57 lines
2.0 KiB
Python
57 lines
2.0 KiB
Python
|
import torch
|
||
|
|
from torch.fx.passes.infra.partitioner import CapabilityBasedPartitioner
|
||
|
|
from torch.fx.passes.operator_support import OperatorSupport
|
||
|
|
from torch.fx.passes.tools_common import CALLABLE_NODE_OPS
|
||
|
|
from torch.fx.passes.fake_tensor_prop import FakeTensorProp
|
||
|
|
from torch.utils import _pytree as pytree
|
||
|
|
|
||
|
|
import operator
|
||
|
|
|
||
|
|
class CudaGraphsSupport(OperatorSupport):
|
||
|
|
# TODO: why is submodules passed here
|
||
|
|
def is_node_supported(self, submodules, node: torch.fx.Node) -> bool:
|
||
|
|
if node.op not in CALLABLE_NODE_OPS:
|
||
|
|
return False
|
||
|
|
|
||
|
|
if node.target in [torch.ops.aten.embedding_dense_backward.default]:
|
||
|
|
return False
|
||
|
|
|
||
|
|
if node.target in [operator.getitem]:
|
||
|
|
return True
|
||
|
|
|
||
|
|
found_not_cuda = False
|
||
|
|
|
||
|
|
def meta_fk(meta):
|
||
|
|
return meta["val"] if "val" in meta else meta["fake_result"]
|
||
|
|
|
||
|
|
def find_not_cuda(t):
|
||
|
|
nonlocal found_not_cuda
|
||
|
|
if isinstance(t, torch.Tensor) and t.device.type != 'cuda':
|
||
|
|
found_not_cuda = True
|
||
|
|
|
||
|
|
for n in node.all_input_nodes:
|
||
|
|
pytree.tree_map_(find_not_cuda, meta_fk(n.meta))
|
||
|
|
|
||
|
|
pytree.tree_map_(find_not_cuda, meta_fk(node.meta))
|
||
|
|
|
||
|
|
# NB: factory function is accounted for because the result would be
|
||
|
|
# cpu or cuda
|
||
|
|
|
||
|
|
return not found_not_cuda
|
||
|
|
|
||
|
|
def partition_cudagraphs(gm, inputs):
|
||
|
|
"""
|
||
|
|
Partition an FX graph into sub-GraphModules that can be validly run under
|
||
|
|
CUDA graphs. For a subgraph to be runnable under CUDA, all of the operations
|
||
|
|
must involve CUDA tensors only/
|
||
|
|
"""
|
||
|
|
|
||
|
|
FakeTensorProp(gm).propagate(*inputs)
|
||
|
|
supported_ops = CudaGraphsSupport()
|
||
|
|
# TODO: single node partition may be wrong due to the pessimization
|
||
|
|
# from copying in and out the data. Check in benchmarks, perhaps
|
||
|
|
partitioner = CapabilityBasedPartitioner(gm, supported_ops, allows_single_node_partition=True)
|
||
|
|
partitions = partitioner.propose_partitions()
|
||
|
|
fused_graph = partitioner.fuse_partitions(partitions)
|
||
|
|
return fused_graph
|