[2/N][torch.compile] make compilation cfg part of vllm cfg (#10383)
Signed-off-by: youkaichao <youkaichao@gmail.com>
This commit is contained in:
@@ -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()
|
||||
|
||||
|
||||
@@ -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()
|
||||
|
||||
@@ -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
|
||||
|
||||
@@ -1,6 +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
|
||||
|
||||
@@ -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)
|
||||
|
||||
@@ -3,6 +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
|
||||
|
||||
@@ -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 = [
|
||||
|
||||
Reference in New Issue
Block a user