]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
overview / pkg / ggml / sources / llama.cpp / commit
summary | shortlog | log | commit | commitdiff | tree
(parent: 1705d43)
graph: guard iswa kq_mask on its own buffer (#24294)
authorPascal <redacted>
Mon, 8 Jun 2026 17:20:28 +0000 (19:20 +0200)
committerGitHub <redacted>
Mon, 8 Jun 2026 17:20:28 +0000 (19:20 +0200)
commita66d50588b105ef81fe6c2debd19440f5a8ab97a
tree5238410fa17b0af0d39525f9cc5e163fde3c7dcctree
parent1705d434f6b3a7bf69763fe9689b790b46253eddcommit | diff
graph: guard iswa kq_mask on its own buffer (#24294)

A SWA-only draft head (e.g. StepFun MTP) leaves the base sub-cache
empty, so its kq_mask buffer stays null and asserts at load. Guard
each mask on its own buffer in set_input and can_reuse, base and swa.

Co-authored-by: Georgi Gerganov <redacted>
src/llama-graph.cpp diff | blob | history
Packaging of ggml-org/llama.cpp