[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()
+1 -1
View File
@@ -3,7 +3,7 @@ from typing import Dict, List, Optional
import pytest
from vllm.compilation.levels import CompilationLevel
from vllm.config import CompilationLevel
from vllm.utils import cuda_device_count_stateless
from ..utils import compare_all_settings
+1 -1
View File
@@ -1,6 +1,6 @@
import pytest
from vllm.compilation.levels import CompilationLevel
from vllm.config import CompilationLevel
from ..utils import fork_new_process_for_each_test
from .utils import TEST_MODELS, check_full_graph_support
+1 -1
View File
@@ -3,10 +3,10 @@ import torch
from compressed_tensors.quantization import FP8_DTYPE
import vllm.envs as envs
from vllm.compilation.config import CompilationConfig
from vllm.compilation.fusion import (FusionPass, find_auto_fn,
find_auto_fn_maybe)
from vllm.compilation.reshapes import RedundantReshapesPass
from vllm.config import CompilationConfig
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
apply_fp8_linear)
+3 -1
View File
@@ -3,6 +3,7 @@ from typing import Optional
import torch
from vllm.compilation.wrapper import TorchCompileWrapperWithCustomDispatcher
from vllm.config import CompilationLevel
class MyMod(torch.nn.Module):
@@ -18,7 +19,8 @@ class MyWrapper(TorchCompileWrapperWithCustomDispatcher):
def __init__(self, model):
self.model = model
compiled_callable = torch.compile(self.forward, backend="eager")
super().__init__(compiled_callable)
super().__init__(compiled_callable,
compilation_level=CompilationLevel.DYNAMO_ONCE)
def forward(self, x: torch.Tensor, cache: Optional[torch.Tensor] = None):
# this is the function to be compiled
+1 -1
View File
@@ -4,7 +4,7 @@ import torch
from tests.quantization.utils import is_quant_method_supported
from vllm import LLM, SamplingParams
from vllm.compilation.levels import CompilationLevel
from vllm.config import CompilationLevel
from vllm.platforms import current_platform
TEST_MODELS = [
+26 -26
View File
@@ -3,11 +3,13 @@ from typing import List
import pytest
from vllm.config import CompilationConfig, VllmConfig
from vllm.model_executor.custom_op import CustomOp
from vllm.model_executor.layers.activation import (GeluAndMul,
ReLUSquaredActivation,
SiluAndMul)
from vllm.model_executor.layers.layernorm import RMSNorm
from vllm.plugins import set_current_vllm_config
# Registered subclass for test
@@ -51,42 +53,40 @@ class Relu3(ReLUSquaredActivation):
])
def test_enabled_ops(env: str, torch_level: int, ops_enabled: List[int],
default_on: bool):
os.environ["VLLM_CUSTOM_OPS"] = env
os.environ["VLLM_TORCH_COMPILE_LEVEL"] = str(torch_level)
vllm_config = VllmConfig(compilation_config=CompilationConfig(
custom_ops=env.split(",")))
with set_current_vllm_config(vllm_config):
assert CustomOp.default_on() == default_on
# Reset default_on (computed once):
CustomOp.default_on.cache_clear()
ops_enabled = [bool(x) for x in ops_enabled]
assert CustomOp.default_on() == default_on
assert RMSNorm(1024).enabled() == ops_enabled[0]
assert CustomOp.op_registry["rms_norm"].enabled() == ops_enabled[0]
ops_enabled = [bool(x) for x in ops_enabled]
assert SiluAndMul().enabled() == ops_enabled[1]
assert CustomOp.op_registry["silu_and_mul"].enabled() == ops_enabled[1]
assert RMSNorm(1024).enabled() == ops_enabled[0]
assert CustomOp.op_registry["rms_norm"].enabled() == ops_enabled[0]
assert GeluAndMul().enabled() == ops_enabled[2]
assert CustomOp.op_registry["gelu_and_mul"].enabled() == ops_enabled[2]
assert SiluAndMul().enabled() == ops_enabled[1]
assert CustomOp.op_registry["silu_and_mul"].enabled() == ops_enabled[1]
# If registered, subclasses should follow their own name
assert Relu3().enabled() == ops_enabled[3]
assert CustomOp.op_registry["relu3"].enabled() == ops_enabled[3]
assert GeluAndMul().enabled() == ops_enabled[2]
assert CustomOp.op_registry["gelu_and_mul"].enabled() == ops_enabled[2]
# Unregistered subclass
class SiluAndMul2(SiluAndMul):
pass
# If registered, subclasses should follow their own name
assert Relu3().enabled() == ops_enabled[3]
assert CustomOp.op_registry["relu3"].enabled() == ops_enabled[3]
# Unregistered subclass
class SiluAndMul2(SiluAndMul):
pass
# Subclasses should not require registration
assert SiluAndMul2().enabled() == SiluAndMul().enabled()
# Subclasses should not require registration
assert SiluAndMul2().enabled() == SiluAndMul().enabled()
@pytest.mark.parametrize(
"env", ["all,none", "all,+rms_norm,all", "+rms_norm,-rms_norm"])
def test_enabled_ops_invalid(env: str):
os.environ["VLLM_CUSTOM_OPS"] = env
CustomOp.default_on.cache_clear()
with pytest.raises(AssertionError):
RMSNorm(1024).enabled()
with pytest.raises(Exception): # noqa
vllm_config = VllmConfig(compilation_config=CompilationConfig(
custom_ops=env.split(",")))
with set_current_vllm_config(vllm_config):
RMSNorm(1024).enabled()
+1 -1
View File
@@ -5,7 +5,7 @@ import tempfile
import depyf
from vllm.compilation.levels import CompilationLevel
from vllm.config import CompilationLevel
# disable custom dispatcher, let Dynamo takes over
# all the control
+1 -1
View File
@@ -1,6 +1,6 @@
import os
from vllm.compilation.levels import CompilationLevel
from vllm.config import CompilationLevel
from ..utils import compare_two_settings