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

Fix GLM4 checkpoints (#38412)

Browse Source 
* fix

* fix

* fix

* fix

* fix

* fix

* test style bot

* Apply style fixes

---------

Co-authored-by: ydshieh <ydshieh@users.noreply.github.com>
Co-authored-by: github-actions[bot] <github-actions[bot]@users.noreply.github.com>
This commit is contained in:
Yih-Dar
2025-05-28 18:40:08 +02:00
committed by GitHub
parent 2872e8bac5
commit 66da700145
4 changed files with 74 additions and 43 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
src/transformers/models/glm4/configuration_glm4.py
View File

@@ -22,7 +22,7 @@ class Glm4Config(PretrainedConfig):
This is the configuration class to store the configuration of a [`Glm4Model`]. It is used to instantiate an Glm4
model according to the specified arguments, defining the model architecture. Instantiating a configuration with the
defaults will yield a similar configuration to that of the Glm4-4-9b-chat.
e.g. [THUDM/glm-4-0414-9b-chat-chat](https://huggingface.co/THUDM/glm-4-0414-9b-chat-chat)
e.g. [THUDM/GLM-4-9B-0414](https://huggingface.co/THUDM/GLM-4-9B-0414)
Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
documentation from [`PretrainedConfig`] for more information.
Args:

4
src/transformers/models/glm4/modeling_glm4.py
View File

@@ -551,8 +551,8 @@ class Glm4ForCausalLM(Glm4PreTrainedModel, GenerationMixin):
```python
>>> from transformers import AutoTokenizer, Glm4ForCausalLM
>>> model = Glm4ForCausalLM.from_pretrained("THUDM/GLM-4-9B-Chat-0414")
>>> tokenizer = AutoTokenizer.from_pretrained("THUDM/GLM-4-9B-Chat-0414")
>>> model = Glm4ForCausalLM.from_pretrained("THUDM/GLM-4-9B-0414")
>>> tokenizer = AutoTokenizer.from_pretrained("THUDM/GLM-4-9B-0414")
>>> prompt = "Hey, are you conscious? Can you talk to me?"
>>> inputs = tokenizer(prompt, return_tensors="pt")

6
src/transformers/models/glm4/modular_glm4.py
View File

@@ -31,7 +31,7 @@ from .modeling_glm4 import Glm4RMSNorm
logger = logging.get_logger(__name__)
_CHECKPOINT_FOR_DOC = "THUDM/GLM-4-9B-Chat-0414"
_CHECKPOINT_FOR_DOC = "THUDM/GLM-4-9B-0414"
class Glm4MLP(Phi3MLP):
@@ -119,8 +119,8 @@ class Glm4ForCausalLM(GlmForCausalLM):
```python
>>> from transformers import AutoTokenizer, Glm4ForCausalLM
>>> model = Glm4ForCausalLM.from_pretrained("THUDM/GLM-4-9B-Chat-0414")
>>> tokenizer = AutoTokenizer.from_pretrained("THUDM/GLM-4-9B-Chat-0414")
>>> model = Glm4ForCausalLM.from_pretrained("THUDM/GLM-4-9B-0414")
>>> tokenizer = AutoTokenizer.from_pretrained("THUDM/GLM-4-9B-0414")
>>> prompt = "Hey, are you conscious? Can you talk to me?"
>>> inputs = tokenizer(prompt, return_tensors="pt")

105
tests/models/glm4/test_modeling_glm4.py
View File

@@ -20,6 +20,8 @@ import pytest
from transformers import AutoModelForCausalLM, AutoTokenizer, Glm4Config, is_torch_available
from transformers.testing_utils import (
Expectations,
cleanup,
require_flash_attn,
require_torch,
require_torch_large_gpu,
@@ -80,113 +82,142 @@ class Glm4ModelTest(CausalLMModelTest, unittest.TestCase):
@require_torch_large_gpu
class Glm4IntegrationTest(unittest.TestCase):
input_text = ["Hello I am doing", "Hi today"]
model_id = "THUDM/glm-4-0414-9b-chat"
revision = "refs/pr/15"
model_id = "THUDM/GLM-4-9B-0414"
def tearDown(self):
cleanup(torch_device, gc_collect=True)
def test_model_9b_fp16(self):
EXPECTED_TEXTS = [
"Hello I am doing a project on the history of the internetSolution:\n\nStep 1: Introduction\nThe history of the",
"Hi today I am going to show you how to make a simple and easy to make a DIY paper flower.",
]
EXPECTED_TEXTS = Expectations(
{
("cuda", 7): [],
("cuda", 8): [
"Hello I am doing a project on the history of the internet and I need to know what the first website was and what",
"Hi today I am going to tell you about the most common disease in the world. This disease is called diabetes",
],
}
)
EXPECTED_TEXT = EXPECTED_TEXTS.get_expectation()
model = AutoModelForCausalLM.from_pretrained(
self.model_id, low_cpu_mem_usage=True, torch_dtype=torch.float16, revision=self.revision
self.model_id, low_cpu_mem_usage=True, torch_dtype=torch.float16
).to(torch_device)
tokenizer = AutoTokenizer.from_pretrained(self.model_id, revision=self.revision)
tokenizer = AutoTokenizer.from_pretrained(self.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_text, EXPECTED_TEXT)
def test_model_9b_bf16(self):
EXPECTED_TEXTS = [
"Hello I am doing a project on the history of the internetSolution:\n\nStep 1: Introduction\nThe history of the",
"Hi today I am going to show you how to make a simple and easy to make a DIY paper flower.",
]
EXPECTED_TEXTS = Expectations(
{
("cuda", 7): [],
("cuda", 8): [
"Hello I am doing a project on the history of the internet and I need to know what the first website was and what",
"Hi today I am going to tell you about the most common disease in the world. This disease is called diabetes",
],
}
)
EXPECTED_TEXT = EXPECTED_TEXTS.get_expectation()
model = AutoModelForCausalLM.from_pretrained(
self.model_id, low_cpu_mem_usage=True, torch_dtype=torch.bfloat16, revision=self.revision
self.model_id, low_cpu_mem_usage=True, torch_dtype=torch.bfloat16
).to(torch_device)
tokenizer = AutoTokenizer.from_pretrained(self.model_id, revision=self.revision)
tokenizer = AutoTokenizer.from_pretrained(self.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_text, EXPECTED_TEXT)
def test_model_9b_eager(self):
EXPECTED_TEXTS = [
"Hello I am doing a project on the history of the internetSolution:\n\nStep 1: Introduction\nThe history of the",
"Hi today I am going to show you how to make a simple and easy to make a DIY paper flower.",
]
EXPECTED_TEXTS = Expectations(
{
("cuda", 7): [],
("cuda", 8): [
"Hello I am doing a project on the history of the internet and I need to know what the first website was and what",
"Hi today I am going to tell you about the most common disease in the world. This disease is called diabetes",
],
}
)
EXPECTED_TEXT = EXPECTED_TEXTS.get_expectation()
model = AutoModelForCausalLM.from_pretrained(
self.model_id,
low_cpu_mem_usage=True,
torch_dtype=torch.bfloat16,
attn_implementation="eager",
revision=self.revision,
)
model.to(torch_device)
tokenizer = AutoTokenizer.from_pretrained(self.model_id, revision=self.revision)
tokenizer = AutoTokenizer.from_pretrained(self.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_text, EXPECTED_TEXT)
@require_torch_sdpa
def test_model_9b_sdpa(self):
EXPECTED_TEXTS = [
"Hello I am doing a project on the history of the internetSolution:\n\nStep 1: Introduction\nThe history of the",
"Hi today I am going to show you how to make a simple and easy to make a DIY paper flower.",
]
EXPECTED_TEXTS = Expectations(
{
("cuda", 7): [],
("cuda", 8): [
"Hello I am doing a project on the history of the internet and I need to know what the first website was and what",
"Hi today I am going to tell you about the most common disease in the world. This disease is called diabetes",
],
}
)
EXPECTED_TEXT = EXPECTED_TEXTS.get_expectation()
model = AutoModelForCausalLM.from_pretrained(
self.model_id,
low_cpu_mem_usage=True,
torch_dtype=torch.bfloat16,
attn_implementation="sdpa",
revision=self.revision,
)
model.to(torch_device)
tokenizer = AutoTokenizer.from_pretrained(self.model_id, revision=self.revision)
tokenizer = AutoTokenizer.from_pretrained(self.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_text, EXPECTED_TEXT)
@require_flash_attn
@pytest.mark.flash_attn_test
def test_model_9b_flash_attn(self):
EXPECTED_TEXTS = [
"Hello I am doing a project on the history of the internetSolution:\n\nStep 1: Introduction\nThe history of the",
"Hi today I am going to show you how to make a simple and easy to make a DIY paper flower.",
]
EXPECTED_TEXTS = Expectations(
{
("cuda", 7): [],
("cuda", 8): [
"Hello I am doing a project on the history of the internet and I need to know what the first website was and what",
"Hi today I am going to tell you about the most common disease in the world. This disease is called diabetes",
],
}
)
EXPECTED_TEXT = EXPECTED_TEXTS.get_expectation()
model = AutoModelForCausalLM.from_pretrained(
self.model_id,
low_cpu_mem_usage=True,
torch_dtype=torch.bfloat16,
attn_implementation="flash_attention_2",
revision=self.revision,
)
model.to(torch_device)
tokenizer = AutoTokenizer.from_pretrained(self.model_id, revision=self.revision)
tokenizer = AutoTokenizer.from_pretrained(self.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_text, EXPECTED_TEXT)