SYCL: Flash attention (-fa) produces corrupted output on Intel Arc iGPU (Arrow Lake Xe2)

[jzupnick]

Environment

• Hardware: Intel Arrow Lake-H system with integrated Arc Xe2 graphics
• OS: Ubuntu 24.04
• Runtime: IPEX-LLM XPU inference container
• Backend: SYCL / Level Zero
• oneAPI: Bundled with the container

Note: intel/ipex-llm was archived on 2026-01-28, so filing here against the upstream SYCL backend.

Description

When enabling flash attention (-fa / --flash-attn) with llama-server on integrated Intel Arc Xe2 graphics via the SYCL backend, the model loads and serves requests without an error or warning, but generated content is corrupted.

Without -fa, the same model, prompt, and configuration produce correct output.

Reproduction

llama-server
--model qwen2.5-14b-instruct-q4_0.gguf
-ngl 999 -c 16384
--jinja -b 2048 -ub 512
-fa
--cont-batching -np 1

Then send a standard chat-completions request to the local server.

Expected

Coherent model output.

Actual

The server returns HTTP 200 with valid JSON, but the generated content is corrupted, including random tokens, broken Unicode, and unrelated text fragments.

Removing -fa produces correct output with the otherwise same configuration.

Related behavior

• Quantized KV cache using q8_0 also segfaults on this SYCL backend and may depend on the same flash-attention path.
• Mixed f16/q8_0 KV cache configuration also segfaults.
• The -fa flag is accepted without warning even when output becomes corrupted.

Suggestion

1. If flash attention is unsupported or unreliable for this SYCL/Xe2 configuration, reject -fa or emit a clear warning rather than silently producing incorrect output.
2. Similarly, reject incompatible quantized KV cache configurations when the required flash-attention path is not functional.

This may affect other Intel Xe2/SYCL configurations as well, but I have not tested them.

[NeoZhangJianyu]

@jzupnick
IPEX-LLM is another project which is close source project.
The llama.cpp included in it is also close source.
We don't support it.

Since this project is end of life, I don't who can help you.

Thank you!