SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

v4.39 deprecations 🧼 (#29492)

Browse Source
This commit is contained in:
Joao Gante
2024-03-07 10:44:43 +00:00
committed by GitHub
parent 979fccc90f
commit ffe60fdcd6
14 changed files with 9 additions and 400 deletions
Download Patch File Download Diff File Expand all files Collapse all files

128
tests/generation/test_utils.py
View File

@@ -52,7 +52,6 @@ if is_torch_available():
GPT2Tokenizer,
ImageGPTForCausalImageModeling,
SpeechEncoderDecoderModel,
top_k_top_p_filtering,
)
from transformers.cache_utils import DynamicCache
from transformers.generation import (
@@ -2345,133 +2344,6 @@ class GenerationTesterMixin:
@require_torch
class UtilsFunctionsTest(unittest.TestCase):
# tests whether the top_k_top_p function behaves as expected
def test_top_k_top_p_filtering(self):
logits = torch.tensor(
[
[
8.2220991, # 3rd highest value; idx. 0
-0.5620044,
5.23229752,
4.0386393,
-6.8798378,
-0.54785802,
-3.2012153,
2.92777176,
1.88171953,
7.35341276,
8.43207833, # 2nd highest value; idx. 10
-9.85711836,
-5.96209236,
-1.13039161,
-7.1115294,
-0.8369633,
-5.3186408,
7.06427407,
0.81369344,
-0.82023817,
-5.9179796,
0.58813443,
-6.99778438,
4.71551189,
-0.18771637,
7.44020759, # 4th highest value; idx. 25
9.38450987, # 1st highest value; idx. 26
2.12662941,
-9.32562038,
2.35652522,
], # cummulative prob of 4 highest values <= 0.6
[
0.58425518,
4.53139238,
-5.57510464,
-6.28030699,
-7.19529503,
-4.02122551,
1.39337037,
-6.06707057,
1.59480517,
-9.643119,
0.03907799,
0.67231762,
-8.88206726,
6.27115922, # 4th highest value; idx. 13
2.28520723,
4.82767506,
4.30421368,
8.8275313, # 2nd highest value; idx. 17
5.44029958,
-4.4735794,
7.38579536, # 3rd highest value; idx. 20
-2.91051663,
2.61946077,
-2.5674762,
-9.48959302,
-4.02922645,
-1.35416918,
9.67702323, # 1st highest value; idx. 27
-5.89478553,
1.85370467,
], # cummulative prob of 4 highest values <= 0.6
],
dtype=torch.float,
device=torch_device,
)
non_inf_expected_idx = torch.tensor(
[[0, 0], [0, 10], [0, 25], [0, 26], [1, 13], [1, 17], [1, 20], [1, 27]],
dtype=torch.long,
device=torch_device,
) # expected non filtered idx as noted above
non_inf_expected_output = torch.tensor(
[
8.2221,
8.4321,
7.4402,
9.3845,
6.2712,
8.8275,
7.3858,
9.6770,
], # expected non filtered values as noted above
dtype=torch.float,
device=torch_device,
)
output = top_k_top_p_filtering(logits, top_k=10, top_p=0.6, min_tokens_to_keep=4)
non_inf_output = output[output != -float("inf")].to(device=torch_device)
non_inf_idx = (output != -float("inf")).nonzero().to(device=torch_device)
self.assertTrue(torch.allclose(non_inf_expected_output, non_inf_output, atol=1e-12))
self.assertTrue(torch.all(torch.eq(non_inf_expected_idx, non_inf_idx)))
# tests whether the function uses filter_value instead of default -inf
def test_top_k_top_p_filtering_with_filter_value(self):
logits = torch.tensor(
[
[
1,
1,
1,
0.99, # get filtered by top-p filtering
0.98, # get filtered by top-k filtering
]
],
dtype=torch.float,
device=torch_device,
)
expected_output = torch.tensor(
[[1, 1, 1, 0, 0]],
dtype=torch.float,
device=torch_device,
)
output = top_k_top_p_filtering(logits, top_k=4, top_p=0.5, filter_value=0.0)
self.assertTrue(torch.allclose(expected_output, output, atol=1e-12))
def test_speculative_sampling(self):
# assume vocab size 10, input length 5 + 3 generated candidates
candidate_input_ids = torch.tensor([[8, 0, 3, 9, 8, 1, 4, 5]]) # input tokens