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

enable finegrained_fp8 and granite_speech cases on XPU (#38036)

Browse Source 
* enable finegrained_fp8 cases on XPU

Signed-off-by: Yao Matrix <matrix.yao@intel.com>

* fix style

Signed-off-by: Yao Matrix <matrix.yao@intel.com>

* change back to auto

Signed-off-by: Yao Matrix <matrix.yao@intel.com>

* rename per comments

Signed-off-by: Matrix Yao <matrix.yao@intel.com>

---------

Signed-off-by: Yao Matrix <matrix.yao@intel.com>
Signed-off-by: Matrix Yao <matrix.yao@intel.com>
Co-authored-by: Marc Sun <57196510+SunMarc@users.noreply.github.com>
This commit is contained in:
Yao Matrix
2025-05-14 16:58:40 +08:00
committed by GitHub
parent b311a3f506
commit 9b5ce556aa
4 changed files with 42 additions and 34 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
tests/models/granite_speech/test_processor_granite_speech.py
View File

@@ -23,8 +23,9 @@ from parameterized import parameterized
from transformers import AutoTokenizer, GPT2TokenizerFast
from transformers.testing_utils import (
require_torch,
require_torch_gpu,
require_torch_accelerator,
require_torchaudio,
torch_device,
)
from transformers.utils import is_torchaudio_available
@@ -195,7 +196,7 @@ class GraniteSpeechProcessorTest(unittest.TestCase):
assert num_calculated_features == [90, 171]
assert sum(num_expected_features) == num_audio_tokens
@require_torch_gpu
@require_torch_accelerator
def test_device_override(self):
"""Ensure that we regardless of the processing device, the tensors
produced are on the CPU.
@@ -214,7 +215,7 @@ class GraniteSpeechProcessorTest(unittest.TestCase):
text=f"{processor.audio_token} Can you transcribe this audio?",
audio=wav,
return_tensors="pt",
device="cuda",
device=torch_device,
)
assert inputs["input_features"].device.type == "cpu"

34
tests/quantization/finegrained_fp8/test_fp8.py
View File

@@ -18,11 +18,13 @@ import unittest
from transformers import AutoConfig, AutoModelForCausalLM, AutoTokenizer, FineGrainedFP8Config, OPTForCausalLM
from transformers.testing_utils import (
backend_empty_cache,
require_accelerate,
require_read_token,
require_torch_gpu,
require_torch_multi_gpu,
require_torch_accelerator,
require_torch_multi_accelerator,
slow,
torch_device,
)
from transformers.utils import is_accelerate_available, is_torch_available
@@ -34,7 +36,7 @@ if is_accelerate_available():
from accelerate import init_empty_weights
@require_torch_gpu
@require_torch_accelerator
class FineGrainedFP8ConfigTest(unittest.TestCase):
def test_to_dict(self):
"""
@@ -60,13 +62,13 @@ class FineGrainedFP8ConfigTest(unittest.TestCase):
@slow
@require_accelerate
@require_read_token
@require_torch_gpu
@require_torch_accelerator
class FP8QuantizerTest(unittest.TestCase):
model_name = "meta-llama/Llama-3.2-1B"
input_text = "Once upon a time"
max_new_tokens = 10
EXPECTED_OUTPUT = "Once upon a time, there was a man who was very rich."
device_map = "cuda"
device_map = torch_device
offload_device_map = {
"model.embed_tokens": 0,
"model.layers.0": 0,
@@ -103,7 +105,7 @@ class FP8QuantizerTest(unittest.TestCase):
def tearDown(self):
gc.collect()
torch.cuda.empty_cache()
backend_empty_cache(torch_device)
gc.collect()
def test_quantized_model_conversion(self):
@@ -151,7 +153,8 @@ class FP8QuantizerTest(unittest.TestCase):
input_ids = self.tokenizer(self.input_text, return_tensors="pt").to(self.device_map)
output = self.quantized_model.generate(**input_ids, max_new_tokens=self.max_new_tokens, do_sample=False)
self.assertEqual(self.tokenizer.decode(output[0], skip_special_tokens=True), self.EXPECTED_OUTPUT)
output_tokens = self.tokenizer.decode(output[0], skip_special_tokens=True)
self.assertEqual(output_tokens, self.EXPECTED_OUTPUT)
def test_save_pretrained(self):
"""
@@ -188,11 +191,12 @@ class FP8QuantizerTest(unittest.TestCase):
)
self.assertEqual(quantized_model.config.quantization_config.weight_block_size, (32, 32))
@require_torch_multi_gpu
def test_quantized_model_multi_gpu(self):
@require_torch_multi_accelerator
def test_quantized_model_multi_accelerator(self):
"""
Simple test that checks if the quantized model is working properly with multiple GPUs
set CUDA_VISIBLE_DEVICES=0,1 if you have more than 2 GPUs
Simple test that checks if the quantized model is working properly with multiple accelerators
set CUDA_VISIBLE_DEVICES=0,1 if you have more than 2 GPUs; or set ZE_AFFINITY_MASK=0,1 if you
have more than 2 XPUs.
"""
input_ids = self.tokenizer(self.input_text, return_tensors="pt").to(self.device_map)
quantization_config = FineGrainedFP8Config()
@@ -204,8 +208,8 @@ class FP8QuantizerTest(unittest.TestCase):
output = quantized_model.generate(**input_ids, max_new_tokens=self.max_new_tokens, do_sample=False)
self.assertEqual(self.tokenizer.decode(output[0], skip_special_tokens=True), self.EXPECTED_OUTPUT)
@require_torch_multi_gpu
def test_save_pretrained_multi_gpu(self):
@require_torch_multi_accelerator
def test_save_pretrained_multi_accelerators(self):
"""
Simple test that checks if the quantized model is working properly after being saved and loaded
"""
@@ -245,9 +249,9 @@ class FP8QuantizerTest(unittest.TestCase):
self.assertEqual(self.tokenizer.decode(output[0], skip_special_tokens=True), self.EXPECTED_OUTPUT)
@require_torch_gpu
@require_torch_accelerator
class FP8LinearTest(unittest.TestCase):
device = "cuda"
device = torch_device
@unittest.skipIf(
torch.cuda.is_available() and torch.cuda.get_device_capability()[0] < 9,