[Model] Consolidate ViTs attention implementation without mask (#10893)

Signed-off-by: Isotr0py <2037008807@qq.com>
This commit is contained in:
Isotr0py
2024-12-04 18:11:08 +00:00
committed by GitHub
parent 01d079fd8e
commit 10398b4706
9 changed files with 107 additions and 224 deletions
+4 -41
View File
@@ -4,11 +4,10 @@ from typing import Iterable, Optional, Set, Tuple, Union
import torch
import torch.nn as nn
import torch.nn.functional as F
from PIL import Image
from transformers import Blip2VisionConfig, BlipVisionConfig
from vllm.attention.selector import _Backend
from vllm.attention.layer import MultiHeadAttention
from vllm.config import ModelConfig
from vllm.distributed import divide, get_tensor_model_parallel_world_size
from vllm.inputs import DecoderOnlyInputs, token_inputs
@@ -22,8 +21,6 @@ from vllm.multimodal.utils import (cached_get_tokenizer,
repeat_and_pad_placeholder_tokens)
from vllm.sequence import SequenceData
from .utils import get_vit_attn_backend
def get_blip_patch_grid_length(*, image_size: int, patch_size: int) -> int:
assert image_size % patch_size == 0
@@ -205,11 +202,8 @@ class BlipAttention(nn.Module):
self.tp_size = get_tensor_model_parallel_world_size()
self.num_heads_per_partition = divide(self.num_heads, self.tp_size)
# Detect attention implementation.
self.attn_backend = get_vit_attn_backend(support_fa=False)
if self.attn_backend not in {_Backend.TORCH_SDPA, _Backend.XFORMERS}:
raise RuntimeError(
f"BLIP does not support {self.attn_backend} backend now.")
self.attn = MultiHeadAttention(self.num_heads_per_partition,
self.head_dim, self.scale)
def _shape(self, tensor: torch.Tensor, seq_len: int, bsz: int):
return tensor.view(bsz, seq_len, self.num_heads,
@@ -220,41 +214,10 @@ class BlipAttention(nn.Module):
hidden_states: torch.Tensor,
):
"""Input shape: Batch x Time x Channel"""
bsz, tgt_len, _ = hidden_states.size()
qkv_states, _ = self.qkv(hidden_states)
query_states, key_states, value_states = qkv_states.chunk(3, dim=-1)
query_states = query_states.view(bsz, tgt_len,
self.num_heads_per_partition,
self.head_dim)
key_states = key_states.view(bsz, tgt_len,
self.num_heads_per_partition,
self.head_dim)
value_states = value_states.view(bsz, tgt_len,
self.num_heads_per_partition,
self.head_dim)
if self.attn_backend == _Backend.XFORMERS:
from xformers import ops as xops
out = xops.memory_efficient_attention_forward(query_states,
key_states,
value_states,
p=self.dropout,
scale=self.scale)
elif self.attn_backend == _Backend.TORCH_SDPA:
query_states, key_states, value_states = (x.transpose(1, 2)
for x in (query_states,
key_states,
value_states))
out = F.scaled_dot_product_attention(query_states,
key_states,
value_states,
dropout_p=self.dropout,
scale=self.scale)
out = out.transpose(1, 2)
out = out.view(bsz, tgt_len, -1)
out = self.attn(query_states, key_states, value_states)
attn_output, _ = self.projection(out)
return attn_output, None