Spaces:
Running
Running
Bram Vanroy
commited on
Commit
β’
51df785
1
Parent(s):
1184fa3
add CUDA support
Browse files
utils.py
CHANGED
@@ -2,8 +2,9 @@ from typing import Tuple
|
|
2 |
|
3 |
import streamlit as st
|
4 |
|
|
|
5 |
from torch.quantization import quantize_dynamic
|
6 |
-
from torch import nn, qint8
|
7 |
from torch.nn import Parameter
|
8 |
from transformers import PreTrainedModel, PreTrainedTokenizer
|
9 |
from optimum.bettertransformer import BetterTransformer
|
@@ -14,17 +15,19 @@ from transformers import MBartForConditionalGeneration
|
|
14 |
|
15 |
st_hash_funcs = {PreTrainedModel: lambda model: model.name_or_path,
|
16 |
PreTrainedTokenizer: lambda tokenizer: tokenizer.name_or_path,
|
17 |
-
Parameter: lambda
|
|
|
18 |
|
19 |
|
20 |
@st.cache(show_spinner=False, hash_funcs=st_hash_funcs, allow_output_mutation=True)
|
21 |
-
def get_resources(multilingual: bool, quantize: bool = True) -> Tuple[MBartForConditionalGeneration, AMRMBartTokenizer, AMRLogitsProcessor]:
|
22 |
"""Get the relevant model, tokenizer and logits_processor. The loaded model depends on whether the multilingual
|
23 |
model is requested, or not. If not, an English-only model is loaded. The model can be optionally quantized
|
24 |
for better performance.
|
25 |
|
26 |
:param multilingual: whether or not to load the multilingual model. If not, loads the English-only model
|
27 |
:param quantize: whether to quantize the model with PyTorch's 'quantize_dynamic'
|
|
|
28 |
:return: the loaded model, tokenizer, and logits processor
|
29 |
"""
|
30 |
if multilingual:
|
@@ -38,7 +41,9 @@ def get_resources(multilingual: bool, quantize: bool = True) -> Tuple[MBartForCo
|
|
38 |
model = BetterTransformer.transform(model, keep_original_model=False)
|
39 |
model.resize_token_embeddings(len(tokenizer))
|
40 |
|
41 |
-
if
|
|
|
|
|
42 |
model = quantize_dynamic(model, {nn.Linear, nn.Dropout, nn.LayerNorm}, dtype=qint8)
|
43 |
|
44 |
logits_processor = AMRLogitsProcessor(tokenizer, model.config.max_length)
|
@@ -60,7 +65,8 @@ def translate(text: str, src_lang: str, model: MBartForConditionalGeneration, to
|
|
60 |
"""
|
61 |
tokenizer.src_lang = LANGUAGES[src_lang]
|
62 |
encoded = tokenizer(text, return_tensors="pt")
|
63 |
-
|
|
|
64 |
return tokenizer.decode_and_fix(generated)[0]
|
65 |
|
66 |
|
|
|
2 |
|
3 |
import streamlit as st
|
4 |
|
5 |
+
import torch
|
6 |
from torch.quantization import quantize_dynamic
|
7 |
+
from torch import nn, qint8, Tensor
|
8 |
from torch.nn import Parameter
|
9 |
from transformers import PreTrainedModel, PreTrainedTokenizer
|
10 |
from optimum.bettertransformer import BetterTransformer
|
|
|
15 |
|
16 |
st_hash_funcs = {PreTrainedModel: lambda model: model.name_or_path,
|
17 |
PreTrainedTokenizer: lambda tokenizer: tokenizer.name_or_path,
|
18 |
+
Parameter: lambda parameter: parameter.data,
|
19 |
+
Tensor: lambda tensor: tensor.cpu()}
|
20 |
|
21 |
|
22 |
@st.cache(show_spinner=False, hash_funcs=st_hash_funcs, allow_output_mutation=True)
|
23 |
+
def get_resources(multilingual: bool, quantize: bool = True, no_cuda: bool = False) -> Tuple[MBartForConditionalGeneration, AMRMBartTokenizer, AMRLogitsProcessor]:
|
24 |
"""Get the relevant model, tokenizer and logits_processor. The loaded model depends on whether the multilingual
|
25 |
model is requested, or not. If not, an English-only model is loaded. The model can be optionally quantized
|
26 |
for better performance.
|
27 |
|
28 |
:param multilingual: whether or not to load the multilingual model. If not, loads the English-only model
|
29 |
:param quantize: whether to quantize the model with PyTorch's 'quantize_dynamic'
|
30 |
+
:param no_cuda: whether to disable CUDA, even if it is available
|
31 |
:return: the loaded model, tokenizer, and logits processor
|
32 |
"""
|
33 |
if multilingual:
|
|
|
41 |
model = BetterTransformer.transform(model, keep_original_model=False)
|
42 |
model.resize_token_embeddings(len(tokenizer))
|
43 |
|
44 |
+
if torch.cuda.is_available() and not no_cuda:
|
45 |
+
model = model.to("cuda")
|
46 |
+
elif quantize: # Quantization not supported on CUDA
|
47 |
model = quantize_dynamic(model, {nn.Linear, nn.Dropout, nn.LayerNorm}, dtype=qint8)
|
48 |
|
49 |
logits_processor = AMRLogitsProcessor(tokenizer, model.config.max_length)
|
|
|
65 |
"""
|
66 |
tokenizer.src_lang = LANGUAGES[src_lang]
|
67 |
encoded = tokenizer(text, return_tensors="pt")
|
68 |
+
encoded = {k: v.to(model.device) for k, v in encoded.items()}
|
69 |
+
generated = model.generate(**encoded, **gen_kwargs).cpu()
|
70 |
return tokenizer.decode_and_fix(generated)[0]
|
71 |
|
72 |
|