Add vision_batch_size to avoid cuda OOM
Browse files- configuration_minicpm.py +3 -1
- modeling_minicpmv.py +27 -25
configuration_minicpm.py
CHANGED
@@ -69,6 +69,7 @@ class MiniCPMVConfig(Qwen2Config):
|
|
69 |
slice_config=None,
|
70 |
vision_config=None,
|
71 |
use_image_id=True,
|
|
|
72 |
**kwargs,
|
73 |
):
|
74 |
self.use_cache = use_cache
|
@@ -77,6 +78,7 @@ class MiniCPMVConfig(Qwen2Config):
|
|
77 |
self.drop_vision_last_layer = drop_vision_last_layer
|
78 |
self.batch_vision_input = batch_vision_input
|
79 |
self.use_image_id = use_image_id
|
|
|
80 |
|
81 |
if slice_config is None:
|
82 |
self.slice_config = MiniCPMVSliceConfig(max_slice_nums=1)
|
@@ -95,4 +97,4 @@ class MiniCPMVConfig(Qwen2Config):
|
|
95 |
|
96 |
self.patch_size = self.vision_config.patch_size
|
97 |
|
98 |
-
super().__init__(**kwargs)
|
|
|
69 |
slice_config=None,
|
70 |
vision_config=None,
|
71 |
use_image_id=True,
|
72 |
+
vision_batch_size=16,
|
73 |
**kwargs,
|
74 |
):
|
75 |
self.use_cache = use_cache
|
|
|
78 |
self.drop_vision_last_layer = drop_vision_last_layer
|
79 |
self.batch_vision_input = batch_vision_input
|
80 |
self.use_image_id = use_image_id
|
81 |
+
self.vision_batch_size = vision_batch_size
|
82 |
|
83 |
if slice_config is None:
|
84 |
self.slice_config = MiniCPMVSliceConfig(max_slice_nums=1)
|
|
|
97 |
|
98 |
self.patch_size = self.vision_config.patch_size
|
99 |
|
100 |
+
super().__init__(**kwargs)
|
modeling_minicpmv.py
CHANGED
@@ -92,31 +92,30 @@ class MiniCPMV(MiniCPMVPreTrainedModel):
|
|
92 |
tgt_sizes = [tgt_size for tgt_size in tgt_sizes if isinstance(tgt_size, torch.Tensor)]
|
93 |
tgt_sizes = torch.vstack(tgt_sizes).type(torch.int32)
|
94 |
|
95 |
-
|
96 |
-
|
97 |
-
|
98 |
-
|
99 |
-
|
100 |
-
|
101 |
-
|
102 |
-
|
103 |
-
|
104 |
-
|
105 |
-
|
106 |
-
|
107 |
-
|
108 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
109 |
else:
|
110 |
-
|
111 |
-
|
112 |
-
for single_tgt_size, single_pixel_values in zip(tgt_sizes, all_pixel_values):
|
113 |
-
single_pixel_values = single_pixel_values.unsqueeze(0)
|
114 |
-
B, L, _ = single_pixel_values.shape
|
115 |
-
single_pixel_values = single_pixel_values.permute(0, 2, 1).reshape(B, 3, -1, L)
|
116 |
-
single_vision_embedding = self.vpm(single_pixel_values.type(dtype), tgt_sizes=single_tgt_size.unsqueeze(0)).last_hidden_state
|
117 |
-
single_vision_embedding = self.resampler(single_vision_embedding, single_tgt_size.unsqueeze(0))
|
118 |
-
vision_embedding.append(single_vision_embedding)
|
119 |
-
vision_embedding = torch.vstack(vision_embedding)
|
120 |
|
121 |
start = 0
|
122 |
for pixel_values in pixel_values_list:
|
@@ -273,7 +272,7 @@ class MiniCPMV(MiniCPMVPreTrainedModel):
|
|
273 |
tokenizer,
|
274 |
processor=None,
|
275 |
vision_hidden_states=None,
|
276 |
-
max_new_tokens=
|
277 |
min_new_tokens=0,
|
278 |
sampling=True,
|
279 |
max_inp_length=8192,
|
@@ -292,6 +291,9 @@ class MiniCPMV(MiniCPMVPreTrainedModel):
|
|
292 |
|
293 |
if batched is False:
|
294 |
images_list, msgs_list = [images_list], [msgs_list]
|
|
|
|
|
|
|
295 |
assert len(images_list) == len(msgs_list), "The batch dim of images_list and msgs_list should be the same."
|
296 |
|
297 |
if processor is None:
|
|
|
92 |
tgt_sizes = [tgt_size for tgt_size in tgt_sizes if isinstance(tgt_size, torch.Tensor)]
|
93 |
tgt_sizes = torch.vstack(tgt_sizes).type(torch.int32)
|
94 |
|
95 |
+
max_patches = torch.max(tgt_sizes[:, 0] * tgt_sizes[:, 1])
|
96 |
+
|
97 |
+
all_pixel_values = torch.nn.utils.rnn.pad_sequence(all_pixel_values, batch_first=True,
|
98 |
+
padding_value=0.0)
|
99 |
+
B, L, _ = all_pixel_values.shape
|
100 |
+
all_pixel_values = all_pixel_values.permute(0, 2, 1).reshape(B, 3, -1, L)
|
101 |
+
|
102 |
+
patch_attn_mask = torch.zeros((B, 1, max_patches), dtype=torch.bool, device=device)
|
103 |
+
for i in range(B):
|
104 |
+
patch_attn_mask[i, 0, :tgt_sizes[i][0] * tgt_sizes[i][1]] = True
|
105 |
+
|
106 |
+
vision_batch_size = self.config.vision_batch_size
|
107 |
+
all_pixel_values = all_pixel_values.type(dtype)
|
108 |
+
if B > vision_batch_size:
|
109 |
+
hs = []
|
110 |
+
for i in range(0, B, vision_batch_size):
|
111 |
+
start_idx = i
|
112 |
+
end_idx = i + vision_batch_size
|
113 |
+
tmp_hs = self.vpm(all_pixel_values[start_idx:end_idx], patch_attention_mask=patch_attn_mask[start_idx:end_idx], tgt_sizes=tgt_sizes[start_idx:end_idx]).last_hidden_state
|
114 |
+
hs.append(tmp_hs)
|
115 |
+
vision_embedding = torch.cat(hs, dim=0)
|
116 |
else:
|
117 |
+
vision_embedding = self.vpm(all_pixel_values, patch_attention_mask=patch_attn_mask, tgt_sizes=tgt_sizes).last_hidden_state
|
118 |
+
vision_embedding = self.resampler(vision_embedding, tgt_sizes)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
119 |
|
120 |
start = 0
|
121 |
for pixel_values in pixel_values_list:
|
|
|
272 |
tokenizer,
|
273 |
processor=None,
|
274 |
vision_hidden_states=None,
|
275 |
+
max_new_tokens=2048,
|
276 |
min_new_tokens=0,
|
277 |
sampling=True,
|
278 |
max_inp_length=8192,
|
|
|
291 |
|
292 |
if batched is False:
|
293 |
images_list, msgs_list = [images_list], [msgs_list]
|
294 |
+
else:
|
295 |
+
assert images_list is None, "Please integrate image to msgs when using batch inference."
|
296 |
+
images_list = [None] * len(msgs_list)
|
297 |
assert len(images_list) == len(msgs_list), "The batch dim of images_list and msgs_list should be the same."
|
298 |
|
299 |
if processor is None:
|