[2/N][torch.compile] make compilation cfg part of vllm cfg (#10383)

Signed-off-by: youkaichao <youkaichao@gmail.com>
This commit is contained in:
youkaichao
2024-11-16 18:02:14 -08:00
committed by GitHub
parent 661a34fd4f
commit 4fd9375028
27 changed files with 359 additions and 283 deletions
+5 -3
View File
@@ -11,8 +11,8 @@ from torch.library import Library
from vllm.compilation.compile_context import set_compile_context
from vllm.compilation.counter import compilation_counter
from vllm.compilation.decorators import support_torch_compile
from vllm.compilation.levels import CompilationLevel
from vllm.config import VllmConfig
from vllm.config import CompilationLevel, VllmConfig
from vllm.plugins import set_current_vllm_config
from vllm.utils import direct_register_custom_op
global_counter = 0
@@ -82,7 +82,9 @@ def test_simple_piecewise_compile():
os.environ["VLLM_TORCH_COMPILE_CONFIG"] = config
os.environ["VLLM_TORCH_COMPILE_LEVEL"] = str(CompilationLevel.PIECEWISE)
model = SillyModel(vllm_config=VllmConfig(), prefix='')
vllm_config = VllmConfig()
with set_current_vllm_config(vllm_config):
model = SillyModel(vllm_config=vllm_config, prefix='')
inputs = torch.randn(100).cuda()
+12 -10
View File
@@ -15,12 +15,10 @@ from torch import nn
from torch.library import Library
from vllm.compilation.compile_context import set_compile_context
from vllm.compilation.config import CompilationConfig
from vllm.compilation.counter import compilation_counter
from vllm.compilation.decorators import support_torch_compile
from vllm.compilation.levels import CompilationLevel
from vllm.config import VllmConfig
from vllm.plugins import set_compilation_config
from vllm.config import CompilationConfig, CompilationLevel, VllmConfig
from vllm.plugins import set_compilation_config, set_current_vllm_config
from vllm.utils import direct_register_custom_op
# create a library to hold the custom op
@@ -272,9 +270,11 @@ def run_model(llama_config,
CompilationLevel.NO_COMPILATION)
set_compilation_config(None)
model = LlamaModel(config=llama_config,
vllm_config=VllmConfig(),
prefix="").eval().cuda()
vllm_config = VllmConfig()
with set_current_vllm_config(vllm_config):
model = LlamaModel(config=llama_config,
vllm_config=vllm_config,
prefix="").eval().cuda()
B = 16 # max batch size
input_ids = torch.randint(0, llama_config.vocab_size, (B, )).cuda()
@@ -395,9 +395,11 @@ def benchmark():
else:
set_compilation_config(None)
model = LlamaModel(config=llama_config,
vllm_config=VllmConfig(),
prefix="").eval().cuda().to(torch.bfloat16)
vllm_config = VllmConfig()
with set_current_vllm_config(vllm_config):
model = LlamaModel(config=llama_config,
vllm_config=vllm_config,
prefix="").eval().cuda().to(torch.bfloat16)
B = 256 # max batch size
input_ids = torch.randint(0, llama_config.vocab_size, (B, )).cuda()