Poor performance of quantized models after torch.compile() #113019
Comments
|
@leslie-fang-intel @jgong5 @Xia-Weiwen could you take a look? |
jerryzh168
added
oncall: quantization
|
Sure, I will take a look soon. |
|
Hi @AlexKoff88, thanks for reporting the issue. You need to put the pytorch/torch/_inductor/compile_fx.py Line 1172 in 7715b47 Here is the script for your reference https://gist.github.com/leslie-fang-intel/0e9d7bcd222e7a9a4db0e7c5fb4a88bc. And another suggestion is, since the warm up run of |
|
Hi @jingxu10, could you kindly help to modify the tutorial: https://pytorch.org/tutorials/prototype/pt2e_quant_ptq_x86_inductor.html to tell user put the |
|
Hi @leslie-fang-intel, thanks for the prompt response. I am able to see the speedup from quantization on Xeon for batch size 50 after applying your suggestions. However, the situation is still different for batch size 1. I guess this is the runtime issue. |
|
@AlexKoff88 glad to hear that. What's the speedup at BS50 and BS1 specifically? |
|
BS=50 BS=1 |
Notice in your report script, you count the warmup run into the elapsed time, have you excluded it with above performance data? Here is the the performance data I got with 4 cores and BS 1: |
|
Yes, I excluded the warmup. The difference is that I use the all the cores available but the performance is much different from what I got. Here is the new script: import time
import torch
import torch.ao.quantization.quantizer.x86_inductor_quantizer as xiq
import torchvision.models as models
from torch._export import capture_pre_autograd_graph
from torch.ao.quantization.quantize_pt2e import convert_pt2e, prepare_pt2e
from torch.ao.quantization.quantizer.x86_inductor_quantizer import X86InductorQuantizer
print(torch._dynamo.list_backends())
# Create the Eager Model
model_name = "resnet18"
model = models.__dict__[model_name](pretrained=True)
# Set the model to eval mode
model = model.eval()
# Create the data, using the dummy data here as an example
traced_bs = 1
x = torch.randn(traced_bs, 3, 224, 224).contiguous(memory_format=torch.channels_last)
example_inputs = (x,)
# Capture the FX Graph to be quantized
with torch.no_grad():
# if you are using the PyTorch nightlies or building from source with the pytorch master,
# use the API of `capture_pre_autograd_graph`
# Note 1: `capture_pre_autograd_graph` is also a short-term API, it will be updated to use the official `torch.export` API when that is ready.
exported_model = capture_pre_autograd_graph(model, example_inputs)
# Note 2: if you are using the PyTorch 2.1 release binary or building from source with the PyTorch 2.1 release branch,
# please use the API of `torch._dynamo.export` to capture the FX Graph.
# exported_model, guards = torch._dynamo.export(
# model,
# *copy.deepcopy(example_inputs),
# aten_graph=True,
# )
with torch.no_grad():
optimized_model = torch.compile(model)
# Running some benchmark
res = optimized_model(*example_inputs) # warmap
iters = 100
count = 0
start_t = time.time()
with torch.no_grad():
for i in range(iters):
res = optimized_model(*example_inputs)
count += 1
print("fp32 elapsed: ", time.time() - start_t, "count: ", count)
quantizer = X86InductorQuantizer()
quantizer.set_global(xiq.get_default_x86_inductor_quantization_config())
prepared_model = prepare_pt2e(exported_model, quantizer)
prepared_model(*example_inputs)
converted_model = convert_pt2e(prepared_model)
with torch.no_grad():
optimized_model = torch.compile(converted_model)
# Running some benchmark
res = optimized_model(*example_inputs) # warmap
count = 0
start_t = time.time()
with torch.no_grad():
for i in range(iters):
res = optimized_model(*example_inputs)
count += 1
print("int8 elapsed: ", time.time() - start_t, "count: ", count) |
|
Hi @AlexKoff88, based on your script:
|
|
thanks, @leslie-fang-intel. Everything works now. I think it makes sense to prepare e2e example that shows the correct flow. Otherwise, it is easy to get incorrect results. I also wonder if it is possible to avoid using |
|
Glad to hear that @AlexKoff88. I will close this issue now. Feel free to let me know if any other question. |
馃悰 Describe the bug
I tried to benchmark the code from the post-training quantization tutorial and noticed significant performance degradation for quantized models vs. non-quantized ones.
Here is the script I used:
Here is the output I got:
The story is similar for batch size 50.
I used 4th Gen Intel Xeon CPU and torch-nightly.
Here is the command that I used to run this script:
TORCHINDUCTOR_FREEZING=1 python pytorch_compile.py.I wonder what the problem is?
Error logs
Unexpected performance results.
Minified repro
from math import inf
import torch
from torch import tensor, device
import torch.fx as fx
import torch._dynamo
from torch._dynamo.testing import rand_strided
from torch._dynamo.debug_utils import run_fwd_maybe_bwd
import torch._dynamo.config
import torch._inductor.config
import torch._functorch.config
torch._dynamo.config.specialize_int = True
torch._dynamo.config.automatic_dynamic_shapes = False
torch._dynamo.config.capture_scalar_outputs = True
torch._dynamo.config.capture_dynamic_output_shape_ops = True
torch._dynamo.config.allow_rnn = True
from torch.nn import *
class Repro(torch.nn.Module):
def init(self):
super().init()
self.L__self___conv1 = Conv2d(3, 64, kernel_size=(7, 7), stride=(2, 2), padding=(3, 3), bias=False)
self.L__self___bn1 = BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.L__self___relu = ReLU(inplace=True)
self.L__self___maxpool = MaxPool2d(kernel_size=3, stride=2, padding=1, dilation=1, ceil_mode=False)
self.getattr_L__self___layer1___0___conv1 = Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
self.getattr_L__self___layer1___0___bn1 = BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.getattr_L__self___layer1___0___relu = ReLU(inplace=True)
self.getattr_L__self___layer1___0___conv2 = Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
self.getattr_L__self___layer1___0___bn2 = BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.getattr_L__self___layer1___1___conv1 = Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
self.getattr_L__self___layer1___1___bn1 = BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.getattr_L__self___layer1___1___relu = ReLU(inplace=True)
self.getattr_L__self___layer1___1___conv2 = Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
self.getattr_L__self___layer1___1___bn2 = BatchNorm2d(64, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.getattr_L__self___layer2___0___conv1 = Conv2d(64, 128, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
self.getattr_L__self___layer2___0___bn1 = BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.getattr_L__self___layer2___0___relu = ReLU(inplace=True)
self.getattr_L__self___layer2___0___conv2 = Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
self.getattr_L__self___layer2___0___bn2 = BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.getattr_L__self___layer2___0___downsample_0 = Conv2d(64, 128, kernel_size=(1, 1), stride=(2, 2), bias=False)
self.getattr_L__self___layer2___0___downsample_1 = BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.getattr_L__self___layer2___1___conv1 = Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
self.getattr_L__self___layer2___1___bn1 = BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.getattr_L__self___layer2___1___relu = ReLU(inplace=True)
self.getattr_L__self___layer2___1___conv2 = Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
self.getattr_L__self___layer2___1___bn2 = BatchNorm2d(128, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.getattr_L__self___layer3___0___conv1 = Conv2d(128, 256, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
self.getattr_L__self___layer3___0___bn1 = BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.getattr_L__self___layer3___0___relu = ReLU(inplace=True)
self.getattr_L__self___layer3___0___conv2 = Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
self.getattr_L__self___layer3___0___bn2 = BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.getattr_L__self___layer3___0___downsample_0 = Conv2d(128, 256, kernel_size=(1, 1), stride=(2, 2), bias=False)
self.getattr_L__self___layer3___0___downsample_1 = BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.getattr_L__self___layer3___1___conv1 = Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
self.getattr_L__self___layer3___1___bn1 = BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.getattr_L__self___layer3___1___relu = ReLU(inplace=True)
self.getattr_L__self___layer3___1___conv2 = Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
self.getattr_L__self___layer3___1___bn2 = BatchNorm2d(256, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.getattr_L__self___layer4___0___conv1 = Conv2d(256, 512, kernel_size=(3, 3), stride=(2, 2), padding=(1, 1), bias=False)
self.getattr_L__self___layer4___0___bn1 = BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.getattr_L__self___layer4___0___relu = ReLU(inplace=True)
self.getattr_L__self___layer4___0___conv2 = Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
self.getattr_L__self___layer4___0___bn2 = BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.getattr_L__self___layer4___0___downsample_0 = Conv2d(256, 512, kernel_size=(1, 1), stride=(2, 2), bias=False)
self.getattr_L__self___layer4___0___downsample_1 = BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.getattr_L__self___layer4___1___conv1 = Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
self.getattr_L__self___layer4___1___bn1 = BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.getattr_L__self___layer4___1___relu = ReLU(inplace=True)
self.getattr_L__self___layer4___1___conv2 = Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1), bias=False)
self.getattr_L__self___layer4___1___bn2 = BatchNorm2d(512, eps=1e-05, momentum=0.1, affine=True, track_running_stats=True)
self.L__self___avgpool = AdaptiveAvgPool2d(output_size=(1, 1))
self.L__self___fc = Linear(in_features=512, out_features=1000, bias=True)
mod = Repro()
def load_args(reader):
buf0 = reader.storage('aee8c913ace6d6ec115d6e3b12687aadf94e8582', 30105600)
reader.tensor(buf0, (50, 3, 224, 224), (150528, 1, 672, 3), is_leaf=True) # L_x_
load_args._version = 0
if name == 'main':
from torch._dynamo.repro.after_dynamo import run_repro
run_repro(mod, load_args, accuracy=False, command='minify',
save_dir='./torch_compile_debug/run_2023_11_06_10_51_32_497128-pid_3389471/minifier/checkpoints', autocast=False, backend=None)
Versions
PyTorch version: 2.2.0.dev20231106+cpu
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A
OS: Ubuntu 20.04.6 LTS (x86_64)
GCC version: (Ubuntu 9.4.0-1ubuntu1~20.04.2) 9.4.0
Clang version: Could not collect
CMake version: version 3.27.5
Libc version: glibc-2.35
Python version: 3.8.10 (default, May 26 2023, 14:05:08) [GCC 9.4.0] (64-bit runtime)
Python platform: Linux-5.17.0-1033-oem-x86_64-with-glibc2.29
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: 143
Model name: Intel(R) Xeon(R) Gold 6430L
Stepping: 7
CPU MHz: 2670.194
CPU max MHz: 3400.0000
CPU min MHz: 800.0000
BogoMIPS: 3800.00
L1d cache: 3 MiB
L1i cache: 2 MiB
L2 cache: 128 MiB
L3 cache: 120 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 Mmio stale data: Not affected
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl
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 tsc_known_freq 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 cat_l2 cdp_l3 invpcid_single cdp_l2 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 avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts hwp hwp_act_window hwp_epp hwp_pkg_req avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
Versions of relevant libraries:
[pip3] numpy==1.24.1
[pip3] torch==2.2.0.dev20231106+cpu
[pip3] torchaudio==2.2.0.dev20231105+cpu
[pip3] torchvision==0.17.0.dev20231105+cpu
[conda] Could not collect
cc @jerryzh168 @jianyuh @raghuramank100 @jamesr66a @vkuzo @jgong5 @Xia-Weiwen @leslie-fang-intel @ezyang @msaroufim @wconstab @bdhirsh @anijain2305 @zou3519
The text was updated successfully, but these errors were encountered: