Vulkan backend inference is non-deterministic with stabilityai/sd-vae-ft-mse exported via ExecuTorch (no quantization used) #15344
Description
🐛 Describe the bug
Summary
I exported stabilityai/sd-vae-ft-mse (Stable Diffusion VAE) with ExecuTorch and run it with the Vulkan backend on Android.
Even though I did not enable any quantization, the model’s reconstruction output changes between runs with the same input. I initially suspected FP16 casting or a quantization side effect, but the original Hugging Face model is not FP16-only, and I didn’t apply quantization in export or runtime.
Additionally:
I modified the model so that the attention in the middle block is replaced with a skip (bypass).
In a Python (PyTorch) environment, this modified model does not exhibit non-determinism and never outputs all-zeros; it only shows a ≈0.1 dB performance drop (e.g., PSNR) compared to the original.
I’d like to confirm:
Whether the Vulkan backend may implicitly cast FP32 → FP16 at any point, and
Why the decoding output flips between all-zeros and a normal-looking image across runs.
Model
Base model: stabilityai/sd-vae-ft-mse
Modification: middle block attention replaced with a skip/bypass.
Confirmed on Hugging Face that default weights are not forced to FP16.
Exported with ExecuTorch; no quantization enabled.
What I did
Load an image normalized to [-1, 1] and feed it into the exported VAE encoder.
Compute reparameterization latents (mean/std logged; also sample via uniform→quantile mapping).
Feed latents into the decoder.
Repeatedly run the exact same pipeline (same inputs, same seeds/config).
Expected behavior
Deterministic (or at least stable) reconstruction given the same inputs, with no hidden quantization if I didn’t request it.
Actual behavior
Run A: The reconstruction is all zeros (min=max=mean=std=0).
Run B: The reconstruction looks normal (non-zero stats), using the exact same inputs and code path.
Encoder/latent stats are stable across runs; the divergence appears at/after decoding.
Logs
Two back-to-back runs with identical inputs (trimmed; timestamps differ).
You can see encoder input and latent stats match, but the final reconstruction differs:
Notes / Questions
I did not enable ExecuTorch quantization or any PTQ/QAT recipes.
Is the Vulkan delegate performing implicit FP16 execution or down-casting on devices with FP16-favored paths?
Are there known non-deterministic kernels or uninitialized buffer issues in the Vulkan backend that could produce an all-zero output intermittently on decoder passes?
Any recommended flags to force FP32 end-to-end (or to disable FP16 fast-math/relaxed precision) when using Vulkan with ExecuTorch?
Versions
Collecting environment information...
PyTorch version: 2.10.0.dev20251015+cpu
Is debug build: False
CUDA used to build PyTorch: Could not collect
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: 10.0.0-4ubuntu1
CMake version: version 3.31.6
Libc version: glibc-2.31
Python version: 3.10.19 | packaged by conda-forge | (main, Oct 13 2025, 14:08:27) [GCC 14.3.0] (64-bit runtime)
Python platform: Linux-5.15.0-139-generic-x86_64-with-glibc2.31
Is CUDA available: False
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: N/A
GPU models and configuration: GPU 0: NVIDIA RTX A6000
Nvidia driver version: 570.133.20
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_adv_infer.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_adv_train.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_infer.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_train.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so.8.9.6
/usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so.8.9.6
Is XPU available: False
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: 43 bits physical, 48 bits virtual
CPU(s): 96
On-line CPU(s) list: 0-95
Thread(s) per core: 1
Core(s) per socket: 48
Socket(s): 2
NUMA node(s): 2
Vendor ID: AuthenticAMD
CPU family: 23
Model: 49
Model name: AMD EPYC 7642 48-Core Processor
Stepping: 0
Frequency boost: enabled
CPU MHz: 1500.000
CPU max MHz: 2300.0000
CPU min MHz: 1500.0000
BogoMIPS: 4600.34
Virtualization: AMD-V
L1d cache: 3 MiB
L1i cache: 3 MiB
L2 cache: 48 MiB
L3 cache: 512 MiB
NUMA node0 CPU(s): 0-47
NUMA node1 CPU(s): 48-95
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 Reg file data sampling: Not affected
Vulnerability Retbleed: Mitigation; untrained return thunk; SMT disabled
Vulnerability Spec rstack overflow: Mitigation; SMT disabled
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; Retpolines; IBPB conditional; STIBP disabled; RSB filling; PBRSB-eIBRS Not affected; BHI Not affected
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 mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 sse4_1 sse4_2 movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip rdpid overflow_recov succor smca sme sev sev_es
Versions of relevant libraries:
[pip3] executorch==1.1.0a0+4421558
[pip3] numpy==2.2.6
[pip3] pytorch_tokenizers==0.1.0
[pip3] torch==2.10.0.dev20251015+cpu
[pip3] torchao==0.14.0+git01849b2b1
[pip3] torchaudio==2.8.0.dev20251015+cpu
[pip3] torchdata==0.11.0
[pip3] torchsr==1.0.4
[pip3] torchtune==0.6.1
[pip3] torchvision==0.25.0.dev20251015+cpu
[conda] executorch 1.1.0a0+4421558 pypi_0 pypi
[conda] numpy 2.2.6 pypi_0 pypi
[conda] pytorch-tokenizers 0.1.0 pypi_0 pypi
[conda] torch 2.10.0.dev20251015+cpu pypi_0 pypi
[conda] torchao 0.14.0+git01849b2b1 pypi_0 pypi
[conda] torchaudio 2.8.0.dev20251015+cpu pypi_0 pypi
[conda] torchdata 0.11.0 pypi_0 pypi
[conda] torchsr 1.0.4 pypi_0 pypi
[conda] torchtune 0.6.1 pypi_0 pypi
[conda] torchvision 0.25.0.dev20251015+cpu pypi_0 pypi