[CI/BUILD] enable intel queue for longer CPU tests (#4113)

This commit is contained in:
Yuan
2024-06-03 10:39:50 -07:00
committed by GitHub
parent cbb2f59cc8
commit cafb8e06c5
11 changed files with 138 additions and 90 deletions
+23 -13
View File
@@ -18,6 +18,7 @@ from vllm.logger import init_logger
from vllm.multimodal import MultiModalData
from vllm.multimodal.image import ImageFeatureData, ImagePixelData
from vllm.sequence import SampleLogprobs
from vllm.utils import is_cpu
logger = init_logger(__name__)
@@ -58,7 +59,8 @@ def cleanup():
with contextlib.suppress(AssertionError):
torch.distributed.destroy_process_group()
gc.collect()
torch.cuda.empty_cache()
if not is_cpu():
torch.cuda.empty_cache()
@pytest.fixture()
@@ -151,6 +153,12 @@ _EMBEDDING_MODELS = [
class HfRunner:
def wrap_device(self, input: any):
if not is_cpu():
return input.to("cuda")
else:
return input.to("cpu")
def __init__(
self,
model_name: str,
@@ -164,16 +172,18 @@ class HfRunner:
if model_name in _EMBEDDING_MODELS:
# Lazy init required for AMD CI
from sentence_transformers import SentenceTransformer
self.model = SentenceTransformer(
model_name,
device="cpu",
).to(dtype=torch_dtype).cuda()
self.model = self.wrap_device(
SentenceTransformer(
model_name,
device="cpu",
).to(dtype=torch_dtype))
else:
self.model = AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch_dtype,
trust_remote_code=True,
).cuda()
self.model = self.wrap_device(
AutoModelForCausalLM.from_pretrained(
model_name,
torch_dtype=torch_dtype,
trust_remote_code=True,
))
self.tokenizer = AutoTokenizer.from_pretrained(
model_name,
@@ -214,7 +224,7 @@ class HfRunner:
inputs = self.processor(**processor_kwargs)
output_ids = self.model.generate(
**inputs.to("cuda"),
**self.wrap_device(inputs),
use_cache=True,
**kwargs,
)
@@ -271,7 +281,7 @@ class HfRunner:
for prompt in prompts:
input_ids = self.tokenizer(prompt, return_tensors="pt").input_ids
output = self.model.generate(
input_ids.cuda(),
self.wrap_device(input_ids),
use_cache=True,
do_sample=False,
max_new_tokens=max_tokens,
@@ -306,7 +316,7 @@ class HfRunner:
for prompt in prompts:
input_ids = self.tokenizer(prompt, return_tensors="pt").input_ids
output = self.model.generate(
input_ids.cuda(),
self.wrap_device(input_ids),
use_cache=True,
do_sample=False,
max_new_tokens=max_tokens,
+7 -4
View File
@@ -8,10 +8,13 @@ import torch
from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
capability = torch.cuda.get_device_capability()
capability = capability[0] * 10 + capability[1]
aqlm_not_supported = (capability <
QUANTIZATION_METHODS["aqlm"].get_min_capability())
aqlm_not_supported = True
if torch.cuda.is_available():
capability = torch.cuda.get_device_capability()
capability = capability[0] * 10 + capability[1]
aqlm_not_supported = (capability <
QUANTIZATION_METHODS["aqlm"].get_min_capability())
# In this test we hardcode prompts and generations for the model so we don't
# need to require the AQLM package as a dependency
+8 -2
View File
@@ -5,6 +5,7 @@ This tests bigger models and use half precision.
Run `pytest tests/models/test_big_models.py`.
"""
import pytest
import torch
MODELS = [
"meta-llama/Llama-2-7b-hf",
@@ -16,9 +17,14 @@ MODELS = [
# "Qwen/Qwen1.5-0.5B" # Broken,
]
#TODO: remove this after CPU float16 support ready
target_dtype = "float"
if torch.cuda.is_available():
target_dtype = "half"
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("dtype", [target_dtype])
@pytest.mark.parametrize("max_tokens", [32])
def test_models(
hf_runner,
@@ -46,7 +52,7 @@ def test_models(
@pytest.mark.parametrize("model", MODELS)
@pytest.mark.parametrize("dtype", ["half"])
@pytest.mark.parametrize("dtype", [target_dtype])
def test_model_print(
vllm_runner,
model: str,
+7 -4
View File
@@ -67,10 +67,13 @@ EXPECTED_STRS_MAP = {
},
}
capability = torch.cuda.get_device_capability()
capability = capability[0] * 10 + capability[1]
fp8_not_supported = (capability <
QUANTIZATION_METHODS["fp8"].get_min_capability())
fp8_not_supported = True
if torch.cuda.is_available():
capability = torch.cuda.get_device_capability()
capability = capability[0] * 10 + capability[1]
fp8_not_supported = (capability <
QUANTIZATION_METHODS["fp8"].get_min_capability())
@pytest.mark.skipif(fp8_not_supported,
+7 -4
View File
@@ -22,10 +22,13 @@ os.environ["TOKENIZERS_PARALLELISM"] = "true"
MAX_MODEL_LEN = 1024
capability = torch.cuda.get_device_capability()
capability = capability[0] * 10 + capability[1]
gptq_marlin_not_supported = (
capability < QUANTIZATION_METHODS["gptq_marlin"].get_min_capability())
gptq_marlin_not_supported = True
if torch.cuda.is_available():
capability = torch.cuda.get_device_capability()
capability = capability[0] * 10 + capability[1]
gptq_marlin_not_supported = (
capability < QUANTIZATION_METHODS["gptq_marlin"].get_min_capability())
MODELS = [
# act_order==False, group_size=channelwise
+7 -4
View File
@@ -14,10 +14,13 @@ import torch
from tests.models.utils import check_logprobs_close
from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
capability = torch.cuda.get_device_capability()
capability = capability[0] * 10 + capability[1]
marlin_not_supported = (capability <
QUANTIZATION_METHODS["marlin"].get_min_capability())
marlin_not_supported = True
if torch.cuda.is_available():
capability = torch.cuda.get_device_capability()
capability = capability[0] * 10 + capability[1]
marlin_not_supported = (
capability < QUANTIZATION_METHODS["marlin"].get_min_capability())
@dataclass
+7 -4
View File
@@ -19,10 +19,13 @@ from vllm.model_executor.layers.quantization import QUANTIZATION_METHODS
from .utils import check_logprobs_close
capability = torch.cuda.get_device_capability()
capability = capability[0] * 10 + capability[1]
marlin_not_supported = (capability <
QUANTIZATION_METHODS["marlin"].get_min_capability())
marlin_not_supported = True
if torch.cuda.is_available():
capability = torch.cuda.get_device_capability()
capability = capability[0] * 10 + capability[1]
marlin_not_supported = (
capability < QUANTIZATION_METHODS["marlin"].get_min_capability())
@dataclass