feat: support batch infer and optimize processor
Browse files- modeling_valley.py +43 -7
- preprocessor_config.json +0 -20
- processing_valley.py +9 -4
modeling_valley.py
CHANGED
@@ -17,7 +17,7 @@ import numpy as np
|
|
17 |
from torch import nn
|
18 |
from torch.nn import CrossEntropyLoss
|
19 |
from abc import ABC, abstractmethod
|
20 |
-
from typing import List, Optional, Tuple, Union
|
21 |
from transformers.modeling_outputs import CausalLMOutputWithPast
|
22 |
from transformers import AutoConfig, AutoModelForCausalLM, Qwen2Config, Qwen2ForCausalLM, Qwen2Model
|
23 |
|
@@ -39,7 +39,7 @@ class ValleyMetaModel:
|
|
39 |
else:
|
40 |
self.vision_tower = build_vision_tower(config, delay_load=False)
|
41 |
# Build Projector
|
42 |
-
if hasattr(config, "mm_projector_type"):
|
43 |
self.mm_projector = build_vision_projector(config)
|
44 |
|
45 |
def get_vision_tower(self):
|
@@ -114,6 +114,15 @@ class ValleyMetaForCausalLM(ABC):
|
|
114 |
|
115 |
return image_features
|
116 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
117 |
|
118 |
def prepare_inputs_labels_for_multimodal(
|
119 |
self, input_ids, position_ids, attention_mask, past_key_values, labels, images,
|
@@ -128,7 +137,6 @@ class ValleyMetaForCausalLM(ABC):
|
|
128 |
dtype=attention_mask.dtype,
|
129 |
device=attention_mask.device
|
130 |
)), dim=1)
|
131 |
-
position_ids = torch.sum(attention_mask, dim=1).unsqueeze(-1) - 1
|
132 |
return input_ids, position_ids, attention_mask, past_key_values, None, labels
|
133 |
|
134 |
# Step1: Get image embedings
|
@@ -355,8 +363,7 @@ class ValleyMetaForCausalLM(ABC):
|
|
355 |
|
356 |
for i, (cur_new_embed, cur_new_labels, cur_attention_mask) in enumerate(zip(new_input_embeds, new_labels, new_attention_mask)):
|
357 |
cur_len = cur_new_embed.shape[0]
|
358 |
-
|
359 |
-
if not self.training and not getattr(self, "right_padding", None):
|
360 |
new_input_embeds_padded.append(torch.cat((
|
361 |
torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device),
|
362 |
cur_new_embed
|
@@ -366,7 +373,6 @@ class ValleyMetaForCausalLM(ABC):
|
|
366 |
new_attention_mask_padded[i, -cur_len:] = cur_attention_mask
|
367 |
position_ids[i, -cur_len:] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
|
368 |
|
369 |
-
# Left padding while training
|
370 |
else:
|
371 |
new_input_embeds_padded.append(torch.cat((
|
372 |
cur_new_embed,
|
@@ -404,6 +410,33 @@ class ValleyQwen2ForCausalLM(Qwen2ForCausalLM, ValleyMetaForCausalLM):
|
|
404 |
def get_model(self):
|
405 |
return self.model
|
406 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
407 |
def forward(
|
408 |
self,
|
409 |
input_ids: torch.LongTensor = None,
|
@@ -481,7 +514,7 @@ class ValleyQwen2ForCausalLM(Qwen2ForCausalLM, ValleyMetaForCausalLM):
|
|
481 |
output = (logits,) + outputs[1:]
|
482 |
return (loss,) + output if loss is not None else output
|
483 |
|
484 |
-
|
485 |
loss=loss,
|
486 |
logits=logits,
|
487 |
past_key_values=outputs.past_key_values,
|
@@ -489,6 +522,9 @@ class ValleyQwen2ForCausalLM(Qwen2ForCausalLM, ValleyMetaForCausalLM):
|
|
489 |
attentions=outputs.attentions,
|
490 |
)
|
491 |
|
|
|
|
|
|
|
492 |
def prepare_inputs_for_generation(
|
493 |
self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
|
494 |
):
|
|
|
17 |
from torch import nn
|
18 |
from torch.nn import CrossEntropyLoss
|
19 |
from abc import ABC, abstractmethod
|
20 |
+
from typing import List, Optional, Tuple, Union, Dict, Any
|
21 |
from transformers.modeling_outputs import CausalLMOutputWithPast
|
22 |
from transformers import AutoConfig, AutoModelForCausalLM, Qwen2Config, Qwen2ForCausalLM, Qwen2Model
|
23 |
|
|
|
39 |
else:
|
40 |
self.vision_tower = build_vision_tower(config, delay_load=False)
|
41 |
# Build Projector
|
42 |
+
if hasattr(config, "mm_projector_type") and not getattr(config, "only_navit", False):
|
43 |
self.mm_projector = build_vision_projector(config)
|
44 |
|
45 |
def get_vision_tower(self):
|
|
|
114 |
|
115 |
return image_features
|
116 |
|
117 |
+
def get_padding_method(self):
|
118 |
+
right_padding = getattr(self, 'right_padding', None)
|
119 |
+
# if right_padding flag is setted, ignore training flag.
|
120 |
+
if right_padding is not None:
|
121 |
+
method = 'right' if right_padding else 'left'
|
122 |
+
# in the other way, use training flag to determine the padding method.
|
123 |
+
method = 'right' if self.training else 'left'
|
124 |
+
|
125 |
+
return method
|
126 |
|
127 |
def prepare_inputs_labels_for_multimodal(
|
128 |
self, input_ids, position_ids, attention_mask, past_key_values, labels, images,
|
|
|
137 |
dtype=attention_mask.dtype,
|
138 |
device=attention_mask.device
|
139 |
)), dim=1)
|
|
|
140 |
return input_ids, position_ids, attention_mask, past_key_values, None, labels
|
141 |
|
142 |
# Step1: Get image embedings
|
|
|
363 |
|
364 |
for i, (cur_new_embed, cur_new_labels, cur_attention_mask) in enumerate(zip(new_input_embeds, new_labels, new_attention_mask)):
|
365 |
cur_len = cur_new_embed.shape[0]
|
366 |
+
if self.get_padding_method() == 'left':
|
|
|
367 |
new_input_embeds_padded.append(torch.cat((
|
368 |
torch.zeros((max_len - cur_len, cur_new_embed.shape[1]), dtype=cur_new_embed.dtype, device=cur_new_embed.device),
|
369 |
cur_new_embed
|
|
|
373 |
new_attention_mask_padded[i, -cur_len:] = cur_attention_mask
|
374 |
position_ids[i, -cur_len:] = torch.arange(0, cur_len, dtype=position_ids.dtype, device=position_ids.device)
|
375 |
|
|
|
376 |
else:
|
377 |
new_input_embeds_padded.append(torch.cat((
|
378 |
cur_new_embed,
|
|
|
410 |
def get_model(self):
|
411 |
return self.model
|
412 |
|
413 |
+
def _update_model_kwargs_for_generation(
|
414 |
+
self,
|
415 |
+
outputs: CausalLMOutputWithPast,
|
416 |
+
model_kwargs: Dict[str, Any],
|
417 |
+
is_encoder_decoder: bool = False,
|
418 |
+
num_new_tokens: int = 1,
|
419 |
+
) -> Dict[str, Any]:
|
420 |
+
new_model_kwargs = super()._update_model_kwargs_for_generation(
|
421 |
+
outputs,
|
422 |
+
model_kwargs,
|
423 |
+
is_encoder_decoder,
|
424 |
+
num_new_tokens
|
425 |
+
)
|
426 |
+
"""
|
427 |
+
Set model_kwargs["attention_mask"] to the expanded `attention_mask` in
|
428 |
+
the `prepare_inputs_labels_for_multimodal` function to ensure the
|
429 |
+
correctness of the generate behavior when `use_cache` is enabled.
|
430 |
+
"""
|
431 |
+
if not is_encoder_decoder:
|
432 |
+
if "attention_mask" in new_model_kwargs:
|
433 |
+
attention_mask = outputs.attention_mask
|
434 |
+
new_model_kwargs["attention_mask"] = torch.cat(
|
435 |
+
[attention_mask, attention_mask.new_ones((attention_mask.shape[0], 1))], dim=-1
|
436 |
+
)
|
437 |
+
return new_model_kwargs
|
438 |
+
|
439 |
+
|
440 |
def forward(
|
441 |
self,
|
442 |
input_ids: torch.LongTensor = None,
|
|
|
514 |
output = (logits,) + outputs[1:]
|
515 |
return (loss,) + output if loss is not None else output
|
516 |
|
517 |
+
res = CausalLMOutputWithPast(
|
518 |
loss=loss,
|
519 |
logits=logits,
|
520 |
past_key_values=outputs.past_key_values,
|
|
|
522 |
attentions=outputs.attentions,
|
523 |
)
|
524 |
|
525 |
+
res.attention_mask = attention_mask
|
526 |
+
return res
|
527 |
+
|
528 |
def prepare_inputs_for_generation(
|
529 |
self, input_ids, past_key_values=None, attention_mask=None, inputs_embeds=None, **kwargs
|
530 |
):
|
preprocessor_config.json
CHANGED
@@ -2,25 +2,5 @@
|
|
2 |
"processor_class": "ValleyProcessor",
|
3 |
"auto_map": {
|
4 |
"AutoProcessor": "processing_valley.ValleyProcessor"
|
5 |
-
},
|
6 |
-
"min_pixels": 1,
|
7 |
-
"qwen2vl_processor_config": {
|
8 |
-
"min_pixels": 3136,
|
9 |
-
"max_pixels": 12845056,
|
10 |
-
"patch_size": 14,
|
11 |
-
"temporal_patch_size": 2,
|
12 |
-
"merge_size": 2,
|
13 |
-
"image_mean": [
|
14 |
-
0.48145466,
|
15 |
-
0.4578275,
|
16 |
-
0.40821073
|
17 |
-
],
|
18 |
-
"image_std": [
|
19 |
-
0.26862954,
|
20 |
-
0.26130258,
|
21 |
-
0.27577711
|
22 |
-
],
|
23 |
-
"image_processor_type": "Qwen2VLImageProcessor",
|
24 |
-
"processor_class": "Qwen2VLProcessor"
|
25 |
}
|
26 |
}
|
|
|
2 |
"processor_class": "ValleyProcessor",
|
3 |
"auto_map": {
|
4 |
"AutoProcessor": "processing_valley.ValleyProcessor"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
5 |
}
|
6 |
}
|
processing_valley.py
CHANGED
@@ -88,10 +88,15 @@ class ValleyProcessor(ProcessorMixin):
|
|
88 |
self.siglip_image_processor = SiglipImageProcessor.from_dict(siglip_processor_config)
|
89 |
self.qwen2vl_image_processor = Qwen2VLImageProcessor.from_dict(
|
90 |
qwen2vl_processor_config,
|
91 |
-
max_pixels=kwargs.get("max_pixels", 1280*28*28),
|
92 |
-
min_pixels=kwargs.get("min_pixels", 4*28*28)
|
93 |
)
|
94 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
95 |
self.anyres = kwargs.get("anyres", True)
|
96 |
self.grid_pinpoints = kwargs.get("grid_pinpoints", "(1x1),...,(3x3)")
|
97 |
self.only_crop_single_image = kwargs.get("only_crop_single_image", True)
|
@@ -259,7 +264,7 @@ class ValleyProcessor(ProcessorMixin):
|
|
259 |
return input_ids
|
260 |
|
261 |
|
262 |
-
def __call__(self, messages, inference=True) -> BatchFeature:
|
263 |
# Deal with images
|
264 |
if "images" not in messages or not messages["images"] or not messages["images"][0]:
|
265 |
images = [self.black_img]
|
|
|
88 |
self.siglip_image_processor = SiglipImageProcessor.from_dict(siglip_processor_config)
|
89 |
self.qwen2vl_image_processor = Qwen2VLImageProcessor.from_dict(
|
90 |
qwen2vl_processor_config,
|
|
|
|
|
91 |
)
|
92 |
+
|
93 |
+
max_pixels = kwargs.get("max_pixels", None)
|
94 |
+
min_pixels = kwargs.get("min_pixels", None)
|
95 |
+
if max_pixels:
|
96 |
+
self.qwen2vl_image_processor.max_pixels = max_pixels
|
97 |
+
if min_pixels:
|
98 |
+
self.qwen2vl_image_processor.min_pixels = min_pixels
|
99 |
+
|
100 |
self.anyres = kwargs.get("anyres", True)
|
101 |
self.grid_pinpoints = kwargs.get("grid_pinpoints", "(1x1),...,(3x3)")
|
102 |
self.only_crop_single_image = kwargs.get("only_crop_single_image", True)
|
|
|
264 |
return input_ids
|
265 |
|
266 |
|
267 |
+
def __call__(self, messages, inference=True, **kwargs) -> BatchFeature:
|
268 |
# Deal with images
|
269 |
if "images" not in messages or not messages["images"] or not messages["images"][0]:
|
270 |
images = [self.black_img]
|