[torch.compile] Inductor code caching fix (#10273)
Signed-off-by: luka <luka@neuralmagic.com> Signed-off-by: Luka Govedic <luka.govedic@gmail.com>
This commit is contained in:
@@ -1,7 +1,9 @@
|
||||
from copy import deepcopy
|
||||
from typing import Callable
|
||||
from typing import Callable, Union
|
||||
|
||||
import torch
|
||||
from torch import fx
|
||||
|
||||
from vllm.compilation.inductor_pass import InductorPass
|
||||
|
||||
|
||||
class TestBackend:
|
||||
@@ -11,19 +13,21 @@ class TestBackend:
|
||||
It also saves the graph before and after the custom passes for inspection.
|
||||
"""
|
||||
|
||||
def __init__(self, *args: Callable[[torch.fx.Graph], None]):
|
||||
self.custom_passes = args
|
||||
def __init__(self, *passes: Union[InductorPass, Callable[[fx.Graph],
|
||||
None]]):
|
||||
self.custom_passes = list(passes)
|
||||
from torch._inductor import config
|
||||
self.current_config = config.shallow_copy_dict()
|
||||
self.current_config['force_disable_caches'] = True
|
||||
self.current_config['post_grad_custom_post_pass'] = self.post_pass
|
||||
|
||||
def __call__(self, graph: torch.fx.GraphModule, example_inputs):
|
||||
def __call__(self, graph: fx.GraphModule, example_inputs):
|
||||
from torch._inductor.compile_fx import compile_fx
|
||||
return compile_fx(graph,
|
||||
example_inputs,
|
||||
config_patches=self.current_config)
|
||||
|
||||
def post_pass(self, graph: torch.fx.Graph):
|
||||
def post_pass(self, graph: fx.Graph):
|
||||
self.graph_pre_pass = deepcopy(graph)
|
||||
for pass_ in self.custom_passes:
|
||||
pass_(graph)
|
||||
|
||||
Reference in New Issue
Block a user