]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
overview / pkg / ggml / sources / llama.cpp / commit
summary | shortlog | log | commit | commitdiff | tree
(parent: 97bdd26)
convert_hf : faster lazy safetensors (#8482)
authorcompilade <redacted>
Tue, 16 Jul 2024 03:13:10 +0000 (23:13 -0400)
committerGitHub <redacted>
Tue, 16 Jul 2024 03:13:10 +0000 (23:13 -0400)
commit7acfd4e8d55082c1b597dfc3ffe04fb5d530c6dc
tree3f07adaca1f91617dab2542de5c404a28e3ff2e9tree
parent97bdd26eee11fe109dec00de75690ceef61c03f2commit | diff
convert_hf : faster lazy safetensors (#8482)

* convert_hf : faster lazy safetensors

This makes '--dry-run' much, much faster.

* convert_hf : fix memory leak in lazy MoE conversion

The '_lazy' queue was sometimes self-referential,
which caused reference cycles of objects old enough
to avoid garbage collection until potential memory exhaustion.
convert_hf_to_gguf.py diff | blob | history
gguf-py/gguf/lazy.py diff | blob | history
gguf-py/gguf/tensor_mapping.py diff | blob | history
Packaging of ggml-org/llama.cpp