When I parallelized the Conv1D layer (from transformers library) with DTensor and torch.compile with the following code, the error occurs. It seems that view operator is not compiled properly when the input tensor is Dtensor wrapping FakeTensor. What I am trying is to do this using real tensor instead of fake tensor. Does anyone know how to do this? or is there any other workaround?
update: I found out how to trace with real tensor. But still I want to do this with fake tensor.
from transformers import Conv1D
import torch.nn as nn
import torch
from torch.distributed._tensor import DeviceMesh, Shard, distribute_tensor
mesh = DeviceMesh("cpu", list(range(2)))
def my_fn(gm, inputs):
print(gm.graph)
return gm
model = Conv1D(8, 16)
# parallelize with DTensor
model.weight = nn.Parameter(distribute_tensor(model.weight, mesh, [Shard(1)]))
model.bias = nn.Parameter(distribute_tensor(model.bias, mesh, [Shard(0)]))
input = distribute_tensor(torch.randn(4, 16), mesh, [Shard(1)])
torch.compile(model, backend=my_fn)(input)
Error message
(pytorch2) root@sunghwanshim-cpu-0:~/pytorch# torchrun --nnodes 1 --nproc-per-node 2 dtensor_test2.py
[2023-10-25 12:04:23,314] torch.distributed.run: [WARNING]
[2023-10-25 12:04:23,314] torch.distributed.run: [WARNING] *****************************************
[2023-10-25 12:04:23,314] torch.distributed.run: [WARNING] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed.
[2023-10-25 12:04:23,314] torch.distributed.run: [WARNING] *****************************************
Traceback (most recent call last):
File "dtensor_test2.py", line 19, in <module>
torch.compile(model, backend=my_fn)(input)
File "/root/pytorch/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
File "/root/pytorch/torch/nn/modules/module.py", line 1527, in _call_impl
Traceback (most recent call last):
File "dtensor_test2.py", line 19, in <module>
return forward_call(*args, **kwargs)
File "/root/pytorch/torch/_dynamo/eval_frame.py", line 338, in _fn
torch.compile(model, backend=my_fn)(input)
File "/root/pytorch/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
return fn(*args, **kwargs)
File "/root/pytorch/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
return self._call_impl(*args, **kwargs)
return self._call_impl(*args, **kwargs)
File "/root/pytorch/torch/nn/modules/module.py", line 1527, in _call_impl
File "/root/pytorch/torch/nn/modules/module.py", line 1527, in _call_impl
return forward_call(*args, **kwargs)
File "/root/pytorch/torch/_dynamo/eval_frame.py", line 500, in catch_errors
return forward_call(*args, **kwargs)
File "/root/pytorch/torch/_dynamo/eval_frame.py", line 338, in _fn
return fn(*args, **kwargs)
File "/root/pytorch/torch/nn/modules/module.py", line 1518, in _wrapped_call_impl
return callback(frame, cache_entry, hooks, frame_state)
File "/root/pytorch/torch/_dynamo/convert_frame.py", line 634, in _convert_frame
result = inner_convert(frame, cache_entry, hooks, frame_state)
File "/root/pytorch/torch/_dynamo/convert_frame.py", line 140, in _fn
return fn(*args, **kwargs)
File "/root/pytorch/torch/_dynamo/convert_frame.py", line 382, in _convert_frame_assert
return self._call_impl(*args, **kwargs)
File "/root/pytorch/torch/nn/modules/module.py", line 1527, in _call_impl
return _compile(
File "/root/pytorch/torch/_dynamo/convert_frame.py", line 562, in _compile
guarded_code = compile_inner(code, one_graph, hooks, transform)
File "/root/pytorch/torch/_dynamo/utils.py", line 189, in time_wrapper
r = func(*args, **kwargs)
File "/root/pytorch/torch/_dynamo/convert_frame.py", line 484, in compile_inner
return forward_call(*args, **kwargs)
File "/root/pytorch/torch/_dynamo/eval_frame.py", line 500, in catch_errors
out_code = transform_code_object(code, transform)
File "/root/pytorch/torch/_dynamo/bytecode_transformation.py", line 1028, in transform_code_object
return callback(frame, cache_entry, hooks, frame_state)
File "/root/pytorch/torch/_dynamo/convert_frame.py", line 634, in _convert_frame
result = inner_convert(frame, cache_entry, hooks, frame_state)
File "/root/pytorch/torch/_dynamo/convert_frame.py", line 140, in _fn
transformations(instructions, code_options)
File "/root/pytorch/torch/_dynamo/convert_frame.py", line 451, in transform
return fn(*args, **kwargs)
File "/root/pytorch/torch/_dynamo/convert_frame.py", line 382, in _convert_frame_assert
tracer.run()
File "/root/pytorch/torch/_dynamo/symbolic_convert.py", line 2088, in run
return _compile(
File "/root/pytorch/torch/_dynamo/convert_frame.py", line 562, in _compile
guarded_code = compile_inner(code, one_graph, hooks, transform)
File "/root/pytorch/torch/_dynamo/utils.py", line 189, in time_wrapper
r = func(*args, **kwargs)
File "/root/pytorch/torch/_dynamo/convert_frame.py", line 484, in compile_inner
out_code = transform_code_object(code, transform)
File "/root/pytorch/torch/_dynamo/bytecode_transformation.py", line 1028, in transform_code_object
super().run()
File "/root/pytorch/torch/_dynamo/symbolic_convert.py", line 728, in run
transformations(instructions, code_options)and self.step()
File "/root/pytorch/torch/_dynamo/convert_frame.py", line 451, in transform
File "/root/pytorch/torch/_dynamo/symbolic_convert.py", line 691, in step
tracer.run()
File "/root/pytorch/torch/_dynamo/symbolic_convert.py", line 2088, in run
getattr(self, inst.opname)(inst)
File "/root/pytorch/torch/_dynamo/symbolic_convert.py", line 392, in wrapper
return inner_fn(self, inst)
File "/root/pytorch/torch/_dynamo/symbolic_convert.py", line 1119, in CALL_FUNCTION
self.call_function(fn, args, {})super().run()
File "/root/pytorch/torch/_dynamo/symbolic_convert.py", line 565, in call_function
File "/root/pytorch/torch/_dynamo/symbolic_convert.py", line 728, in run
self.push(fn.call_function(self, args, kwargs))
File "/root/pytorch/torch/_dynamo/variables/misc.py", line 594, in call_function
and self.step()
File "/root/pytorch/torch/_dynamo/symbolic_convert.py", line 691, in step
getattr(self, inst.opname)(inst)
File "/root/pytorch/torch/_dynamo/symbolic_convert.py", line 392, in wrapper
return self.obj.call_method(tx, self.name, args, kwargs).add_options(self)
File "/root/pytorch/torch/_dynamo/variables/tensor.py", line 652, in call_method
return inner_fn(self, inst)
File "/root/pytorch/torch/_dynamo/symbolic_convert.py", line 1119, in CALL_FUNCTION
return wrap_fx_proxy(
File "/root/pytorch/torch/_dynamo/variables/builder.py", line 1207, in wrap_fx_proxy
self.call_function(fn, args, {})
File "/root/pytorch/torch/_dynamo/symbolic_convert.py", line 565, in call_function
return wrap_fx_proxy_cls(
File "/root/pytorch/torch/_dynamo/variables/builder.py", line 1294, in wrap_fx_proxy_cls
self.push(fn.call_function(self, args, kwargs))
File "/root/pytorch/torch/_dynamo/variables/misc.py", line 594, in call_function
return self.obj.call_method(tx, self.name, args, kwargs).add_options(self)
File "/root/pytorch/torch/_dynamo/variables/tensor.py", line 652, in call_method
example_value = get_fake_value(proxy.node, tx)
File "/root/pytorch/torch/_dynamo/utils.py", line 1381, in get_fake_value
return wrap_fx_proxy(
File "/root/pytorch/torch/_dynamo/variables/builder.py", line 1207, in wrap_fx_proxy
raise TorchRuntimeError(str(e)).with_traceback(e.__traceback__) from None
File "/root/pytorch/torch/_dynamo/utils.py", line 1342, in get_fake_value
return wrap_fx_proxy_cls(
File "/root/pytorch/torch/_dynamo/variables/builder.py", line 1294, in wrap_fx_proxy_cls
return wrap_fake_exception(
File "/root/pytorch/torch/_dynamo/utils.py", line 917, in wrap_fake_exception
example_value = get_fake_value(proxy.node, tx)
File "/root/pytorch/torch/_dynamo/utils.py", line 1381, in get_fake_value
return fn()
File "/root/pytorch/torch/_dynamo/utils.py", line 1343, in <lambda>
raise TorchRuntimeError(str(e)).with_traceback(e.__traceback__) from None
File "/root/pytorch/torch/_dynamo/utils.py", line 1342, in get_fake_value
lambda: run_node(tx.output, node, args, kwargs, nnmodule)
File "/root/pytorch/torch/_dynamo/utils.py", line 1415, in run_node
return wrap_fake_exception(
File "/root/pytorch/torch/_dynamo/utils.py", line 917, in wrap_fake_exception
raise RuntimeError(fn_str + str(e)).with_traceback(e.__traceback__) from e
File "/root/pytorch/torch/_dynamo/utils.py", line 1404, in run_node
return fn()
File "/root/pytorch/torch/_dynamo/utils.py", line 1343, in <lambda>
return getattr(args[0], node.target)(*args[1:], **kwargs)
File "/root/pytorch/torch/_tensor.py", line 1386, in __torch_function__
lambda: run_node(tx.output, node, args, kwargs, nnmodule)
File "/root/pytorch/torch/_dynamo/utils.py", line 1415, in run_node
ret = func(*args, **kwargs)
File "/root/pytorch/torch/distributed/_tensor/api.py", line 241, in __torch_dispatch__
raise RuntimeError(fn_str + str(e)).with_traceback(e.__traceback__) from ereturn op_dispatch.operator_dispatch(
File "/root/pytorch/torch/_dynamo/utils.py", line 1404, in run_node
File "/root/pytorch/torch/distributed/_tensor/dispatch.py", line 109, in operator_dispatch
out, _, _ = _operator_dispatch(op_call, args, kwargs, sharding_propagator)
File "/root/pytorch/torch/distributed/_tensor/dispatch.py", line 183, in _operator_dispatch
sharding_propagator.propagate(op_info)
File "/root/pytorch/torch/distributed/_tensor/sharding_prop.py", line 57, in propagate
output_sharding = self.propagate_op_sharding(op_overload, op_info.schema)
File "/root/pytorch/torch/distributed/_tensor/sharding_prop.py", line 159, in propagate_op_sharding
return getattr(args[0], node.target)(*args[1:], **kwargs)
File "/root/pytorch/torch/_tensor.py", line 1386, in __torch_function__
f"Sharding propagation failed on op {op_overload}.\n"
File "/root/pytorch/torch/distributed/_tensor/op_schema.py", line 174, in __repr__
f"OpSchema(func_schema={self.func_schema},"
File "/root/miniconda3/envs/pytorch2/lib/python3.8/dataclasses.py", line 368, in wrapper
result = user_function(self)
File "<string>", line 3, in __repr__
ret = func(*args, **kwargs) File "/root/pytorch/torch/distributed/_tensor/device_mesh.py", line 259, in __repr__
File "/root/pytorch/torch/distributed/_tensor/api.py", line 241, in __torch_dispatch__
return op_dispatch.operator_dispatch(return f"DeviceMesh:({self.mesh.tolist()})"
File "/root/pytorch/torch/distributed/_tensor/dispatch.py", line 109, in operator_dispatch
File "/root/pytorch/torch/utils/_stats.py", line 20, in wrapper
return fn(*args, **kwargs)
out, _, _ = _operator_dispatch(op_call, args, kwargs, sharding_propagator)
File "/root/pytorch/torch/_subclasses/fake_tensor.py", line 1290, in __torch_dispatch__
File "/root/pytorch/torch/distributed/_tensor/dispatch.py", line 183, in _operator_dispatch
sharding_propagator.propagate(op_info)
File "/root/pytorch/torch/distributed/_tensor/sharding_prop.py", line 57, in propagate
output_sharding = self.propagate_op_sharding(op_overload, op_info.schema)
File "/root/pytorch/torch/distributed/_tensor/sharding_prop.py", line 159, in propagate_op_sharding
f"Sharding propagation failed on op {op_overload}.\n"
File "/root/pytorch/torch/distributed/_tensor/op_schema.py", line 174, in __repr__
return self.dispatch(func, types, args, kwargs)
File "/root/pytorch/torch/_subclasses/fake_tensor.py", line 1421, in dispatch
f"OpSchema(func_schema={self.func_schema},"
File "/root/miniconda3/envs/pytorch2/lib/python3.8/dataclasses.py", line 368, in wrapper
result = user_function(self)
File "<string>", line 3, in __repr__
File "/root/pytorch/torch/distributed/_tensor/device_mesh.py", line 259, in __repr__
return f"DeviceMesh:({self.mesh.tolist()})") = self.validate_and_convert_non_fake_tensors(func, converter, args, kwargs)
File "/root/pytorch/torch/utils/_stats.py", line 20, in wrapper
File "/root/pytorch/torch/_subclasses/fake_tensor.py", line 1642, in validate_and_convert_non_fake_tensors
return fn(*args, **kwargs)
File "/root/pytorch/torch/_subclasses/fake_tensor.py", line 1290, in __torch_dispatch__
args, kwargs = tree_map_only(return self.dispatch(func, types, args, kwargs)
File "/root/pytorch/torch/utils/_pytree.py", line 362, in tree_map_only
File "/root/pytorch/torch/_subclasses/fake_tensor.py", line 1421, in dispatch
return tree_map(map_only(ty)(fn), pytree)
File "/root/pytorch/torch/utils/_pytree.py", line 292, in tree_map
return tree_unflatten([fn(i) for i in flat_args], spec)
File "/root/pytorch/torch/utils/_pytree.py", line 292, in <listcomp>
) = self.validate_and_convert_non_fake_tensors(func, converter, args, kwargs)
File "/root/pytorch/torch/_subclasses/fake_tensor.py", line 1642, in validate_and_convert_non_fake_tensors
return tree_unflatten([fn(i) for i in flat_args], spec)
File "/root/pytorch/torch/utils/_pytree.py", line 343, in inner
return f(x)
File "/root/pytorch/torch/_subclasses/fake_tensor.py", line 1632, in validate
args, kwargs = tree_map_only(
File "/root/pytorch/torch/utils/_pytree.py", line 362, in tree_map_only
return tree_map(map_only(ty)(fn), pytree)
File "/root/pytorch/torch/utils/_pytree.py", line 292, in tree_map
raise Exception(
torch._dynamo.exc.TorchRuntimeError: Failed running call_method view(*(DTensor(local_tensor=FakeTensor(..., size=(4, 8)), device_mesh=DeviceMesh:([0, 1]), placements=(Shard(dim=1),)), -1, 16), **{}):
Please convert all Tensors to FakeTensors first or instantiate FakeTensorMode with 'allow_non_fake_inputs'. Found in aten.resolve_conj.default(tensor([...], size=(2,), dtype=torch.int32))
from user code:
File "/root/miniconda3/envs/pytorch2/lib/python3.8/site-packages/transformers/pytorch_utils.py", line 108, in forward
x = torch.addmm(self.bias, x.view(-1, x.size(-1)), self.weight)
Set TORCH_LOGS="+dynamo" and TORCHDYNAMO_VERBOSE=1 for more information
You can suppress this exception and fall back to eager by setting:
import torch._dynamo
torch._dynamo.config.suppress_errors = True
return tree_unflatten([fn(i) for i in flat_args], spec)
File "/root/pytorch/torch/utils/_pytree.py", line 292, in <listcomp>
return tree_unflatten([fn(i) for i in flat_args], spec)
File "/root/pytorch/torch/utils/_pytree.py", line 343, in inner
return f(x)
File "/root/pytorch/torch/_subclasses/fake_tensor.py", line 1632, in validate
raise Exception(
torch._dynamo.exc.TorchRuntimeError: Failed running call_method view(*(DTensor(local_tensor=FakeTensor(..., size=(4, 8)), device_mesh=DeviceMesh:([0, 1]), placements=(Shard(dim=1),)), -1, 16), **{}):
Please convert all Tensors to FakeTensors first or instantiate FakeTensorMode with 'allow_non_fake_inputs'. Found in aten.resolve_conj.default(tensor([...], size=(2,), dtype=torch.int32))
from user code:
File "/root/miniconda3/envs/pytorch2/lib/python3.8/site-packages/transformers/pytorch_utils.py", line 108, in forward
x = torch.addmm(self.bias, x.view(-1, x.size(-1)), self.weight)
Set TORCH_LOGS="+dynamo" and TORCHDYNAMO_VERBOSE=1 for more information
You can suppress this exception and fall back to eager by setting:
import torch._dynamo
torch._dynamo.config.suppress_errors = True
[2023-10-25 12:04:28,336] torch.distributed.elastic.multiprocessing.api: [ERROR] failed (exitcode: 1) local_rank: 0 (pid: 137293) of binary: /root/miniconda3/envs/pytorch2/bin/python
Traceback (most recent call last):
File "/root/miniconda3/envs/pytorch2/bin/torchrun", line 33, in <module>
sys.exit(load_entry_point('torch', 'console_scripts', 'torchrun')())
File "/root/pytorch/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 346, in wrapper
return f(*args, **kwargs)
File "/root/pytorch/torch/distributed/run.py", line 806, in main
run(args)
File "/root/pytorch/torch/distributed/run.py", line 797, in run
elastic_launch(
File "/root/pytorch/torch/distributed/launcher/api.py", line 134, in __call__
return launch_agent(self._config, self._entrypoint, list(args))
File "/root/pytorch/torch/distributed/launcher/api.py", line 264, in launch_agent
raise ChildFailedError(
torch.distributed.elastic.multiprocessing.errors.ChildFailedError:
============================================================
dtensor_test2.py FAILED
------------------------------------------------------------
Failures:
[1]:
time : 2023-10-25_12:04:28
host : sunghwanshim-cpu-0
rank : 1 (local_rank: 1)
exitcode : 1 (pid: 137294)
error_file: <N/A>
traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
------------------------------------------------------------
Root Cause (first observed failure):
[0]:
time : 2023-10-25_12:04:28
host : sunghwanshim-cpu-0
rank : 0 (local_rank: 0)
exitcode : 1 (pid: 137293)
error_file: <N/A>
traceback : To enable traceback see: https://pytorch.org/docs/stable/elastic/errors.html
============================================================
Versions
PyTorch version: 2.2.0a0+gite4f3e54
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A
OS: Debian GNU/Linux 11 (bullseye) (x86_64)
GCC version: (Debian 10.2.1-6) 10.2.1 20210110
Clang version: 11.0.1-2
CMake version: version 3.26.4
Libc version: glibc-2.31
Python version: 3.8.18 (default, Sep 11 2023, 13:40:15) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-5.15.0-87-generic-x86_64-with-glibc2.17
Is CUDA available: False
CUDA runtime version: No CUDA
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: No CUDA
Nvidia driver version: No CUDA
cuDNN version: No CUDA
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Byte Order: Little Endian
Address sizes: 46 bits physical, 48 bits virtual
CPU(s): 192
On-line CPU(s) list: 0-191
Thread(s) per core: 2
Core(s) per socket: 24
Socket(s): 4
NUMA node(s): 4
Vendor ID: GenuineIntel
CPU family: 6
Model: 85
Model name: Intel(R) Xeon(R) Gold 6348H CPU @ 2.30GHz
Stepping: 11
CPU MHz: 1000.000
CPU max MHz: 4200.0000
CPU min MHz: 1000.0000
BogoMIPS: 4600.00
Virtualization: VT-x
L1d cache: 3 MiB
L1i cache: 3 MiB
L2 cache: 96 MiB
L3 cache: 132 MiB
NUMA node0 CPU(s): 0-23,96-119
NUMA node1 CPU(s): 24-47,120-143
NUMA node2 CPU(s): 48-71,144-167
NUMA node3 CPU(s): 72-95,168-191
Vulnerability Gather data sampling: Mitigation; Microcode
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Mitigation; Clear CPU buffers; SMT vulnerable
Vulnerability Retbleed: Mitigation; Enhanced IBRS
Vulnerability Spec rstack overflow: Not affected
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Enhanced IBRS, IBPB conditional, RSB filling, PBRSB-eIBRS SW sequence
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cdp_l3 invpcid_single intel_ppin ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm mpx rdt_a avx512f avx512dq rdseed adx smap clflushopt clwb intel_pt avx512cd avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local avx512_bf16 dtherm ida arat pln pts pku ospke avx512_vnni md_clear flush_l1d arch_capabilities
Versions of relevant libraries:
[pip3] numpy==1.24.4
[pip3] optree==0.9.2
[pip3] torch==2.2.0a0+gite4f3e54
[conda] mkl 2023.1.0 h213fc3f_46343
[conda] mkl-include 2023.1.0 h06a4308_46343
[conda] numpy 1.24.4 pypi_0 pypi
[conda] optree 0.9.2 pypi_0 pypi
[conda] torch 2.2.0a0+gite4f3e54 dev_0 <develop>