haixuantao
commited on
Commit
•
be75a6c
1
Parent(s):
c826c43
Adding fused_quantized fix
Browse files- operators/utils.py +13 -1
- tests/test_idefics2.py +25 -4
operators/utils.py
CHANGED
@@ -24,7 +24,7 @@ def speak(text):
|
|
24 |
|
25 |
speak("hello")
|
26 |
|
27 |
-
MODE = "
|
28 |
DEVICE = "cuda"
|
29 |
PROCESSOR = AutoProcessor.from_pretrained("HuggingFaceM4/idefics2-tfrm-compatible")
|
30 |
BAD_WORDS_IDS = PROCESSOR.tokenizer(
|
@@ -70,6 +70,17 @@ else:
|
|
70 |
raise ValueError("Unknown mode")
|
71 |
|
72 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
73 |
def ask_vlm(image, instruction):
|
74 |
prompts = [
|
75 |
"User:",
|
@@ -87,6 +98,7 @@ def ask_vlm(image, instruction):
|
|
87 |
generated_texts = PROCESSOR.batch_decode(generated_ids, skip_special_tokens=True)
|
88 |
|
89 |
text = generated_texts[0].split("\nAssistant: ")[1]
|
|
|
90 |
speak(text)
|
91 |
return text
|
92 |
|
|
|
24 |
|
25 |
speak("hello")
|
26 |
|
27 |
+
MODE = "fused_quantized"
|
28 |
DEVICE = "cuda"
|
29 |
PROCESSOR = AutoProcessor.from_pretrained("HuggingFaceM4/idefics2-tfrm-compatible")
|
30 |
BAD_WORDS_IDS = PROCESSOR.tokenizer(
|
|
|
70 |
raise ValueError("Unknown mode")
|
71 |
|
72 |
|
73 |
+
def reset_awq_cache(model):
|
74 |
+
"""
|
75 |
+
Simple method to reset the AWQ fused modules cache
|
76 |
+
"""
|
77 |
+
from awq.modules.fused.attn import QuantAttentionFused
|
78 |
+
|
79 |
+
for name, module in model.named_modules():
|
80 |
+
if isinstance(module, QuantAttentionFused):
|
81 |
+
module.start_pos = 0
|
82 |
+
|
83 |
+
|
84 |
def ask_vlm(image, instruction):
|
85 |
prompts = [
|
86 |
"User:",
|
|
|
98 |
generated_texts = PROCESSOR.batch_decode(generated_ids, skip_special_tokens=True)
|
99 |
|
100 |
text = generated_texts[0].split("\nAssistant: ")[1]
|
101 |
+
reset_awq_cache(model)
|
102 |
speak(text)
|
103 |
return text
|
104 |
|
tests/test_idefics2.py
CHANGED
@@ -3,10 +3,16 @@ import torch
|
|
3 |
from PIL import Image
|
4 |
from io import BytesIO
|
5 |
|
6 |
-
from transformers import
|
|
|
|
|
|
|
|
|
7 |
import awq_ext
|
8 |
|
9 |
-
|
|
|
|
|
10 |
DEVICE = "cuda"
|
11 |
PROCESSOR = AutoProcessor.from_pretrained("HuggingFaceM4/idefics2-tfrm-compatible")
|
12 |
BAD_WORDS_IDS = PROCESSOR.tokenizer(
|
@@ -78,7 +84,19 @@ image1 = download_image(
|
|
78 |
)
|
79 |
|
80 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
81 |
def ask_vlm(image, instruction):
|
|
|
82 |
prompts = [
|
83 |
"User:",
|
84 |
image,
|
@@ -93,17 +111,20 @@ def ask_vlm(image, instruction):
|
|
93 |
max_new_tokens=100,
|
94 |
)
|
95 |
generated_texts = PROCESSOR.batch_decode(generated_ids, skip_special_tokens=True)
|
|
|
96 |
return generated_texts
|
97 |
|
98 |
|
99 |
-
import time
|
100 |
-
|
101 |
now = time.time()
|
|
|
102 |
print(ask_vlm(image1, "What is this?")[0].split("\nAssistant: ")[1])
|
103 |
|
104 |
print("resp:", time.time() - now)
|
105 |
import time
|
106 |
|
|
|
107 |
now = time.time()
|
108 |
|
109 |
print(ask_vlm(image1, "What is this?")[0].split("\nAssistant: ")[1])
|
|
|
|
|
|
3 |
from PIL import Image
|
4 |
from io import BytesIO
|
5 |
|
6 |
+
from transformers import (
|
7 |
+
AutoProcessor,
|
8 |
+
AutoModelForVision2Seq,
|
9 |
+
AwqConfig,
|
10 |
+
)
|
11 |
import awq_ext
|
12 |
|
13 |
+
import time
|
14 |
+
|
15 |
+
MODE = "fused_quantized"
|
16 |
DEVICE = "cuda"
|
17 |
PROCESSOR = AutoProcessor.from_pretrained("HuggingFaceM4/idefics2-tfrm-compatible")
|
18 |
BAD_WORDS_IDS = PROCESSOR.tokenizer(
|
|
|
84 |
)
|
85 |
|
86 |
|
87 |
+
def reset_awq_cache(model):
|
88 |
+
"""
|
89 |
+
Simple method to reset the AWQ fused modules cache
|
90 |
+
"""
|
91 |
+
from awq.modules.fused.attn import QuantAttentionFused
|
92 |
+
|
93 |
+
for name, module in model.named_modules():
|
94 |
+
if isinstance(module, QuantAttentionFused):
|
95 |
+
module.start_pos = 0
|
96 |
+
|
97 |
+
|
98 |
def ask_vlm(image, instruction):
|
99 |
+
global model
|
100 |
prompts = [
|
101 |
"User:",
|
102 |
image,
|
|
|
111 |
max_new_tokens=100,
|
112 |
)
|
113 |
generated_texts = PROCESSOR.batch_decode(generated_ids, skip_special_tokens=True)
|
114 |
+
reset_awq_cache(model)
|
115 |
return generated_texts
|
116 |
|
117 |
|
|
|
|
|
118 |
now = time.time()
|
119 |
+
|
120 |
print(ask_vlm(image1, "What is this?")[0].split("\nAssistant: ")[1])
|
121 |
|
122 |
print("resp:", time.time() - now)
|
123 |
import time
|
124 |
|
125 |
+
|
126 |
now = time.time()
|
127 |
|
128 |
print(ask_vlm(image1, "What is this?")[0].split("\nAssistant: ")[1])
|
129 |
+
|
130 |
+
print("resp:", time.time() - now)
|