59 lines
1.8 KiB
Python
59 lines
1.8 KiB
Python
"""
|
|
Copyright 2023-2024 SGLang Team
|
|
Licensed under the Apache License, Version 2.0 (the "License");
|
|
you may not use this file except in compliance with the License.
|
|
You may obtain a copy of the License at
|
|
|
|
http://www.apache.org/licenses/LICENSE-2.0
|
|
|
|
Unless required by applicable law or agreed to in writing, software
|
|
distributed under the License is distributed on an "AS IS" BASIS,
|
|
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
See the License for the specific language governing permissions and
|
|
limitations under the License.
|
|
"""
|
|
|
|
"""Radix attention."""
|
|
|
|
from torch import nn
|
|
|
|
from sglang.srt.model_executor.forward_batch_info import InputMetadata
|
|
|
|
|
|
class RadixAttention(nn.Module):
|
|
"""
|
|
The attention layer implementation.
|
|
"""
|
|
|
|
def __init__(
|
|
self,
|
|
num_heads: int,
|
|
head_dim: int,
|
|
scaling: float,
|
|
num_kv_heads: int,
|
|
layer_id: int,
|
|
sliding_window_size: int = -1,
|
|
logit_cap: float = 0.0,
|
|
v_head_dim: int = -1,
|
|
):
|
|
super().__init__()
|
|
self.tp_q_head_num = num_heads
|
|
self.tp_k_head_num = num_kv_heads
|
|
self.tp_v_head_num = num_kv_heads
|
|
self.head_dim = head_dim
|
|
self.qk_head_dim = head_dim
|
|
self.v_head_dim = v_head_dim if v_head_dim != -1 else head_dim
|
|
self.scaling = scaling
|
|
self.layer_id = layer_id
|
|
self.logit_cap = logit_cap
|
|
self.sliding_window_size = sliding_window_size or -1
|
|
|
|
def forward(self, q, k, v, input_metadata: InputMetadata):
|
|
if k is not None:
|
|
# For cross-layer sharing, kv can be None
|
|
assert v is not None
|
|
k = k.view(-1, self.tp_k_head_num, self.qk_head_dim)
|
|
v = v.view(-1, self.tp_v_head_num, self.v_head_dim)
|
|
|
|
return input_metadata.attn_backend.forward(q, k, v, self, input_metadata)
|