4 ms·
Huh, cool. I guess that makes a lot of sense with all the success the quantization people have been having. So am I misunderstanding "Tensor type F32 · I32 · B
by gpm 6mo ago
Huh, cool. I guess that makes a lot of sense with all the success the quantization people have been having.
So am I misunderstanding "Tensor type F32 · I32 · BF16" or is it just tagged wrong?
- rockinghigh 6mo agoThe MoE experts are quantized to int4, all other weights like the shared expert weights are excluded from quantization and use bf16.
- liuliu 6mo agoI32 are 8 4-bit value packed into one int32.