Avoiding CUDA Memory limit by rebatching inputs
Browse files
app.py
CHANGED
@@ -13,9 +13,7 @@ quantization_config = BitsAndBytesConfig(
|
|
13 |
)
|
14 |
|
15 |
embedder = SentenceTransformer('all-mpnet-base-v2')
|
16 |
-
|
17 |
model_id = "llava-hf/llava-1.5-7b-hf"
|
18 |
-
|
19 |
processor = AutoProcessor.from_pretrained(model_id)
|
20 |
model = LlavaForConditionalGeneration.from_pretrained(
|
21 |
model_id,
|
@@ -25,24 +23,79 @@ model = LlavaForConditionalGeneration.from_pretrained(
|
|
25 |
low_cpu_mem_usage=True
|
26 |
)
|
27 |
|
|
|
28 |
|
29 |
-
def text_to_image(image, prompt, duplications:
|
30 |
prompt = f'USER: <image>\n{prompt}\nASSISTANT:'
|
31 |
|
32 |
image_batch = [image]
|
33 |
prompt_batch = [prompt]
|
34 |
-
for _ in range(duplications):
|
35 |
image_batch.append(deepcopy(image))
|
36 |
prompt_batch.append(prompt)
|
37 |
|
38 |
-
inputs = processor(prompt_batch, images=image_batch, padding=True, return_tensors="pt")
|
39 |
-
|
40 |
-
|
41 |
-
|
42 |
-
|
43 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
44 |
|
45 |
-
return
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
46 |
|
47 |
|
48 |
demo = gr.Interface(
|
|
|
13 |
)
|
14 |
|
15 |
embedder = SentenceTransformer('all-mpnet-base-v2')
|
|
|
16 |
model_id = "llava-hf/llava-1.5-7b-hf"
|
|
|
17 |
processor = AutoProcessor.from_pretrained(model_id)
|
18 |
model = LlavaForConditionalGeneration.from_pretrained(
|
19 |
model_id,
|
|
|
23 |
low_cpu_mem_usage=True
|
24 |
)
|
25 |
|
26 |
+
MAXIMUM_PIXEL_VALUES = 3725568
|
27 |
|
28 |
+
def text_to_image(image, prompt, duplications: float):
|
29 |
prompt = f'USER: <image>\n{prompt}\nASSISTANT:'
|
30 |
|
31 |
image_batch = [image]
|
32 |
prompt_batch = [prompt]
|
33 |
+
for _ in range(int(duplications)):
|
34 |
image_batch.append(deepcopy(image))
|
35 |
prompt_batch.append(prompt)
|
36 |
|
37 |
+
inputs = processor(prompt_batch, images=image_batch, padding=True, return_tensors="pt")
|
38 |
+
|
39 |
+
batched_inputs :list[dict[str, torch.Tensor]] = list()
|
40 |
+
if inputs['pixel_values'].flatten().shape[0] > MAXIMUM_PIXEL_VALUES:
|
41 |
+
batch = dict(input_ids=list(), attention_mask=list(), pixel_values=list())
|
42 |
+
i = 0
|
43 |
+
while i < len(inputs['pixel_values']):
|
44 |
+
batch['input_ids'].append(inputs['input_ids'][i])
|
45 |
+
batch['attention_mask'].append(inputs['attention_mask'][i])
|
46 |
+
batch['pixel_values'].append(inputs['pixel_values'][i])
|
47 |
+
|
48 |
+
if torch.cat(batch['pixel_values'], dim=0).flatten().shape[0] > MAXIMUM_PIXEL_VALUES:
|
49 |
+
print(f'[{i}/{len(inputs["pixel_values"])}] - Reached max pixel values for batch prediction on T4 '
|
50 |
+
f'16GB GPU. Will split in more batches')
|
51 |
+
# Remove the last added image because it's too big to process
|
52 |
+
batch['input_ids'].pop()
|
53 |
+
batch['attention_mask'].pop()
|
54 |
+
batch['pixel_values'].pop()
|
55 |
+
|
56 |
+
# transform lists to tensors
|
57 |
+
batch['input_ids'] = torch.cat(batch['input_ids'], dim=0)
|
58 |
+
batch['attention_mask'] = torch.cat(batch['input_ids'], dim=0)
|
59 |
+
batch['pixel_values'] = torch.cat(batch['input_ids'], dim=0)
|
60 |
+
|
61 |
+
# Add to the batched_inputs
|
62 |
+
batched_inputs.append(batch)
|
63 |
+
else:
|
64 |
+
i += 1
|
65 |
+
|
66 |
+
maurice_description = list()
|
67 |
+
maurice_embeddings = list()
|
68 |
+
for batch in batched_inputs:
|
69 |
+
# Load on device
|
70 |
+
batch['input_ids'].to(model.device)
|
71 |
+
batch['attention_mask'].to(model.device)
|
72 |
+
batch['pixel_values'].to(model.device)
|
73 |
+
output = model.generate(**inputs, max_new_tokens=500, temperature=0.3)
|
74 |
+
|
75 |
+
# Unload GPU
|
76 |
+
batch['input_ids'].to('cpu')
|
77 |
+
batch['attention_mask'].to('cpu')
|
78 |
+
batch['pixel_values'].to('cpu')
|
79 |
+
|
80 |
+
generated_text = processor.batch_decode(output, skip_special_tokens=True)
|
81 |
+
output = output.to('cpu')
|
82 |
+
|
83 |
+
for text in generated_text:
|
84 |
+
text_output = text.split("ASSISTANT:")[-1]
|
85 |
+
text_embeddings = embedder.encode(text_output).to('cpu')
|
86 |
+
maurice_description.append(text_output)
|
87 |
+
maurice_embeddings.append(text_embeddings)
|
88 |
|
89 |
+
return '\n---\n'.join(maurice_description), dict(text_embeddings=maurice_embeddings)
|
90 |
+
# inputs = inputs.to(model.device)
|
91 |
+
# print()
|
92 |
+
# output = model.generate(**inputs, max_new_tokens=500, temperature=0.3)
|
93 |
+
# generated_text = processor.batch_decode(output, skip_special_tokens=True)
|
94 |
+
# text = generated_text.pop()
|
95 |
+
# text_output = text.split("ASSISTANT:")[-1]
|
96 |
+
# text_embeddings = embedder.encode(text_output)
|
97 |
+
#
|
98 |
+
# return text_output, dict(text_embeddings=text_embeddings)
|
99 |
|
100 |
|
101 |
demo = gr.Interface(
|