import ipex_ops as ops

    if residual is not None:
        ops.fused_add_rms_norm(
            x,
            residual,
            self.weight.data,
            self.variance_epsilon,
        )
        return x, residual
    return ops.rms_norm(
        x,
        self.weight.data,
        self.variance_epsilon,
    )

dispatch_cuda_rmsnorm_func 

dispatch_cuda_rmsnorm_func(add_residual: bool)
Source code in vllm/model_executor/layers/layernorm.py 
77
78
79
80
81
82
83
84
85
def dispatch_cuda_rmsnorm_func(add_residual: bool):
    if add_residual:
        if is_rocm_aiter_rmsnorm_enabled():
            return rocm_aiter_fused_add_rms_norm
        return fused_add_rms_norm

    if is_rocm_aiter_rmsnorm_enabled():
        return rocm_aiter_rms_norm
    return rms_norm

fused_add_rms_norm 

fused_add_rms_norm(
    x: Tensor,
    residual: Tensor,
    weight: Tensor,
    variance_epsilon: float,
) -> tuple[Tensor, Tensor]
Source code in vllm/model_executor/layers/layernorm.py 
33
34
35
36
37
38
39
40
41
42
43
def fused_add_rms_norm(
        x: torch.Tensor, residual: torch.Tensor, weight: torch.Tensor,
        variance_epsilon: float) -> tuple[torch.Tensor, torch.Tensor]:
    from vllm import _custom_ops as ops
    ops.fused_add_rms_norm(
        x,
        residual,
        weight,
        variance_epsilon,
    )
    return x, residual

is_rocm_aiter_rmsnorm_enabled 

is_rocm_aiter_rmsnorm_enabled() -> bool
Source code in vllm/model_executor/layers/layernorm.py 
14
15
16
17
def is_rocm_aiter_rmsnorm_enabled() -> bool:
    return current_platform.is_rocm() \
        and envs.VLLM_ROCM_USE_AITER_RMSNORM \
        and envs.VLLM_ROCM_USE_AITER

rms_norm 

rms_norm(
    x: Tensor, weight: Tensor, variance_epsilon: float
) -> Tensor
Source code in vllm/model_executor/layers/layernorm.py 
20
21
22
23
24
25
26
27
28
29
30
def rms_norm(x: torch.Tensor, weight: torch.Tensor,
             variance_epsilon: float) -> torch.Tensor:
    from vllm import _custom_ops as ops
    out = torch.empty_like(x)
    ops.rms_norm(
        out,
        x,
        weight,
        variance_epsilon,
    )
    return out

rocm_aiter_fused_add_rms_norm 

rocm_aiter_fused_add_rms_norm(
    x: Tensor,
    residual: Tensor,
    weight: Tensor,
    variance_epsilon: float,
) -> tuple[Tensor, Tensor]
Source code in vllm/model_executor/layers/layernorm.py 
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
def rocm_aiter_fused_add_rms_norm(
        x: torch.Tensor, residual: torch.Tensor, weight: torch.Tensor,
        variance_epsilon: float) -> tuple[torch.Tensor, torch.Tensor]:

    import aiter as rocm_aiter

    residual_out = torch.empty_like(residual)
    output = torch.empty_like(x)
    rocm_aiter.rmsnorm2d_fwd_with_add(
        output,  # output
        x,  # input
        residual,  # residual input
        residual_out,  # residual output
        weight,
        variance_epsilon,
    )
    return output, residual_out

rocm_aiter_rms_norm 

rocm_aiter_rms_norm(
    x: Tensor, weight: Tensor, variance_epsilon: float
) -> Tensor
Source code in vllm/model_executor/layers/layernorm.py 
46
47
48
49
50
51
52
53
54
55
def rocm_aiter_rms_norm(x: torch.Tensor, weight: torch.Tensor,
                        variance_epsilon: float) -> torch.Tensor:
    import aiter as rocm_aiter
    if x.dim() > 2:
        x_original_shape = x.shape
        x = x.reshape(-1, x_original_shape[-1])
        x = rocm_aiter.rms_norm(x, weight, variance_epsilon)
        return x.reshape(x_original_shape)

    return rocm_aiter.rms_norm(x, weight, variance_epsilon)