llama.cpp issue #24168: SYCL gibberish output and ggml_sycl_op_mul_mat crash on hybrid models with Arc Pro B60

Fields

id

sources/noncuda-llamacpp-24168-sycl-b60-hybrid-crash

URL

https://github.com/ggml-org/llama.cpp/issues/24168

Publisher

ggml-org/llama.cpp

Kind

forum

Published

2026-06-05

Accessed

2026-10-03

notes

2x Intel Arc Pro B60 (24GB each, no XeLink), Level Zero 1.15.38308+1, Ryzen 7 7700. qwen3next / qwen35 architectures (Qwen3-Coder-Next Q3_K_XL and Q4_K_M) crash in ggml_sycl_op_mul_mat or return empty/gibberish on builds around b9479, while build 8477 (2026-03-23) works. A regression window was pinpointed by bisect rather than by a single bad commit.

Notes

2x Intel Arc Pro B60 (24GB each, no XeLink), Level Zero 1.15.38308+1, Ryzen 7 7700. qwen3next / qwen35 architectures (Qwen3-Coder-Next Q3_K_XL and Q4_K_M) crash in ggml_sycl_op_mul_mat or return empty/gibberish on builds around b9479, while build 8477 (2026-03-23) works. A regression window was pinpointed by bisect rather than by a single bad commit.

Cross-references

This record references no other records.