Qeravio
Canonical AI event

Qualcomm's Li He updates OpenCL for speculative decoding in llama.cpp

The update includes optimizations for speculative decoding and stops writing zeros into padded activation slots.

15 Sept 20261 verified claims1 sources1 observations
What happened

The official source reports this update: b10988: opencl: choose the MoE expert matmul by batch size for speculative decoding/MTP.

Why it matters

This official update documents a development concerning b10988: opencl: choose the MoE expert matmul by batch size for speculative decoding/MTP. Its practical significance depends on the scope and evidence stated by the source.

What to watch next

Read the official source update and verify its stated scope, evidence and timing before acting on it.

Connected knowledge

Entities affected by this event

Llama Cppsubject
Continue this topic
infrastructureHash-cache now only stores weights in RPC protocol.15 Sept 2026infrastructureCUDA 12 and 13 support row-contiguous SUM_ROWS in llama.cpp.15 Sept 2026infrastructureLlama.cpp moves build_arch_graph() function definitions after graph template specializations.15 Sept 2026infrastructureOpenVINO updates GPU MoE inference and stateful decode performance.15 Sept 2026infrastructureLlama.cpp adds generic ssm_scan to OpenCL15 Sept 2026infrastructureHIP fattn-mma now uses fp32 accumulation on MFMA devices.14 Sept 2026
Evidence trail

Sources behind the event

llama.cpp official releasesb10988: opencl: choose the MoE expert matmul by batch size for speculative decoding/MTP
Editorial presentation

Read the public story

Open story