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

enable utils test cases on XPU (#38005)

Browse Source 
* enable utils test cases on XPU

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

* fix style

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

* Update tests/utils/test_skip_decorators.py

Co-authored-by: Ilyas Moutawwakil <57442720+IlyasMoutawwakil@users.noreply.github.com>

* fix comment

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

---------

Signed-off-by: Yao Matrix <matrix.yao@intel.com>
Co-authored-by: Yih-Dar <2521628+ydshieh@users.noreply.github.com>
Co-authored-by: Ilyas Moutawwakil <57442720+IlyasMoutawwakil@users.noreply.github.com>
This commit is contained in:
Yao Matrix
2025-05-09 14:45:01 +08:00
committed by GitHub
parent 1dfad4beb2
commit a72cb31434
4 changed files with 37 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
tests/utils/test_cache_utils.py
View File

@@ -28,6 +28,7 @@ from transformers.testing_utils import (
require_torch,
require_torch_accelerator,
require_torch_gpu,
require_torch_multi_accelerator,
require_torch_multi_gpu,
slow,
torch_device,
@@ -355,7 +356,7 @@ class CacheHardIntegrationTest(unittest.TestCase):
self.assertIsInstance(gen_out.past_key_values, DynamicCache) # sanity check
@parameterized.expand([("eager"), ("sdpa")])
@require_torch_gpu
@require_torch_accelerator
@slow
def test_static_cache_greedy_decoding_pad_left(self, attn_implementation):
"""Tests that different cache implementations work well with eager and SDPA inference"""
@@ -436,7 +437,7 @@ class CacheHardIntegrationTest(unittest.TestCase):
offloaded_peak_memory = torch_accelerator_module.max_memory_allocated(device)
self.assertTrue(offloaded_peak_memory < original_peak_memory)
@require_torch_gpu
@require_torch_accelerator
@slow
def test_cache_copy(self):
"""Tests that we can manually set a cache, copy, and reuse it for generation"""
@@ -444,14 +445,14 @@ class CacheHardIntegrationTest(unittest.TestCase):
# lazy init of cache layers
model_name = "microsoft/Phi-3-mini-4k-instruct"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map="cuda", torch_dtype=torch.bfloat16)
model = AutoModelForCausalLM.from_pretrained(model_name, device_map=torch_device, torch_dtype=torch.bfloat16)
prompt_cache = StaticCache(
config=model.config, max_batch_size=1, max_cache_len=1024, device="cuda", dtype=torch.bfloat16
config=model.config, max_batch_size=1, max_cache_len=1024, device=torch_device, dtype=torch.bfloat16
)
INITIAL_PROMPT = "You are a helpful assistant. "
inputs_initial_prompt = tokenizer(INITIAL_PROMPT, return_tensors="pt").to("cuda")
inputs_initial_prompt = tokenizer(INITIAL_PROMPT, return_tensors="pt").to(torch_device)
# This is the common prompt cached, we need to run forward without grad to be able to copy
with torch.no_grad():
prompt_cache = model(**inputs_initial_prompt, past_key_values=prompt_cache).past_key_values
@@ -459,7 +460,7 @@ class CacheHardIntegrationTest(unittest.TestCase):
prompts = ["Help me to write a blogpost about travelling.", "What is the capital of France?"]
responses = []
for prompt in prompts:
new_inputs = tokenizer(INITIAL_PROMPT + prompt, return_tensors="pt").to("cuda")
new_inputs = tokenizer(INITIAL_PROMPT + prompt, return_tensors="pt").to(torch_device)
past_key_values = copy.deepcopy(prompt_cache)
outputs = model.generate(
**new_inputs, past_key_values=past_key_values, max_new_tokens=40, disable_compile=True
@@ -474,6 +475,7 @@ class CacheHardIntegrationTest(unittest.TestCase):
"You are a helpful assistant. What is the capital of France?\n\n\n## Response:Paris is the capital "
"of France.\n\n\n\n\n\n\n<|endoftext|>",
]
self.assertEqual(responses, EXPECTED_DECODED_TEXT)
@require_torch_multi_gpu
@@ -526,11 +528,11 @@ class CacheHardIntegrationTest(unittest.TestCase):
model.generate(**inputs, max_new_tokens=2, cache_implementation="static")
self.assertNotIn("cuda", cap.err.lower())
@require_torch_multi_gpu
@require_torch_multi_accelerator
@slow
@require_read_token
def test_static_cache_multi_gpu(self):
"""Regression test for #35164: static cache with multi-gpu"""
def test_static_cache_multi_accelerator(self):
"""Regression test for #35164: static cache with multi-accelerator"""
model_id = "google/gemma-2-2b-it"
tokenizer = AutoTokenizer.from_pretrained(model_id)