frozenleaves/pytorch
1
0
Fork 0
mirror of https://github.com/pytorch/pytorch.git synced 2025-10-20 12:54:11 +08:00
Code Issues Packages Projects Releases Wiki Activity
Files
481a57bc37ffb90a670e63a7bdbfb68314190943
pytorch/torch/testing/_internal/hop_db.py
eellison 481a57bc37 Support torch.compile rng selective activation checkpointing with cudagraph (#146878) 
TODO:
- [x]  Add handling for when forward is invoked multiple times without invoking backward, so that the fwd/backward states are out of sync
- [x] Update rng state initialization to take from correct device
- [x]  Tests
- [x] handling of retain_graph
- [x] respect fallback random

Fix for https://github.com/pytorch/pytorch/issues/130123.

Updates the aot_eager and cudagraph compilation of `run_and_save_rng_state` to use the new mechanism added by https://github.com/pytorch/pytorch/pull/114068 for CUDAGraph safe rng states.

We have a pair of rng states for the fwd and backward respectively. In both forward and backward the rng op will get run with `graphsafe_run_with_rng_state` which takes in RNG state and it hooks onto the current RNG generator before running the operator. The rng states for fwd/backward are initialized with the same value. We ensure that for any given run of the forward, the corresponding backward run will have the same rng states for the op as was observed in the forward.

```
 ===== Forward graph 1 =====
 /data/users/eellison/pytorch/torch/fx/_lazy_graph_module.py class GraphModule(torch.nn.Module):
    def forward(self, primals_1: "f32[4, 4][4, 1]cuda:0", primals_2: "f32[4, 4][4, 1]cuda:0", fwd_rng_state_0):
        sin: "f32[4, 4][4, 1]cuda:0" = torch.ops.aten.sin.default(primals_1)

        # No stacktrace found for following nodes
        graphsafe_run_with_rng_state = torch.ops.higher_order.graphsafe_run_with_rng_state(torch.ops.aten.rand.default, [4, 4], dtype = torch.float32, device = device(type='cuda', index=0), pin_memory = False, rng_state = fwd_rng_state_0);  fwd_rng_state_0 = None
        ...

 ===== Backward graph 1 =====
    def forward(self, primals_1: "f32[4, 4][4, 1]cuda:0", primals_2: "f32[4, 4][4, 1]cuda:0", tangents_1: "f32[4, 4][4, 1]cuda:0", bwd_rng_state_0):
        sin: "f32[4, 4][4, 1]cuda:0" = torch.ops.aten.sin.default(primals_1)

        # No stacktrace found for following nodes
        graphsafe_run_with_rng_state = torch.ops.higher_order.graphsafe_run_with_rng_state(torch.ops.aten.rand.default, [4, 4], dtype = torch.float32, device = device(type='cuda', index=0), pin_memory = False, rng_state = bwd_rng_state_0);  bwd_rng_state_0 = None
```

There is some extra complication when a user either calls backward with retain_graph, or calls the backward in a different order as they called the forward. If a user has state fwd_rng_state0, bwd_rng_state0 and calls:
- fwd0: fwd_rng_state0 -> fwd_rng_state1
- fwd1: fwd_rng_state1 -> fwd_rng_state2
- bwd1
- bwd0

Then naively, when bwd1 is invoked the bwd rng states would not be equal to the same states that were observed in fwd1. I added handling of this in the aot runtime wrappers to detect pending backward invocations, and the current position of the bwd rng states, and to update when necesssary.

Other notes:

Because nodes which appear later in the forward appear earlier in the backward, we need a separate rng state for each operator. If we reused the rng across ops, the forward and backward would be run with different rng states. I.e., not applied in the same order.

Questions for reviewers:

This does change numerics, bc the rng of the op is now taken from the input rng state instead of whatever the rng would be midway through running the graph. Technically, we only need this for cuda graph. But, I'd prefer to not have a rng divergence just for cudagraph. I am making it respect `fallback_random`.

Edit: decided to apply to non cudagraphs as well, so long as fallback_random is not set

I'm initializing the rng states by cloning the current state. If you had something like 5 different rands in the model with the same shape, theyd all get the same value. This doesn't seem great. I could use some other initialization scheme like taking seed from graph position, or etc etc. Not sure. Let me know thoughts.

Edit: updated to be taken from randint()

Update: initializing rng states from torch.randint..

Pull Request resolved: https://github.com/pytorch/pytorch/pull/146878
Approved by: https://github.com/anijain2305, https://github.com/bdhirsh
2025-02-28 00:47:03 +00:00

432 lines
14 KiB
Python
Raw Blame History

# mypy: ignore-errors
import functools
import unittest
import torch
from functorch.experimental.control_flow import map
from torch.nn.attention.flex_attention import _create_empty_block_mask, flex_attention
from torch.testing import make_tensor
from torch.testing._internal.common_device_type import onlyCUDA
from torch.testing._internal.common_dtype import all_types_and, custom_types
from torch.testing._internal.opinfo.core import DecorateInfo, OpInfo, SampleInput
from torch._higher_order_ops.invoke_subgraph import mark_compile_region
from torch._higher_order_ops import InvokeQuant, invoke_quant_packed
def sample_inputs_map(opinfo, device, dtype, requires_grad, **kwargs):
make_arg = functools.partial(
make_tensor, device=device, dtype=dtype, requires_grad=requires_grad
)
yield SampleInput(
[make_arg(2, 2, 2, low=0.1, high=2), make_arg(2, 2, 2, low=0.1, high=2)],
args=(make_arg(1, low=0.1, high=2), make_arg(1, low=0.1, high=2)),
)
def inner_f(x, y0, y1):
return [x[0].cos().add_(1.0) * y0, (x[1] + y1.sin()).cos_().view(x[1].size())]
def simple_map(xs, y0, y1):
def f(x, y0, y1):
return inner_f(x, y0, y1)
return map(f, xs, y0, y1)
def nested_map(xs, y0, y1):
def f1(xx, y0, y1):
def f2(x, y0, y1):
return inner_f(x, y0, y1)
return map(f2, xx, y0, y1)
return map(f1, xs, y0, y1)
def triple_nested_map(xs, y0, y1):
def f0(xs, y0, y1):
def f1(xx, y0, y1):
def f2(x, y0, y1):
return inner_f(x, y0, y1)
return map(f2, xx, y0, y1)
return map(f1, xs, y0, y1)
return map(f0, xs, y0, y1)
# PLEASE DON'T ADD ANYTHING NEW TO THIS LIST,
# and do add an OpInfo for your HOP.
# The OpInfo lets us do automated testing for the HOP to check that
# your HOP will work correctly with PyTorch!
#
# Your new HOP may fail some automated testing. That's OK. If you don't
# care about certain features (like torch.export), it's fine to xfail those
# failing tests. It is less fine to xfail a more critical check (like checking
# if torch.compile works with your HOP, or if your HOP has a docstring).
# If you don't know if a test is fine to xfail, please ask.
#
# There are legitimate reasons why something cannot be added to this list
# (e.g. it uses executorch which is not in PyTorch). If that's the case then
# please leave a comment.
FIXME_hop_that_doesnt_have_opinfo_test_allowlist = [
"custom_function_call",
"autograd_function_apply",
"run_and_save_rng_state",
"run_with_rng_state",
"graphsafe_run_with_rng_state",
"out_dtype",
"trace_wrapped",
'tag_activation_checkpoint',
'executorch_call_delegate',
'wrap',
'wrap_with_set_grad_enabled',
'auto_functionalized_v2',
'associative_scan',
'flat_apply', # is WIP, doesn't pass any of the tests yet
'wrap_with_autocast',
'wrap_activation_checkpoint',
'run_const_graph',
'auto_functionalized',
"map", # T183144629
"map_impl",
"with_effects",
"strict_mode",
"_export_tracepoint",
"call_torchbind",
"triton_kernel_wrapper_mutation",
"triton_kernel_wrapper_functional",
"hints_wrapper",
"foreach_map",
"aoti_call_delegate",
]
torch.library.define(
"testlib::mutating_custom_op",
"(Tensor(a!) x, Tensor(b!) z) -> (Tensor, Tensor, Tensor)",
tags=torch.Tag.pt2_compliant_tag,
)
@torch.library.impl("testlib::mutating_custom_op", "cpu")
def foo_impl_cpu(x, z):
x.add_(5)
z.add_(5)
return x, z, x + z
@torch.library.impl("testlib::mutating_custom_op", "cuda")
def foo_impl_cuda(x, z):
x.add_(5)
z.add_(5)
return x, z, x + z
@torch.library.register_fake("testlib::mutating_custom_op")
def foo_impl_abstract(x, z):
return x, z, x + z
def sample_inputs_cond(opinfo, device, dtype, requires_grad, **kwargs):
make_arg = functools.partial(
make_tensor, device=device, dtype=dtype, requires_grad=requires_grad
)
yield SampleInput(make_arg(2, 2, 2, low=0.1, high=2))
def simple_cond(x):
return torch.cond(x.sum() > 2, lambda x: (x.cos(),), lambda x: (x.sin(),), [x])
def sample_inputs_invoke_subgraph(opinfo, device, dtype, requires_grad, **kwargs):
make_arg = functools.partial(
make_tensor, device=device, dtype=dtype, requires_grad=requires_grad
)
yield SampleInput(make_arg(2, 2, 2, low=0.1, high=2))
@mark_compile_region
def fn_for_invoke_subgraph(x):
return torch.sin(x)
def simple_invoke_subgraph(x):
return fn_for_invoke_subgraph(x)
def sample_inputs_auto_functionalize(opinfo, device, dtype, requires_grad, **kwargs):
make_arg = functools.partial(
make_tensor, device=device, dtype=dtype, requires_grad=False
)
yield SampleInput(
make_arg(2, 2, 2, low=0.1, high=2), make_arg(2, 2, 2, low=0.1, high=2)
)
def simple_auto_functionalize(x, z):
return torch.ops.testlib.mutating_custom_op(x, z)
def sample_inputs_flex_attention(opinfo, device, dtype, requires_grad, **kwargs):
make_arg = functools.partial(
make_tensor, device=device, dtype=dtype, requires_grad=requires_grad
)
def score_mod(score, b, h, m, n):
return score + h
q, k, v = (make_arg(2, 2, 128, 8, low=0.1, high=2) for _ in range(3))
block_mask = _create_empty_block_mask(q, k)
yield SampleInput(q, k, v, score_mod, block_mask)
def sample_inputs_while_loop(opinfo, device, dtype, requires_grad, **kwargs):
make_arg = functools.partial(
make_tensor, device=device, dtype=dtype, requires_grad=False
)
yield SampleInput(
torch.tensor(3),
make_arg(2, 3, 4, low=0.1, high=2),
)
def simple_while_loop(iter_t, x):
def cond_fn(iter_t, x):
return iter_t > 0
def body_fn(iter_t, x):
return iter_t - 1, x.cos()
return torch._higher_order_ops.while_loop(cond_fn, body_fn, (iter_t, x))
def sample_inputs_scan(opinfo, device, dtype, requires_grad, **kwargs):
make_arg = functools.partial(
make_tensor, device=device, dtype=dtype, requires_grad=requires_grad
)
yield SampleInput(
make_arg(2, 2, low=0.1, high=2),
make_arg(2, 2, 2, low=0.1, high=2),
)
def simple_scan(init, xs):
def combine_fn(carry, x):
result = carry @ x + x
return result, carry.clone()
return torch._higher_order_ops.scan(combine_fn, init, xs)
quant_tracer = InvokeQuant()
def simple_invoke_quant(x):
def fn(x, y):
return (torch.sin(x) * y,)
return quant_tracer(fn, x, x)[0] * 2.
def simple_invoke_quant_packed(x):
def fn(x):
return (torch.sin(x),)
return invoke_quant_packed(fn, x)[0] * 2.
hop_db = [
OpInfo(
name="scan",
variant_test_name="simple",
op=simple_scan,
sample_inputs_func=sample_inputs_scan,
dtypes=all_types_and(torch.bool, torch.half),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
supports_autograd=False,
# "torch.compile with aot_autograd does not currently support double backward."
supports_gradgrad=False,
),
OpInfo(
name="invoke_subgraph",
variant_test_name="simple",
op=simple_invoke_subgraph,
sample_inputs_func=sample_inputs_invoke_subgraph,
dtypes=all_types_and(torch.bool, torch.half),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
supports_autograd=True,
# "torch.compile with aot_autograd does not currently support double backward."
supports_gradgrad=False,
),
OpInfo(
name="map",
variant_test_name="simple",
op=simple_map,
sample_inputs_func=sample_inputs_map,
dtypes=all_types_and(torch.bool, torch.half),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
),
OpInfo(
name="map",
variant_test_name="nested",
op=nested_map,
sample_inputs_func=sample_inputs_map,
dtypes=all_types_and(torch.bool, torch.half),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
),
OpInfo(
name="map",
variant_test_name="triple_nested",
op=triple_nested_map,
sample_inputs_func=sample_inputs_map,
dtypes=all_types_and(torch.bool, torch.half),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
),
OpInfo(
name="cond",
variant_test_name="simple",
op=simple_cond,
sample_inputs_func=sample_inputs_cond,
dtypes=all_types_and(torch.bool, torch.half),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
supports_autograd=True,
# "torch.compile with aot_autograd does not currently support double backward."
supports_gradgrad=False,
),
OpInfo(
name="invoke_quant",
variant_test_name="simple",
op=simple_invoke_quant,
sample_inputs_func=sample_inputs_invoke_subgraph,
dtypes=all_types_and(torch.bool, torch.half),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
supports_autograd=True,
# "torch.compile with aot_autograd does not currently support double backward."
skips=(
DecorateInfo(unittest.expectedFailure, "TestHOP", "test_aot_export"),
DecorateInfo(
unittest.expectedFailure, "TestHOP", "test_pre_dispatch_export"
),
DecorateInfo(unittest.expectedFailure, "TestHOP", "test_serialize_export"),
DecorateInfo(unittest.expectedFailure, "TestHOP", "test_retrace_export"),
),
# "torch.compile with aot_autograd does not currently support double backward."
supports_gradgrad=False,
),
OpInfo(
name="invoke_quant_packed",
variant_test_name="simple",
op=simple_invoke_quant_packed,
sample_inputs_func=sample_inputs_invoke_subgraph,
dtypes=all_types_and(torch.bool, torch.half),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
supports_autograd=True,
# "torch.compile with aot_autograd does not currently support double backward."
supports_gradgrad=False,
),
OpInfo(
name="while_loop",
variant_test_name="simple",
op=simple_while_loop,
sample_inputs_func=sample_inputs_while_loop,
dtypes=all_types_and(torch.bool, torch.half),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
supports_autograd=False,
),
OpInfo(
name="auto_functionalize",
variant_test_name="simple",
op=simple_auto_functionalize,
sample_inputs_func=sample_inputs_auto_functionalize,
dtypes=all_types_and(torch.bool, torch.half),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
supports_autograd=False,
),
OpInfo(
name="flex_attention",
variant_test_name="simple",
op=flex_attention,
sample_inputs_func=sample_inputs_flex_attention,
dtypes=custom_types(torch.float16, torch.float32),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
skips=(
DecorateInfo(unittest.expectedFailure, "TestHOP", "test_aot_export"),
DecorateInfo(
unittest.expectedFailure, "TestHOP", "test_pre_dispatch_export"
),
DecorateInfo(unittest.expectedFailure, "TestHOP", "test_serialize_export"),
DecorateInfo(unittest.expectedFailure, "TestHOP", "test_retrace_export"),
),
decorators=[onlyCUDA],
),
OpInfo(
name="flex_attention_backward",
variant_test_name="simple",
op=flex_attention,
sample_inputs_func=sample_inputs_flex_attention,
dtypes=custom_types(torch.float16, torch.float32),
supports_out=False,
check_batched_grad=False,
check_batched_gradgrad=False,
check_batched_forward_grad=False,
check_inplace_batched_forward_grad=False,
skips=(
DecorateInfo(unittest.expectedFailure, "TestHOP", "test_aot_export"),
DecorateInfo(
unittest.expectedFailure, "TestHOP", "test_pre_dispatch_export"
),
DecorateInfo(unittest.expectedFailure, "TestHOP", "test_serialize_export"),
DecorateInfo(unittest.expectedFailure, "TestHOP", "test_retrace_export"),
),
decorators=[onlyCUDA],
),
]
Reference in New Issue View Git Blame Copy Permalink