mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-07 00:21:07 +01:00
e071d72f3c
Hides debug/experimental backends by default. Before: ``` torch._dynamo.list_backends() ['aot_eager', 'aot_eager_decomp_partition', 'aot_torchxla_trace_once', 'aot_torchxla_trivial', 'aot_ts', 'aot_ts_nvfuser', 'cudagraphs', 'dynamo_accuracy_minifier_backend', 'dynamo_minifier_backend', 'eager', 'inductor', 'ipex', 'nvprims_aten', 'nvprims_nvfuser', 'onnxrt', 'tensorrt', 'torchxla_trace_once', 'torchxla_trivial', 'ts', 'tvm'] ``` After: ``` torch._dynamo.list_backends() ['aot_ts_nvfuser', 'cudagraphs', 'inductor', 'ipex', 'nvprims_nvfuser', 'onnxrt', 'tensorrt', 'tvm'] ``` Fixes https://github.com/pytorch/pytorch/issues/93733 Pull Request resolved: https://github.com/pytorch/pytorch/pull/93878 Approved by: https://github.com/voznesenskym
57 lines
1.7 KiB
Python
57 lines
1.7 KiB
Python
import functools
|
|
from importlib import import_module
|
|
|
|
from functorch.compile import min_cut_rematerialization_partition, nop
|
|
|
|
import torch
|
|
from torch._functorch.compilers import ts_compile
|
|
from .common import aot_autograd
|
|
from .registry import register_debug_backend as register_backend
|
|
|
|
"""
|
|
This file contains TorchDynamo backends intended for debugging uses.
|
|
"""
|
|
|
|
|
|
@register_backend
|
|
def eager(gm, fake_tensor_inputs):
|
|
return gm
|
|
|
|
|
|
@register_backend(name="ts")
|
|
def torchscript(gm, fake_tensor_inputs):
|
|
return torch.jit.script(gm)
|
|
|
|
|
|
# Useful for debugging purpose
|
|
# aot_eager uses AOT Autograd backend with nop compiler. It is helpful in debugging.
|
|
aot_eager = aot_autograd(fw_compiler=nop)
|
|
register_backend(name="aot_eager", compiler_fn=aot_eager)
|
|
|
|
|
|
# Uses TorchInductor AOT Autograd decomps and partitioner to isolate aot vs
|
|
# inductor problems.
|
|
# aot_eager_decomp_partition just replaces the inductor compiler with nop to help
|
|
# isolate inductor vs aot_eager errors
|
|
aot_eager_decomp_partition = aot_autograd(
|
|
# these are taken from memory_efficient_fusion()
|
|
fw_compiler=nop,
|
|
bw_compiler=nop,
|
|
# NB: lambda here is to delay import of inductor
|
|
decompositions=lambda: import_module(
|
|
"torch._inductor.compile_fx"
|
|
).select_decomp_table(),
|
|
partition_fn=functools.partial(
|
|
min_cut_rematerialization_partition, compiler="inductor"
|
|
),
|
|
)
|
|
register_backend(
|
|
name="aot_eager_decomp_partition", compiler_fn=aot_eager_decomp_partition
|
|
)
|
|
|
|
# AOT Autograd with torchscript backend. Default partitioner.
|
|
# aot_ts uses torchscript backend. We can use this with both nnc and nvfuser
|
|
# by using the relevant fuser with torch.jit.fuser(...)
|
|
aot_ts = aot_autograd(fw_compiler=ts_compile)
|
|
register_backend(name="aot_ts", compiler_fn=aot_ts)
|