Intel Arc GPU

[PlanteAmigor]

PyTorch XPU backward pass crash with Transformer/SDPA on Intel Arc iGPU

Environment:

CPU: Intel Core Ultra 9 285H (Meteor Lake)

GPU: Intel Arc iGPU (8 Xe-core, shared memory, 128GB DDR5)

OS: Linux (Ubuntu 24.04)

PyTorch: 2.12.0+xpu

Intel oneAPI XPU driver: latest

Problem:
I'm experiencing a crash during the backward pass of nn.TransformerEncoderLayer (or F.scaled_dot_product_attention) when running on Intel XPU. The forward pass works fine, but loss.backward() crashes with memory allocation errors or segfaults.

Minimal repro:
python

import torch
import torch.nn as nn

m = nn.TransformerEncoderLayer(2048, 16, batch_first=True).to('xpu')
x = torch.randn(8, 512, 2048, device='xpu')
m(x).sum().backward() # crash

Error message (varies each run, values like -7.9e16 to -5.0e17, looks like integer overflow):
text

RuntimeError: Trying to create tensor with negative dimension -79243236477491020: [-79243236477491020]

Sometimes also:

IndexError: select(): index -1 out of range for tensor of size [0] at dimension 0

In severe cases (e.g., when AMP BF16 is enabled), the entire system freezes and requires a hard reboot — the GPU driver itself crashes, not just the Python process.

Observations:

Same code runs perfectly on CPU (device='cpu').

CNN operations (Conv2d, Linear, BatchNorm) work fine on XPU — only attention backward triggers this.

Forward pass is always fine, only loss.backward() crashes.

Not always reproducible with tiny models (batch=2, hidden=512), but almost guaranteed with larger sizes (batch=8, hidden=2048).

System freeze (driver crash) happens with AMP BF16 enabled.

Things I've tried that didn't help:

Replacing nn.MultiheadAttention with F.scaled_dot_product_attention

AMP BF16 (made it worse — system freeze)

Periodic torch.xpu.empty_cache() + gc.collect() (delays but doesn't prevent)

torch.xpu.synchronize() before/after backward

Question:
Is this a known PyTorch XPU backend bug, an Intel oneAPI driver issue, or something wrong with my setup? Any known fixes or workarounds would be greatly appreciated.