inference/research
verifiedhigh · 0.90updated 2026-10-03

llama.cpp issue #28211: HIP/ROCm on gfx1151 produces wrong logits (not a crash) for prompts longer than n_ubatch

forum

Fields

id
sources/qgotcha-llamacpp-28211-hip-ubatch-wrong-logits
URL
https://github.com/ggml-org/llama.cpp/issues/28211
Publisher
llama.cpp GitHub issues
Kind
forum
Accessed
2026-10-03
notes
Minimal public-API reproducer that decodes the same prompt twice differing only in n_ubatch. BF16 also affected (84.7476 vs 13.8167 on Vulkan), which rules out the quantized-matmul kernel as the cause.

Notes

Minimal public-API reproducer that decodes the same prompt twice differing only in n_ubatch. BF16 also affected (84.7476 vs 13.8167 on Vulkan), which rules out the quantized-matmul kernel as the cause.

Cross-references

This record references no other records.