The ARM CPU path for i2_s (1.58-bit ternary) produces constant, prompt-independent output (@@@@) on AArch64, while the x86 AVX2 path is correct and Metal aborts on tensor type 36. This affects builds using the pinned 3rdparty/llama.cpp submodule (isHuangXin/llama.cpp @ release-bitnet-embedding-0.6b-270m).
Root cause is in the ARM fallback of ggml_vec_dot_i2_i8_s in ggml/src/ggml-cpu/quants.c: it unpacked the 2-bit weights with a sequential layout instead of the interleaved layout the dequantizer and AVX2 path use, remapped the ternary codes instead of keeping the raw 0/1/2 the caller subtracts against, and only handled one of the two vec-dot calling conventions. Confirmed by reverting one file at a time on a clean tree (reverting quants.c brings the constant output back).
Fix (2 files, quants.c + ggml-cpu.c) submitted upstream to the pinned fork:
isHuangXin/llama.cpp#6
Tested on Apple M2, same i2_s GGUF before and after: coherent prompt-dependent output restored (5 of 5 distinct prompts distinct), about 30 tok/s generation vs about 2 tok/s on the scalar path, and 48 of 48 exact-integer parity against an independent scalar reference. This lines up with the open issues #468 (to_float UB) and #585 (i2_s layout).
Credit: the degeneracy was first isolated in a two-node verification pass by John Reed (github.com/jar9, Meridian North). Once the submodule pin picks up the fix by commit, the BitNet ARM CPU path is restored.
The ARM CPU path for i2_s (1.58-bit ternary) produces constant, prompt-independent output (
@@@@) on AArch64, while the x86 AVX2 path is correct and Metal aborts on tensor type 36. This affects builds using the pinned3rdparty/llama.cppsubmodule (isHuangXin/llama.cpp @release-bitnet-embedding-0.6b-270m).Root cause is in the ARM fallback of
ggml_vec_dot_i2_i8_singgml/src/ggml-cpu/quants.c: it unpacked the 2-bit weights with a sequential layout instead of the interleaved layout the dequantizer and AVX2 path use, remapped the ternary codes instead of keeping the raw 0/1/2 the caller subtracts against, and only handled one of the two vec-dot calling conventions. Confirmed by reverting one file at a time on a clean tree (reverting quants.c brings the constant output back).Fix (2 files,
quants.c+ggml-cpu.c) submitted upstream to the pinned fork:isHuangXin/llama.cpp#6
Tested on Apple M2, same i2_s GGUF before and after: coherent prompt-dependent output restored (5 of 5 distinct prompts distinct), about 30 tok/s generation vs about 2 tok/s on the scalar path, and 48 of 48 exact-integer parity against an independent scalar reference. This lines up with the open issues #468 (to_float UB) and #585 (i2_s layout).
Credit: the degeneracy was first isolated in a two-node verification pass by John Reed (github.com/jar9, Meridian North). Once the submodule pin picks up the fix by commit, the BitNet ARM CPU path is restored.