]> git.djapps.eu Git - pkg/ggml/sources/ggml/commit
CANN: Resolve soft_max precision issue (llama/15730)
authorhipudding <redacted>
Tue, 2 Sep 2025 09:12:37 +0000 (17:12 +0800)
committerGeorgi Gerganov <redacted>
Fri, 5 Sep 2025 09:54:11 +0000 (12:54 +0300)
commit0db3a9b20a8d6c945fcf96f1ea87e8107ce910a7
treeb22845f70170b36b445bee5513365fbbe4456a68
parent3f53ddab61ff8cedca002b65bc6f340903872050
CANN: Resolve soft_max precision issue (llama/15730)

Previously, the slope tensor was set to fp16 to improve efficiency.
While this worked correctly in FA, it caused precision issues in soft_max.
This change applies different data types for different operators
to balance both accuracy and performance.
src/ggml-cann/aclnn_ops.cpp