[2.5.0+] Running a compiled model under FlopCounterMode regresses its performance #140909
Description
🐛 Describe the bug
Invoking a compiled model under FlopCounterMode context results in a slower compiled model.
If we run our benchmark before the model is instrumented with FlopCounterMode, then ms_per_iter records a decent latency, benefitting from compilation.
If we run our benchmark after the model is instrumented with FlopCounterMode, then ms_per_iter records a slower time, equal to a non-compiled model.
Here's a pseudocode illustration of what I mean:
from typing import Callable
from torch.utils.flop_counter import FlopCounterMode
from triton.testing import do_bench
def get_flops(f: Callable[[], None]) -> None:
+ ms_per_iter: float = do_bench(f)
flop_counter = FlopCounterMode(display=True)
with flop_counter:
f()
- ms_per_iter: float = do_bench(f)
flop_count: int = flop_counter.get_total_flops()
iters_per_second = 1e3/ms_per_iter
flops: float = iters_per_second * flop_count
print(f"{flops / 1e12} TF/s")
model: Callable[[LongTensor, FloatTensor]
model_c = torch.compile(model)
input_ids = torch.ones(8, 512, dtype=torch.long, device='cuda')
get_flops(lambda: model_c(input_ids)I shared this finding last month on Twitter:
https://x.com/Birchlabs/status/1847369302976188819
I noticed this problem when I upgraded from torch 2.4.x to 2.5.0.
As in, I had to modify my script (benchmark first, count flops later) to get fast FLOP/s on 2.5.0.
I've created a repro:
https://gist.github.com/Birch-san/b661d5e6812559280438a43ae4ff89ff
Fast mode:
python -m scripts.bench_repro --ckpt xxlbenchmarking SDPA...
Module FLOP % Total
------------------------------------------- -------- ---------
SDPAAttn 365.072B 100.00%
- aten.mm 343.597B 94.12%
- aten._scaled_dot_product_cudnn_attention 21.475B 5.88%
SDPAAttn.o_proj 85.899B 23.53%
- aten.mm 85.899B 23.53%
SDPAAttn.qkv_proj 257.698B 70.59%
- aten.mm 257.698B 70.59%
370.9 TFLOP/s
benchmarking SDPA (compiled)...
Module FLOP % Total
-------------------------------------------- -------- ---------
OptimizedModule 365.072B 100.00%
- aten.mm 343.597B 94.12%
- aten._scaled_dot_product_cudnn_attention 21.475B 5.88%
OptimizedModule._orig_mod 365.072B 100.00%
- aten.mm 343.597B 94.12%
- aten._scaled_dot_product_cudnn_attention 21.475B 5.88%
450.8 TFLOP/s
Slow mode (reproduces torch 2.5.0+ bug):
python -m scripts.bench_repro --ckpt xxl --count-flops-earlybenchmarking SDPA...
Module FLOP % Total
------------------------------------------- -------- ---------
SDPAAttn 365.072B 100.00%
- aten.mm 343.597B 94.12%
- aten._scaled_dot_product_cudnn_attention 21.475B 5.88%
SDPAAttn.o_proj 85.899B 23.53%
- aten.mm 85.899B 23.53%
SDPAAttn.qkv_proj 257.698B 70.59%
- aten.mm 257.698B 70.59%
371.6 TFLOP/s
benchmarking SDPA (compiled)...
Module FLOP % Total
-------------------------------------------- -------- ---------
OptimizedModule 365.072B 100.00%
- aten.mm 343.597B 94.12%
- aten._scaled_dot_product_cudnn_attention 21.475B 5.88%
OptimizedModule._orig_mod 365.072B 100.00%
- aten.mm 343.597B 94.12%
- aten._scaled_dot_product_cudnn_attention 21.475B 5.88%
372.2 TFLOP/s
Versions
Collecting environment information...
PyTorch version: 2.5.0
Is debug build: False
CUDA used to build PyTorch: 12.6
ROCM used to build PyTorch: N/A
OS: Ubuntu 22.04.5 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: version 3.30.5
Libc version: glibc-2.35
Python version: 3.10.12 (main, Sep 11 2024, 15:47:36) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-6.5.13-650-3434-22042-coreweave-1-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: 12.6.77
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration:
GPU 0: NVIDIA H100 80GB HBM3
GPU 1: NVIDIA H100 80GB HBM3
Nvidia driver version: 535.216.01
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.9.3.0
/usr/lib/x86_64-linux-gnu/libcudnn_adv.so.9.3.0
/usr/lib/x86_64-linux-gnu/libcudnn_cnn.so.9.3.0
/usr/lib/x86_64-linux-gnu/libcudnn_engines_precompiled.so.9.3.0
/usr/lib/x86_64-linux-gnu/libcudnn_engines_runtime_compiled.so.9.3.0
/usr/lib/x86_64-linux-gnu/libcudnn_graph.so.9.3.0
/usr/lib/x86_64-linux-gnu/libcudnn_heuristic.so.9.3.0
/usr/lib/x86_64-linux-gnu/libcudnn_ops.so.9.3.0
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
Address sizes: 46 bits physical, 57 bits virtual
Byte Order: Little Endian
CPU(s): 128
On-line CPU(s) list: 0-127
Vendor ID: GenuineIntel
Model name: Intel(R) Xeon(R) Platinum 8462Y+
CPU family: 6
Model: 143
Thread(s) per core: 2
Core(s) per socket: 32
Socket(s): 2
Stepping: 8
CPU max MHz: 4100.0000
CPU min MHz: 800.0000
BogoMIPS: 5600.00
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 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 cat_l2 cdp_l3 invpcid_single cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow flexpriority ept vpid ept_ad 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 hfi vnmi 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 ibt amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities
Virtualization: VT-x
L1d cache: 3 MiB (64 instances)
L1i cache: 2 MiB (64 instances)
L2 cache: 128 MiB (64 instances)
L3 cache: 120 MiB (2 instances)
NUMA node(s): 2
NUMA node0 CPU(s): 0,2,4,6,8,10,12,14,16,18,20,22,24,26,28,30,32,34,36,38,40,42,44,46,48,50,52,54,56,58,60,62,64,66,68,70,72,74,76,78,80,82,84,86,88,90,92,94,96,98,100,102,104,106,108,110,112,114,116,118,120,122,124,126
NUMA node1 CPU(s): 1,3,5,7,9,11,13,15,17,19,21,23,25,27,29,31,33,35,37,39,41,43,45,47,49,51,53,55,57,59,61,63,65,67,69,71,73,75,77,79,81,83,85,87,89,91,93,95,97,99,101,103,105,107,109,111,113,115,117,119,121,123,125,127
Vulnerability Gather data sampling: Not affected
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Retbleed: Not affected
Vulnerability Spec rstack overflow: 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 / Automatic IBRS; IBPB conditional; RSB filling; PBRSB-eIBRS SW sequence; BHI BHI_DIS_S
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
Versions of relevant libraries:
[pip3] clip-anytorch==2.6.0
[pip3] dctorch==0.1.2
[pip3] DISTS-pytorch==0.1
[pip3] lovely-numpy==0.2.13
[pip3] mypy-extensions==1.0.0
[pip3] natten==0.17.2.dev0+torch250cu126
[pip3] numpy==1.26.4
[pip3] open_clip_torch==2.29.0
[pip3] pytorch-lightning==2.4.0
[pip3] torch==2.5.0
[pip3] torchaudio==2.5.0
[pip3] torchdiffeq==0.2.4
[pip3] torchmetrics==1.5.1
[pip3] torchsde==0.2.6
[pip3] torchvision==0.20.0
[pip3] triton==3.1.0
[pip3] welford-torch==0.2.4
[conda] Could not collect
cc @Chillee @ezyang @zou3519 @albanD @samdow @chauhang @penguinwu @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @amjames