[torch.compile] Stop lazily compiling (#35472)

Signed-off-by: Richard Zou <zou3519@gmail.com>
This commit is contained in:
Richard Zou
2026-03-04 12:13:17 -08:00
committed by GitHub
parent 138d891d7f
commit 5569f5218d
7 changed files with 177 additions and 150 deletions
+126 -90
View File
@@ -5,7 +5,6 @@ import dataclasses
import io
import json
import pickle
import time
from collections.abc import Callable
from pickle import Pickler
from typing import Any
@@ -16,7 +15,6 @@ from torch._inductor.runtime.triton_heuristics import CachingAutotuner
from torch._logging._internal import trace_structured
from vllm.compilation.backends import VllmBackend
from vllm.compilation.monitor import end_monitoring_torch_compile
from vllm.config import VllmConfig
from vllm.config.utils import Range
from vllm.logger import init_logger
@@ -24,6 +22,55 @@ from vllm.logger import init_logger
logger = init_logger(__name__)
def get_fake_args_from_graph(graph: fx.GraphModule) -> list[Any]:
"""Get fake args directly from graph placeholder nodes."""
fake_args = []
for node in graph.graph.nodes:
if node.op == "placeholder":
fake_args.append(node.meta["example_value"])
else:
break
return fake_args
def create_concrete_args(graph: fx.GraphModule, size: int) -> list[Any]:
"""Create example inputs with symbolic dims replaced by a concrete size.
Used for single-size eager compilation where we need concrete-shaped
inputs but don't have real runtime tensors yet.
"""
from torch._prims_common import compute_required_storage_length
from torch.fx.experimental.symbolic_shapes import is_symbolic
def concretize(sym_val: Any) -> int:
"""Replace all symbolic variables in a SymInt expression with size."""
if not is_symbolic(sym_val):
return int(sym_val)
expr = sym_val.node.expr
return int(expr.subs({s: size for s in expr.free_symbols}))
args: list[Any] = []
for node in graph.graph.nodes:
if node.op != "placeholder":
break
val = node.meta["example_value"]
if isinstance(val, torch.SymInt):
args.append(concretize(val))
elif isinstance(val, torch.Tensor):
new_shape = tuple(concretize(d) for d in val.shape)
new_strides = tuple(concretize(s) for s in val.stride())
new_storage_offset = concretize(val.storage_offset())
needed_size = compute_required_storage_length(
new_shape, new_strides, new_storage_offset
)
t = torch.empty(needed_size, dtype=val.dtype, device=val.device)
t = t.as_strided(new_shape, new_strides, new_storage_offset)
args.append(t)
else:
args.append(val)
return args
@dataclasses.dataclass
class RangeEntry:
compile_range: Range
@@ -109,10 +156,6 @@ class PiecewiseBackend:
# the entries for ranges that we need to either
self.range_entries: dict[Range, RangeEntry] = {}
# to_be_compiled_ranges tracks the remaining ranges to compile,
# and updates during the compilation process, so we need to copy it
self.to_be_compiled_ranges: set[Range] = set(self.compile_ranges)
# We only keep compilation management inside this class directly.
if self.compile_sizes is not None:
for size in self.compile_sizes:
@@ -129,7 +172,6 @@ class PiecewiseBackend:
self.range_entries[range] = RangeEntry(
compile_range=range,
)
self.to_be_compiled_ranges.add(range)
for range in self.compile_ranges:
self.range_entries[range] = RangeEntry(
@@ -139,12 +181,10 @@ class PiecewiseBackend:
# Track whether we've logged the graph for this subgraph (only log once)
self._graph_logged = False
# get the on_compilation_complete callback from context...
# PiecewiseBackend is created during the first call,
# which is when the context is set (see compilation/decorators.py)
from vllm.compilation.backends import _on_compilation_complete_callback
self.on_compilation_complete = _on_compilation_complete_callback.get()
if self.graph is not None:
self.compile_all_ranges()
else:
self.load_all_ranges()
def get_compiled_graph_wrapper(
self, compiled_graph: Callable[..., Any]
@@ -161,25 +201,6 @@ class PiecewiseBackend:
return compiled_graph_wrapper
def check_for_ending_compilation(self) -> None:
if self.is_last_graph and not self.to_be_compiled_ranges:
# no specific sizes to compile
# save the hash of the inductor graph for the next run
time_before_saving = time.perf_counter()
self.vllm_backend.compiler_manager.save_to_file()
elapsed = time.perf_counter() - time_before_saving
if elapsed > 1:
logger.info_once(
"Saved compiler manager cache in %.2f seconds.",
elapsed,
scope="local",
)
end_monitoring_torch_compile(self.vllm_config)
# Call the completion callback (e.g., to save AOT compiled function)
if self.on_compilation_complete is not None:
self.on_compilation_complete()
def to_bytes(self) -> dict[str, bytes]:
class StandaloneCompiledArtifactsPickler(Pickler):
def reducer_override(self, obj: object) -> Any:
@@ -216,27 +237,54 @@ class PiecewiseBackend:
return out
def _fakify_args(self, args: tuple[Any, ...]) -> list[Any]:
# We need to pass fake example_inputs, otherwise torch.compile
# will fakify the example_inputs potentially causing some non dynamic
# dimension to be be duck shaped to other existing shapes that have hints
# matching their values.
# This is problem because it can lead to unintended specializations!
# if the new wrongly dynamic dim is specialized
# it will force specializing the whole shape
# torch.compile probably should not accept
# non fake tensors as example inputs!
# See issue https://github.com/vllm-project/vllm/issues/27899
fake_example_inputs = []
assert self.graph is not None
for node in self.graph.graph.nodes:
# All place holders come first
if node.op == "placeholder":
fake_example_inputs.append(node.meta["example_value"])
def compile_all_ranges(self) -> None:
"""Compile all range entries for this piecewise subgraph up front."""
assert self.graph is not None, (
"Cannot compile without a graph. "
"When loading from cache/AOT artifacts, "
"compile_all_ranges should not be called."
)
for range_entry in self.range_entries.values():
if range_entry.compiled:
continue
self._log_compile_start(range_entry.compile_range)
if range_entry.compile_range.is_single_size():
args_list = create_concrete_args(
self.graph, range_entry.compile_range.start
)
else:
break
assert len(fake_example_inputs) == len(args)
return fake_example_inputs
args_list = get_fake_args_from_graph(self.graph)
# TODO(https://github.com/vllm-project/vllm/issues/35766)
# Can we remove strict_autograd_cache and
# force_non_lazy_backward_lowering overrides?
# I added them explicitly because this is what they are
# set to before the refactor
# (https://github.com/vllm-project/vllm/pull/35472).
# They affect the aotautograd cache key computation
# but they shouldn't have any effect on the actual
# compilation.
config_patches = dict(
bundled_autograd_cache=True,
strict_autograd_cache=False,
)
if hasattr(torch._functorch.config, "force_non_lazy_backward_lowering"):
config_patches["force_non_lazy_backward_lowering"] = False
with torch._functorch.config.patch(**config_patches):
range_entry.runnable = self.vllm_backend.compiler_manager.compile(
self.graph,
args_list,
self.vllm_backend.inductor_config,
self.compilation_config,
compile_range=range_entry.compile_range,
graph_index=self.piecewise_compile_index,
num_graphs=self.total_piecewise_compiles,
)
range_entry.compiled = True
def _log_compile_start(self, compile_range: Range):
"""Log compilation event for TORCH_TRACE/tlparse."""
@@ -277,44 +325,29 @@ class PiecewiseBackend:
payload_fn=lambda: self.graph.print_readable(print_output=False),
)
def _maybe_compile_for_range_entry(
self, range_entry: RangeEntry, args: tuple[Any, ...]
) -> Any:
if not range_entry.compiled:
if self.compiled_runnables is not None:
range_entry.runnable = self.get_compiled_graph_wrapper(
self.compiled_runnables[str(range_entry.compile_range)]
)
else:
self._log_compile_start(range_entry.compile_range)
# args are real arguments
# fakify for range, real args for concrete size.
# For concrete size, we clear the shape env in
# compiler_manager.compile() so no need to fakify.
args_list = (
self._fakify_args(args)
if not range_entry.compile_range.is_single_size()
else list(args)
)
with (
torch._functorch.config.patch("bundled_autograd_cache", True),
):
range_entry.runnable = self.vllm_backend.compiler_manager.compile(
self.graph,
args_list,
self.vllm_backend.inductor_config,
self.compilation_config,
compile_range=range_entry.compile_range,
graph_index=self.piecewise_compile_index,
num_graphs=self.total_piecewise_compiles,
)
def load_all_ranges(self) -> None:
"""Load all pre-compiled runnables for this piecewise subgraph.
Called during warm start to wrap all cached compiled_runnables
into range_entry.runnable up front, analogous to compile_all_ranges()
for the cold start path.
"""
assert self.compiled_runnables is not None, (
"load_all_ranges should only be called when compiled_runnables "
"is set (warm start / cache loading path)."
)
for range_entry in self.range_entries.values():
if range_entry.compiled:
continue
key = str(range_entry.compile_range)
assert key in self.compiled_runnables, (
f"Missing compiled runnable for range {range_entry.compile_range}. "
f"Available keys: {list(self.compiled_runnables.keys())}"
)
range_entry.runnable = self.get_compiled_graph_wrapper(
self.compiled_runnables[key]
)
range_entry.compiled = True
self.to_be_compiled_ranges.remove(range_entry.compile_range)
self.check_for_ending_compilation()
def _find_range_for_shape(self, runtime_shape: int) -> RangeEntry | None:
# First we try to find the range entry for the concrete compile size
@@ -338,6 +371,9 @@ class PiecewiseBackend:
assert range_entry is not None, (
f"Shape: {runtime_shape} out of considered ranges: {self.compile_ranges}"
)
self._maybe_compile_for_range_entry(range_entry, args)
assert range_entry.compiled, (
"All ranges should be compiled or loaded up front in "
"PiecewiseBackend.__init__. "
f"range_entry={range_entry.compile_range}"
)
return range_entry.runnable(*args)