[Frontend][torch.compile] CompilationConfig Overhaul (#20283): name change compilation level to compilation mode, deprecation compilation level (#26355)
Signed-off-by: morrison-turnansky <mturnans@redhat.com> Signed-off-by: Morrison Turnansky <mturnans@redhat.com> Co-authored-by: Luka Govedič <ProExpertProg@users.noreply.github.com>
This commit is contained in:
co-authored by
Luka Govedič
parent
e66d787bce
commit
96b9aa5aa0
@@ -56,7 +56,7 @@ def make_compiler(compilation_config: CompilationConfig) -> CompilerInterface:
|
||||
return InductorAdaptor()
|
||||
else:
|
||||
assert compilation_config.backend == "eager", (
|
||||
"Custom backends not supported with CompilationLevel.PIECEWISE"
|
||||
"Custom backends not supported with CompilationMode.VLLM_COMPILE"
|
||||
)
|
||||
|
||||
logger.debug("Using EagerAdaptor")
|
||||
@@ -481,7 +481,7 @@ def set_model_tag(tag: str):
|
||||
|
||||
class VllmBackend:
|
||||
"""The compilation backend for `torch.compile` with vLLM.
|
||||
It is used for compilation level of `CompilationLevel.PIECEWISE`,
|
||||
It is used for compilation mode of `CompilationMode.VLLM_COMPILE`,
|
||||
where we customize the compilation.
|
||||
|
||||
The major work of this backend is to split the graph into
|
||||
|
||||
@@ -575,7 +575,7 @@ class InductorAdaptor(CompilerInterface):
|
||||
|
||||
Because it is re-entrant, we always set it (even if entering via Dynamo
|
||||
and the context was already entered). We might want to revisit if it
|
||||
should be set at a different level of compilation.
|
||||
should be set at a different mode of compilation.
|
||||
|
||||
This is likely a bug in PyTorch: public APIs should not rely on
|
||||
manually setting up internal contexts. But we also rely on non-public
|
||||
|
||||
@@ -27,8 +27,8 @@ class CompilationCounter:
|
||||
num_cache_entries_updated: int = 0
|
||||
# The number of standalone_compile compiled artifacts saved
|
||||
num_compiled_artifacts_saved: int = 0
|
||||
# Number of times a model was loaded with CompilationLevel.DYNAMO_AS_IS
|
||||
dynamo_as_is_count: int = 0
|
||||
# Number of times a model was loaded with CompilationMode.STOCK_TORCH_COMPILE
|
||||
stock_torch_compile_count: int = 0
|
||||
|
||||
def clone(self) -> "CompilationCounter":
|
||||
return copy.deepcopy(self)
|
||||
|
||||
@@ -18,7 +18,7 @@ from torch._dynamo.symbolic_convert import InliningInstructionTranslator
|
||||
import vllm.envs as envs
|
||||
from vllm.compilation.counter import compilation_counter
|
||||
from vllm.compilation.wrapper import TorchCompileWrapperWithCustomDispatcher
|
||||
from vllm.config import CompilationLevel, VllmConfig, set_current_vllm_config
|
||||
from vllm.config import CompilationMode, VllmConfig, set_current_vllm_config
|
||||
from vllm.logger import init_logger
|
||||
from vllm.sequence import IntermediateTensors
|
||||
from vllm.utils import resolve_obj_by_qualname, supports_dynamo
|
||||
@@ -233,11 +233,11 @@ def _support_torch_compile(
|
||||
old_init(self, vllm_config=vllm_config, prefix=prefix, **kwargs)
|
||||
self.vllm_config = vllm_config
|
||||
enable_compile = enable_if is None or enable_if(vllm_config)
|
||||
# for CompilationLevel.DYNAMO_AS_IS , the upper level model runner
|
||||
# for CompilationMode.STOCK_TORCH_COMPILE , the upper level model runner
|
||||
# will handle the compilation, so we don't need to do anything here.
|
||||
self.do_not_compile = (
|
||||
vllm_config.compilation_config.level
|
||||
in [CompilationLevel.NO_COMPILATION, CompilationLevel.DYNAMO_AS_IS]
|
||||
vllm_config.compilation_config.mode
|
||||
in [CompilationMode.NONE, CompilationMode.STOCK_TORCH_COMPILE]
|
||||
or not supports_dynamo()
|
||||
or _should_ignore_torch_compile(self.__class__)
|
||||
or not enable_compile
|
||||
@@ -247,7 +247,7 @@ def _support_torch_compile(
|
||||
|
||||
compilation_counter.num_models_seen += 1
|
||||
TorchCompileWrapperWithCustomDispatcher.__init__(
|
||||
self, compilation_level=vllm_config.compilation_config.level
|
||||
self, compilation_mode=vllm_config.compilation_config.mode
|
||||
)
|
||||
|
||||
cls.__init__ = __init__
|
||||
|
||||
@@ -3,7 +3,7 @@
|
||||
|
||||
import time
|
||||
|
||||
from vllm.config import CompilationConfig, CompilationLevel, VllmConfig
|
||||
from vllm.config import CompilationConfig, CompilationMode, VllmConfig
|
||||
from vllm.logger import init_logger
|
||||
|
||||
logger = init_logger(__name__)
|
||||
@@ -18,7 +18,7 @@ def start_monitoring_torch_compile(vllm_config: VllmConfig):
|
||||
|
||||
compilation_config: CompilationConfig = vllm_config.compilation_config
|
||||
path = vllm_config.compile_debug_dump_path()
|
||||
if compilation_config.level == CompilationLevel.PIECEWISE and path:
|
||||
if compilation_config.mode == CompilationMode.VLLM_COMPILE and path:
|
||||
import depyf
|
||||
|
||||
path.mkdir(parents=True, exist_ok=True)
|
||||
@@ -29,7 +29,7 @@ def start_monitoring_torch_compile(vllm_config: VllmConfig):
|
||||
|
||||
def end_monitoring_torch_compile(vllm_config: VllmConfig):
|
||||
compilation_config: CompilationConfig = vllm_config.compilation_config
|
||||
if compilation_config.level == CompilationLevel.PIECEWISE:
|
||||
if compilation_config.mode == CompilationMode.VLLM_COMPILE:
|
||||
logger.info(
|
||||
"torch.compile takes %.2f s in total", compilation_config.compilation_time
|
||||
)
|
||||
|
||||
@@ -11,7 +11,7 @@ from types import CodeType
|
||||
import torch
|
||||
|
||||
import vllm.envs as envs
|
||||
from vllm.config import CompilationLevel, CUDAGraphMode, get_current_vllm_config
|
||||
from vllm.config import CompilationMode, CUDAGraphMode, get_current_vllm_config
|
||||
from vllm.logger import init_logger
|
||||
|
||||
logger = init_logger(__name__)
|
||||
@@ -31,7 +31,7 @@ class TorchCompileWrapperWithCustomDispatcher:
|
||||
"""
|
||||
|
||||
def __init__(
|
||||
self, compiled_callable: Callable | None = None, compilation_level: int = 0
|
||||
self, compiled_callable: Callable | None = None, compilation_mode: int = 0
|
||||
):
|
||||
vllm_config = get_current_vllm_config()
|
||||
self.vllm_config = vllm_config
|
||||
@@ -72,7 +72,7 @@ class TorchCompileWrapperWithCustomDispatcher:
|
||||
# subclasses can use this to switch between the custom dispatcher
|
||||
# and the default Dynamo guard mechanism.
|
||||
self.use_custom_dispatcher: bool = (
|
||||
compilation_level >= CompilationLevel.DYNAMO_ONCE
|
||||
compilation_mode >= CompilationMode.DYNAMO_TRACE_ONCE
|
||||
)
|
||||
|
||||
def aot_compile(self, *args, **kwargs):
|
||||
@@ -85,7 +85,7 @@ class TorchCompileWrapperWithCustomDispatcher:
|
||||
return self.compiled_callable.aot_compile((args, kwargs))
|
||||
|
||||
def __call__(self, *args, **kwargs):
|
||||
"""Implement the dispatch logic here, beyond the torch.compile level.
|
||||
"""Implement the dispatch logic here, beyond the torch.compile mode.
|
||||
NOTE: this function can have additional arguments beyond the forward
|
||||
method, for directly dispatching to the compiled code.
|
||||
"""
|
||||
|
||||
Reference in New Issue
Block a user