llama.cpp issue #18808: totally broken for agentic use on Intel dGPUs (Arc Pro B50 Battlemage)
forum
Fields
id
sources/gotcha-llamacpp-issue-18808-intel-dgpu-agentic-stallPublisher
llama.cpp GitHub issues
Kind
forum
Published
2026-01-13
Accessed
2026-10-03
notes
Documents enabling Flash Attention dropping pp from ~300 to ~15-20 tok/s, and SYCL black-screen/OOM at 8k context.
Notes
Documents enabling Flash Attention dropping pp from ~300 to ~15-20 tok/s, and SYCL black-screen/OOM at 8k context.
Cross-references
This record references no other records.