XNNPACK weights cache crashes on null pointer

[jgibson2]

Bug

ExecuTorch crashes with SIGSEGV on Android when loading a dynamically quantized model containing bias-less Conv2d layers. The crash occurs during XNNWeightsCache::look_up_or_insert which calls memcmp on a null pointer from the xnn_weights_cache_look_up_key struct.

The same model runs correctly on the macOS Python runtime, suggesting the weights cache is either disabled or uses a different code path there.

Root Cause

When creating a xnn_create_convolution2d_nhwc_qd8_f32_qc8w op (dynamic int8 quantization with per-channel weights), the XNNPACK weights cache builds a lookup key that includes pointers to the raw kernel and bias buffers. For bias-less convolutions, the bias pointer is null. The cache then calls memcmp on this null pointer, causing a segfault.

Stack Trace (Android, aarch64)

Fatal signal 11 (SIGSEGV), code 1 (SEGV_MAPERR), fault addr 0x0

#00 pc 0x693cc   libc.so                (__memcmp_aarch64+12)
#01 pc 0x36be70  libexecutorch_jni.so   (XNNWeightsCache::look_up_or_insert +92)
#02 pc 0x410c68  libexecutorch_jni.so   [XNNPACK weights cache internal]
#03 pc 0x40c99c  libexecutorch_jni.so   [XNNPACK weights packing]
#04 pc 0x40d070  libexecutorch_jni.so   (xnn_create_convolution2d_nhwc_qd8_f32_qc8w +452)
#05 pc 0x409ae4  libexecutorch_jni.so
#06 pc 0x3d8a08  libexecutorch_jni.so
#07 pc 0x368e88  libexecutorch_jni.so   (XNNCompiler::compileModel +1168)
#08 pc 0x36ac8c  libexecutorch_jni.so   (XnnpackBackend::init +340)

Reproduction

1. Export any model with bias-less Conv2d layers (e.g., MobileNetV4 backbone — 379 out of 381 convs have no bias)
2. Apply dynamic quantization:

   from executorch.backends.xnnpack.quantizer.xnnpack_quantizer import (
       XNNPACKQuantizer, get_symmetric_quantization_config,
   )
   quantizer = XNNPACKQuantizer()
   quantizer.set_global(get_symmetric_quantization_config(is_per_channel=True, is_dynamic=True))

3. Lower with XnnpackPartitioner(per_op_mode=True) and export to .pte
4. Load and execute on Android via libexecutorch_jni.so

The model runs fine on macOS Python ExecuTorch runtime.

Workaround

Add zero bias to all bias-less Conv2d modules before export:

for module in model.modules():
    if isinstance(module, torch.nn.Conv2d) and module.bias is None:
        module.bias = torch.nn.Parameter(torch.zeros(module.out_channels))

This is semantically identical (zero bias is a no-op) but ensures the weights cache lookup key always has a non-null bias pointer.

Suggested Fix

XNNWeightsCache::look_up_or_insert should check for null bias pointer before calling memcmp, or the xnn_weights_cache_look_up_key construction should handle the bias-less case (e.g., zero-length memcmp or skip bias comparison).

Environment

• ExecuTorch 1.1.0 (Android JNI build)
• PyTorch 2.9.1 (export)
• Android aarch64
• Dynamic quantization with per-channel int8 weights (qd8_f32_qc8w ops)

[jgibson2]

Updated Root Cause

After further investigation, the null pointer is not from a bias-less convolution. The crash is in XNNWeightsCache::look_up_or_insert:

void* saved_ptr = context->offset_to_addr(context, offset);  // returns nullptr
if (0 == memcmp(ptr, saved_ptr, size)) {  // SEGV

offset_to_addr returns packed_data_ptrs_[offset] which can be nullptr if delete_packed_data previously nulled out that slot. look_up still returns the stale offset because the name-to-metadata mapping is inconsistent, and there's no null check on saved_ptr before the memcmp.

This is a cache staleness bug triggered when per_op_mode=True creates many small delegate calls that share weights. Adding zero bias to bias-less Conv2d layers does not fix it.

The crash specifically occurs with xnn_create_convolution2d_nhwc_qd8_f32_qc8w (dynamic int8 + per-channel weights) on Android libexecutorch_jni.so. The macOS Python runtime does not reproduce, likely because the weights cache code path differs.

Suggested fix: Add a null check on saved_ptr in look_up_or_insert:

void* saved_ptr = context->offset_to_addr(context, offset);
if (saved_ptr != nullptr && 0 == memcmp(ptr, saved_ptr, size)) {
    return offset;
}

[jgibson2]

Confirmed this also reproduces on ExecuTorch 1.2.0 Android prebuilt. The null check fix on saved_ptr in XNNWeightsCache::look_up_or_insert is the only viable path — per-tensor quantization fails at export with missing out variants, and per-channel crashes at runtime on Android.

[kirklandsign]

Does https://github.com/pytorch/executorch/pull/18295/changes fix the issue? Can you try on main branch?

[jgibson2]

Does https://github.com/pytorch/executorch/pull/18295/changes fix the issue? Can you try on main branch?

Are there prebuilt AARs for main? Happy to give that a try if so. Otherwise it will require a bit more effort.