Added helper functions from version e978c4c9b177e8d1f36deeec20edb18377dc2ff7
Browse filesAdded helper functions from an older version of the MiniCPM-V-2.5 repository which are still compatible with the tokenizer configuration used tot rain this model.
- configuration_minicpm.py +113 -0
- modeling_minicpmv.py +702 -0
- resampler.py +813 -0
configuration_minicpm.py
ADDED
@@ -0,0 +1,113 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
# coding=utf-8
|
2 |
+
# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved.
|
3 |
+
#
|
4 |
+
# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
|
5 |
+
# and OPT implementations in this library. It has been modified from its
|
6 |
+
# original forms to accommodate minor architectural differences compared
|
7 |
+
# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
|
8 |
+
#
|
9 |
+
# Licensed under the Apache License, Version 2.0 (the "License");
|
10 |
+
# you may not use this file except in compliance with the License.
|
11 |
+
# You may obtain a copy of the License at
|
12 |
+
#
|
13 |
+
# http://www.apache.org/licenses/LICENSE-2.0
|
14 |
+
#
|
15 |
+
# Unless required by applicable law or agreed to in writing, software
|
16 |
+
# distributed under the License is distributed on an "AS IS" BASIS,
|
17 |
+
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
18 |
+
# See the License for the specific language governing permissions and
|
19 |
+
# limitations under the License.
|
20 |
+
""" MiniCPM model configuration"""
|
21 |
+
import os
|
22 |
+
from typing import Union
|
23 |
+
|
24 |
+
from transformers.utils import logging
|
25 |
+
from transformers import LlamaConfig, PretrainedConfig
|
26 |
+
from transformers.models.idefics2.modeling_idefics2 import Idefics2VisionConfig
|
27 |
+
|
28 |
+
logger = logging.get_logger(__name__)
|
29 |
+
|
30 |
+
|
31 |
+
class MiniCPMVSliceConfig(PretrainedConfig):
|
32 |
+
model_type = "minicpmv"
|
33 |
+
|
34 |
+
def __init__(
|
35 |
+
self,
|
36 |
+
patch_size=14,
|
37 |
+
max_slice_nums=9,
|
38 |
+
scale_resolution=448,
|
39 |
+
**kwargs,
|
40 |
+
):
|
41 |
+
super().__init__(**kwargs)
|
42 |
+
self.patch_size = patch_size
|
43 |
+
self.max_slice_nums = max_slice_nums
|
44 |
+
self.scale_resolution = scale_resolution
|
45 |
+
|
46 |
+
@classmethod
|
47 |
+
def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
|
48 |
+
cls._set_token_in_kwargs(kwargs)
|
49 |
+
|
50 |
+
config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
|
51 |
+
|
52 |
+
if config_dict.get("model_type") == "minicpmv":
|
53 |
+
config_dict = config_dict["slice_config"]
|
54 |
+
|
55 |
+
if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
|
56 |
+
logger.warning(
|
57 |
+
f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
|
58 |
+
f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
|
59 |
+
)
|
60 |
+
|
61 |
+
return cls.from_dict(config_dict, **kwargs)
|
62 |
+
|
63 |
+
|
64 |
+
|
65 |
+
class MiniCPMVConfig(LlamaConfig):
|
66 |
+
model_type = "minicpmv"
|
67 |
+
keys_to_ignore_at_inference = ["past_key_values"]
|
68 |
+
|
69 |
+
default_vision_config = {
|
70 |
+
"hidden_size": 1152,
|
71 |
+
"image_size": 980,
|
72 |
+
"intermediate_size": 4304,
|
73 |
+
"model_type": "idefics2",
|
74 |
+
"num_attention_heads": 16,
|
75 |
+
"num_hidden_layers": 27,
|
76 |
+
"patch_size": 14,
|
77 |
+
}
|
78 |
+
|
79 |
+
def __init__(
|
80 |
+
self,
|
81 |
+
use_cache=True,
|
82 |
+
query_num=64,
|
83 |
+
image_size=448,
|
84 |
+
drop_vision_last_layer=True,
|
85 |
+
batch_vision_input=True,
|
86 |
+
slice_config=None,
|
87 |
+
vision_config=None,
|
88 |
+
**kwargs,
|
89 |
+
):
|
90 |
+
self.use_cache = use_cache
|
91 |
+
self.query_num = query_num
|
92 |
+
self.image_size = image_size
|
93 |
+
self.drop_vision_last_layer = drop_vision_last_layer
|
94 |
+
self.batch_vision_input = batch_vision_input
|
95 |
+
|
96 |
+
if slice_config is None:
|
97 |
+
self.slice_config = MiniCPMVSliceConfig(max_slice_nums=1)
|
98 |
+
else:
|
99 |
+
self.slice_config = MiniCPMVSliceConfig(**slice_config)
|
100 |
+
self.slice_mode = True
|
101 |
+
|
102 |
+
# same as HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit
|
103 |
+
if vision_config is None:
|
104 |
+
self.vision_config = Idefics2VisionConfig(**self.default_vision_config)
|
105 |
+
logger.info("vision_config is None, using default vision config")
|
106 |
+
elif isinstance(vision_config, dict):
|
107 |
+
self.vision_config = Idefics2VisionConfig(**vision_config)
|
108 |
+
elif isinstance(vision_config, Idefics2VisionConfig):
|
109 |
+
self.vision_config = vision_config
|
110 |
+
|
111 |
+
self.patch_size = self.vision_config.patch_size
|
112 |
+
|
113 |
+
super().__init__(**kwargs)
|
modeling_minicpmv.py
ADDED
@@ -0,0 +1,702 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import math
|
2 |
+
from typing import List, Optional
|
3 |
+
import json
|
4 |
+
import torch
|
5 |
+
import torchvision
|
6 |
+
from threading import Thread
|
7 |
+
from copy import deepcopy
|
8 |
+
from PIL import Image
|
9 |
+
from torchvision import transforms
|
10 |
+
from transformers import LlamaTokenizer, LlamaPreTrainedModel, LlamaForCausalLM, AutoModel, PreTrainedTokenizerFast, TextIteratorStreamer
|
11 |
+
from transformers.models.idefics2.modeling_idefics2 import Idefics2VisionTransformer
|
12 |
+
|
13 |
+
from .configuration_minicpm import MiniCPMVConfig
|
14 |
+
from .resampler import Resampler
|
15 |
+
|
16 |
+
IMAGENET_INCEPTION_MEAN = (0.5, 0.5, 0.5) # timm.data.IMAGENET_INCEPTION_MEAN
|
17 |
+
IMAGENET_INCEPTION_STD = (0.5, 0.5, 0.5) # timm.data.IMAGENET_INCEPTION_STD
|
18 |
+
|
19 |
+
class MiniCPMVPreTrainedModel(LlamaPreTrainedModel):
|
20 |
+
config_class = MiniCPMVConfig
|
21 |
+
|
22 |
+
|
23 |
+
class MiniCPMV(MiniCPMVPreTrainedModel):
|
24 |
+
def __init__(self, config):
|
25 |
+
super().__init__(config)
|
26 |
+
|
27 |
+
self.llm = LlamaForCausalLM(config)
|
28 |
+
self.vpm = self.init_vision_module()
|
29 |
+
self.vision_dim = self.vpm.embed_dim
|
30 |
+
self.embed_dim = self.llm.config.hidden_size
|
31 |
+
self.resampler = self.init_resampler(self.embed_dim, self.vision_dim)
|
32 |
+
self.transform = self.init_transform()
|
33 |
+
|
34 |
+
def init_vision_module(self):
|
35 |
+
# same as HuggingFaceM4/siglip-so400m-14-980-flash-attn2-navit
|
36 |
+
model = Idefics2VisionTransformer(self.config.vision_config)
|
37 |
+
if self.config.drop_vision_last_layer:
|
38 |
+
model.encoder.layers = model.encoder.layers[:-1]
|
39 |
+
|
40 |
+
setattr(model, 'embed_dim', model.embeddings.embed_dim)
|
41 |
+
setattr(model, 'patch_size', model.embeddings.patch_size)
|
42 |
+
|
43 |
+
return model
|
44 |
+
|
45 |
+
def init_resampler(self, embed_dim, vision_dim,):
|
46 |
+
return Resampler(
|
47 |
+
num_queries=self.config.query_num,
|
48 |
+
embed_dim=embed_dim,
|
49 |
+
num_heads=embed_dim // 128,
|
50 |
+
kv_dim=vision_dim,
|
51 |
+
adaptive=True,
|
52 |
+
)
|
53 |
+
|
54 |
+
def init_transform(self):
|
55 |
+
return transforms.Compose(
|
56 |
+
[
|
57 |
+
transforms.ToTensor(),
|
58 |
+
transforms.Normalize(
|
59 |
+
mean=IMAGENET_INCEPTION_MEAN, std=IMAGENET_INCEPTION_STD
|
60 |
+
),
|
61 |
+
]
|
62 |
+
)
|
63 |
+
|
64 |
+
def get_input_embeddings(self):
|
65 |
+
return self.llm.get_input_embeddings()
|
66 |
+
|
67 |
+
def set_input_embeddings(self, value):
|
68 |
+
self.llm.embed_tokens = value
|
69 |
+
|
70 |
+
def get_vllm_embedding(self, data):
|
71 |
+
if 'vision_hidden_states' not in data:
|
72 |
+
dtype = self.llm.model.embed_tokens.weight.dtype
|
73 |
+
device = self.llm.model.embed_tokens.weight.device
|
74 |
+
tgt_sizes = data['tgt_sizes']
|
75 |
+
pixel_values_list = data['pixel_values']
|
76 |
+
vision_hidden_states = []
|
77 |
+
all_pixel_values = []
|
78 |
+
img_cnt = []
|
79 |
+
for pixel_values in pixel_values_list:
|
80 |
+
img_cnt.append(len(pixel_values))
|
81 |
+
all_pixel_values.extend([i.flatten(end_dim=1).permute(1, 0) for i in pixel_values])
|
82 |
+
|
83 |
+
# exist image
|
84 |
+
if all_pixel_values:
|
85 |
+
tgt_sizes = torch.vstack(tgt_sizes).type(torch.int32)
|
86 |
+
|
87 |
+
if self.config.batch_vision_input:
|
88 |
+
max_patches = torch.max(tgt_sizes[:, 0] * tgt_sizes[:, 1])
|
89 |
+
|
90 |
+
all_pixel_values = torch.nn.utils.rnn.pad_sequence(all_pixel_values, batch_first=True,
|
91 |
+
padding_value=0.0)
|
92 |
+
B, L, _ = all_pixel_values.shape
|
93 |
+
all_pixel_values = all_pixel_values.permute(0, 2, 1).reshape(B, 3, -1, L)
|
94 |
+
|
95 |
+
patch_attn_mask = torch.zeros((B, 1, max_patches), dtype=torch.bool, device=device)
|
96 |
+
for i in range(B):
|
97 |
+
patch_attn_mask[i, :tgt_sizes[i][0] * tgt_sizes[i][1]] = True
|
98 |
+
|
99 |
+
vision_embedding = self.vpm(all_pixel_values.type(dtype), patch_attention_mask=patch_attn_mask).last_hidden_state
|
100 |
+
vision_embedding = self.resampler(vision_embedding, tgt_sizes)
|
101 |
+
else:
|
102 |
+
# get vision_embedding foreach
|
103 |
+
vision_embedding = []
|
104 |
+
for single_tgt_size, single_pixel_values in zip(tgt_sizes, all_pixel_values):
|
105 |
+
single_pixel_values = single_pixel_values.unsqueeze(0)
|
106 |
+
B, L, _ = single_pixel_values.shape
|
107 |
+
single_pixel_values = single_pixel_values.permute(0, 2, 1).reshape(B, 3, -1, L)
|
108 |
+
single_vision_embedding = self.vpm(single_pixel_values.type(dtype)).last_hidden_state
|
109 |
+
single_vision_embedding = self.resampler(single_vision_embedding, single_tgt_size.unsqueeze(0))
|
110 |
+
|
111 |
+
vision_embedding.append(single_vision_embedding)
|
112 |
+
vision_embedding = torch.vstack(vision_embedding)
|
113 |
+
|
114 |
+
start = 0
|
115 |
+
for pixel_values in pixel_values_list:
|
116 |
+
img_cnt = len(pixel_values)
|
117 |
+
if img_cnt > 0:
|
118 |
+
vision_hidden_states.append(vision_embedding[start: start + img_cnt])
|
119 |
+
start += img_cnt
|
120 |
+
else:
|
121 |
+
vision_hidden_states.append([])
|
122 |
+
else: # no image
|
123 |
+
if self.training:
|
124 |
+
dummy_image = torch.zeros(
|
125 |
+
(1, 3, 224, 224),
|
126 |
+
device=device, dtype=dtype
|
127 |
+
)
|
128 |
+
tgt_sizes = torch.Tensor([[(224 // self.config.patch_size), math.ceil(224 / self.config.patch_size)]]).type(torch.int32)
|
129 |
+
dummy_feature = self.resampler(self.vpm(dummy_image).last_hidden_state, tgt_sizes)
|
130 |
+
else:
|
131 |
+
dummy_feature = []
|
132 |
+
for _ in range(len(pixel_values_list)):
|
133 |
+
vision_hidden_states.append(dummy_feature)
|
134 |
+
|
135 |
+
else:
|
136 |
+
vision_hidden_states = data['vision_hidden_states']
|
137 |
+
|
138 |
+
if hasattr(self.llm.config, 'scale_emb'):
|
139 |
+
vllm_embedding = self.llm.model.embed_tokens(data['input_ids']) * self.llm.config.scale_emb
|
140 |
+
else:
|
141 |
+
vllm_embedding = self.llm.model.embed_tokens(data['input_ids'])
|
142 |
+
|
143 |
+
vision_hidden_states = [i.type(vllm_embedding.dtype) if isinstance(
|
144 |
+
i, torch.Tensor) else i for i in vision_hidden_states]
|
145 |
+
|
146 |
+
bs = len(data['input_ids'])
|
147 |
+
for i in range(bs):
|
148 |
+
cur_vs_hs = vision_hidden_states[i]
|
149 |
+
if len(cur_vs_hs) > 0:
|
150 |
+
cur_vllm_emb = vllm_embedding[i]
|
151 |
+
cur_image_bound = data['image_bound'][i]
|
152 |
+
if len(cur_image_bound) > 0:
|
153 |
+
image_indices = torch.stack(
|
154 |
+
[torch.arange(r[0], r[1], dtype=torch.long) for r in cur_image_bound]
|
155 |
+
).to(vllm_embedding.device)
|
156 |
+
cur_vllm_emb.scatter_(0, image_indices.view(-1, 1).repeat(1, cur_vllm_emb.shape[-1]),
|
157 |
+
cur_vs_hs.view(-1, cur_vs_hs.shape[-1]))
|
158 |
+
elif self.training:
|
159 |
+
cur_vllm_emb += cur_vs_hs[0].mean() * 0
|
160 |
+
|
161 |
+
return vllm_embedding, vision_hidden_states
|
162 |
+
|
163 |
+
def forward(self, data, **kwargs):
|
164 |
+
vllm_embedding, vision_hidden_states = self.get_vllm_embedding(data)
|
165 |
+
position_ids = data["position_ids"]
|
166 |
+
if position_ids.dtype != torch.int64:
|
167 |
+
position_ids = position_ids.long()
|
168 |
+
|
169 |
+
return self.llm(
|
170 |
+
input_ids=None,
|
171 |
+
position_ids=position_ids,
|
172 |
+
inputs_embeds=vllm_embedding,
|
173 |
+
**kwargs
|
174 |
+
)
|
175 |
+
|
176 |
+
def _convert_to_tensors(
|
177 |
+
self, tokenizer, input_ids, max_inp_length: Optional[int] = None
|
178 |
+
):
|
179 |
+
if max_inp_length is not None:
|
180 |
+
input_ids = input_ids[:max_inp_length]
|
181 |
+
input_ids = torch.tensor(input_ids, dtype=torch.int32)
|
182 |
+
|
183 |
+
image_start_tokens = torch.where(input_ids == tokenizer.im_start_id)[0]
|
184 |
+
# 跳过 im_start
|
185 |
+
image_start_tokens += 1
|
186 |
+
image_end_tokens = torch.where(input_ids == tokenizer.im_end_id)[0]
|
187 |
+
valid_image_nums = max(len(image_start_tokens), len(image_end_tokens))
|
188 |
+
image_bound = torch.hstack(
|
189 |
+
[
|
190 |
+
image_start_tokens[:valid_image_nums].unsqueeze(-1),
|
191 |
+
image_end_tokens[:valid_image_nums].unsqueeze(-1),
|
192 |
+
]
|
193 |
+
)
|
194 |
+
|
195 |
+
model_input = {}
|
196 |
+
model_input["input_ids"] = input_ids.unsqueeze(0).to(self.device)
|
197 |
+
model_input["image_bound"] = image_bound
|
198 |
+
|
199 |
+
return model_input
|
200 |
+
|
201 |
+
def _process_list(
|
202 |
+
self, tokenizer, input_id_list, max_inp_length: Optional[int] = None
|
203 |
+
):
|
204 |
+
pad_keys = ["input_ids"]
|
205 |
+
input_tensors = []
|
206 |
+
for input_ids in input_id_list:
|
207 |
+
input_tensors.append(
|
208 |
+
self._convert_to_tensors(tokenizer, input_ids, max_inp_length)
|
209 |
+
)
|
210 |
+
padded = {}
|
211 |
+
for key in pad_keys:
|
212 |
+
padded[key] = pad(input_tensors, key, padding_side="left").to(self.device)
|
213 |
+
padded["image_bound"] = [i["image_bound"] for i in input_tensors]
|
214 |
+
return padded
|
215 |
+
|
216 |
+
def _decode(self, inputs_embeds, tokenizer, **kwargs):
|
217 |
+
terminators = [
|
218 |
+
tokenizer.eos_token_id,
|
219 |
+
tokenizer.convert_tokens_to_ids("<|eot_id|>")
|
220 |
+
]
|
221 |
+
output = self.llm.generate(
|
222 |
+
inputs_embeds=inputs_embeds,
|
223 |
+
pad_token_id=0,
|
224 |
+
eos_token_id=terminators,
|
225 |
+
**kwargs
|
226 |
+
)
|
227 |
+
return self._decode_text(output, tokenizer)
|
228 |
+
|
229 |
+
def _decode_stream(self, inputs_embeds, tokenizer, **kwargs):
|
230 |
+
terminators = [
|
231 |
+
tokenizer.eos_token_id,
|
232 |
+
tokenizer.convert_tokens_to_ids("<|eot_id|>")
|
233 |
+
]
|
234 |
+
streamer = TextIteratorStreamer(tokenizer=tokenizer)
|
235 |
+
generation_kwargs = {
|
236 |
+
'inputs_embeds': inputs_embeds,
|
237 |
+
'pad_token_id': 0,
|
238 |
+
'eos_token_id': terminators,
|
239 |
+
'streamer': streamer
|
240 |
+
}
|
241 |
+
generation_kwargs.update(kwargs)
|
242 |
+
|
243 |
+
thread = Thread(target=self.llm.generate, kwargs=generation_kwargs)
|
244 |
+
thread.start()
|
245 |
+
|
246 |
+
return streamer
|
247 |
+
|
248 |
+
def _decode_text(self, result_ids, tokenizer):
|
249 |
+
result_text = []
|
250 |
+
for result in result_ids:
|
251 |
+
result = result[result != 0]
|
252 |
+
if result[0] == tokenizer.bos_id:
|
253 |
+
result = result[1:]
|
254 |
+
if result[-1] == tokenizer.eos_id or result[-1] == tokenizer.eot_id:
|
255 |
+
result = result[:-1]
|
256 |
+
result_text.append(tokenizer.decode(result).strip())
|
257 |
+
return result_text
|
258 |
+
|
259 |
+
def slice_image(self, image):
|
260 |
+
return slice_image(
|
261 |
+
image,
|
262 |
+
self.config.slice_config.max_slice_nums,
|
263 |
+
self.config.slice_config.scale_resolution,
|
264 |
+
self.config.slice_config.patch_size,
|
265 |
+
)
|
266 |
+
|
267 |
+
def get_slice_image_placeholder(self, image, tokenizer):
|
268 |
+
image_placeholder = (
|
269 |
+
tokenizer.im_start
|
270 |
+
+ tokenizer.unk_token * self.config.query_num
|
271 |
+
+ tokenizer.im_end
|
272 |
+
)
|
273 |
+
|
274 |
+
slice_images = []
|
275 |
+
|
276 |
+
source_image, patches, best_grid = slice_image(
|
277 |
+
image,
|
278 |
+
self.config.slice_config.max_slice_nums,
|
279 |
+
self.config.slice_config.scale_resolution,
|
280 |
+
self.config.slice_config.patch_size,
|
281 |
+
)
|
282 |
+
|
283 |
+
slice_images.append(source_image)
|
284 |
+
final_placeholder = image_placeholder
|
285 |
+
|
286 |
+
if len(patches) > 0:
|
287 |
+
for i in range(len(patches)):
|
288 |
+
for j in range(len(patches[0])):
|
289 |
+
slice_images.append(patches[i][j])
|
290 |
+
|
291 |
+
final_placeholder += get_grid_placeholder(
|
292 |
+
tokenizer, best_grid, self.config.query_num
|
293 |
+
)
|
294 |
+
|
295 |
+
return slice_images, final_placeholder
|
296 |
+
|
297 |
+
def reshape_by_patch(self, image_tensor):
|
298 |
+
"""
|
299 |
+
:param image_tensor: shape [3, H, W]
|
300 |
+
:param patch_size:
|
301 |
+
:return: [3, patch_size, HW/patch_size]
|
302 |
+
"""
|
303 |
+
patch_size = self.config.patch_size
|
304 |
+
patches = torch.nn.functional.unfold(
|
305 |
+
image_tensor,
|
306 |
+
(patch_size, patch_size),
|
307 |
+
stride=(patch_size, patch_size)
|
308 |
+
)
|
309 |
+
|
310 |
+
patches = patches.reshape(image_tensor.size(0), patch_size, patch_size, -1)
|
311 |
+
patches = patches.permute(0, 1, 3, 2).reshape(image_tensor.size(0), patch_size, -1)
|
312 |
+
return patches
|
313 |
+
|
314 |
+
def generate(
|
315 |
+
self,
|
316 |
+
input_id_list=None,
|
317 |
+
img_list=None,
|
318 |
+
tgt_sizes=None,
|
319 |
+
tokenizer=None,
|
320 |
+
max_inp_length: Optional[int] = None,
|
321 |
+
vision_hidden_states=None,
|
322 |
+
return_vision_hidden_states=False,
|
323 |
+
stream=False,
|
324 |
+
**kwargs
|
325 |
+
):
|
326 |
+
|
327 |
+
assert input_id_list is not None
|
328 |
+
bs = len(input_id_list)
|
329 |
+
if img_list == None:
|
330 |
+
img_list = [[] for i in range(bs)]
|
331 |
+
assert bs == len(img_list)
|
332 |
+
|
333 |
+
model_inputs = self._process_list(tokenizer, input_id_list, max_inp_length)
|
334 |
+
|
335 |
+
if vision_hidden_states is None:
|
336 |
+
pixel_values = []
|
337 |
+
for i in range(bs):
|
338 |
+
img_inps = []
|
339 |
+
for img in img_list[i]:
|
340 |
+
img_inps.append(img.to(self.device))
|
341 |
+
if img_inps:
|
342 |
+
pixel_values.append(img_inps)
|
343 |
+
else:
|
344 |
+
pixel_values.append([])
|
345 |
+
model_inputs["pixel_values"] = pixel_values
|
346 |
+
model_inputs['tgt_sizes'] = tgt_sizes
|
347 |
+
else:
|
348 |
+
model_inputs["vision_hidden_states"] = vision_hidden_states
|
349 |
+
|
350 |
+
with torch.inference_mode():
|
351 |
+
(
|
352 |
+
model_inputs["inputs_embeds"],
|
353 |
+
vision_hidden_states,
|
354 |
+
) = self.get_vllm_embedding(model_inputs)
|
355 |
+
|
356 |
+
if stream:
|
357 |
+
result = self._decode_stream(model_inputs["inputs_embeds"], tokenizer, **kwargs)
|
358 |
+
else:
|
359 |
+
result = self._decode(model_inputs["inputs_embeds"], tokenizer, **kwargs)
|
360 |
+
|
361 |
+
if return_vision_hidden_states:
|
362 |
+
return result, vision_hidden_states
|
363 |
+
|
364 |
+
return result
|
365 |
+
|
366 |
+
def chat(
|
367 |
+
self,
|
368 |
+
image,
|
369 |
+
msgs,
|
370 |
+
tokenizer,
|
371 |
+
vision_hidden_states=None,
|
372 |
+
max_new_tokens=1024,
|
373 |
+
sampling=True,
|
374 |
+
max_inp_length=2048,
|
375 |
+
system_prompt='',
|
376 |
+
stream=False,
|
377 |
+
**kwargs
|
378 |
+
):
|
379 |
+
if isinstance(msgs, str):
|
380 |
+
msgs = json.loads(msgs)
|
381 |
+
|
382 |
+
copy_msgs = deepcopy(msgs)
|
383 |
+
assert len(copy_msgs) > 0, 'msgs is empty'
|
384 |
+
assert sampling or not stream, 'if use stream mode, make sure sampling=True'
|
385 |
+
|
386 |
+
if image is not None and isinstance(copy_msgs[0]['content'], str):
|
387 |
+
copy_msgs[0]['content'] = [image, copy_msgs[0]['content']]
|
388 |
+
|
389 |
+
images = []
|
390 |
+
tgt_sizes = []
|
391 |
+
for i, msg in enumerate(copy_msgs):
|
392 |
+
role = msg["role"]
|
393 |
+
content = msg["content"]
|
394 |
+
assert role in ["user", "assistant"]
|
395 |
+
if i == 0:
|
396 |
+
assert role == "user", "The role of first msg should be user"
|
397 |
+
if isinstance(content, str):
|
398 |
+
content = [content]
|
399 |
+
|
400 |
+
cur_msgs = []
|
401 |
+
for c in content:
|
402 |
+
if isinstance(c, Image.Image):
|
403 |
+
image = c
|
404 |
+
if self.config.slice_mode:
|
405 |
+
slice_images, image_placeholder = self.get_slice_image_placeholder(
|
406 |
+
image, tokenizer
|
407 |
+
)
|
408 |
+
cur_msgs.append(image_placeholder)
|
409 |
+
for slice_image in slice_images:
|
410 |
+
slice_image = self.transform(slice_image)
|
411 |
+
H, W = slice_image.shape[1:]
|
412 |
+
images.append(self.reshape_by_patch(slice_image))
|
413 |
+
tgt_sizes.append(torch.Tensor([H // self.config.patch_size, W // self.config.patch_size]).type(torch.int32))
|
414 |
+
else:
|
415 |
+
images.append(self.transform(image))
|
416 |
+
cur_msgs.append(
|
417 |
+
tokenizer.im_start
|
418 |
+
+ tokenizer.unk_token * self.config.query_num
|
419 |
+
+ tokenizer.im_end
|
420 |
+
)
|
421 |
+
elif isinstance(c, str):
|
422 |
+
cur_msgs.append(c)
|
423 |
+
|
424 |
+
|
425 |
+
msg['content'] = '\n'.join(cur_msgs)
|
426 |
+
if tgt_sizes:
|
427 |
+
tgt_sizes = torch.vstack(tgt_sizes)
|
428 |
+
|
429 |
+
if system_prompt:
|
430 |
+
sys_msg = {'role': 'system', 'content': system_prompt}
|
431 |
+
copy_msgs = [sys_msg] + copy_msgs
|
432 |
+
|
433 |
+
input_ids = tokenizer.apply_chat_template(copy_msgs, tokenize=True, add_generation_prompt=False)
|
434 |
+
|
435 |
+
if sampling:
|
436 |
+
generation_config = {
|
437 |
+
"top_p": 0.8,
|
438 |
+
"top_k": 100,
|
439 |
+
"temperature": 0.7,
|
440 |
+
"do_sample": True,
|
441 |
+
"repetition_penalty": 1.05
|
442 |
+
}
|
443 |
+
else:
|
444 |
+
generation_config = {
|
445 |
+
"num_beams": 3,
|
446 |
+
"repetition_penalty": 1.2,
|
447 |
+
}
|
448 |
+
|
449 |
+
generation_config.update(
|
450 |
+
(k, kwargs[k]) for k in generation_config.keys() & kwargs.keys()
|
451 |
+
)
|
452 |
+
|
453 |
+
with torch.inference_mode():
|
454 |
+
res, vision_hidden_states = self.generate(
|
455 |
+
input_id_list=[input_ids],
|
456 |
+
max_inp_length=max_inp_length,
|
457 |
+
img_list=[images],
|
458 |
+
tgt_sizes=[tgt_sizes],
|
459 |
+
tokenizer=tokenizer,
|
460 |
+
max_new_tokens=max_new_tokens,
|
461 |
+
vision_hidden_states=vision_hidden_states,
|
462 |
+
return_vision_hidden_states=True,
|
463 |
+
stream=stream,
|
464 |
+
**generation_config
|
465 |
+
)
|
466 |
+
|
467 |
+
if stream:
|
468 |
+
def stream_gen():
|
469 |
+
for text in res:
|
470 |
+
text = text.replace(tokenizer.eot_token, '').replace(tokenizer.eos_token, '')
|
471 |
+
yield text
|
472 |
+
return stream_gen()
|
473 |
+
|
474 |
+
else:
|
475 |
+
answer = res[0]
|
476 |
+
return answer
|
477 |
+
|
478 |
+
|
479 |
+
class PreTrainedTokenizerFastWrapper(PreTrainedTokenizerFast):
|
480 |
+
def __init__(self, **kwargs):
|
481 |
+
super().__init__(**kwargs)
|
482 |
+
self.eot_token = "<|eot_id|>"
|
483 |
+
self.im_start = "<image>"
|
484 |
+
self.im_end = "</image>"
|
485 |
+
self.ref_start = "<ref>"
|
486 |
+
self.ref_end = "</ref>"
|
487 |
+
self.box_start = "<box>"
|
488 |
+
self.box_end = "</box>"
|
489 |
+
self.quad_start = "<quad>"
|
490 |
+
self.quad_end = "</quad>"
|
491 |
+
self.slice_start = "<slice>"
|
492 |
+
self.slice_end = "</slice>"
|
493 |
+
|
494 |
+
@property
|
495 |
+
def eos_id(self):
|
496 |
+
return self.eos_token_id
|
497 |
+
|
498 |
+
@property
|
499 |
+
def bos_id(self):
|
500 |
+
return self.bos_token_id
|
501 |
+
|
502 |
+
@property
|
503 |
+
def unk_id(self):
|
504 |
+
return self.unk_token_id
|
505 |
+
|
506 |
+
@property
|
507 |
+
def eot_id(self):
|
508 |
+
return self.convert_tokens_to_ids(self.eot_token)
|
509 |
+
|
510 |
+
@property
|
511 |
+
def im_start_id(self):
|
512 |
+
return self.convert_tokens_to_ids(self.im_start)
|
513 |
+
|
514 |
+
@property
|
515 |
+
def im_end_id(self):
|
516 |
+
return self.convert_tokens_to_ids(self.im_end)
|
517 |
+
|
518 |
+
@staticmethod
|
519 |
+
def escape(text: str) -> str:
|
520 |
+
return text
|
521 |
+
|
522 |
+
@staticmethod
|
523 |
+
def unescape(text: str) -> str:
|
524 |
+
return text
|
525 |
+
|
526 |
+
|
527 |
+
def pad(orig_items, key, max_length=None, padding_value=0, padding_side="left"):
|
528 |
+
items = []
|
529 |
+
if isinstance(orig_items[0][key], list):
|
530 |
+
assert isinstance(orig_items[0][key][0], torch.Tensor)
|
531 |
+
for it in orig_items:
|
532 |
+
for tr in it[key]:
|
533 |
+
items.append({key: tr})
|
534 |
+
else:
|
535 |
+
assert isinstance(orig_items[0][key], torch.Tensor)
|
536 |
+
items = orig_items
|
537 |
+
|
538 |
+
batch_size = len(items)
|
539 |
+
shape = items[0][key].shape
|
540 |
+
dim = len(shape)
|
541 |
+
assert dim <= 3
|
542 |
+
if max_length is None:
|
543 |
+
max_length = 0
|
544 |
+
max_length = max(max_length, max(item[key].shape[-1] for item in items))
|
545 |
+
min_length = min(item[key].shape[-1] for item in items)
|
546 |
+
dtype = items[0][key].dtype
|
547 |
+
|
548 |
+
if dim == 1:
|
549 |
+
return torch.cat([item[key] for item in items], dim=0)
|
550 |
+
elif dim == 2:
|
551 |
+
if max_length == min_length:
|
552 |
+
return torch.cat([item[key] for item in items], dim=0)
|
553 |
+
tensor = torch.zeros((batch_size, max_length), dtype=dtype) + padding_value
|
554 |
+
else:
|
555 |
+
tensor = (
|
556 |
+
torch.zeros((batch_size, max_length, shape[-1]), dtype=dtype)
|
557 |
+
+ padding_value
|
558 |
+
)
|
559 |
+
|
560 |
+
for i, item in enumerate(items):
|
561 |
+
if dim == 2:
|
562 |
+
if padding_side == "left":
|
563 |
+
tensor[i, -len(item[key][0]) :] = item[key][0].clone()
|
564 |
+
else:
|
565 |
+
tensor[i, : len(item[key][0])] = item[key][0].clone()
|
566 |
+
elif dim == 3:
|
567 |
+
if padding_side == "left":
|
568 |
+
tensor[i, -len(item[key][0]) :, :] = item[key][0].clone()
|
569 |
+
else:
|
570 |
+
tensor[i, : len(item[key][0]), :] = item[key][0].clone()
|
571 |
+
|
572 |
+
return tensor
|
573 |
+
|
574 |
+
|
575 |
+
def slice_image(
|
576 |
+
image, max_slice_nums=9, scale_resolution=448, patch_size=14, never_split=False
|
577 |
+
):
|
578 |
+
original_size = image.size
|
579 |
+
original_width, original_height = original_size
|
580 |
+
log_ratio = math.log(original_width / original_height)
|
581 |
+
ratio = original_width * original_height / (scale_resolution * scale_resolution)
|
582 |
+
multiple = min(math.ceil(ratio), max_slice_nums)
|
583 |
+
|
584 |
+
source_image = None
|
585 |
+
best_grid = None
|
586 |
+
patches = []
|
587 |
+
|
588 |
+
if multiple <= 1 or never_split:
|
589 |
+
# dont need to slice, upsample
|
590 |
+
best_size = find_best_resize(
|
591 |
+
original_size, scale_resolution, patch_size, allow_upscale=True
|
592 |
+
)
|
593 |
+
source_image = image.resize(best_size, Image.Resampling.BICUBIC)
|
594 |
+
else:
|
595 |
+
candidate_split_grids_nums = []
|
596 |
+
for i in [multiple - 1, multiple, multiple + 1]:
|
597 |
+
if i == 1 or i > max_slice_nums:
|
598 |
+
continue
|
599 |
+
candidate_split_grids_nums.append(i)
|
600 |
+
|
601 |
+
# source image, down-sampling and ensure divided by patch_size
|
602 |
+
best_resize = find_best_resize(original_size, scale_resolution, patch_size)
|
603 |
+
source_image = image.copy().resize(best_resize, Image.Resampling.BICUBIC)
|
604 |
+
candidate_grids = []
|
605 |
+
|
606 |
+
# find best grid
|
607 |
+
for split_grids_nums in candidate_split_grids_nums:
|
608 |
+
m = 1
|
609 |
+
while m <= split_grids_nums:
|
610 |
+
if split_grids_nums % m == 0:
|
611 |
+
candidate_grids.append([m, split_grids_nums // m])
|
612 |
+
m += 1
|
613 |
+
|
614 |
+
best_grid = [1, 1]
|
615 |
+
min_error = float("inf")
|
616 |
+
for grid in candidate_grids:
|
617 |
+
error = abs(log_ratio - math.log(grid[0] / grid[1]))
|
618 |
+
if error < min_error:
|
619 |
+
best_grid = grid
|
620 |
+
min_error = error
|
621 |
+
|
622 |
+
refine_size = get_refine_size(
|
623 |
+
original_size, best_grid, scale_resolution, patch_size, allow_upscale=True
|
624 |
+
)
|
625 |
+
|
626 |
+
refine_image = image.resize(refine_size, Image.Resampling.BICUBIC)
|
627 |
+
patches = split_to_patches(refine_image, best_grid)
|
628 |
+
|
629 |
+
return source_image, patches, best_grid
|
630 |
+
|
631 |
+
|
632 |
+
def ensure_divide(length, patch_size):
|
633 |
+
return max(round(length / patch_size) * patch_size, patch_size)
|
634 |
+
|
635 |
+
|
636 |
+
def find_best_resize(original_size, scale_resolution, patch_size, allow_upscale=False):
|
637 |
+
width, height = original_size
|
638 |
+
if (width * height > scale_resolution * scale_resolution) or allow_upscale:
|
639 |
+
r = width / height
|
640 |
+
height = int(scale_resolution / math.sqrt(r))
|
641 |
+
width = int(height * r)
|
642 |
+
best_width = ensure_divide(width, patch_size)
|
643 |
+
best_height = ensure_divide(height, patch_size)
|
644 |
+
return (best_width, best_height)
|
645 |
+
|
646 |
+
|
647 |
+
def get_refine_size(
|
648 |
+
original_size, grid, scale_resolution, patch_size, allow_upscale=False
|
649 |
+
):
|
650 |
+
width, height = original_size
|
651 |
+
grid_x, grid_y = grid
|
652 |
+
|
653 |
+
refine_width = ensure_divide(width, grid_x)
|
654 |
+
refine_height = ensure_divide(height, grid_y)
|
655 |
+
|
656 |
+
grid_width = refine_width / grid_x
|
657 |
+
grid_height = refine_height / grid_y
|
658 |
+
|
659 |
+
best_grid_size = find_best_resize(
|
660 |
+
(grid_width, grid_height),
|
661 |
+
scale_resolution,
|
662 |
+
patch_size,
|
663 |
+
allow_upscale=allow_upscale,
|
664 |
+
)
|
665 |
+
|
666 |
+
refine_size = (best_grid_size[0] * grid_x, best_grid_size[1] * grid_y)
|
667 |
+
|
668 |
+
return refine_size
|
669 |
+
|
670 |
+
|
671 |
+
def split_to_patches(image, grid):
|
672 |
+
patches = []
|
673 |
+
width, height = image.size
|
674 |
+
grid_x = int(width / grid[0])
|
675 |
+
grid_y = int(height / grid[1])
|
676 |
+
|
677 |
+
for i in range(0, height, grid_y):
|
678 |
+
images = []
|
679 |
+
for j in range(0, width, grid_x):
|
680 |
+
box = (j, i, j + grid_x, i + grid_y)
|
681 |
+
patch = image.crop(box)
|
682 |
+
images.append(patch)
|
683 |
+
patches.append(images)
|
684 |
+
|
685 |
+
return patches
|
686 |
+
|
687 |
+
|
688 |
+
def get_grid_placeholder(tokenizer, grid, query_num):
|
689 |
+
image_placeholder = (
|
690 |
+
tokenizer.im_start + tokenizer.unk_token * query_num + tokenizer.im_end
|
691 |
+
)
|
692 |
+
|
693 |
+
cols = grid[0]
|
694 |
+
rows = grid[1]
|
695 |
+
slices = []
|
696 |
+
for i in range(rows):
|
697 |
+
lines = []
|
698 |
+
for j in range(cols):
|
699 |
+
lines.append(image_placeholder)
|
700 |
+
slices.append("".join(lines))
|
701 |
+
slice_placeholder = tokenizer.slice_start + "\n".join(slices) + tokenizer.slice_end
|
702 |
+
return slice_placeholder
|
resampler.py
ADDED
@@ -0,0 +1,813 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from functools import partial
|
2 |
+
import numpy as np
|
3 |
+
import warnings
|
4 |
+
from typing import Optional, Tuple
|
5 |
+
import torch
|
6 |
+
from torch import nn
|
7 |
+
from torch import Tensor
|
8 |
+
import deepspeed
|
9 |
+
import torch.nn.functional as F
|
10 |
+
from torch.nn.functional import *
|
11 |
+
from torch.nn.modules.activation import *
|
12 |
+
from torch.nn.init import trunc_normal_
|
13 |
+
from torch.nn.init import constant_, xavier_normal_, xavier_uniform_
|
14 |
+
from transformers import PreTrainedModel
|
15 |
+
from transformers.integrations import is_deepspeed_zero3_enabled
|
16 |
+
|
17 |
+
def get_2d_sincos_pos_embed(embed_dim, image_size):
|
18 |
+
"""
|
19 |
+
image_size: image_size or (image_height, image_width)
|
20 |
+
return:
|
21 |
+
pos_embed: [image_height, image_width, embed_dim]
|
22 |
+
"""
|
23 |
+
if isinstance(image_size, int):
|
24 |
+
grid_h_size, grid_w_size = image_size, image_size
|
25 |
+
else:
|
26 |
+
grid_h_size, grid_w_size = image_size[0], image_size[1]
|
27 |
+
|
28 |
+
grid_h = np.arange(grid_h_size, dtype=np.float32)
|
29 |
+
grid_w = np.arange(grid_w_size, dtype=np.float32)
|
30 |
+
grid = np.meshgrid(grid_w, grid_h) # here w goes first
|
31 |
+
grid = np.stack(grid, axis=0)
|
32 |
+
|
33 |
+
pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
|
34 |
+
return pos_embed
|
35 |
+
|
36 |
+
|
37 |
+
def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
|
38 |
+
assert embed_dim % 2 == 0
|
39 |
+
|
40 |
+
# use half of dimensions to encode grid_h
|
41 |
+
emb_h = get_1d_sincos_pos_embed_from_grid_new(embed_dim // 2, grid[0]) # (H, W, D/2)
|
42 |
+
emb_w = get_1d_sincos_pos_embed_from_grid_new(embed_dim // 2, grid[1]) # (H, W, D/2)
|
43 |
+
|
44 |
+
emb = np.concatenate([emb_h, emb_w], axis=-1) # (H, W, D)
|
45 |
+
return emb
|
46 |
+
|
47 |
+
|
48 |
+
def get_1d_sincos_pos_embed_from_grid_new(embed_dim, pos):
|
49 |
+
"""
|
50 |
+
embed_dim: output dimension for each position
|
51 |
+
pos: a list of positions to be encoded: size (H, W)
|
52 |
+
out: (H, W, D)
|
53 |
+
"""
|
54 |
+
assert embed_dim % 2 == 0
|
55 |
+
omega = np.arange(embed_dim // 2, dtype=np.float32)
|
56 |
+
omega /= embed_dim / 2.
|
57 |
+
omega = 1. / 10000 ** omega # (D/2,)
|
58 |
+
|
59 |
+
out = np.einsum('hw,d->hwd', pos, omega) # (H, W, D/2), outer product
|
60 |
+
|
61 |
+
emb_sin = np.sin(out) # (H, W, D/2)
|
62 |
+
emb_cos = np.cos(out) # (H, W, D/2)
|
63 |
+
|
64 |
+
emb = np.concatenate([emb_sin, emb_cos], axis=-1) # (H, W, D)
|
65 |
+
return emb
|
66 |
+
|
67 |
+
|
68 |
+
class Resampler(nn.Module):
|
69 |
+
"""
|
70 |
+
A 2D perceiver-resampler network with one cross attention layers by
|
71 |
+
given learnable queries and 2d sincos pos_emb
|
72 |
+
Outputs:
|
73 |
+
A tensor with the shape of (batch_size, num_queries, embed_dim)
|
74 |
+
"""
|
75 |
+
|
76 |
+
def __init__(
|
77 |
+
self,
|
78 |
+
num_queries,
|
79 |
+
embed_dim,
|
80 |
+
num_heads,
|
81 |
+
kv_dim=None,
|
82 |
+
norm_layer=partial(nn.LayerNorm, eps=1e-6),
|
83 |
+
adaptive=False,
|
84 |
+
max_size=(70, 70),
|
85 |
+
):
|
86 |
+
super().__init__()
|
87 |
+
self.num_queries = num_queries
|
88 |
+
self.embed_dim = embed_dim
|
89 |
+
self.num_heads = num_heads
|
90 |
+
self.adaptive = adaptive
|
91 |
+
self.max_size = max_size
|
92 |
+
|
93 |
+
self.query = nn.Parameter(torch.zeros(self.num_queries, embed_dim))
|
94 |
+
|
95 |
+
if kv_dim is not None and kv_dim != embed_dim:
|
96 |
+
self.kv_proj = nn.Linear(kv_dim, embed_dim, bias=False)
|
97 |
+
else:
|
98 |
+
self.kv_proj = nn.Identity()
|
99 |
+
|
100 |
+
self.attn = MultiheadAttention(embed_dim, num_heads)
|
101 |
+
self.ln_q = norm_layer(embed_dim)
|
102 |
+
self.ln_kv = norm_layer(embed_dim)
|
103 |
+
|
104 |
+
self.ln_post = norm_layer(embed_dim)
|
105 |
+
self.proj = nn.Parameter((embed_dim ** -0.5) * torch.randn(embed_dim, embed_dim))
|
106 |
+
|
107 |
+
self._set_2d_pos_cache(self.max_size)
|
108 |
+
|
109 |
+
def _set_2d_pos_cache(self, max_size, device='cpu'):
|
110 |
+
if is_deepspeed_zero3_enabled():
|
111 |
+
device='cuda'
|
112 |
+
pos_embed = torch.from_numpy(get_2d_sincos_pos_embed(self.embed_dim, max_size)).float().to(device)
|
113 |
+
self.register_buffer("pos_embed", pos_embed, persistent=False)
|
114 |
+
|
115 |
+
def _adjust_pos_cache(self, tgt_sizes, device):
|
116 |
+
max_h = torch.max(tgt_sizes[:, 0])
|
117 |
+
max_w = torch.max(tgt_sizes[:, 1])
|
118 |
+
if max_h > self.max_size[0] or max_w > self.max_size[1]:
|
119 |
+
self.max_size = [max(max_h, self.max_size[0]), max(max_w, self.max_size[1])]
|
120 |
+
self._set_2d_pos_cache(self.max_size, device)
|
121 |
+
|
122 |
+
def _init_weights(self, m):
|
123 |
+
if isinstance(m, nn.Linear):
|
124 |
+
trunc_normal_(m.weight, std=.02)
|
125 |
+
if isinstance(m, nn.Linear) and m.bias is not None:
|
126 |
+
nn.init.constant_(m.bias, 0)
|
127 |
+
elif isinstance(m, nn.LayerNorm):
|
128 |
+
nn.init.constant_(m.bias, 0)
|
129 |
+
nn.init.constant_(m.weight, 1.0)
|
130 |
+
|
131 |
+
def forward(self, x, tgt_sizes=None):
|
132 |
+
assert x.shape[0] == tgt_sizes.shape[0]
|
133 |
+
bs = x.shape[0]
|
134 |
+
|
135 |
+
device = x.device
|
136 |
+
dtype = x.dtype
|
137 |
+
|
138 |
+
patch_len = tgt_sizes[:, 0] * tgt_sizes[:, 1]
|
139 |
+
|
140 |
+
self._adjust_pos_cache(tgt_sizes, device=device)
|
141 |
+
|
142 |
+
max_patch_len = torch.max(patch_len)
|
143 |
+
key_padding_mask = torch.zeros((bs, max_patch_len), dtype=torch.bool, device=device)
|
144 |
+
|
145 |
+
pos_embed = []
|
146 |
+
for i in range(bs):
|
147 |
+
tgt_h, tgt_w = tgt_sizes[i]
|
148 |
+
pos_embed.append(self.pos_embed[:tgt_h, :tgt_w, :].reshape((tgt_h * tgt_w, -1)).to(dtype)) # patches * D
|
149 |
+
key_padding_mask[i, patch_len[i]:] = True
|
150 |
+
|
151 |
+
pos_embed = torch.nn.utils.rnn.pad_sequence(
|
152 |
+
pos_embed, batch_first=True, padding_value=0.0).permute(1, 0, 2) # BLD => L * B * D
|
153 |
+
|
154 |
+
x = self.kv_proj(x) # B * L * D
|
155 |
+
x = self.ln_kv(x).permute(1, 0, 2) # L * B * D
|
156 |
+
|
157 |
+
q = self.ln_q(self.query) # Q * D
|
158 |
+
|
159 |
+
out = self.attn(
|
160 |
+
self._repeat(q, bs), # Q * B * D
|
161 |
+
x + pos_embed, # L * B * D + L * B * D
|
162 |
+
x,
|
163 |
+
key_padding_mask=key_padding_mask)[0]
|
164 |
+
# out: Q * B * D
|
165 |
+
x = out.permute(1, 0, 2) # B * Q * D
|
166 |
+
|
167 |
+
x = self.ln_post(x)
|
168 |
+
x = x @ self.proj
|
169 |
+
return x
|
170 |
+
|
171 |
+
def _repeat(self, query, N: int):
|
172 |
+
return query.unsqueeze(1).repeat(1, N, 1)
|
173 |
+
|
174 |
+
|
175 |
+
class MultiheadAttention(nn.MultiheadAttention):
|
176 |
+
def __init__(self, embed_dim, num_heads, dropout=0., bias=True, add_bias_kv=False,
|
177 |
+
add_zero_attn=False, kdim=None, vdim=None, batch_first=False, device=None, dtype=None):
|
178 |
+
super().__init__(embed_dim, num_heads, dropout, bias, add_bias_kv, add_zero_attn, kdim, vdim, batch_first, device, dtype)
|
179 |
+
|
180 |
+
# rewrite out_proj layer,with nn.Linear
|
181 |
+
self.out_proj = nn.Linear(embed_dim, embed_dim, bias=bias, device=device, dtype=dtype)
|
182 |
+
|
183 |
+
def forward(
|
184 |
+
self,
|
185 |
+
query: Tensor,
|
186 |
+
key: Tensor,
|
187 |
+
value: Tensor,
|
188 |
+
key_padding_mask: Optional[Tensor] = None,
|
189 |
+
need_weights: bool = True,
|
190 |
+
attn_mask: Optional[Tensor] = None,
|
191 |
+
average_attn_weights: bool = True,
|
192 |
+
is_causal : bool = False) -> Tuple[Tensor, Optional[Tensor]]:
|
193 |
+
why_not_fast_path = ''
|
194 |
+
if ((attn_mask is not None and torch.is_floating_point(attn_mask))
|
195 |
+
or (key_padding_mask is not None) and torch.is_floating_point(key_padding_mask)):
|
196 |
+
why_not_fast_path = "floating-point masks are not supported for fast path."
|
197 |
+
|
198 |
+
is_batched = query.dim() == 3
|
199 |
+
|
200 |
+
key_padding_mask = F._canonical_mask(
|
201 |
+
mask=key_padding_mask,
|
202 |
+
mask_name="key_padding_mask",
|
203 |
+
other_type=F._none_or_dtype(attn_mask),
|
204 |
+
other_name="attn_mask",
|
205 |
+
target_type=query.dtype
|
206 |
+
)
|
207 |
+
|
208 |
+
attn_mask = F._canonical_mask(
|
209 |
+
mask=attn_mask,
|
210 |
+
mask_name="attn_mask",
|
211 |
+
other_type=None,
|
212 |
+
other_name="",
|
213 |
+
target_type=query.dtype,
|
214 |
+
check_other=False,
|
215 |
+
)
|
216 |
+
|
217 |
+
|
218 |
+
if not is_batched:
|
219 |
+
why_not_fast_path = f"input not batched; expected query.dim() of 3 but got {query.dim()}"
|
220 |
+
elif query is not key or key is not value:
|
221 |
+
# When lifting this restriction, don't forget to either
|
222 |
+
# enforce that the dtypes all match or test cases where
|
223 |
+
# they don't!
|
224 |
+
why_not_fast_path = "non-self attention was used (query, key, and value are not the same Tensor)"
|
225 |
+
elif self.in_proj_bias is not None and query.dtype != self.in_proj_bias.dtype:
|
226 |
+
why_not_fast_path = f"dtypes of query ({query.dtype}) and self.in_proj_bias ({self.in_proj_bias.dtype}) don't match"
|
227 |
+
elif self.in_proj_weight is None:
|
228 |
+
why_not_fast_path = "in_proj_weight was None"
|
229 |
+
elif query.dtype != self.in_proj_weight.dtype:
|
230 |
+
# this case will fail anyway, but at least they'll get a useful error message.
|
231 |
+
why_not_fast_path = f"dtypes of query ({query.dtype}) and self.in_proj_weight ({self.in_proj_weight.dtype}) don't match"
|
232 |
+
elif self.training:
|
233 |
+
why_not_fast_path = "training is enabled"
|
234 |
+
elif (self.num_heads % 2) != 0:
|
235 |
+
why_not_fast_path = "self.num_heads is not even"
|
236 |
+
elif not self.batch_first:
|
237 |
+
why_not_fast_path = "batch_first was not True"
|
238 |
+
elif self.bias_k is not None:
|
239 |
+
why_not_fast_path = "self.bias_k was not None"
|
240 |
+
elif self.bias_v is not None:
|
241 |
+
why_not_fast_path = "self.bias_v was not None"
|
242 |
+
elif self.add_zero_attn:
|
243 |
+
why_not_fast_path = "add_zero_attn was enabled"
|
244 |
+
elif not self._qkv_same_embed_dim:
|
245 |
+
why_not_fast_path = "_qkv_same_embed_dim was not True"
|
246 |
+
elif query.is_nested and (key_padding_mask is not None or attn_mask is not None):
|
247 |
+
why_not_fast_path = "supplying both src_key_padding_mask and src_mask at the same time \
|
248 |
+
is not supported with NestedTensor input"
|
249 |
+
elif torch.is_autocast_enabled():
|
250 |
+
why_not_fast_path = "autocast is enabled"
|
251 |
+
|
252 |
+
if not why_not_fast_path:
|
253 |
+
tensor_args = (
|
254 |
+
query,
|
255 |
+
key,
|
256 |
+
value,
|
257 |
+
self.in_proj_weight,
|
258 |
+
self.in_proj_bias,
|
259 |
+
self.out_proj.weight,
|
260 |
+
self.out_proj.bias,
|
261 |
+
)
|
262 |
+
# We have to use list comprehensions below because TorchScript does not support
|
263 |
+
# generator expressions.
|
264 |
+
if torch.overrides.has_torch_function(tensor_args):
|
265 |
+
why_not_fast_path = "some Tensor argument has_torch_function"
|
266 |
+
elif _is_make_fx_tracing():
|
267 |
+
why_not_fast_path = "we are running make_fx tracing"
|
268 |
+
elif not all(_check_arg_device(x) for x in tensor_args):
|
269 |
+
why_not_fast_path = ("some Tensor argument's device is neither one of "
|
270 |
+
f"cpu, cuda or {torch.utils.backend_registration._privateuse1_backend_name}")
|
271 |
+
elif torch.is_grad_enabled() and any(_arg_requires_grad(x) for x in tensor_args):
|
272 |
+
why_not_fast_path = ("grad is enabled and at least one of query or the "
|
273 |
+
"input/output projection weights or biases requires_grad")
|
274 |
+
if not why_not_fast_path:
|
275 |
+
merged_mask, mask_type = self.merge_masks(attn_mask, key_padding_mask, query)
|
276 |
+
|
277 |
+
if self.in_proj_bias is not None and self.in_proj_weight is not None:
|
278 |
+
return torch._native_multi_head_attention(
|
279 |
+
query,
|
280 |
+
key,
|
281 |
+
value,
|
282 |
+
self.embed_dim,
|
283 |
+
self.num_heads,
|
284 |
+
self.in_proj_weight,
|
285 |
+
self.in_proj_bias,
|
286 |
+
self.out_proj.weight,
|
287 |
+
self.out_proj.bias,
|
288 |
+
merged_mask,
|
289 |
+
need_weights,
|
290 |
+
average_attn_weights,
|
291 |
+
mask_type)
|
292 |
+
|
293 |
+
any_nested = query.is_nested or key.is_nested or value.is_nested
|
294 |
+
assert not any_nested, ("MultiheadAttention does not support NestedTensor outside of its fast path. " +
|
295 |
+
f"The fast path was not hit because {why_not_fast_path}")
|
296 |
+
|
297 |
+
if self.batch_first and is_batched:
|
298 |
+
# make sure that the transpose op does not affect the "is" property
|
299 |
+
if key is value:
|
300 |
+
if query is key:
|
301 |
+
query = key = value = query.transpose(1, 0)
|
302 |
+
else:
|
303 |
+
query, key = (x.transpose(1, 0) for x in (query, key))
|
304 |
+
value = key
|
305 |
+
else:
|
306 |
+
query, key, value = (x.transpose(1, 0) for x in (query, key, value))
|
307 |
+
|
308 |
+
if not self._qkv_same_embed_dim:
|
309 |
+
attn_output, attn_output_weights = self.multi_head_attention_forward(
|
310 |
+
query, key, value, self.embed_dim, self.num_heads,
|
311 |
+
self.in_proj_weight, self.in_proj_bias,
|
312 |
+
self.bias_k, self.bias_v, self.add_zero_attn,
|
313 |
+
self.dropout, self.out_proj.weight, self.out_proj.bias,
|
314 |
+
training=self.training,
|
315 |
+
key_padding_mask=key_padding_mask, need_weights=need_weights,
|
316 |
+
attn_mask=attn_mask,
|
317 |
+
use_separate_proj_weight=True,
|
318 |
+
q_proj_weight=self.q_proj_weight, k_proj_weight=self.k_proj_weight,
|
319 |
+
v_proj_weight=self.v_proj_weight,
|
320 |
+
average_attn_weights=average_attn_weights,
|
321 |
+
is_causal=is_causal)
|
322 |
+
else:
|
323 |
+
attn_output, attn_output_weights = self.multi_head_attention_forward(
|
324 |
+
query, key, value, self.embed_dim, self.num_heads,
|
325 |
+
self.in_proj_weight, self.in_proj_bias,
|
326 |
+
self.bias_k, self.bias_v, self.add_zero_attn,
|
327 |
+
self.dropout, self.out_proj.weight, self.out_proj.bias,
|
328 |
+
training=self.training,
|
329 |
+
key_padding_mask=key_padding_mask,
|
330 |
+
need_weights=need_weights,
|
331 |
+
attn_mask=attn_mask,
|
332 |
+
average_attn_weights=average_attn_weights,
|
333 |
+
is_causal=is_causal)
|
334 |
+
if self.batch_first and is_batched:
|
335 |
+
return attn_output.transpose(1, 0), attn_output_weights
|
336 |
+
else:
|
337 |
+
return attn_output, attn_output_weights
|
338 |
+
|
339 |
+
def multi_head_attention_forward(
|
340 |
+
self,
|
341 |
+
query: Tensor,
|
342 |
+
key: Tensor,
|
343 |
+
value: Tensor,
|
344 |
+
embed_dim_to_check: int,
|
345 |
+
num_heads: int,
|
346 |
+
in_proj_weight: Optional[Tensor],
|
347 |
+
in_proj_bias: Optional[Tensor],
|
348 |
+
bias_k: Optional[Tensor],
|
349 |
+
bias_v: Optional[Tensor],
|
350 |
+
add_zero_attn: bool,
|
351 |
+
dropout_p: float,
|
352 |
+
out_proj_weight: Tensor,
|
353 |
+
out_proj_bias: Optional[Tensor],
|
354 |
+
training: bool = True,
|
355 |
+
key_padding_mask: Optional[Tensor] = None,
|
356 |
+
need_weights: bool = True,
|
357 |
+
attn_mask: Optional[Tensor] = None,
|
358 |
+
use_separate_proj_weight: bool = False,
|
359 |
+
q_proj_weight: Optional[Tensor] = None,
|
360 |
+
k_proj_weight: Optional[Tensor] = None,
|
361 |
+
v_proj_weight: Optional[Tensor] = None,
|
362 |
+
static_k: Optional[Tensor] = None,
|
363 |
+
static_v: Optional[Tensor] = None,
|
364 |
+
average_attn_weights: bool = True,
|
365 |
+
is_causal: bool = False,
|
366 |
+
) -> Tuple[Tensor, Optional[Tensor]]:
|
367 |
+
tens_ops = (query, key, value, in_proj_weight, in_proj_bias, bias_k, bias_v, out_proj_weight, out_proj_bias)
|
368 |
+
if has_torch_function(tens_ops):
|
369 |
+
return handle_torch_function(
|
370 |
+
multi_head_attention_forward,
|
371 |
+
tens_ops,
|
372 |
+
query,
|
373 |
+
key,
|
374 |
+
value,
|
375 |
+
embed_dim_to_check,
|
376 |
+
num_heads,
|
377 |
+
in_proj_weight,
|
378 |
+
in_proj_bias,
|
379 |
+
bias_k,
|
380 |
+
bias_v,
|
381 |
+
add_zero_attn,
|
382 |
+
dropout_p,
|
383 |
+
out_proj_weight,
|
384 |
+
out_proj_bias,
|
385 |
+
training=training,
|
386 |
+
key_padding_mask=key_padding_mask,
|
387 |
+
need_weights=need_weights,
|
388 |
+
attn_mask=attn_mask,
|
389 |
+
is_causal=is_causal,
|
390 |
+
use_separate_proj_weight=use_separate_proj_weight,
|
391 |
+
q_proj_weight=q_proj_weight,
|
392 |
+
k_proj_weight=k_proj_weight,
|
393 |
+
v_proj_weight=v_proj_weight,
|
394 |
+
static_k=static_k,
|
395 |
+
static_v=static_v,
|
396 |
+
average_attn_weights=average_attn_weights,
|
397 |
+
)
|
398 |
+
|
399 |
+
is_batched = _mha_shape_check(query, key, value, key_padding_mask, attn_mask, num_heads)
|
400 |
+
|
401 |
+
# For unbatched input, we unsqueeze at the expected batch-dim to pretend that the input
|
402 |
+
# is batched, run the computation and before returning squeeze the
|
403 |
+
# batch dimension so that the output doesn't carry this temporary batch dimension.
|
404 |
+
if not is_batched:
|
405 |
+
# unsqueeze if the input is unbatched
|
406 |
+
query = query.unsqueeze(1)
|
407 |
+
key = key.unsqueeze(1)
|
408 |
+
value = value.unsqueeze(1)
|
409 |
+
if key_padding_mask is not None:
|
410 |
+
key_padding_mask = key_padding_mask.unsqueeze(0)
|
411 |
+
|
412 |
+
# set up shape vars
|
413 |
+
tgt_len, bsz, embed_dim = query.shape
|
414 |
+
src_len, _, _ = key.shape
|
415 |
+
|
416 |
+
key_padding_mask = _canonical_mask(
|
417 |
+
mask=key_padding_mask,
|
418 |
+
mask_name="key_padding_mask",
|
419 |
+
other_type=_none_or_dtype(attn_mask),
|
420 |
+
other_name="attn_mask",
|
421 |
+
target_type=query.dtype
|
422 |
+
)
|
423 |
+
|
424 |
+
if is_causal and attn_mask is None:
|
425 |
+
raise RuntimeError(
|
426 |
+
"Need attn_mask if specifying the is_causal hint. "
|
427 |
+
"You may use the Transformer module method "
|
428 |
+
"`generate_square_subsequent_mask` to create this mask."
|
429 |
+
)
|
430 |
+
|
431 |
+
if is_causal and key_padding_mask is None and not need_weights:
|
432 |
+
# when we have a kpm or need weights, we need attn_mask
|
433 |
+
# Otherwise, we use the is_causal hint go as is_causal
|
434 |
+
# indicator to SDPA.
|
435 |
+
attn_mask = None
|
436 |
+
else:
|
437 |
+
attn_mask = _canonical_mask(
|
438 |
+
mask=attn_mask,
|
439 |
+
mask_name="attn_mask",
|
440 |
+
other_type=None,
|
441 |
+
other_name="",
|
442 |
+
target_type=query.dtype,
|
443 |
+
check_other=False,
|
444 |
+
)
|
445 |
+
|
446 |
+
if key_padding_mask is not None:
|
447 |
+
# We have the attn_mask, and use that to merge kpm into it.
|
448 |
+
# Turn off use of is_causal hint, as the merged mask is no
|
449 |
+
# longer causal.
|
450 |
+
is_causal = False
|
451 |
+
|
452 |
+
assert embed_dim == embed_dim_to_check, \
|
453 |
+
f"was expecting embedding dimension of {embed_dim_to_check}, but got {embed_dim}"
|
454 |
+
if isinstance(embed_dim, torch.Tensor):
|
455 |
+
# embed_dim can be a tensor when JIT tracing
|
456 |
+
head_dim = embed_dim.div(num_heads, rounding_mode='trunc')
|
457 |
+
else:
|
458 |
+
head_dim = embed_dim // num_heads
|
459 |
+
assert head_dim * num_heads == embed_dim, f"embed_dim {embed_dim} not divisible by num_heads {num_heads}"
|
460 |
+
if use_separate_proj_weight:
|
461 |
+
# allow MHA to have different embedding dimensions when separate projection weights are used
|
462 |
+
assert key.shape[:2] == value.shape[:2], \
|
463 |
+
f"key's sequence and batch dims {key.shape[:2]} do not match value's {value.shape[:2]}"
|
464 |
+
else:
|
465 |
+
assert key.shape == value.shape, f"key shape {key.shape} does not match value shape {value.shape}"
|
466 |
+
|
467 |
+
#
|
468 |
+
# compute in-projection
|
469 |
+
#
|
470 |
+
if not use_separate_proj_weight:
|
471 |
+
assert in_proj_weight is not None, "use_separate_proj_weight is False but in_proj_weight is None"
|
472 |
+
q, k, v = _in_projection_packed(query, key, value, in_proj_weight, in_proj_bias)
|
473 |
+
else:
|
474 |
+
assert q_proj_weight is not None, "use_separate_proj_weight is True but q_proj_weight is None"
|
475 |
+
assert k_proj_weight is not None, "use_separate_proj_weight is True but k_proj_weight is None"
|
476 |
+
assert v_proj_weight is not None, "use_separate_proj_weight is True but v_proj_weight is None"
|
477 |
+
if in_proj_bias is None:
|
478 |
+
b_q = b_k = b_v = None
|
479 |
+
else:
|
480 |
+
b_q, b_k, b_v = in_proj_bias.chunk(3)
|
481 |
+
q, k, v = _in_projection(query, key, value, q_proj_weight, k_proj_weight, v_proj_weight, b_q, b_k, b_v)
|
482 |
+
|
483 |
+
# prep attention mask
|
484 |
+
|
485 |
+
if attn_mask is not None:
|
486 |
+
# ensure attn_mask's dim is 3
|
487 |
+
if attn_mask.dim() == 2:
|
488 |
+
correct_2d_size = (tgt_len, src_len)
|
489 |
+
if attn_mask.shape != correct_2d_size:
|
490 |
+
raise RuntimeError(f"The shape of the 2D attn_mask is {attn_mask.shape}, but should be {correct_2d_size}.")
|
491 |
+
attn_mask = attn_mask.unsqueeze(0)
|
492 |
+
elif attn_mask.dim() == 3:
|
493 |
+
correct_3d_size = (bsz * num_heads, tgt_len, src_len)
|
494 |
+
if attn_mask.shape != correct_3d_size:
|
495 |
+
raise RuntimeError(f"The shape of the 3D attn_mask is {attn_mask.shape}, but should be {correct_3d_size}.")
|
496 |
+
else:
|
497 |
+
raise RuntimeError(f"attn_mask's dimension {attn_mask.dim()} is not supported")
|
498 |
+
|
499 |
+
# add bias along batch dimension (currently second)
|
500 |
+
if bias_k is not None and bias_v is not None:
|
501 |
+
assert static_k is None, "bias cannot be added to static key."
|
502 |
+
assert static_v is None, "bias cannot be added to static value."
|
503 |
+
k = torch.cat([k, bias_k.repeat(1, bsz, 1)])
|
504 |
+
v = torch.cat([v, bias_v.repeat(1, bsz, 1)])
|
505 |
+
if attn_mask is not None:
|
506 |
+
attn_mask = pad(attn_mask, (0, 1))
|
507 |
+
if key_padding_mask is not None:
|
508 |
+
key_padding_mask = pad(key_padding_mask, (0, 1))
|
509 |
+
else:
|
510 |
+
assert bias_k is None
|
511 |
+
assert bias_v is None
|
512 |
+
|
513 |
+
#
|
514 |
+
# reshape q, k, v for multihead attention and make em batch first
|
515 |
+
#
|
516 |
+
q = q.view(tgt_len, bsz * num_heads, head_dim).transpose(0, 1)
|
517 |
+
if static_k is None:
|
518 |
+
k = k.view(k.shape[0], bsz * num_heads, head_dim).transpose(0, 1)
|
519 |
+
else:
|
520 |
+
# TODO finish disentangling control flow so we don't do in-projections when statics are passed
|
521 |
+
assert static_k.size(0) == bsz * num_heads, \
|
522 |
+
f"expecting static_k.size(0) of {bsz * num_heads}, but got {static_k.size(0)}"
|
523 |
+
assert static_k.size(2) == head_dim, \
|
524 |
+
f"expecting static_k.size(2) of {head_dim}, but got {static_k.size(2)}"
|
525 |
+
k = static_k
|
526 |
+
if static_v is None:
|
527 |
+
v = v.view(v.shape[0], bsz * num_heads, head_dim).transpose(0, 1)
|
528 |
+
else:
|
529 |
+
# TODO finish disentangling control flow so we don't do in-projections when statics are passed
|
530 |
+
assert static_v.size(0) == bsz * num_heads, \
|
531 |
+
f"expecting static_v.size(0) of {bsz * num_heads}, but got {static_v.size(0)}"
|
532 |
+
assert static_v.size(2) == head_dim, \
|
533 |
+
f"expecting static_v.size(2) of {head_dim}, but got {static_v.size(2)}"
|
534 |
+
v = static_v
|
535 |
+
|
536 |
+
# add zero attention along batch dimension (now first)
|
537 |
+
if add_zero_attn:
|
538 |
+
zero_attn_shape = (bsz * num_heads, 1, head_dim)
|
539 |
+
k = torch.cat([k, torch.zeros(zero_attn_shape, dtype=k.dtype, device=k.device)], dim=1)
|
540 |
+
v = torch.cat([v, torch.zeros(zero_attn_shape, dtype=v.dtype, device=v.device)], dim=1)
|
541 |
+
if attn_mask is not None:
|
542 |
+
attn_mask = pad(attn_mask, (0, 1))
|
543 |
+
if key_padding_mask is not None:
|
544 |
+
key_padding_mask = pad(key_padding_mask, (0, 1))
|
545 |
+
|
546 |
+
# update source sequence length after adjustments
|
547 |
+
src_len = k.size(1)
|
548 |
+
|
549 |
+
# merge key padding and attention masks
|
550 |
+
if key_padding_mask is not None:
|
551 |
+
assert key_padding_mask.shape == (bsz, src_len), \
|
552 |
+
f"expecting key_padding_mask shape of {(bsz, src_len)}, but got {key_padding_mask.shape}"
|
553 |
+
key_padding_mask = key_padding_mask.view(bsz, 1, 1, src_len). \
|
554 |
+
expand(-1, num_heads, -1, -1).reshape(bsz * num_heads, 1, src_len)
|
555 |
+
if attn_mask is None:
|
556 |
+
attn_mask = key_padding_mask
|
557 |
+
else:
|
558 |
+
attn_mask = attn_mask + key_padding_mask
|
559 |
+
|
560 |
+
# adjust dropout probability
|
561 |
+
if not training:
|
562 |
+
dropout_p = 0.0
|
563 |
+
|
564 |
+
#
|
565 |
+
# (deep breath) calculate attention and out projection
|
566 |
+
#
|
567 |
+
|
568 |
+
if need_weights:
|
569 |
+
B, Nt, E = q.shape
|
570 |
+
q_scaled = q / math.sqrt(E)
|
571 |
+
|
572 |
+
assert not (is_causal and attn_mask is None), "FIXME: is_causal not implemented for need_weights"
|
573 |
+
|
574 |
+
if attn_mask is not None:
|
575 |
+
attn_output_weights = torch.baddbmm(attn_mask, q_scaled, k.transpose(-2, -1))
|
576 |
+
else:
|
577 |
+
attn_output_weights = torch.bmm(q_scaled, k.transpose(-2, -1))
|
578 |
+
attn_output_weights = softmax(attn_output_weights, dim=-1)
|
579 |
+
if dropout_p > 0.0:
|
580 |
+
attn_output_weights = dropout(attn_output_weights, p=dropout_p)
|
581 |
+
|
582 |
+
attn_output = torch.bmm(attn_output_weights, v)
|
583 |
+
|
584 |
+
attn_output = attn_output.transpose(0, 1).contiguous().view(tgt_len * bsz, embed_dim)
|
585 |
+
attn_output = self.out_proj(attn_output)
|
586 |
+
attn_output = attn_output.view(tgt_len, bsz, attn_output.size(1))
|
587 |
+
|
588 |
+
# optionally average attention weights over heads
|
589 |
+
attn_output_weights = attn_output_weights.view(bsz, num_heads, tgt_len, src_len)
|
590 |
+
if average_attn_weights:
|
591 |
+
attn_output_weights = attn_output_weights.mean(dim=1)
|
592 |
+
|
593 |
+
if not is_batched:
|
594 |
+
# squeeze the output if input was unbatched
|
595 |
+
attn_output = attn_output.squeeze(1)
|
596 |
+
attn_output_weights = attn_output_weights.squeeze(0)
|
597 |
+
return attn_output, attn_output_weights
|
598 |
+
else:
|
599 |
+
# attn_mask can be either (L,S) or (N*num_heads, L, S)
|
600 |
+
# if attn_mask's shape is (1, L, S) we need to unsqueeze to (1, 1, L, S)
|
601 |
+
# in order to match the input for SDPA of (N, num_heads, L, S)
|
602 |
+
if attn_mask is not None:
|
603 |
+
if attn_mask.size(0) == 1 and attn_mask.dim() == 3:
|
604 |
+
attn_mask = attn_mask.unsqueeze(0)
|
605 |
+
else:
|
606 |
+
attn_mask = attn_mask.view(bsz, num_heads, -1, src_len)
|
607 |
+
|
608 |
+
q = q.view(bsz, num_heads, tgt_len, head_dim)
|
609 |
+
k = k.view(bsz, num_heads, src_len, head_dim)
|
610 |
+
v = v.view(bsz, num_heads, src_len, head_dim)
|
611 |
+
|
612 |
+
attn_output = F.scaled_dot_product_attention(q, k, v, attn_mask, dropout_p, is_causal)
|
613 |
+
attn_output = attn_output.permute(2, 0, 1, 3).contiguous().view(bsz * tgt_len, embed_dim)
|
614 |
+
|
615 |
+
attn_output = self.out_proj(attn_output)
|
616 |
+
attn_output = attn_output.view(tgt_len, bsz, attn_output.size(1))
|
617 |
+
if not is_batched:
|
618 |
+
# squeeze the output if input was unbatched
|
619 |
+
attn_output = attn_output.squeeze(1)
|
620 |
+
return attn_output, None
|
621 |
+
|
622 |
+
|
623 |
+
def _mha_shape_check(query: Tensor, key: Tensor, value: Tensor,
|
624 |
+
key_padding_mask: Optional[Tensor], attn_mask: Optional[Tensor], num_heads: int):
|
625 |
+
# Verifies the expected shape for `query, `key`, `value`, `key_padding_mask` and `attn_mask`
|
626 |
+
# and returns if the input is batched or not.
|
627 |
+
# Raises an error if `query` is not 2-D (unbatched) or 3-D (batched) tensor.
|
628 |
+
|
629 |
+
# Shape check.
|
630 |
+
if query.dim() == 3:
|
631 |
+
# Batched Inputs
|
632 |
+
is_batched = True
|
633 |
+
assert key.dim() == 3 and value.dim() == 3, \
|
634 |
+
("For batched (3-D) `query`, expected `key` and `value` to be 3-D"
|
635 |
+
f" but found {key.dim()}-D and {value.dim()}-D tensors respectively")
|
636 |
+
if key_padding_mask is not None:
|
637 |
+
assert key_padding_mask.dim() == 2, \
|
638 |
+
("For batched (3-D) `query`, expected `key_padding_mask` to be `None` or 2-D"
|
639 |
+
f" but found {key_padding_mask.dim()}-D tensor instead")
|
640 |
+
if attn_mask is not None:
|
641 |
+
assert attn_mask.dim() in (2, 3), \
|
642 |
+
("For batched (3-D) `query`, expected `attn_mask` to be `None`, 2-D or 3-D"
|
643 |
+
f" but found {attn_mask.dim()}-D tensor instead")
|
644 |
+
elif query.dim() == 2:
|
645 |
+
# Unbatched Inputs
|
646 |
+
is_batched = False
|
647 |
+
assert key.dim() == 2 and value.dim() == 2, \
|
648 |
+
("For unbatched (2-D) `query`, expected `key` and `value` to be 2-D"
|
649 |
+
f" but found {key.dim()}-D and {value.dim()}-D tensors respectively")
|
650 |
+
|
651 |
+
if key_padding_mask is not None:
|
652 |
+
assert key_padding_mask.dim() == 1, \
|
653 |
+
("For unbatched (2-D) `query`, expected `key_padding_mask` to be `None` or 1-D"
|
654 |
+
f" but found {key_padding_mask.dim()}-D tensor instead")
|
655 |
+
|
656 |
+
if attn_mask is not None:
|
657 |
+
assert attn_mask.dim() in (2, 3), \
|
658 |
+
("For unbatched (2-D) `query`, expected `attn_mask` to be `None`, 2-D or 3-D"
|
659 |
+
f" but found {attn_mask.dim()}-D tensor instead")
|
660 |
+
if attn_mask.dim() == 3:
|
661 |
+
expected_shape = (num_heads, query.shape[0], key.shape[0])
|
662 |
+
assert attn_mask.shape == expected_shape, \
|
663 |
+
(f"Expected `attn_mask` shape to be {expected_shape} but got {attn_mask.shape}")
|
664 |
+
else:
|
665 |
+
raise AssertionError(
|
666 |
+
f"query should be unbatched 2D or batched 3D tensor but received {query.dim()}-D query tensor")
|
667 |
+
|
668 |
+
return is_batched
|
669 |
+
|
670 |
+
|
671 |
+
def _canonical_mask(
|
672 |
+
mask: Optional[Tensor],
|
673 |
+
mask_name: str,
|
674 |
+
other_type: Optional[DType],
|
675 |
+
other_name: str,
|
676 |
+
target_type: DType,
|
677 |
+
check_other: bool = True,
|
678 |
+
) -> Optional[Tensor]:
|
679 |
+
|
680 |
+
if mask is not None:
|
681 |
+
_mask_dtype = mask.dtype
|
682 |
+
_mask_is_float = torch.is_floating_point(mask)
|
683 |
+
if _mask_dtype != torch.bool and not _mask_is_float:
|
684 |
+
raise AssertionError(
|
685 |
+
f"only bool and floating types of {mask_name} are supported")
|
686 |
+
if check_other and other_type is not None:
|
687 |
+
if _mask_dtype != other_type:
|
688 |
+
warnings.warn(
|
689 |
+
f"Support for mismatched {mask_name} and {other_name} "
|
690 |
+
"is deprecated. Use same type for both instead."
|
691 |
+
)
|
692 |
+
if not _mask_is_float:
|
693 |
+
mask = (
|
694 |
+
torch.zeros_like(mask, dtype=target_type)
|
695 |
+
.masked_fill_(mask, float("-inf"))
|
696 |
+
)
|
697 |
+
return mask
|
698 |
+
|
699 |
+
|
700 |
+
def _none_or_dtype(input: Optional[Tensor]) -> Optional[DType]:
|
701 |
+
if input is None:
|
702 |
+
return None
|
703 |
+
elif isinstance(input, torch.Tensor):
|
704 |
+
return input.dtype
|
705 |
+
raise RuntimeError("input to _none_or_dtype() must be None or torch.Tensor")
|
706 |
+
|
707 |
+
def _in_projection_packed(
|
708 |
+
q: Tensor,
|
709 |
+
k: Tensor,
|
710 |
+
v: Tensor,
|
711 |
+
w: Tensor,
|
712 |
+
b: Optional[Tensor] = None,
|
713 |
+
) -> List[Tensor]:
|
714 |
+
r"""
|
715 |
+
Performs the in-projection step of the attention operation, using packed weights.
|
716 |
+
Output is a triple containing projection tensors for query, key and value.
|
717 |
+
Args:
|
718 |
+
q, k, v: query, key and value tensors to be projected. For self-attention,
|
719 |
+
these are typically the same tensor; for encoder-decoder attention,
|
720 |
+
k and v are typically the same tensor. (We take advantage of these
|
721 |
+
identities for performance if they are present.) Regardless, q, k and v
|
722 |
+
must share a common embedding dimension; otherwise their shapes may vary.
|
723 |
+
w: projection weights for q, k and v, packed into a single tensor. Weights
|
724 |
+
are packed along dimension 0, in q, k, v order.
|
725 |
+
b: optional projection biases for q, k and v, packed into a single tensor
|
726 |
+
in q, k, v order.
|
727 |
+
Shape:
|
728 |
+
Inputs:
|
729 |
+
- q: :math:`(..., E)` where E is the embedding dimension
|
730 |
+
- k: :math:`(..., E)` where E is the embedding dimension
|
731 |
+
- v: :math:`(..., E)` where E is the embedding dimension
|
732 |
+
- w: :math:`(E * 3, E)` where E is the embedding dimension
|
733 |
+
- b: :math:`E * 3` where E is the embedding dimension
|
734 |
+
Output:
|
735 |
+
- in output list :math:`[q', k', v']`, each output tensor will have the
|
736 |
+
same shape as the corresponding input tensor.
|
737 |
+
"""
|
738 |
+
E = q.size(-1)
|
739 |
+
if k is v:
|
740 |
+
if q is k:
|
741 |
+
# self-attention
|
742 |
+
proj = linear(q, w, b)
|
743 |
+
# reshape to 3, E and not E, 3 is deliberate for better memory coalescing and keeping same order as chunk()
|
744 |
+
proj = proj.unflatten(-1, (3, E)).unsqueeze(0).transpose(0, -2).squeeze(-2).contiguous()
|
745 |
+
return proj[0], proj[1], proj[2]
|
746 |
+
else:
|
747 |
+
# encoder-decoder attention
|
748 |
+
w_q, w_kv = w.split([E, E * 2])
|
749 |
+
if b is None:
|
750 |
+
b_q = b_kv = None
|
751 |
+
else:
|
752 |
+
b_q, b_kv = b.split([E, E * 2])
|
753 |
+
q_proj = linear(q, w_q, b_q)
|
754 |
+
kv_proj = linear(k, w_kv, b_kv)
|
755 |
+
# reshape to 2, E and not E, 2 is deliberate for better memory coalescing and keeping same order as chunk()
|
756 |
+
kv_proj = kv_proj.unflatten(-1, (2, E)).unsqueeze(0).transpose(0, -2).squeeze(-2).contiguous()
|
757 |
+
return (q_proj, kv_proj[0], kv_proj[1])
|
758 |
+
else:
|
759 |
+
w_q, w_k, w_v = w.chunk(3)
|
760 |
+
if b is None:
|
761 |
+
b_q = b_k = b_v = None
|
762 |
+
else:
|
763 |
+
b_q, b_k, b_v = b.chunk(3)
|
764 |
+
return linear(q, w_q, b_q), linear(k, w_k, b_k), linear(v, w_v, b_v)
|
765 |
+
|
766 |
+
|
767 |
+
def _in_projection(
|
768 |
+
q: Tensor,
|
769 |
+
k: Tensor,
|
770 |
+
v: Tensor,
|
771 |
+
w_q: Tensor,
|
772 |
+
w_k: Tensor,
|
773 |
+
w_v: Tensor,
|
774 |
+
b_q: Optional[Tensor] = None,
|
775 |
+
b_k: Optional[Tensor] = None,
|
776 |
+
b_v: Optional[Tensor] = None,
|
777 |
+
) -> Tuple[Tensor, Tensor, Tensor]:
|
778 |
+
r"""
|
779 |
+
Performs the in-projection step of the attention operation. This is simply
|
780 |
+
a triple of linear projections, with shape constraints on the weights which
|
781 |
+
ensure embedding dimension uniformity in the projected outputs.
|
782 |
+
Output is a triple containing projection tensors for query, key and value.
|
783 |
+
Args:
|
784 |
+
q, k, v: query, key and value tensors to be projected.
|
785 |
+
w_q, w_k, w_v: weights for q, k and v, respectively.
|
786 |
+
b_q, b_k, b_v: optional biases for q, k and v, respectively.
|
787 |
+
Shape:
|
788 |
+
Inputs:
|
789 |
+
- q: :math:`(Qdims..., Eq)` where Eq is the query embedding dimension and Qdims are any
|
790 |
+
number of leading dimensions.
|
791 |
+
- k: :math:`(Kdims..., Ek)` where Ek is the key embedding dimension and Kdims are any
|
792 |
+
number of leading dimensions.
|
793 |
+
- v: :math:`(Vdims..., Ev)` where Ev is the value embedding dimension and Vdims are any
|
794 |
+
number of leading dimensions.
|
795 |
+
- w_q: :math:`(Eq, Eq)`
|
796 |
+
- w_k: :math:`(Eq, Ek)`
|
797 |
+
- w_v: :math:`(Eq, Ev)`
|
798 |
+
- b_q: :math:`(Eq)`
|
799 |
+
- b_k: :math:`(Eq)`
|
800 |
+
- b_v: :math:`(Eq)`
|
801 |
+
Output: in output triple :math:`(q', k', v')`,
|
802 |
+
- q': :math:`[Qdims..., Eq]`
|
803 |
+
- k': :math:`[Kdims..., Eq]`
|
804 |
+
- v': :math:`[Vdims..., Eq]`
|
805 |
+
"""
|
806 |
+
Eq, Ek, Ev = q.size(-1), k.size(-1), v.size(-1)
|
807 |
+
assert w_q.shape == (Eq, Eq), f"expecting query weights shape of {(Eq, Eq)}, but got {w_q.shape}"
|
808 |
+
assert w_k.shape == (Eq, Ek), f"expecting key weights shape of {(Eq, Ek)}, but got {w_k.shape}"
|
809 |
+
assert w_v.shape == (Eq, Ev), f"expecting value weights shape of {(Eq, Ev)}, but got {w_v.shape}"
|
810 |
+
assert b_q is None or b_q.shape == (Eq,), f"expecting query bias shape of {(Eq,)}, but got {b_q.shape}"
|
811 |
+
assert b_k is None or b_k.shape == (Eq,), f"expecting key bias shape of {(Eq,)}, but got {b_k.shape}"
|
812 |
+
assert b_v is None or b_v.shape == (Eq,), f"expecting value bias shape of {(Eq,)}, but got {b_v.shape}"
|
813 |
+
return linear(q, w_q, b_q), linear(k, w_k, b_k), linear(v, w_v, b_v)
|