[Refactor] Remove multiple dead codes (#50285)
Signed-off-by: yewentao256 <zhyanwentao@126.com>
This commit is contained in:
@@ -8,14 +8,8 @@ import torch
|
||||
import torch.nn as nn
|
||||
import torch.nn.functional as F
|
||||
|
||||
from vllm.distributed import (
|
||||
divide,
|
||||
get_tensor_model_parallel_rank,
|
||||
get_tensor_model_parallel_world_size,
|
||||
)
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.custom_op import CustomOp
|
||||
from vllm.model_executor.utils import set_weight_attrs
|
||||
from vllm.platforms import CpuArchEnum, current_platform
|
||||
from vllm.triton_utils import tl, triton
|
||||
from vllm.utils.collection_utils import LazyDict
|
||||
@@ -788,48 +782,6 @@ class XIELU(CustomOp):
|
||||
return self.forward_native(input)
|
||||
|
||||
|
||||
class ScaledActivation(nn.Module):
|
||||
"""An activation function with post-scale parameters.
|
||||
|
||||
This is used for some quantization methods like AWQ.
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
act_module: nn.Module,
|
||||
intermediate_size: int,
|
||||
input_is_parallel: bool = True,
|
||||
params_dtype: torch.dtype | None = None,
|
||||
):
|
||||
super().__init__()
|
||||
self.act = act_module
|
||||
self.input_is_parallel = input_is_parallel
|
||||
if input_is_parallel:
|
||||
tp_size = get_tensor_model_parallel_world_size()
|
||||
intermediate_size_per_partition = divide(intermediate_size, tp_size)
|
||||
else:
|
||||
intermediate_size_per_partition = intermediate_size
|
||||
if params_dtype is None:
|
||||
params_dtype = torch.get_default_dtype()
|
||||
self.scales = nn.Parameter(
|
||||
torch.empty(intermediate_size_per_partition, dtype=params_dtype)
|
||||
)
|
||||
set_weight_attrs(self.scales, {"weight_loader": self.weight_loader})
|
||||
|
||||
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
||||
return self.act(x) / self.scales
|
||||
|
||||
def weight_loader(self, param: nn.Parameter, loaded_weight: torch.Tensor):
|
||||
param_data = param.data
|
||||
if self.input_is_parallel:
|
||||
tp_rank = get_tensor_model_parallel_rank()
|
||||
shard_size = param_data.shape[0]
|
||||
start_idx = tp_rank * shard_size
|
||||
loaded_weight = loaded_weight.narrow(0, start_idx, shard_size)
|
||||
assert param_data.shape == loaded_weight.shape
|
||||
param_data.copy_(loaded_weight)
|
||||
|
||||
|
||||
_ACTIVATION_REGISTRY = LazyDict(
|
||||
{
|
||||
"gelu": lambda: GELU(),
|
||||
|
||||
@@ -155,58 +155,6 @@ class Conv2dLayer(ConvLayerBase):
|
||||
return self._forward_conv(x)
|
||||
|
||||
|
||||
class CausalConv2dLayer(Conv2dLayer):
|
||||
"""
|
||||
A causal version of nn.Conv2d where each location in the 2D matrix would
|
||||
have no access to locations on its right or down
|
||||
All arguments are the same as nn.Conv2d except padding which should be
|
||||
set as None
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
in_channels: int,
|
||||
out_channels: int,
|
||||
kernel_size: int,
|
||||
stride: int,
|
||||
padding: int = 0,
|
||||
dilation: int = 1,
|
||||
groups: int = 1,
|
||||
bias: bool = True,
|
||||
padding_mode: str = "zeros",
|
||||
*,
|
||||
params_dtype: torch.dtype | None = None,
|
||||
) -> None:
|
||||
if padding is not None:
|
||||
raise ValueError(
|
||||
"Argument padding should be set to None for CausalConv2dLayer."
|
||||
)
|
||||
self._left_padding: int = kernel_size - 1
|
||||
self._right_padding: int = stride - 1
|
||||
padding = 0
|
||||
|
||||
super().__init__(
|
||||
in_channels,
|
||||
out_channels,
|
||||
kernel_size,
|
||||
stride,
|
||||
padding,
|
||||
dilation,
|
||||
groups,
|
||||
bias,
|
||||
padding_mode,
|
||||
params_dtype=params_dtype,
|
||||
)
|
||||
|
||||
def forward(
|
||||
self,
|
||||
x: torch.Tensor,
|
||||
) -> torch.Tensor:
|
||||
x = F.pad(x, pad=(self._left_padding, self._right_padding, 0, 0))
|
||||
x = super().forward(x)
|
||||
return x
|
||||
|
||||
|
||||
# --8<-- [start:conv3d]
|
||||
@CustomOp.register("conv3d")
|
||||
class Conv3dLayer(ConvLayerBase):
|
||||
|
||||
@@ -4,7 +4,6 @@
|
||||
from typing import TYPE_CHECKING, Any
|
||||
|
||||
import torch
|
||||
from torch.utils._python_dispatch import TorchDispatchMode
|
||||
|
||||
import vllm.envs as envs
|
||||
import vllm.model_executor.layers.fused_moe.modular_kernel as mk
|
||||
@@ -234,26 +233,6 @@ class Fp8Config(QuantizationConfig):
|
||||
return cache_scale_mapper | QuantizationConfig.get_cache_scale_mapper()
|
||||
|
||||
|
||||
class CopyNumelCounter(TorchDispatchMode):
|
||||
"""
|
||||
Tracks total number of elements modified with `copy_`. Useful for keeping
|
||||
track of weight loading where underlying weights can be arbitrarily
|
||||
transformed (such as with `narrow`) before calling copy.
|
||||
"""
|
||||
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
self.copied_numel = 0
|
||||
|
||||
def __torch_dispatch__(self, func, types, args=(), kwargs=None):
|
||||
if kwargs is None:
|
||||
kwargs = {}
|
||||
out = func(*args, **kwargs)
|
||||
if func == torch.ops.aten.copy_.default:
|
||||
self.copied_numel += args[0].numel()
|
||||
return out
|
||||
|
||||
|
||||
class Fp8LinearMethod(LinearMethodBase):
|
||||
"""Linear method for FP8.
|
||||
Supports loading FP8 checkpoints with static weight scale and
|
||||
|
||||
@@ -25,12 +25,6 @@ MOE_LAYER_ROUTER_GATE_SUFFIXES = {
|
||||
}
|
||||
|
||||
|
||||
def is_layer_moe_router_gate(prefix: str) -> bool:
|
||||
if not prefix:
|
||||
return False
|
||||
return prefix.rsplit(".", 1)[-1] in MOE_LAYER_ROUTER_GATE_SUFFIXES
|
||||
|
||||
|
||||
def get_token_bin_counts_and_mask(
|
||||
tokens: torch.Tensor,
|
||||
vocab_size: int,
|
||||
|
||||
Reference in New Issue
Block a user