saymrwulf/transformers
1
0
Fork 0
mirror of https://github.com/saymrwulf/transformers.git synced 2026-05-14 20:58:08 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Gemma 2: Update slow tests (#31759)

Browse source 
gemma 2 slow tests
This commit is contained in:
Joao Gante 2024-07-03 10:43:44 +01:00 committed by GitHub
parent c1fe12595e
commit ddfaf11926
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
2 changed files with 60 additions and 39 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

20
src/transformers/pipelines/text_generation.py
View file

@ -272,12 +272,17 @@ class TextGenerationPipeline(Pipeline):
max_length=None,
**generate_kwargs,
):
# Only set non-None tokenizer kwargs, so as to rely on the tokenizer's defaults
tokenizer_kwargs = {
"add_special_tokens": add_special_tokens,
"truncation": truncation,
"padding": padding,
"max_length": max_length,
}
tokenizer_kwargs = {key: value for key, value in tokenizer_kwargs.items() if value is not None}
if isinstance(prompt_text, Chat):
# Only set non-None tokenizer kwargs, so as to rely on the tokenizer's defaults
tokenizer_kwargs = {}
for tokenizer_kwarg_name in ["truncation", "padding", "max_length"]:
if locals()[tokenizer_kwarg_name] is not None:
tokenizer_kwargs[tokenizer_kwarg_name] = locals()[tokenizer_kwarg_name]
tokenizer_kwargs.pop("add_special_tokens", None) # ignore add_special_tokens on chats
inputs = self.tokenizer.apply_chat_template(
prompt_text.messages,
add_generation_prompt=True,
@ -286,11 +291,6 @@ class TextGenerationPipeline(Pipeline):
**tokenizer_kwargs,
)
else:
# Only set non-None tokenizer kwargs, so as to rely on the tokenizer's defaults
tokenizer_kwargs = {}
for tokenizer_kwarg_name in ["add_special_tokens", "truncation", "padding", "max_length"]:
if locals()[tokenizer_kwarg_name] is not None:
tokenizer_kwargs[tokenizer_kwarg_name] = locals()[tokenizer_kwarg_name]
inputs = self.tokenizer(prefix + prompt_text, return_tensors=self.framework, **tokenizer_kwargs)
inputs["prompt_text"] = prompt_text

79
tests/models/gemma2/test_modeling_gemma2.py
View file

@ -16,7 +16,7 @@
import unittest
from transformers import AutoModelForCausalLM, AutoTokenizer, Gemma2Config, is_torch_available
from transformers import AutoModelForCausalLM, AutoTokenizer, Gemma2Config, is_torch_available, pipeline
from transformers.testing_utils import (
require_read_token,
require_torch,
@ -102,41 +102,62 @@ class Gemma2IntegrationTest(unittest.TestCase):
cls.cuda_compute_capability_major_version = torch.cuda.get_device_capability()[0]
@require_read_token
def test_model_2b_bf16(self):
def test_model_9b_bf16(self):
model_id = "google/gemma-2-9b"
EXPECTED_TEXTS = [
"<bos>Hello I am doing a project for a class and I am trying to use the <code><a-image></code>",
"<pad><pad><bos>Hi today. So, I'm going to show you how to do a problem from the textbook. So",
"<bos>Hello I am doing a project on the 1918 flu pandemic and I am trying to find out how many",
"<pad><pad><bos>Hi today I'm going to be talking about the history of the United States. The United States of America",
]
model = AutoModelForCausalLM.from_pretrained(
model_id, low_cpu_mem_usage=True, torch_dtype=torch.bfloat16, attn_implementation="eager"
).to(torch_device)
tokenizer = AutoTokenizer.from_pretrained(model_id)
inputs = tokenizer(self.input_text, return_tensors="pt", padding=True).to(torch_device)
output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
output_text = tokenizer.batch_decode(output, skip_special_tokens=False)
self.assertEqual(output_text, EXPECTED_TEXTS)
@require_read_token
def test_model_9b_fp16(self):
model_id = "google/gemma-2-9b"
EXPECTED_TEXTS = [
"<bos>Hello I am doing a project on the 1918 flu pandemic and I am trying to find out how many",
"<pad><pad><bos>Hi today I'm going to be talking about the history of the United States. The United States of America",
]
model = AutoModelForCausalLM.from_pretrained(
model_id, low_cpu_mem_usage=True, torch_dtype=torch.float16, attn_implementation="eager"
).to(torch_device)
tokenizer = AutoTokenizer.from_pretrained(model_id)
inputs = tokenizer(self.input_text, return_tensors="pt", padding=True).to(torch_device)
output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
output_text = tokenizer.batch_decode(output, skip_special_tokens=False)
self.assertEqual(output_text, EXPECTED_TEXTS)
@require_read_token
def test_model_9b_pipeline_bf16(self):
# See https://github.com/huggingface/transformers/pull/31747 -- pipeline was broken for Gemma2 before this PR
model_id = "google/gemma-2-9b"
# EXPECTED_TEXTS should match the same non-pipeline test, minus the special tokens
EXPECTED_TEXTS = [
"Hello I am doing a project on the 1918 flu pandemic and I am trying to find out how many",
"Hi today I'm going to be talking about the history of the United States. The United States of America",
]
model = AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True, torch_dtype=torch.bfloat16).to(
torch_device
)
tokenizer = AutoTokenizer.from_pretrained(model_id)
inputs = tokenizer(self.input_text, return_tensors="pt", padding=True).to(torch_device)
pipe = pipeline("text-generation", model=model, tokenizer=tokenizer)
output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
output_text = tokenizer.batch_decode(output, skip_special_tokens=True)
output = pipe(self.input_text, max_new_tokens=20, do_sample=False, padding=True)
self.assertEqual(output_text, EXPECTED_TEXTS)
@require_read_token
def test_model_2b_fp16(self):
model_id = "google/gemma-2-9b"
EXPECTED_TEXTS = [
"<bos>Hello I am doing a project on the effect of the temperature on the rate of a reaction. I am using a ",
"<pad><pad><bos>Hi today I'm going to be talking about the 1000-4000-",
]
model = AutoModelForCausalLM.from_pretrained(model_id, low_cpu_mem_usage=True, torch_dtype=torch.float16).to(
torch_device
)
tokenizer = AutoTokenizer.from_pretrained(model_id)
inputs = tokenizer(self.input_text, return_tensors="pt", padding=True).to(torch_device)
output = model.generate(**inputs, max_new_tokens=20, do_sample=False)
output_text = tokenizer.batch_decode(output, skip_special_tokens=True)
self.assertEqual(output_text, EXPECTED_TEXTS)
self.assertEqual(output[0][0]["generated_text"], EXPECTED_TEXTS[0])
self.assertEqual(output[1][0]["generated_text"], EXPECTED_TEXTS[1])