Spaces:
Running
on
Zero
Running
on
Zero
Upload infer.py
Browse files
infer.py
CHANGED
@@ -1,24 +1,36 @@
|
|
1 |
-
# modified from https://github.com/XiaomiMiMo/MiMo-VL/tree/main/infer.py
|
2 |
import os
|
3 |
import torch
|
4 |
-
from transformers import
|
5 |
-
|
|
|
|
|
|
|
|
|
6 |
from qwen_vl_utils import process_vision_info
|
7 |
from threading import Thread
|
8 |
|
9 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
10 |
class MiMoVLInfer:
|
11 |
def __init__(self, checkpoint_path, **kwargs):
|
12 |
-
dtype = torch.float16
|
13 |
self.model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
14 |
checkpoint_path,
|
15 |
-
torch_dtype=
|
16 |
device_map={"": "cpu"},
|
|
|
17 |
trust_remote_code=True,
|
18 |
).eval()
|
19 |
self.processor = AutoProcessor.from_pretrained(checkpoint_path, trust_remote_code=True)
|
20 |
self._on_cuda = False
|
21 |
|
|
|
|
|
22 |
def to_device(self, device: str):
|
23 |
if device == "cuda" and not self._on_cuda:
|
24 |
self.model.to("cuda")
|
@@ -30,55 +42,67 @@ class MiMoVLInfer:
|
|
30 |
def __call__(self, inputs: dict, history: list = [], temperature: float = 1.0):
|
31 |
messages = self.construct_messages(inputs)
|
32 |
updated_history = history + messages
|
33 |
-
|
|
|
|
|
|
|
34 |
image_inputs, video_inputs = process_vision_info(updated_history)
|
35 |
|
36 |
model_inputs = self.processor(
|
37 |
-
text=[
|
|
|
|
|
|
|
|
|
38 |
).to(self.model.device)
|
39 |
|
40 |
tokenizer = self.processor.tokenizer
|
41 |
-
streamer = TextIteratorStreamer(
|
|
|
|
|
42 |
|
43 |
-
max_new = int(os.getenv("MAX_NEW_TOKENS", "1024"))
|
44 |
temp = float(temperature or 0.0)
|
45 |
do_sample = temp > 1e-3
|
46 |
-
if do_sample
|
47 |
-
|
48 |
-
|
49 |
-
|
|
|
|
|
|
|
50 |
|
51 |
gen_kwargs = {
|
52 |
-
"max_new_tokens": 1024,
|
53 |
-
"streamer": streamer,
|
54 |
-
"stopping_criteria": StoppingCriteriaList([EosTokenCriteria(eos_token_id=self.model.config.eos_token_id)]),
|
55 |
-
"pad_token_id": self.model.config.eos_token_id,
|
56 |
**model_inputs,
|
57 |
-
|
|
|
|
|
|
|
|
|
58 |
}
|
59 |
|
60 |
thread = Thread(target=self.model.generate, kwargs=gen_kwargs, daemon=True)
|
61 |
thread.start()
|
62 |
-
|
63 |
-
|
64 |
-
|
65 |
-
|
66 |
-
|
67 |
-
|
|
|
68 |
}]
|
69 |
|
70 |
def _is_video_file(self, filename):
|
71 |
return any(filename.lower().endswith(ext) for ext in
|
72 |
-
[
|
73 |
|
74 |
def construct_messages(self, inputs: dict) -> list:
|
75 |
content = []
|
76 |
-
for path in inputs.get(
|
77 |
if self._is_video_file(path):
|
78 |
-
content.append({"type": "video", "video": f
|
79 |
else:
|
80 |
-
content.append({"type": "image", "image": f
|
81 |
-
|
82 |
-
if
|
83 |
-
content.append({"type": "text", "text":
|
84 |
return [{"role": "user", "content": content}]
|
|
|
|
|
1 |
import os
|
2 |
import torch
|
3 |
+
from transformers import (
|
4 |
+
AutoProcessor,
|
5 |
+
Qwen2_5_VLForConditionalGeneration,
|
6 |
+
TextIteratorStreamer,
|
7 |
+
)
|
8 |
+
from transformers.generation.logits_process import LogitsProcessor
|
9 |
from qwen_vl_utils import process_vision_info
|
10 |
from threading import Thread
|
11 |
|
12 |
|
13 |
+
class _NanSafeLogitsProcessor(LogitsProcessor):
|
14 |
+
def __call__(self, input_ids: torch.LongTensor, scores: torch.FloatTensor) -> torch.FloatTensor:
|
15 |
+
scores = torch.nan_to_num(scores, neginf=-1e4, posinf=1e4)
|
16 |
+
scores.clamp_(min=-1e4, max=1e4)
|
17 |
+
return scores
|
18 |
+
|
19 |
+
|
20 |
class MiMoVLInfer:
|
21 |
def __init__(self, checkpoint_path, **kwargs):
|
|
|
22 |
self.model = Qwen2_5_VLForConditionalGeneration.from_pretrained(
|
23 |
checkpoint_path,
|
24 |
+
torch_dtype=torch.float16,
|
25 |
device_map={"": "cpu"},
|
26 |
+
attn_implementation="eager",
|
27 |
trust_remote_code=True,
|
28 |
).eval()
|
29 |
self.processor = AutoProcessor.from_pretrained(checkpoint_path, trust_remote_code=True)
|
30 |
self._on_cuda = False
|
31 |
|
32 |
+
torch.set_float32_matmul_precision("high")
|
33 |
+
|
34 |
def to_device(self, device: str):
|
35 |
if device == "cuda" and not self._on_cuda:
|
36 |
self.model.to("cuda")
|
|
|
42 |
def __call__(self, inputs: dict, history: list = [], temperature: float = 1.0):
|
43 |
messages = self.construct_messages(inputs)
|
44 |
updated_history = history + messages
|
45 |
+
|
46 |
+
prompt = self.processor.apply_chat_template(
|
47 |
+
updated_history, tokenize=False, add_generation_prompt=True
|
48 |
+
)
|
49 |
image_inputs, video_inputs = process_vision_info(updated_history)
|
50 |
|
51 |
model_inputs = self.processor(
|
52 |
+
text=[prompt],
|
53 |
+
images=image_inputs,
|
54 |
+
videos=video_inputs,
|
55 |
+
padding=True,
|
56 |
+
return_tensors="pt",
|
57 |
).to(self.model.device)
|
58 |
|
59 |
tokenizer = self.processor.tokenizer
|
60 |
+
streamer = TextIteratorStreamer(
|
61 |
+
tokenizer, timeout=60.0, skip_prompt=True, skip_special_tokens=True
|
62 |
+
)
|
63 |
|
|
|
64 |
temp = float(temperature or 0.0)
|
65 |
do_sample = temp > 1e-3
|
66 |
+
sampling_args = {"do_sample": False} if not do_sample else {
|
67 |
+
"do_sample": True,
|
68 |
+
"temperature": max(temp, 0.01),
|
69 |
+
"top_p": 0.95,
|
70 |
+
}
|
71 |
+
|
72 |
+
max_new = int(os.getenv("MAX_NEW_TOKENS", "768"))
|
73 |
|
74 |
gen_kwargs = {
|
|
|
|
|
|
|
|
|
75 |
**model_inputs,
|
76 |
+
"max_new_tokens": max_new,
|
77 |
+
"streamer": streamer,
|
78 |
+
"pad_token_id": tokenizer.pad_token_id or tokenizer.eos_token_id,
|
79 |
+
"logits_processor": [_NanSafeLogitsProcessor()],
|
80 |
+
**sampling_args,
|
81 |
}
|
82 |
|
83 |
thread = Thread(target=self.model.generate, kwargs=gen_kwargs, daemon=True)
|
84 |
thread.start()
|
85 |
+
|
86 |
+
partial = ""
|
87 |
+
for chunk in streamer:
|
88 |
+
partial += chunk
|
89 |
+
yield partial, updated_history + [{
|
90 |
+
"role": "assistant",
|
91 |
+
"content": [{"type": "text", "text": partial}]
|
92 |
}]
|
93 |
|
94 |
def _is_video_file(self, filename):
|
95 |
return any(filename.lower().endswith(ext) for ext in
|
96 |
+
[".mp4", ".avi", ".mkv", ".mov", ".wmv", ".flv", ".webm", ".mpeg"])
|
97 |
|
98 |
def construct_messages(self, inputs: dict) -> list:
|
99 |
content = []
|
100 |
+
for path in inputs.get("files", []):
|
101 |
if self._is_video_file(path):
|
102 |
+
content.append({"type": "video", "video": f"file://{path}"})
|
103 |
else:
|
104 |
+
content.append({"type": "image", "image": f"file://{path}"})
|
105 |
+
q = inputs.get("text", "")
|
106 |
+
if q:
|
107 |
+
content.append({"type": "text", "text": q})
|
108 |
return [{"role": "user", "content": content}]
|