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

Update all references to canonical models (#29001)

Browse Source 
* Script & Manual edition

* Update
This commit is contained in:
Lysandre Debut
2024-02-16 08:16:58 +01:00
committed by GitHub
parent 1e402b957d
commit f497f564bb
561 changed files with 2682 additions and 2687 deletions
Download Patch File Download Diff File Expand all files Collapse all files

26
tests/quantization/bnb/test_mixed_int8.py
View File

@@ -42,7 +42,7 @@ from transformers.testing_utils import (
def get_some_linear_layer(model):
if model.config.model_type == "gpt2":
if model.config.model_type == "openai-community/gpt2":
return model.transformer.h[0].mlp.c_fc
return model.transformer.h[0].mlp.dense_4h_to_h
@@ -174,7 +174,7 @@ class MixedInt8Test(BaseMixedInt8Test):
model = OPTForCausalLM(config)
self.assertEqual(get_keys_to_not_convert(model).sort(), ["lm_head", "model.decoder.embed_tokens"].sort())
model_id = "roberta-large"
model_id = "FacebookAI/roberta-large"
config = AutoConfig.from_pretrained(model_id, revision="716877d372b884cad6d419d828bac6c85b3b18d9")
with init_empty_weights():
model = AutoModelForMaskedLM.from_config(config)
@@ -240,7 +240,7 @@ class MixedInt8Test(BaseMixedInt8Test):
quantization_config = BitsAndBytesConfig(load_in_8bit=True, llm_int8_skip_modules=["classifier"])
seq_classification_model = AutoModelForSequenceClassification.from_pretrained(
"roberta-large-mnli", quantization_config=quantization_config
"FacebookAI/roberta-large-mnli", quantization_config=quantization_config
)
self.assertTrue(seq_classification_model.roberta.encoder.layer[0].output.dense.weight.dtype == torch.int8)
self.assertTrue(
@@ -340,7 +340,7 @@ class MixedInt8Test(BaseMixedInt8Test):
r"""
Test whether it is possible to mix both `int8` and `fp32` weights when using `keep_in_fp32_modules` correctly.
"""
model = AutoModelForSeq2SeqLM.from_pretrained("t5-small", load_in_8bit=True, device_map="auto")
model = AutoModelForSeq2SeqLM.from_pretrained("google-t5/t5-small", load_in_8bit=True, device_map="auto")
self.assertTrue(model.decoder.block[0].layer[2].DenseReluDense.wo.weight.dtype == torch.float32)
def test_int8_serialization(self):
@@ -447,7 +447,7 @@ class MixedInt8Test(BaseMixedInt8Test):
class MixedInt8T5Test(unittest.TestCase):
@classmethod
def setUpClass(cls):
cls.model_name = "t5-small"
cls.model_name = "google-t5/t5-small"
cls.dense_act_model_name = "google/flan-t5-small" # flan-t5 uses dense-act instead of dense-relu-dense
cls.tokenizer = AutoTokenizer.from_pretrained(cls.model_name)
cls.input_text = "Translate in German: Hello, my dog is cute"
@@ -463,7 +463,7 @@ class MixedInt8T5Test(unittest.TestCase):
def test_inference_without_keep_in_fp32(self):
r"""
Test whether it is possible to mix both `int8` and `fp32` weights when using `keep_in_fp32_modules` correctly.
`flan-t5-small` uses `T5DenseGatedActDense` whereas `t5-small` uses `T5DenseReluDense`. We need to test
`flan-t5-small` uses `T5DenseGatedActDense` whereas `google-t5/t5-small` uses `T5DenseReluDense`. We need to test
both cases.
"""
from transformers import T5ForConditionalGeneration
@@ -471,7 +471,7 @@ class MixedInt8T5Test(unittest.TestCase):
modules = T5ForConditionalGeneration._keep_in_fp32_modules
T5ForConditionalGeneration._keep_in_fp32_modules = None
# test with `t5-small`
# test with `google-t5/t5-small`
model = T5ForConditionalGeneration.from_pretrained(self.model_name, load_in_8bit=True, device_map="auto")
encoded_input = self.tokenizer(self.input_text, return_tensors="pt").to(0)
_ = model.generate(**encoded_input)
@@ -487,14 +487,14 @@ class MixedInt8T5Test(unittest.TestCase):
def test_inference_with_keep_in_fp32(self):
r"""
Test whether it is possible to mix both `int8` and `fp32` weights when using `keep_in_fp32_modules` correctly.
`flan-t5-small` uses `T5DenseGatedActDense` whereas `t5-small` uses `T5DenseReluDense`. We need to test
`flan-t5-small` uses `T5DenseGatedActDense` whereas `google-t5/t5-small` uses `T5DenseReluDense`. We need to test
both cases.
"""
import bitsandbytes as bnb
from transformers import T5ForConditionalGeneration
# test with `t5-small`
# test with `google-t5/t5-small`
model = T5ForConditionalGeneration.from_pretrained(self.model_name, load_in_8bit=True, device_map="auto")
# there was a bug with decoders - this test checks that it is fixed
@@ -514,14 +514,14 @@ class MixedInt8T5Test(unittest.TestCase):
r"""
Test whether it is possible to mix both `int8` and `fp32` weights when using `keep_in_fp32_modules` correctly on
a serialized model.
`flan-t5-small` uses `T5DenseGatedActDense` whereas `t5-small` uses `T5DenseReluDense`. We need to test
`flan-t5-small` uses `T5DenseGatedActDense` whereas `google-t5/t5-small` uses `T5DenseReluDense`. We need to test
both cases.
"""
import bitsandbytes as bnb
from transformers import T5ForConditionalGeneration
# test with `t5-small`
# test with `google-t5/t5-small`
model = T5ForConditionalGeneration.from_pretrained(self.model_name, load_in_8bit=True, device_map="auto")
with tempfile.TemporaryDirectory() as tmp_dir:
@@ -548,7 +548,7 @@ class MixedInt8ModelClassesTest(BaseMixedInt8Test):
super().setUp()
# model_name
self.model_name = "bigscience/bloom-560m"
self.seq_to_seq_name = "t5-small"
self.seq_to_seq_name = "google-t5/t5-small"
# Different types of model
@@ -842,7 +842,7 @@ class MixedInt8TestTraining(BaseMixedInt8Test):
class MixedInt8GPT2Test(MixedInt8Test):
model_name = "gpt2-xl"
model_name = "openai-community/gpt2-xl"
EXPECTED_RELATIVE_DIFFERENCE = 1.8720077507258357
EXPECTED_OUTPUTS = set()
EXPECTED_OUTPUTS.add("Hello my name is John Doe, and I'm a big fan of")