[release/2.10] Some models inductor performance dropped ~ 10% - 30%

[mengfei25]

🐛 Describe the bug

Inductor latency (2.9/2.10)	Eager latency (2.9/2.10)	Model	Category
0.72610501	1.025510577	hf_T5	torchbench_amp_fp16_training
0.73099026	1.03495716	T5ForConditionalGeneration	huggingface_amp_fp16_training
0.735776757	1.041661775	T5Small	huggingface_amp_fp16_training
0.737908893	1.022629232	hf_T5_large	torchbench_amp_fp16_training
0.73895893	1.101044035	hf_Bert_large	torchbench_amp_fp16_training
0.740390088	1.014260415	hf_T5	torchbench_amp_bf16_training
0.744329785	1.009002422	T5Small	huggingface_amp_bf16_training
0.744863203	1.021090637	hf_T5_base	torchbench_amp_fp16_training
0.745682418	1.020768778	T5ForConditionalGeneration	huggingface_amp_bf16_training
0.756941149	0.997778599	hf_T5_base	torchbench_amp_bf16_training
0.766096062	0.999174179	hf_T5_large	torchbench_amp_bf16_training
0.777368912	1.03257717	RobertaForCausalLM	huggingface_amp_fp16_training
0.781435989	1.067483312	hf_Bert_large	torchbench_amp_bf16_training
0.781447491	1.072060631	hf_Bert	torchbench_amp_fp16_training
0.790601984	1.015495208	nvidia_deeprecommender	torchbench_bfloat16_training
0.792600022	1.029893407	nvidia_deeprecommender	torchbench_float16_training
0.793479222	1.026854581	RobertaForCausalLM	huggingface_amp_bf16_training
0.797416575	1.065162481	hf_Bert	torchbench_amp_bf16_training
0.817952188	1.058941128	hf_DistilBert	torchbench_amp_fp16_training
0.821282005	1.027905051	nvidia_deeprecommender	torchbench_float32_training
0.828940716	0.936525048	cm3leon_generate	torchbench_amp_bf16_inference
0.832581429	1.012749187	cait_m36_384	timm_models_amp_fp16_training
0.832634247	1.019993632	hf_DistilBert	torchbench_amp_bf16_training
0.834760191	1.003724431	MegatronBertForCausalLM	huggingface_amp_bf16_training
0.835253899	1.080106732	BertForMaskedLM	huggingface_amp_fp16_training
0.83819159	0.997527805	cait_m36_384	timm_models_amp_bf16_training
0.838499532	1.033842331	MegatronBertForCausalLM	huggingface_amp_fp16_training
0.839248456	0.917759385	cm3leon_generate	torchbench_bfloat16_inference
0.84319389	1.01168299	MegatronBertForQuestionAnswering	huggingface_amp_bf16_training
0.854570411	1.080008059	BertForMaskedLM	huggingface_amp_bf16_training
0.854633055	1.046595918	MegatronBertForCausalLM	huggingface_float16_training
0.854841305	1.035040232	MegatronBertForQuestionAnswering	huggingface_amp_fp16_training
0.857927257	1.077568317	hf_Bert_large	torchbench_float16_training
0.858227582	1.085429821	nfnet_l0	timm_models_amp_bf16_training
0.860412414	1.045948918	hf_Bert_large	torchbench_bfloat16_training
0.86220791	1.058394598	RobertaForCausalLM	huggingface_float16_training
0.862822993	1.10097243	nfnet_l0	timm_models_amp_fp16_training
0.865090223	1.063423975	hf_Bert	torchbench_bfloat16_training
0.86656158	1.013820158	MegatronBertForCausalLM	huggingface_bfloat16_training
0.867096966	1.063023382	MegatronBertForQuestionAnswering	huggingface_float16_training
0.867672465	1.058526803	MegatronBertForQuestionAnswering	huggingface_bfloat16_training
0.869067316	1.007590129	hf_T5	torchbench_bfloat16_inference
0.869479568	1.032627504	RobertaForCausalLM	huggingface_bfloat16_training
0.870200174	1.035579506	LayoutLMForMaskedLM	huggingface_amp_fp16_training
0.871785239	1.007962757	hf_T5_large	torchbench_bfloat16_training
0.872581665	1.012940341	hf_T5	torchbench_float16_inference
0.875032542	1.084723548	timm_nfnet	torchbench_amp_bf16_training
0.876608108	1.077426002	hf_Bert	torchbench_float16_training
0.876685353	1.00108782	PLBartForConditionalGeneration	huggingface_amp_bf16_training
0.876690161	1.015495415	LayoutLMForMaskedLM	huggingface_amp_bf16_training
0.88035635	1.101150998	dm_nfnet_f0	timm_models_amp_fp16_training
0.882346621	1.021002982	hf_DistilBert	torchbench_bfloat16_training
0.884984897	1.012550166	volo_d1_224	timm_models_amp_bf16_inference
0.888364965	1.102525905	nfnet_l0	timm_models_bfloat16_training
0.889071522	1.097869058	BertForMaskedLM	huggingface_float16_training
0.890791743	1.021974958	LayoutLMForMaskedLM	huggingface_bfloat16_training
0.891880454	1.059520152	LayoutLMForMaskedLM	huggingface_float16_training
0.893603208	1.088030044	BertForQuestionAnswering	huggingface_amp_fp16_training
0.89649246	0.999278129	hf_T5	torchbench_bfloat16_training
0.896948944	1.073180417	tnt_s_patch16_224	timm_models_float16_inference
0.899363278	1.074808448	jx_nest_base	timm_models_amp_fp16_inference

Reproducer

docker run -it --device=/dev/mem --device=/dev/dri --group-add video --privileged --shm-size=8g intelgpu/ubuntu-24.04-lts2:2523.31 bash

apt update
apt install python3-dev python3-venv python3-pip
python3 -m venv myvenv
source myvenv/bin/activate

pip install torch==2.10.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/test/xpu
# pip install torch==2.9.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/xpu

git clone https://github.com/pytorch/pytorch
cd pytorch
git checkout v2.10.0-rc1
git clone https://github.com/intel/torch-xpu-ops
cp torch-xpu-ops/.ci/benchmarks/timm_models_list.txt benchmarks/dynamo/timm_models_list.txt

pip install timm==1.0.19 pandas psutil scipy

python benchmarks/dynamo/timm_models.py --training --bfloat16 -d xpu -n10 --performance --disable-cudagraphs --batch-size 4 --only beit_base_patch16_224 --cold-start-latency --backend=inductor

Versions

View detailed

Collecting environment information...
PyTorch version: 2.10.0+xpu
Is debug build: False
CUDA used to build PyTorch: None
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.5 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04.2) 11.4.0
Clang version: Could not collect
CMake version: version 3.22.1
Libc version: glibc-2.35

Python version: 3.10.19 (main, Oct 10 2025, 01:02:36) [GCC 11.4.0] (64-bit runtime)
Python platform: Linux-5.15.0-153-generic-x86_64-with-glibc2.35
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
Is XPU available: True
XPU used to build PyTorch: 20250301
Intel GPU driver version:

• intel-opencl-icd: 25.18.33578.38-1146~22.04
• libze1: 1.21.9.0-1136~22.04
  Intel GPU models onboard:
  N/A
  Intel GPU models detected:
• [0] _XpuDeviceProperties(name='Intel(R) Data Center GPU Max 1100', platform_name='Intel(R) oneAPI Unified Runtime over Level-Zero', type='gpu', device_id=0xBDA, uuid=8680da0b-2f00-0000-0f00-000000000000, driver_version='1.6.33578+38', total_memory=49136MB, max_compute_units=448, gpu_eu_count=448, gpu_subslice_count=56, max_work_group_size=1024, max_num_sub_groups=64, sub_group_sizes=[16 32], has_fp16=1, has_fp64=1, has_atomic64=1)
• [1] _XpuDeviceProperties(name='Intel(R) Data Center GPU Max 1100', platform_name='Intel(R) oneAPI Unified Runtime over Level-Zero', type='gpu', device_id=0xBDA, uuid=8680da0b-2f00-0000-1600-000000000000, driver_version='1.6.33578+38', total_memory=49136MB, max_compute_units=448, gpu_eu_count=448, gpu_subslice_count=56, max_work_group_size=1024, max_num_sub_groups=64, sub_group_sizes=[16 32], has_fp16=1, has_fp64=1, has_atomic64=1)
• [2] _XpuDeviceProperties(name='Intel(R) Data Center GPU Max 1100', platform_name='Intel(R) oneAPI Unified Runtime over Level-Zero', type='gpu', device_id=0xBDA, uuid=8680da0b-2f00-0000-1a00-000000000000, driver_version='1.6.33578+38', total_memory=49136MB, max_compute_units=448, gpu_eu_count=448, gpu_subslice_count=56, max_work_group_size=1024, max_num_sub_groups=64, sub_group_sizes=[16 32], has_fp16=1, has_fp64=1, has_atomic64=1)
• [3] _XpuDeviceProperties(name='Intel(R) Data Center GPU Max 1100', platform_name='Intel(R) oneAPI Unified Runtime over Level-Zero', type='gpu', device_id=0xBDA, uuid=8680da0b-2f00-0000-1e00-000000000000, driver_version='1.6.33578+38', total_memory=49136MB, max_compute_units=448, gpu_eu_count=448, gpu_subslice_count=56, max_work_group_size=1024, max_num_sub_groups=64, sub_group_sizes=[16 32], has_fp16=1, has_fp64=1, has_atomic64=1)
• [4] _XpuDeviceProperties(name='Intel(R) Data Center GPU Max 1100', platform_name='Intel(R) oneAPI Unified Runtime over Level-Zero', type='gpu', device_id=0xBDA, uuid=8680da0b-2f00-0000-8a00-000000000000, driver_version='1.6.33578+38', total_memory=49136MB, max_compute_units=448, gpu_eu_count=448, gpu_subslice_count=56, max_work_group_size=1024, max_num_sub_groups=64, sub_group_sizes=[16 32], has_fp16=1, has_fp64=1, has_atomic64=1)
• [5] _XpuDeviceProperties(name='Intel(R) Data Center GPU Max 1100', platform_name='Intel(R) oneAPI Unified Runtime over Level-Zero', type='gpu', device_id=0xBDA, uuid=8680da0b-2f00-0000-8e00-000000000000, driver_version='1.6.33578+38', total_memory=49136MB, max_compute_units=448, gpu_eu_count=448, gpu_subslice_count=56, max_work_group_size=1024, max_num_sub_groups=64, sub_group_sizes=[16 32], has_fp16=1, has_fp64=1, has_atomic64=1)
• [6] _XpuDeviceProperties(name='Intel(R) Data Center GPU Max 1100', platform_name='Intel(R) oneAPI Unified Runtime over Level-Zero', type='gpu', device_id=0xBDA, uuid=8680da0b-2f00-0000-c000-000000000000, driver_version='1.6.33578+38', total_memory=49136MB, max_compute_units=448, gpu_eu_count=448, gpu_subslice_count=56, max_work_group_size=1024, max_num_sub_groups=64, sub_group_sizes=[16 32], has_fp16=1, has_fp64=1, has_atomic64=1)
• [7] _XpuDeviceProperties(name='Intel(R) Data Center GPU Max 1100', platform_name='Intel(R) oneAPI Unified Runtime over Level-Zero', type='gpu', device_id=0xBDA, uuid=8680da0b-2f00-0000-c400-000000000000, driver_version='1.6.33578+38', total_memory=49136MB, max_compute_units=448, gpu_eu_count=448, gpu_subslice_count=56, max_work_group_size=1024, max_num_sub_groups=64, sub_group_sizes=[16 32], has_fp16=1, has_fp64=1, has_atomic64=1)
  HIP runtime version: N/A
  MIOpen runtime version: N/A
  Is XNNPACK available: True
  Caching allocator config: N/A

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): 192
On-line CPU(s) list: 0-191
Vendor ID: GenuineIntel
Model name: Intel(R) Xeon(R) Platinum 8468V
CPU family: 6
Model: 143
Thread(s) per core: 2
Core(s) per socket: 48
Socket(s): 2
Stepping: 8
Frequency boost: enabled
CPU max MHz: 2401.0000
CPU min MHz: 800.0000
BogoMIPS: 4800.00
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 intel_ppin cdp_l2 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 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 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
Virtualization: VT-x
L1d cache: 4.5 MiB (96 instances)
L1i cache: 3 MiB (96 instances)
L2 cache: 192 MiB (96 instances)
L3 cache: 195 MiB (2 instances)
NUMA node(s): 2
NUMA node0 CPU(s): 0-47,96-143
NUMA node1 CPU(s): 48-95,144-191
Vulnerability Gather data sampling: Not affected
Vulnerability Indirect target selection: 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 Reg file data sampling: Not affected
Vulnerability Retbleed: Not affected
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 / Automatic IBRS; IBPB disabled; PBRSB-eIBRS SW sequence; BHI BHI_DIS_S
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected

Versions of relevant libraries:
[pip3] dpcpp-cpp-rt==2025.3.1
[pip3] impi-rt==2021.17.0
[pip3] intel-cmplr-lib-rt==2025.3.1
[pip3] intel-cmplr-lib-ur==2025.3.1
[pip3] intel-cmplr-lic-rt==2025.3.1
[pip3] intel-opencl-rt==2025.3.1
[pip3] intel-openmp==2025.3.1
[pip3] intel-pti==0.15.0
[pip3] intel-sycl-rt==2025.3.1
[pip3] mkl==2025.3.0
[pip3] numpy==1.26.4
[pip3] oneccl==2021.17.1
[pip3] oneccl-devel==2021.17.1
[pip3] onemkl-license==2025.3.0
[pip3] onemkl-sycl-blas==2025.3.0
[pip3] onemkl-sycl-dft==2025.3.0
[pip3] onemkl-sycl-lapack==2025.3.0
[pip3] onemkl-sycl-rng==2025.3.0
[pip3] onemkl-sycl-sparse==2025.3.0
[pip3] tbb==2022.3.0
[pip3] tcmlib==1.4.1
[pip3] torch==2.10.0+xpu
[pip3] torchao==0.15.0+git252d1db7
[pip3] torchaudio==2.10.0+xpu
[pip3] torchvision==0.25.0+xpu
[pip3] triton-xpu==3.6.0
[pip3] umf==1.0.2
[conda] Could not collect

[weishi-deng]

triage: known issue fixed by pytorch/pytorch#169257

[jianyizh]

from mengfei:

Model	Category	Inductor latency (2.9/bf16 atomic fallback PR+ reduction heuristic fix)
nvidia_deeprecommender	torchbench_bfloat16_training	0.7861918
nvidia_deeprecommender	torchbench_float16_training	0.794994846
cm3leon_generate	torchbench_amp_bf16_inference	0.815268698
nvidia_deeprecommender	torchbench_float32_training	0.827939869
cait_m36_384	timm_models_amp_fp16_training	0.828954833
cm3leon_generate	torchbench_bfloat16_inference	0.831794045
nfnet_l0	timm_models_amp_bf16_training	0.845073016
cait_m36_384	timm_models_amp_bf16_training	0.845101388
nfnet_l0	timm_models_bfloat16_training	0.861476288
timm_nfnet	torchbench_amp_bf16_training	0.863070355
nfnet_l0	timm_models_amp_fp16_training	0.866064097
hf_T5	torchbench_bfloat16_inference	0.87158018
hf_T5	torchbench_float16_inference	0.875738528
dm_nfnet_f0	timm_models_amp_fp16_training	0.876728921
tnt_s_patch16_224	timm_models_float16_inference	0.887588623
basic_gnn_sage	torchbench_amp_bf16_inference	0.895359177
jx_nest_base	timm_models_amp_fp16_inference	0.895739762

[jianyizh]

nvidia_deeprecommender, cait_m36_384, nfnet_l0: cannot reproduce on 1550
cm3leon_generate: #2223

[weishi-deng]

cait_m36_384 training: intel/intel-xpu-backend-for-triton#5750

[jianyizh]

@weishi-deng cait_m36_384 has 2 regressions on 1100, pytorch/pytorch@6a2bd1f and online softmax (TORCHINDUCTOR_ONLINE_SOFTMAX), please have a check

[weishi-deng]

cait_m36_384: fusion changed by pytorch/pytorch@6a2bd1f which changes the triton kernel from 1 pass "triton_red_fused__softmax__to_copy__unsafe_view_add_clone_permute_8.run" to 2-pass triton_red_fused__softmax__to_copy__unsafe_view_add_permute_8 + triton_poi_fused__softmax__to_copy__unsafe_view_add_clone_permute_9. fusion failed with memory "no shared data" for these 2-pass kernels.