cheesyFishes
commited on
update handling of init args
Browse files- custom_st.py +16 -5
custom_st.py
CHANGED
@@ -22,6 +22,8 @@ class Transformer(nn.Module):
|
|
22 |
min_pixels: int = 1 * 28 * 28,
|
23 |
dimension: int = 2048,
|
24 |
max_seq_length: Optional[int] = None,
|
|
|
|
|
25 |
cache_dir: Optional[str] = None,
|
26 |
device: str = 'cuda:0',
|
27 |
**kwargs,
|
@@ -34,6 +36,17 @@ class Transformer(nn.Module):
|
|
34 |
self.min_pixels = min_pixels
|
35 |
self.max_seq_length = max_seq_length
|
36 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
37 |
# Initialize model
|
38 |
try:
|
39 |
self.model = Qwen2VLForConditionalGeneration.from_pretrained(
|
@@ -42,7 +55,7 @@ class Transformer(nn.Module):
|
|
42 |
torch_dtype=torch.bfloat16,
|
43 |
device_map=device,
|
44 |
cache_dir=cache_dir,
|
45 |
-
**
|
46 |
).eval()
|
47 |
except (ImportError, ValueError) as e:
|
48 |
print(f"Flash attention not available, falling back to default attention: {e}")
|
@@ -51,15 +64,13 @@ class Transformer(nn.Module):
|
|
51 |
torch_dtype=torch.bfloat16,
|
52 |
device_map=device,
|
53 |
cache_dir=cache_dir,
|
54 |
-
**
|
55 |
).eval()
|
56 |
|
57 |
# Initialize processor
|
58 |
self.processor = AutoProcessor.from_pretrained(
|
59 |
processor_name_or_path or model_name_or_path,
|
60 |
-
|
61 |
-
max_pixels=max_pixels,
|
62 |
-
cache_dir=cache_dir
|
63 |
)
|
64 |
|
65 |
# Set padding sides
|
|
|
22 |
min_pixels: int = 1 * 28 * 28,
|
23 |
dimension: int = 2048,
|
24 |
max_seq_length: Optional[int] = None,
|
25 |
+
model_args: Optional[Dict[str, Any]] = None,
|
26 |
+
processor_args: Optional[Dict[str, Any]] = None,
|
27 |
cache_dir: Optional[str] = None,
|
28 |
device: str = 'cuda:0',
|
29 |
**kwargs,
|
|
|
36 |
self.min_pixels = min_pixels
|
37 |
self.max_seq_length = max_seq_length
|
38 |
|
39 |
+
# Handle args
|
40 |
+
model_kwargs = model_args or {}
|
41 |
+
model_kwargs.update(kwargs)
|
42 |
+
|
43 |
+
processor_kwargs = processor_args or {}
|
44 |
+
processor_kwargs.update({
|
45 |
+
'min_pixels': min_pixels,
|
46 |
+
'max_pixels': max_pixels,
|
47 |
+
'cache_dir': cache_dir
|
48 |
+
})
|
49 |
+
|
50 |
# Initialize model
|
51 |
try:
|
52 |
self.model = Qwen2VLForConditionalGeneration.from_pretrained(
|
|
|
55 |
torch_dtype=torch.bfloat16,
|
56 |
device_map=device,
|
57 |
cache_dir=cache_dir,
|
58 |
+
**model_kwargs
|
59 |
).eval()
|
60 |
except (ImportError, ValueError) as e:
|
61 |
print(f"Flash attention not available, falling back to default attention: {e}")
|
|
|
64 |
torch_dtype=torch.bfloat16,
|
65 |
device_map=device,
|
66 |
cache_dir=cache_dir,
|
67 |
+
**model_kwargs
|
68 |
).eval()
|
69 |
|
70 |
# Initialize processor
|
71 |
self.processor = AutoProcessor.from_pretrained(
|
72 |
processor_name_or_path or model_name_or_path,
|
73 |
+
**processor_kwargs
|
|
|
|
|
74 |
)
|
75 |
|
76 |
# Set padding sides
|