]> git.djapps.eu Git - pkg/ggml/sources/ggml/commit
vulkan: fix NaN issue in flash attention shader (llama/12776)
authorJeff Bolz <redacted>
Sun, 6 Apr 2025 09:03:47 +0000 (04:03 -0500)
committerGeorgi Gerganov <redacted>
Tue, 8 Apr 2025 08:47:46 +0000 (11:47 +0300)
commitb272d9865e64ed2e45f745d6cc5e5f76a737d553
tree68ac6189445aea0b79b671ff2f6b09e1f1f18a60
parentfe9f9730dc9b8514503ea3796601528e6e047fc5
vulkan: fix NaN issue in flash attention shader (llama/12776)

Use -FLT_MAX/2 rather than -inf as the initial value for computing the maximum.
src/ggml-vulkan/vulkan-shaders/flash_attn_cm2.comp