Regression: `capture_pre_autograd_graph` does not support empty args and kwargs only anymore

[Xia-Weiwen]

🐛 Describe the bug

A regression about capture_pre_autograd_graph is found to be introduced by #110222 .
capture_pre_autograd_graph is for graph capture and is now used for, for example, quantization with TorchInductor. It takes the original model and example inputs, positional args and/or kwargs.
Before #110222, we can use kwargs only to capture the graph for models like BERT, i.e., use () for args.
Example script (bert.py):

import torch
from torch._export import capture_pre_autograd_graph
from transformers import BertTokenizer, BertModel, AutoConfig

model_id = 'bert-base-uncased'
config = AutoConfig.from_pretrained(model_id, return_dict=False)
tokenizer = BertTokenizer.from_pretrained(model_id, return_dict=False)
model = BertModel.from_pretrained(model_id, config=config)

text = "Replace me by any text you'd like."
encoded_input = tokenizer(text, return_tensors='pt')

with torch.no_grad():
    captured_model = capture_pre_autograd_graph(model, (), kwargs=encoded_input)
    print('Success')

After this PR was merged, this practice will result in error:

Traceback (most recent call last):
  File "/home/weiwen/workspace/test/bert.py", line 40, in <module>
    captured_model = capture_pre_autograd_graph(model, (), kwargs=encoded_input)
  File "/home/weiwen/workspace/pytorch-fork/torch/_export/__init__.py", line 388, in capture_pre_autograd_graph
    range_constraints, equality_constraints = _process_constraints(m, 0, flat_args)
  File "/home/weiwen/workspace/pytorch-fork/torch/_export/exported_program.py", line 358, in _process_constraints
    example_input = example_inputs[i - num_lifted_params_buffers]
IndexError: list index out of range

Versions

PyTorch version: 2.2.0a0+git8f5fead
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: Ubuntu 20.04.3 LTS (x86_64)
GCC version: (Ubuntu 11.1.0-1ubuntu1~20.04) 11.1.0
Clang version: Could not collect
CMake version: version 3.22.1
Libc version: glibc-2.31

Python version: 3.9.16 (main, May 15 2023, 23:46:34) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-5.11.0-27-generic-x86_64-with-glibc2.31
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: 52 bits physical, 57 bits virtual
CPU(s): 128
On-line CPU(s) list: 0-127
Thread(s) per core: 2
Core(s) per socket: 32
Socket(s): 2
NUMA node(s): 2
Vendor ID: GenuineIntel
CPU family: 6
Model: 106
Model name: Intel(R) Xeon(R) Platinum 8358 CPU @ 2.60GHz
Stepping: 6
CPU MHz: 2600.000
CPU max MHz: 3400.0000
CPU min MHz: 800.0000
BogoMIPS: 5200.00
L1d cache: 3 MiB
L1i cache: 2 MiB
L2 cache: 80 MiB
L3 cache: 96 MiB
NUMA node0 CPU(s): 0-31,64-95
NUMA node1 CPU(s): 32-63,96-127
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: 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
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 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 invpcid_single ssbd mba ibrs ibpb stibp ibrs_enhanced fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect wbnoinvd dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req avx512vbmi umip pku ospke avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid fsrm md_clear pconfig flush_l1d arch_capabilities

Versions of relevant libraries:
[pip3] bert-pytorch==0.0.1a4
[pip3] clip-anytorch==2.5.0
[pip3] CoCa-pytorch==0.0.7
[pip3] dalle2-pytorch==1.10.5
[pip3] ema-pytorch==0.1.4
[pip3] flake8==3.8.2
[pip3] flake8-comprehensions==3.12.0
[pip3] flake8-executable==2.1.3
[pip3] flake8-logging-format==0.9.0
[pip3] flake8-pyi==23.3.1
[pip3] flake8-simplify==0.19.3
[pip3] functorch==1.14.0a0+408bcf1
[pip3] intel-extension-for-pytorch==2.2.0+git53af6a7
[pip3] mypy==1.4.1
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.22.3
[pip3] onnx==1.13.0
[pip3] pytorch-transformers==1.2.0
[pip3] pytorch-warmup==0.1.1
[pip3] pytorch-widedeep==0.3.7
[pip3] rotary-embedding-torch==0.2.1
[pip3] torch==2.2.0a0+gitfda9412
[pip3] torch-fidelity==0.3.0
[pip3] torch-struct==0.5
[pip3] torchaudio==2.0.0.dev20230219+cpu
[pip3] torchdata==0.6.0a0+2d4f145
[pip3] torchdynamo==1.13.0.dev0
[pip3] torchmetrics==0.11.1
[pip3] torchrec-nightly==2023.2.1
[pip3] torchtext==0.15.0a0+5b21235
[pip3] torchvision==0.16.0a0+0d75d9e
[pip3] triton==2.0.0
[pip3] vector-quantize-pytorch==1.0.0
[conda] bert-pytorch 0.0.1a4 dev_0
[conda] blas 1.0 mkl
[conda] clip-anytorch 2.5.0 pypi_0 pypi
[conda] coca-pytorch 0.0.7 pypi_0 pypi
[conda] dalle2-pytorch 1.10.5 pypi_0 pypi
[conda] ema-pytorch 0.1.4 pypi_0 pypi
[conda] functorch 1.14.0a0+408bcf1 pypi_0 pypi
[conda] intel-extension-for-pytorch 2.2.0+git53af6a7 dev_0
[conda] mkl 2021.4.0 h06a4308_640
[conda] mkl-include 2023.2.0 pypi_0 pypi
[conda] mkl-service 2.4.0 py39h7f8727e_0
[conda] mkl-static 2023.2.0 pypi_0 pypi
[conda] mkl_fft 1.3.1 py39hd3c417c_0
[conda] mkl_random 1.2.2 py39h51133e4_0
[conda] numpy 1.24.3 pypi_0 pypi
[conda] numpy-base 1.22.3 py39hf524024_0
[conda] pytorch-transformers 1.2.0 pypi_0 pypi
[conda] pytorch-warmup 0.1.1 pypi_0 pypi
[conda] rotary-embedding-torch 0.2.1 pypi_0 pypi
[conda] torch 2.2.0a0+gitfda9412 dev_0
[conda] torch-fidelity 0.3.0 pypi_0 pypi
[conda] torch-struct 0.5 pypi_0 pypi
[conda] torchaudio 2.0.0.dev20230219+cpu pypi_0 pypi
[conda] torchdata 0.6.0a0+2d4f145 pypi_0 pypi
[conda] torchdynamo 1.13.0.dev0 dev_0
[conda] torchmetrics 0.11.1 pypi_0 pypi
[conda] torchrec-nightly 2023.2.1 pypi_0 pypi
[conda] torchtext 0.15.0a0+5b21235 pypi_0 pypi
[conda] torchvision 0.16.0a0+0d75d9e pypi_0 pypi
[conda] triton 2.0.0 pypi_0 pypi
[conda] vector-quantize-pytorch 1.0.0 pypi_0 pypi

cc @ezyang @msaroufim @wconstab @bdhirsh @anijain2305 @zou3519 @avikchaudhuri @gmagogsfm @zhxchen17 @tugsbayasgalan @angelayi @suo @ydwu4

[leslie-fang-intel]

• _process_constraints expects example_inputs is a list of tensors
  

  pytorch/torch/_export/exported_program.py

  Lines 345 to 349 in c41a32a

  	def _process_constraints(
  	graph_module: torch.fx.GraphModule,
  	num_lifted_params_buffers: int,
  	example_inputs: List[torch.Tensor],
  	) -> Tuple[Dict[sympy.Symbol, ValueRanges], List[Tuple[InputDim, InputDim]]]:

• However, in the example provided, we still see dict after pytree.tree_flatten
  

  pytorch/torch/_export/__init__.py

  Line 388 in c41a32a

  flat_args, _ = pytree.tree_flatten(combine_args_kwargs(args, kwargs))

  
  Here, flat_args is: [{'input_ids': tensor([[ 101, 5672, 2033, 2011, 2151, 3793, 2017, 1005, 1040, 2066, 1012, 102]]), 'token_type_ids': tensor([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]), 'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]])}] which looks like doesn't meet the assumption of _process_constraints

[zhxchen17]

@Xia-Weiwen currently we don't provide compatibility guarantee on capture_pre_autograd_graph(), instead we will support the same thing through torch.export.export very soon and delete this function.

[Xia-Weiwen]

@zhxchen17 Thanks for the heads-up. capture_pre_autograd_graph is used in the current quantization flow. Looks like we need to switch to torch.export.export in the future. Please let us know when it is ready. Thanks.

CC @leslie-fang-intel

[zhxchen17]

no problem! cc @tugsbayasgalan who's actually working on this

[Xia-Weiwen]

Close as capture_pre_autograd_graph is deprecated.