]> git.djapps.eu Git - pkg/ggml/sources/llama.cpp/commit
overview / pkg / ggml / sources / llama.cpp / commit
summary | shortlog | log | commit | commitdiff | tree
(parent: 86a9c79)
llama-quant : exclude i32 ffn_gate_tid2eid routing table from quantization (#25787)
authorYash Raj Pandey <redacted>
Sat, 18 Jul 2026 11:43:18 +0000 (07:43 -0400)
committerGitHub <redacted>
Sat, 18 Jul 2026 11:43:18 +0000 (13:43 +0200)
commit4937ca83f4f3da63004943fe05d8aa4f0217d238
tree81ccafdb9926a3276c28fdb6fbdeb4bf8b528aaftree
parent86a9c79f866799eb0e7e89c03578ccfbcc5d808ecommit | diff
llama-quant : exclude i32 ffn_gate_tid2eid routing table from quantization (#25787)

DeepSeek-V4's ffn_gate_tid2eid tensor is an i32 token-id -> expert-id
index table, not weights. It was never added to the name-based
exclusion list alongside ffn_gate_inp.weight, so llama-quantize tries
to quantize it and fails since i32 cannot convert to a float type.

Fixes ggml-org/llama.cpp#25754

Signed-off-by: Yash Raj Pandey <redacted>
src/llama-quant.cpp diff | blob | history
Packaging of ggml-org/llama.cpp