465 lines
15 KiB
Python
465 lines
15 KiB
Python
# Copied and adapted from: https://github.com/hao-ai-lab/FastVideo
|
|
|
|
# SPDX-License-Identifier: Apache-2.0
|
|
# Adapted from vllm: https://github.com/vllm-project/vllm/blob/v0.7.3/vllm/model_executor/layers/layernorm.py
|
|
"""Custom normalization layers."""
|
|
from typing import Optional, Tuple, Union
|
|
|
|
import torch
|
|
import torch.nn as nn
|
|
import torch.nn.functional as F
|
|
from sgl_kernel import fused_add_rmsnorm, rmsnorm
|
|
|
|
from sglang.jit_kernel.norm import can_use_fused_inplace_qknorm, fused_inplace_qknorm
|
|
from sglang.multimodal_gen.runtime.layers.custom_op import CustomOp
|
|
from sglang.multimodal_gen.runtime.layers.triton_ops import (
|
|
fuse_scale_shift_kernel,
|
|
norm_infer,
|
|
rms_norm_fn,
|
|
)
|
|
from sglang.multimodal_gen.runtime.utils.common import get_bool_env_var
|
|
|
|
|
|
# Copied and adapted from sglang
|
|
@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,
|
|
dtype: torch.dtype = torch.float32,
|
|
var_hidden_size: Optional[int] = None,
|
|
) -> None:
|
|
super().__init__()
|
|
self.weight = nn.Parameter(torch.ones(hidden_size))
|
|
self.variance_epsilon = eps
|
|
self.hidden_size = hidden_size
|
|
self.variance_size_override = (
|
|
None if var_hidden_size == hidden_size else var_hidden_size
|
|
)
|
|
if get_bool_env_var("SGLANG_ENABLE_DETERMINISTIC_INFERENCE"):
|
|
self._forward_method = self.forward_native
|
|
|
|
def forward_triton(self, x: torch.Tensor, residual: Optional[torch.Tensor] = None):
|
|
return rms_norm_fn(
|
|
x, self.weight, bias=None, residual=residual, eps=self.variance_epsilon
|
|
)
|
|
|
|
def forward_cuda(
|
|
self,
|
|
x: torch.Tensor,
|
|
residual: Optional[torch.Tensor] = None,
|
|
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
|
shape = x.shape
|
|
x = x.reshape(-1, shape[-1])
|
|
if residual is not None:
|
|
residual_shape = residual.shape
|
|
residual = residual.view(-1, shape[-1])
|
|
|
|
if x.dtype == torch.float:
|
|
# fp32
|
|
out = self.forward_triton(x, residual)
|
|
elif self.variance_size_override is not None:
|
|
return self.forward_native(x, residual)
|
|
elif residual is not None:
|
|
fused_add_rmsnorm(x, residual, self.weight.data, self.variance_epsilon)
|
|
return x.view(shape), residual.view(residual_shape)
|
|
else:
|
|
out = rmsnorm(x, self.weight.data, self.variance_epsilon)
|
|
out = out.view(shape)
|
|
return out
|
|
|
|
def forward_native(
|
|
self,
|
|
x: torch.Tensor,
|
|
residual: Optional[torch.Tensor] = None,
|
|
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
|
if not x.is_contiguous():
|
|
x = x.contiguous()
|
|
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 * self.weight).to(orig_dtype)
|
|
if residual is None:
|
|
return x
|
|
else:
|
|
return x, residual
|
|
|
|
def forward_cpu(
|
|
self,
|
|
x: torch.Tensor,
|
|
residual: Optional[torch.Tensor] = None,
|
|
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
|
return self.forward_native(x, residual)
|
|
|
|
def forward_hip(
|
|
self,
|
|
x: torch.Tensor,
|
|
residual: Optional[torch.Tensor] = None,
|
|
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
|
# ROCm builds of sgl-kernel do not expose rmsnorm custom ops yet.
|
|
return self.forward_native(x, residual)
|
|
|
|
def extra_repr(self) -> str:
|
|
s = f"hidden_size={self.weight.data.size(0)}"
|
|
s += f", eps={self.variance_epsilon}"
|
|
return s
|
|
|
|
|
|
# Copied and adapted from sglang
|
|
@CustomOp.register("layer_norm")
|
|
class LayerNorm(CustomOp):
|
|
def __init__(
|
|
self,
|
|
hidden_size: int,
|
|
eps=1e-5,
|
|
bias: bool = True,
|
|
elementwise_affine=True,
|
|
device=None,
|
|
dtype=None,
|
|
) -> None:
|
|
super().__init__()
|
|
self.eps = eps
|
|
factory_kwargs = {"device": device, "dtype": dtype}
|
|
self.hidden_size = hidden_size
|
|
if elementwise_affine:
|
|
self.weight = torch.nn.Parameter(torch.empty(hidden_size, **factory_kwargs))
|
|
self.bias = (
|
|
torch.nn.Parameter(torch.empty(hidden_size, **factory_kwargs))
|
|
if bias
|
|
else None
|
|
)
|
|
else:
|
|
self.register_parameter("weight", None)
|
|
self.register_parameter("bias", None)
|
|
# Lazy cache for ones vector (not a registered buffer to avoid FSDP/meta issues)
|
|
self._weight_fallback_cache = None
|
|
|
|
def _get_weight_fallback(self, x: torch.Tensor) -> torch.Tensor:
|
|
wf = getattr(self, "_weight_fallback_cache", None)
|
|
if (
|
|
wf is None
|
|
or wf.device != x.device
|
|
or wf.dtype != x.dtype
|
|
or wf.numel() != self.hidden_size
|
|
):
|
|
wf = torch.ones(self.hidden_size, device=x.device, dtype=x.dtype)
|
|
self._weight_fallback_cache = wf
|
|
return wf
|
|
|
|
def forward_triton(self, x: torch.Tensor):
|
|
# Fast inference kernel without residual/dropout branches
|
|
return norm_infer(
|
|
x.view(-1, self.hidden_size),
|
|
self.weight,
|
|
self.bias,
|
|
eps=self.eps,
|
|
is_rms_norm=False,
|
|
).view(x.shape)
|
|
|
|
def forward_cuda(
|
|
self,
|
|
x: torch.Tensor,
|
|
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
|
shape = x.shape
|
|
x = x.view(-1, self.hidden_size)
|
|
return self.forward_triton(x).view(shape)
|
|
|
|
@torch.compile(backend="inductor")
|
|
def forward_native(
|
|
self,
|
|
x: torch.Tensor,
|
|
residual: Optional[torch.Tensor] = None,
|
|
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
|
input_dtype = x.dtype
|
|
mean = x.mean(-1, keepdim=True)
|
|
variance = (x - mean).pow(2).mean(-1, keepdim=True)
|
|
x = (x - mean) * torch.rsqrt(variance + self.eps)
|
|
if self.weight is not None:
|
|
x = self.weight * x
|
|
# if no affine, this is a no-op
|
|
if self.bias is not None:
|
|
x = x + self.bias
|
|
return x.to(input_dtype)
|
|
|
|
def forward_cpu(
|
|
self,
|
|
x: torch.Tensor,
|
|
residual: Optional[torch.Tensor] = None,
|
|
) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
|
|
return self.forward_native(x, residual)
|
|
|
|
def extra_repr(self) -> str:
|
|
s = f"hidden_size={self.weight.data.size(0)}"
|
|
s += f", eps={self.variance_epsilon}"
|
|
return s
|
|
|
|
|
|
class ScaleResidual(nn.Module):
|
|
"""
|
|
Applies gated residual connection.
|
|
"""
|
|
|
|
def __init__(self, prefix: str = ""):
|
|
super().__init__()
|
|
|
|
def forward(
|
|
self, residual: torch.Tensor, x: torch.Tensor, gate: torch.Tensor
|
|
) -> torch.Tensor:
|
|
"""Apply gated residual connection."""
|
|
# x.shape: [batch_size, seq_len, inner_dim]
|
|
if gate.dim() == 4:
|
|
# gate.shape: [batch_size, num_frames, 1, inner_dim]
|
|
num_frames = gate.shape[1]
|
|
frame_seqlen = x.shape[1] // num_frames
|
|
return residual + (
|
|
x.unflatten(dim=1, sizes=(num_frames, frame_seqlen)) * gate
|
|
).flatten(1, 2)
|
|
else:
|
|
# gate.shape: [batch_size, 1, inner_dim]
|
|
return residual + x * gate
|
|
|
|
|
|
# adapted from Diffusers: https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/normalization.py
|
|
# NOTE(will): Needed to match behavior of diffusers and wan2.1 even while using
|
|
# FSDP's MixedPrecisionPolicy
|
|
class FP32LayerNorm(nn.LayerNorm):
|
|
def forward(self, inputs: torch.Tensor) -> torch.Tensor:
|
|
origin_dtype = inputs.dtype
|
|
return F.layer_norm(
|
|
inputs.float(),
|
|
self.normalized_shape,
|
|
self.weight.float() if self.weight is not None else None,
|
|
self.bias.float() if self.bias is not None else None,
|
|
self.eps,
|
|
).to(origin_dtype)
|
|
|
|
|
|
class ScaleResidualLayerNormScaleShift(nn.Module):
|
|
"""
|
|
Fused operation that combines:
|
|
1. Gated residual connection
|
|
2. LayerNorm
|
|
3. Scale and shift operations
|
|
|
|
This reduces memory bandwidth by combining memory-bound operations.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
hidden_size: int,
|
|
norm_type: str = "rms",
|
|
eps: float = 1e-6,
|
|
elementwise_affine: bool = False,
|
|
dtype: torch.dtype = torch.float32,
|
|
compute_dtype: torch.dtype | None = None,
|
|
prefix: str = "",
|
|
):
|
|
super().__init__()
|
|
if norm_type == "rms":
|
|
self.norm = RMSNorm(
|
|
hidden_size, has_weight=elementwise_affine, eps=eps, dtype=dtype
|
|
)
|
|
elif norm_type == "layer":
|
|
if compute_dtype == torch.float32:
|
|
self.norm = FP32LayerNorm(
|
|
hidden_size, elementwise_affine=elementwise_affine, eps=eps
|
|
)
|
|
else:
|
|
self.norm = LayerNorm(
|
|
hidden_size,
|
|
elementwise_affine=elementwise_affine,
|
|
eps=eps,
|
|
dtype=dtype,
|
|
)
|
|
else:
|
|
raise NotImplementedError(f"Norm type {norm_type} not implemented")
|
|
|
|
def forward(
|
|
self,
|
|
residual: torch.Tensor,
|
|
x: torch.Tensor,
|
|
gate: torch.Tensor | int,
|
|
shift: torch.Tensor,
|
|
scale: torch.Tensor,
|
|
) -> tuple[torch.Tensor, torch.Tensor]:
|
|
"""
|
|
Apply gated residual connection, followed by layernorm and
|
|
scale/shift in a single fused operation.
|
|
|
|
Returns:
|
|
Tuple containing:
|
|
- normalized and modulated output of shape: [batch_size, seq_len, inner_dim]
|
|
- residual value (value after residual connection
|
|
but before normalization)
|
|
"""
|
|
# x.shape: [batch_size, seq_len, inner_dim]
|
|
# Apply residual connection with gating
|
|
if isinstance(gate, int):
|
|
# used by cross-attention, should be 1
|
|
assert gate == 1
|
|
residual_output = residual + x
|
|
elif isinstance(gate, torch.Tensor):
|
|
if gate.dim() == 4:
|
|
# gate.shape: [batch_size, num_frames, 1, inner_dim]
|
|
num_frames = gate.shape[1]
|
|
frame_seqlen = x.shape[1] // num_frames
|
|
residual_output = residual + (
|
|
x.unflatten(dim=1, sizes=(num_frames, frame_seqlen)) * gate
|
|
).flatten(1, 2)
|
|
else:
|
|
# used by bidirectional self attention
|
|
# gate.shape: [batch_size, 1, inner_dim]
|
|
residual_output = residual + x * gate
|
|
else:
|
|
raise ValueError(f"Gate type {type(gate)} not supported")
|
|
# residual_output.shape: [batch_size, seq_len, inner_dim]
|
|
|
|
# Apply normalization
|
|
normalized = self.norm(residual_output)
|
|
|
|
# modulated = fused_scale_shift(
|
|
# normalized,
|
|
# scale,
|
|
# shift,
|
|
# )
|
|
modulated = fuse_scale_shift_kernel(
|
|
normalized,
|
|
scale,
|
|
shift,
|
|
)
|
|
return modulated, residual_output
|
|
|
|
|
|
class LayerNormScaleShift(nn.Module):
|
|
"""
|
|
Fused operation that combines LayerNorm with scale and shift operations.
|
|
This reduces memory bandwidth by combining memory-bound operations.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
hidden_size: int,
|
|
norm_type: str = "rms",
|
|
eps: float = 1e-6,
|
|
elementwise_affine: bool = False,
|
|
dtype: torch.dtype = torch.float32,
|
|
compute_dtype: torch.dtype | None = None,
|
|
prefix: str = "",
|
|
):
|
|
super().__init__()
|
|
self.compute_dtype = compute_dtype
|
|
if norm_type == "rms":
|
|
self.norm = RMSNorm(hidden_size, has_weight=elementwise_affine, eps=eps)
|
|
elif norm_type == "layer":
|
|
if self.compute_dtype == torch.float32:
|
|
self.norm = FP32LayerNorm(
|
|
hidden_size, elementwise_affine=elementwise_affine, eps=eps
|
|
)
|
|
else:
|
|
self.norm = nn.LayerNorm(
|
|
hidden_size,
|
|
elementwise_affine=elementwise_affine,
|
|
eps=eps,
|
|
dtype=dtype,
|
|
)
|
|
else:
|
|
raise NotImplementedError(f"Norm type {norm_type} not implemented")
|
|
|
|
def forward(
|
|
self, x: torch.Tensor, shift: torch.Tensor, scale: torch.Tensor
|
|
) -> torch.Tensor:
|
|
"""Apply ln followed by scale and shift in a single fused operation."""
|
|
# x.shape: [batch_size, seq_len, inner_dim]
|
|
normalized = self.norm(x)
|
|
if self.compute_dtype == torch.float32:
|
|
normalized = normalized.float()
|
|
|
|
if scale.dim() == 4:
|
|
# scale.shape: [batch_size, num_frames, 1, inner_dim]
|
|
num_frames = scale.shape[1]
|
|
frame_seqlen = normalized.shape[1] // num_frames
|
|
output = (
|
|
normalized.unflatten(dim=1, sizes=(num_frames, frame_seqlen))
|
|
* (1.0 + scale)
|
|
+ shift
|
|
).flatten(1, 2)
|
|
else:
|
|
# scale.shape: [batch_size, 1, inner_dim]
|
|
# shift.shape: [batch_size, 1, inner_dim]
|
|
output = normalized * (1.0 + scale) + shift
|
|
|
|
if self.compute_dtype == torch.float32:
|
|
output = output.to(x.dtype)
|
|
|
|
return output
|
|
|
|
|
|
def apply_qk_norm(
|
|
q: torch.Tensor,
|
|
k: torch.Tensor,
|
|
q_norm: "RMSNorm",
|
|
k_norm: "RMSNorm",
|
|
head_dim: int,
|
|
allow_inplace: bool = True,
|
|
) -> Tuple[torch.Tensor, torch.Tensor]:
|
|
"""Apply QK normalization for query and key tensors.
|
|
|
|
Minimal multimodal_gen-only implementation: only the JIT fused inplace
|
|
QK-norm kernel path is supported (no fallback).
|
|
"""
|
|
|
|
batch_size = q.size(0)
|
|
q_eps = q_norm.variance_epsilon
|
|
k_eps = k_norm.variance_epsilon
|
|
# Only try fused path on CUDA and when it won't introduce implicit copies.
|
|
if (
|
|
q.is_cuda
|
|
and allow_inplace
|
|
and (q_eps == k_eps)
|
|
and can_use_fused_inplace_qknorm(head_dim)
|
|
):
|
|
fused_inplace_qknorm(
|
|
q=q.view(batch_size, -1, head_dim),
|
|
k=k.view(batch_size, -1, head_dim),
|
|
q_weight=q_norm.weight,
|
|
k_weight=k_norm.weight,
|
|
head_dim=head_dim,
|
|
eps=q_eps,
|
|
)
|
|
return q, k
|
|
|
|
raise RuntimeError(
|
|
"apply_qk_norm: fused inplace QK-norm is not applicable "
|
|
"(expected CUDA, contiguous q/k, matching eps, and supported head_dim)"
|
|
)
|