Spaces:
Paused
Paused
x-lai
commited on
Commit
·
968fffb
1
Parent(s):
0146331
support 4bit and 8bit inference
Browse filesFormer-commit-id: 23930126323a0effb75929a5cc88c75c0d7bfbc2
- README.md +6 -1
- chat.py +21 -4
- model/LISA.py +71 -50
- model/llava/model/llava.py +2 -0
- model/segment_anything/modeling/image_encoder.py +7 -2
README.md
CHANGED
@@ -53,10 +53,15 @@ To chat with [LISA-13B-llama2-v0](https://huggingface.co/xinlai/LISA-13B-llama2-
|
|
53 |
```
|
54 |
CUDA_VISIBLE_DEVICES=0 python3 chat.py --version='xinlai/LISA-13B-llama2-v0'
|
55 |
```
|
56 |
-
To use `
|
57 |
```
|
58 |
CUDA_VISIBLE_DEVICES=0 python3 chat.py --version='xinlai/LISA-13B-llama2-v0' --precision='bf16'
|
59 |
```
|
|
|
|
|
|
|
|
|
|
|
60 |
|
61 |
After that, input the text prompt and then the image path. For example,
|
62 |
```
|
|
|
53 |
```
|
54 |
CUDA_VISIBLE_DEVICES=0 python3 chat.py --version='xinlai/LISA-13B-llama2-v0'
|
55 |
```
|
56 |
+
To use `bf16` or `fp16` data type for inference:
|
57 |
```
|
58 |
CUDA_VISIBLE_DEVICES=0 python3 chat.py --version='xinlai/LISA-13B-llama2-v0' --precision='bf16'
|
59 |
```
|
60 |
+
To use `8bit` or `4bit` data type for inference:
|
61 |
+
```
|
62 |
+
CUDA_VISIBLE_DEVICES=0 python3 chat.py --version='xinlai/LISA-13B-llama2-v0' --precision='fp16' --load_in_8bit
|
63 |
+
CUDA_VISIBLE_DEVICES=0 python3 chat.py --version='xinlai/LISA-13B-llama2-v0' --precision='fp16' --load_in_4bit
|
64 |
+
```
|
65 |
|
66 |
After that, input the text prompt and then the image path. For example,
|
67 |
```
|
chat.py
CHANGED
@@ -17,19 +17,22 @@ def parse_args(args):
|
|
17 |
parser = argparse.ArgumentParser(description='LISA chat')
|
18 |
parser.add_argument('--version', default='xinlai/LISA-13B-llama2-v0')
|
19 |
parser.add_argument('--vis_save_path', default='./vis_output', type=str)
|
20 |
-
parser.add_argument('--precision', default='bf16', type=str, choices=['fp32', 'bf16'], help="precision for inference")
|
21 |
parser.add_argument('--image-size', default=1024, type=int, help='image size')
|
22 |
parser.add_argument('--model-max-length', default=512, type=int)
|
23 |
parser.add_argument('--lora-r', default=-1, type=int)
|
24 |
parser.add_argument('--vision-tower', default='openai/clip-vit-large-patch14', type=str)
|
25 |
parser.add_argument('--local-rank', default=0, type=int, help='node rank')
|
|
|
|
|
26 |
return parser.parse_args(args)
|
27 |
|
28 |
|
29 |
def preprocess(x,
|
30 |
-
|
31 |
-
|
32 |
-
|
|
|
33 |
"""Normalize pixel values and pad to a square input."""
|
34 |
# Normalize colors
|
35 |
x = (x - pixel_mean) / pixel_std
|
@@ -65,6 +68,8 @@ def main(args):
|
|
65 |
args.version,
|
66 |
args.lora_r,
|
67 |
args.precision,
|
|
|
|
|
68 |
)
|
69 |
|
70 |
weight = {}
|
@@ -76,6 +81,14 @@ def main(args):
|
|
76 |
|
77 |
if args.precision == 'bf16':
|
78 |
model = model.bfloat16().cuda()
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
79 |
else:
|
80 |
model = model.float().cuda()
|
81 |
|
@@ -113,12 +126,16 @@ def main(args):
|
|
113 |
original_size_list = [image.shape[:2]]
|
114 |
if args.precision == 'bf16':
|
115 |
images_clip = clip_image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0].unsqueeze(0).cuda().bfloat16()
|
|
|
|
|
116 |
else:
|
117 |
images_clip = clip_image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0].unsqueeze(0).cuda().float()
|
118 |
images = transform.apply_image(image)
|
119 |
resize_list = [images.shape[:2]]
|
120 |
if args.precision == 'bf16':
|
121 |
images = preprocess(torch.from_numpy(images).permute(2,0,1).contiguous()).unsqueeze(0).cuda().bfloat16()
|
|
|
|
|
122 |
else:
|
123 |
images = preprocess(torch.from_numpy(images).permute(2,0,1).contiguous()).unsqueeze(0).cuda().float()
|
124 |
|
|
|
17 |
parser = argparse.ArgumentParser(description='LISA chat')
|
18 |
parser.add_argument('--version', default='xinlai/LISA-13B-llama2-v0')
|
19 |
parser.add_argument('--vis_save_path', default='./vis_output', type=str)
|
20 |
+
parser.add_argument('--precision', default='bf16', type=str, choices=['fp32', 'bf16', 'fp16'], help="precision for inference")
|
21 |
parser.add_argument('--image-size', default=1024, type=int, help='image size')
|
22 |
parser.add_argument('--model-max-length', default=512, type=int)
|
23 |
parser.add_argument('--lora-r', default=-1, type=int)
|
24 |
parser.add_argument('--vision-tower', default='openai/clip-vit-large-patch14', type=str)
|
25 |
parser.add_argument('--local-rank', default=0, type=int, help='node rank')
|
26 |
+
parser.add_argument('--load_in_8bit', action='store_true', default=False)
|
27 |
+
parser.add_argument('--load_in_4bit', action='store_true', default=False)
|
28 |
return parser.parse_args(args)
|
29 |
|
30 |
|
31 |
def preprocess(x,
|
32 |
+
pixel_mean=torch.Tensor([123.675, 116.28, 103.53]).view(-1, 1, 1),
|
33 |
+
pixel_std=torch.Tensor([58.395, 57.12, 57.375]).view(-1, 1, 1),
|
34 |
+
img_size=1024
|
35 |
+
) -> torch.Tensor:
|
36 |
"""Normalize pixel values and pad to a square input."""
|
37 |
# Normalize colors
|
38 |
x = (x - pixel_mean) / pixel_std
|
|
|
68 |
args.version,
|
69 |
args.lora_r,
|
70 |
args.precision,
|
71 |
+
load_in_8bit=args.load_in_8bit,
|
72 |
+
load_in_4bit=args.load_in_4bit,
|
73 |
)
|
74 |
|
75 |
weight = {}
|
|
|
81 |
|
82 |
if args.precision == 'bf16':
|
83 |
model = model.bfloat16().cuda()
|
84 |
+
elif args.precision == 'fp16':
|
85 |
+
import deepspeed
|
86 |
+
model_engine = deepspeed.init_inference(model=model,
|
87 |
+
dtype=torch.half,
|
88 |
+
replace_with_kernel_inject=True,
|
89 |
+
replace_method="auto",
|
90 |
+
)
|
91 |
+
model = model_engine.module
|
92 |
else:
|
93 |
model = model.float().cuda()
|
94 |
|
|
|
126 |
original_size_list = [image.shape[:2]]
|
127 |
if args.precision == 'bf16':
|
128 |
images_clip = clip_image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0].unsqueeze(0).cuda().bfloat16()
|
129 |
+
elif args.precision == 'fp16':
|
130 |
+
images_clip = clip_image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0].unsqueeze(0).cuda().half()
|
131 |
else:
|
132 |
images_clip = clip_image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0].unsqueeze(0).cuda().float()
|
133 |
images = transform.apply_image(image)
|
134 |
resize_list = [images.shape[:2]]
|
135 |
if args.precision == 'bf16':
|
136 |
images = preprocess(torch.from_numpy(images).permute(2,0,1).contiguous()).unsqueeze(0).cuda().bfloat16()
|
137 |
+
elif args.precision == 'fp16':
|
138 |
+
images = preprocess(torch.from_numpy(images).permute(2,0,1).contiguous()).unsqueeze(0).cuda().half()
|
139 |
else:
|
140 |
images = preprocess(torch.from_numpy(images).permute(2,0,1).contiguous()).unsqueeze(0).cuda().float()
|
141 |
|
model/LISA.py
CHANGED
@@ -9,7 +9,7 @@ import torch.nn as nn
|
|
9 |
import torch.nn.functional as F
|
10 |
import transformers
|
11 |
|
12 |
-
from transformers import LlamaForCausalLM, CLIPVisionModel
|
13 |
from peft import (
|
14 |
LoraConfig,
|
15 |
get_peft_model,
|
@@ -49,6 +49,8 @@ class LISA(nn.Module):
|
|
49 |
llm_version,
|
50 |
lora_r,
|
51 |
precision,
|
|
|
|
|
52 |
lora_target_modules=['q_proj', 'v_proj'],
|
53 |
lora_alpha=16,
|
54 |
lora_dropout=0.05,
|
@@ -69,6 +71,20 @@ class LISA(nn.Module):
|
|
69 |
num_new_tokens = tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
|
70 |
if precision == "bf16":
|
71 |
self.lm = LlavaLlamaForCausalLM.from_pretrained(llm_version, torch_dtype=torch.bfloat16, cache_dir=None, low_cpu_mem_usage=True)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
72 |
else:
|
73 |
self.lm = LlavaLlamaForCausalLM.from_pretrained(llm_version, torch_dtype=torch.float32, cache_dir=None, low_cpu_mem_usage=True)
|
74 |
|
@@ -85,6 +101,8 @@ class LISA(nn.Module):
|
|
85 |
if vision_tower.device.type == 'meta':
|
86 |
if precision == 'bf16':
|
87 |
vision_tower = CLIPVisionModel.from_pretrained(vision_tower.config._name_or_path, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True).cuda(local_rank)
|
|
|
|
|
88 |
else:
|
89 |
vision_tower = CLIPVisionModel.from_pretrained(vision_tower.config._name_or_path, torch_dtype=torch.float32, low_cpu_mem_usage=True).cuda(local_rank)
|
90 |
self.lm.get_model().vision_tower[0] = vision_tower
|
@@ -92,6 +110,8 @@ class LISA(nn.Module):
|
|
92 |
|
93 |
if precision == "bf16":
|
94 |
vision_tower.to(device='cuda', dtype=torch.bfloat16)
|
|
|
|
|
95 |
else:
|
96 |
vision_tower.to(device='cuda', dtype=torch.float32)
|
97 |
|
@@ -135,58 +155,59 @@ class LISA(nn.Module):
|
|
135 |
|
136 |
def evaluate(self, images_clip, images, input_ids, resize_list, original_size_list, max_new_tokens=32, tokenizer=None):
|
137 |
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
-
|
142 |
-
seg_token_mask = (output_ids[:, 1:] == self.seg_token_idx)
|
143 |
|
144 |
-
|
145 |
-
last_output_logit = None
|
146 |
-
hidden_states = []
|
147 |
|
148 |
-
|
149 |
-
|
|
|
150 |
|
151 |
-
|
152 |
-
|
153 |
-
|
154 |
-
|
155 |
-
|
156 |
-
seg_token_offset = torch.cat([torch.zeros(1).long().cuda(), seg_token_offset], dim=0)
|
157 |
-
|
158 |
-
pred_embeddings_ = []
|
159 |
-
for i in range(len(seg_token_offset)-1):
|
160 |
-
start_i, end_i = seg_token_offset[i], seg_token_offset[i+1]
|
161 |
-
pred_embeddings_.append(pred_embeddings[start_i: end_i])
|
162 |
-
pred_embeddings = pred_embeddings_
|
163 |
-
|
164 |
-
image_embeddings = self.get_visual_embs(images)
|
165 |
-
|
166 |
-
multimask_output = False
|
167 |
-
pred_masks = []
|
168 |
-
for i in range(len(pred_embeddings)):
|
169 |
-
sparse_embeddings, dense_embeddings = self.visual_model.prompt_encoder(
|
170 |
-
points=None,
|
171 |
-
boxes=None,
|
172 |
-
masks=None,
|
173 |
-
text_embeds=pred_embeddings[i].unsqueeze(1),
|
174 |
-
)
|
175 |
-
|
176 |
-
sparse_embeddings = sparse_embeddings.to(pred_embeddings[i].dtype)
|
177 |
-
low_res_masks, iou_predictions = self.visual_model.mask_decoder(
|
178 |
-
image_embeddings=image_embeddings[i].unsqueeze(0),
|
179 |
-
image_pe=self.visual_model.prompt_encoder.get_dense_pe(),
|
180 |
-
sparse_prompt_embeddings=sparse_embeddings,
|
181 |
-
dense_prompt_embeddings=dense_embeddings,
|
182 |
-
multimask_output=multimask_output,
|
183 |
-
)
|
184 |
|
185 |
-
|
186 |
-
|
187 |
-
|
188 |
-
|
189 |
-
|
190 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
191 |
|
192 |
return output_ids, pred_masks
|
|
|
9 |
import torch.nn.functional as F
|
10 |
import transformers
|
11 |
|
12 |
+
from transformers import LlamaForCausalLM, CLIPVisionModel, BitsAndBytesConfig
|
13 |
from peft import (
|
14 |
LoraConfig,
|
15 |
get_peft_model,
|
|
|
49 |
llm_version,
|
50 |
lora_r,
|
51 |
precision,
|
52 |
+
load_in_4bit=False,
|
53 |
+
load_in_8bit=False,
|
54 |
lora_target_modules=['q_proj', 'v_proj'],
|
55 |
lora_alpha=16,
|
56 |
lora_dropout=0.05,
|
|
|
71 |
num_new_tokens = tokenizer.add_tokens([DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN], special_tokens=True)
|
72 |
if precision == "bf16":
|
73 |
self.lm = LlavaLlamaForCausalLM.from_pretrained(llm_version, torch_dtype=torch.bfloat16, cache_dir=None, low_cpu_mem_usage=True)
|
74 |
+
elif precision == "fp16":
|
75 |
+
if load_in_4bit:
|
76 |
+
self.lm = LlavaLlamaForCausalLM.from_pretrained(llm_version, load_in_4bit=True, cache_dir=None, low_cpu_mem_usage=True, device_map='auto',
|
77 |
+
quantization_config=BitsAndBytesConfig(
|
78 |
+
load_in_4bit=True,
|
79 |
+
bnb_4bit_compute_dtype=torch.float16,
|
80 |
+
bnb_4bit_use_double_quant=True,
|
81 |
+
bnb_4bit_quant_type='nf4'
|
82 |
+
)
|
83 |
+
)
|
84 |
+
elif load_in_8bit:
|
85 |
+
self.lm = LlavaLlamaForCausalLM.from_pretrained(llm_version, load_in_8bit=True, cache_dir=None, low_cpu_mem_usage=True, device_map='auto')
|
86 |
+
else:
|
87 |
+
self.lm = LlavaLlamaForCausalLM.from_pretrained(llm_version, torch_dtype=torch.half, cache_dir=None, low_cpu_mem_usage=True)
|
88 |
else:
|
89 |
self.lm = LlavaLlamaForCausalLM.from_pretrained(llm_version, torch_dtype=torch.float32, cache_dir=None, low_cpu_mem_usage=True)
|
90 |
|
|
|
101 |
if vision_tower.device.type == 'meta':
|
102 |
if precision == 'bf16':
|
103 |
vision_tower = CLIPVisionModel.from_pretrained(vision_tower.config._name_or_path, torch_dtype=torch.bfloat16, low_cpu_mem_usage=True).cuda(local_rank)
|
104 |
+
elif precision == 'fp16':
|
105 |
+
vision_tower = CLIPVisionModel.from_pretrained(vision_tower.config._name_or_path, torch_dtype=torch.half, low_cpu_mem_usage=True).cuda(local_rank)
|
106 |
else:
|
107 |
vision_tower = CLIPVisionModel.from_pretrained(vision_tower.config._name_or_path, torch_dtype=torch.float32, low_cpu_mem_usage=True).cuda(local_rank)
|
108 |
self.lm.get_model().vision_tower[0] = vision_tower
|
|
|
110 |
|
111 |
if precision == "bf16":
|
112 |
vision_tower.to(device='cuda', dtype=torch.bfloat16)
|
113 |
+
elif precision == "fp16":
|
114 |
+
vision_tower.to(device='cuda', dtype=torch.half)
|
115 |
else:
|
116 |
vision_tower.to(device='cuda', dtype=torch.float32)
|
117 |
|
|
|
155 |
|
156 |
def evaluate(self, images_clip, images, input_ids, resize_list, original_size_list, max_new_tokens=32, tokenizer=None):
|
157 |
|
158 |
+
with torch.no_grad():
|
159 |
+
outputs = self.lm.generate(images=images_clip, input_ids=input_ids, max_new_tokens=max_new_tokens, num_beams=1, output_hidden_states=True, return_dict_in_generate=True)
|
160 |
+
output_hidden_states = outputs.hidden_states[-1]
|
161 |
+
output_ids = outputs.sequences
|
|
|
162 |
|
163 |
+
seg_token_mask = (output_ids[:, 1:] == self.seg_token_idx)
|
|
|
|
|
164 |
|
165 |
+
last_embedding = None
|
166 |
+
last_output_logit = None
|
167 |
+
hidden_states = []
|
168 |
|
169 |
+
assert len(self.text_hidden_fcs) == 1
|
170 |
+
hidden_states.append(self.text_hidden_fcs[0](output_hidden_states))
|
171 |
+
|
172 |
+
last_hidden_state = torch.stack(hidden_states, dim=-1).sum(dim=-1)
|
173 |
+
pred_embeddings = last_hidden_state[seg_token_mask]
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
174 |
|
175 |
+
seg_token_counts = seg_token_mask.int().sum(-1) #[bs, ]
|
176 |
+
seg_token_offset = seg_token_counts.cumsum(-1)
|
177 |
+
seg_token_offset = torch.cat([torch.zeros(1).long().cuda(), seg_token_offset], dim=0)
|
178 |
+
|
179 |
+
pred_embeddings_ = []
|
180 |
+
for i in range(len(seg_token_offset)-1):
|
181 |
+
start_i, end_i = seg_token_offset[i], seg_token_offset[i+1]
|
182 |
+
pred_embeddings_.append(pred_embeddings[start_i: end_i])
|
183 |
+
pred_embeddings = pred_embeddings_
|
184 |
+
|
185 |
+
image_embeddings = self.get_visual_embs(images)
|
186 |
+
|
187 |
+
multimask_output = False
|
188 |
+
pred_masks = []
|
189 |
+
for i in range(len(pred_embeddings)):
|
190 |
+
sparse_embeddings, dense_embeddings = self.visual_model.prompt_encoder(
|
191 |
+
points=None,
|
192 |
+
boxes=None,
|
193 |
+
masks=None,
|
194 |
+
text_embeds=pred_embeddings[i].unsqueeze(1),
|
195 |
+
)
|
196 |
+
|
197 |
+
sparse_embeddings = sparse_embeddings.to(pred_embeddings[i].dtype)
|
198 |
+
low_res_masks, iou_predictions = self.visual_model.mask_decoder(
|
199 |
+
image_embeddings=image_embeddings[i].unsqueeze(0),
|
200 |
+
image_pe=self.visual_model.prompt_encoder.get_dense_pe(),
|
201 |
+
sparse_prompt_embeddings=sparse_embeddings,
|
202 |
+
dense_prompt_embeddings=dense_embeddings,
|
203 |
+
multimask_output=multimask_output,
|
204 |
+
)
|
205 |
+
|
206 |
+
pred_mask = self.visual_model.postprocess_masks(
|
207 |
+
low_res_masks,
|
208 |
+
input_size=resize_list[i],
|
209 |
+
original_size=original_size_list[i],
|
210 |
+
)
|
211 |
+
pred_masks.append(pred_mask[:, 0])
|
212 |
|
213 |
return output_ids, pred_masks
|
model/llava/model/llava.py
CHANGED
@@ -63,6 +63,8 @@ class LlavaLlamaModel(LlamaModel):
|
|
63 |
vision_tower.requires_grad_(False)
|
64 |
if precision == 'bf16':
|
65 |
vision_tower = vision_tower.to(torch.bfloat16)
|
|
|
|
|
66 |
else:
|
67 |
vision_tower = vision_tower.to(torch.float32)
|
68 |
|
|
|
63 |
vision_tower.requires_grad_(False)
|
64 |
if precision == 'bf16':
|
65 |
vision_tower = vision_tower.to(torch.bfloat16)
|
66 |
+
elif precision == 'fp16':
|
67 |
+
vision_tower = vision_tower.to(torch.half)
|
68 |
else:
|
69 |
vision_tower = vision_tower.to(torch.float32)
|
70 |
|
model/segment_anything/modeling/image_encoder.py
CHANGED
@@ -114,8 +114,13 @@ class ImageEncoderViT(nn.Module):
|
|
114 |
for blk in self.blocks:
|
115 |
x = blk(x)
|
116 |
|
117 |
-
|
118 |
-
|
|
|
|
|
|
|
|
|
|
|
119 |
return x
|
120 |
|
121 |
|
|
|
114 |
for blk in self.blocks:
|
115 |
x = blk(x)
|
116 |
|
117 |
+
dtype = x.dtype
|
118 |
+
if dtype == torch.float16: # prevent overflow
|
119 |
+
with torch.autocast(device_type='cuda', dtype=torch.float32):
|
120 |
+
x = self.neck(x.permute(0, 3, 1, 2))
|
121 |
+
x = x.to(dtype)
|
122 |
+
else:
|
123 |
+
x = self.neck(x.permute(0, 3, 1, 2))
|
124 |
return x
|
125 |
|
126 |
|