vulkan: dequantize iq4_xs 4 at a time (#20657)
This commit is contained in:
@@ -554,7 +554,7 @@ void matmul_shaders(bool fp16, MatMulIdType matmul_id_type, bool coopmat, bool c
|
||||
std::string load_vec_quant = "2";
|
||||
if ((tname == "q4_0") || (tname == "q4_1") || (tname == "q5_1") || (tname == "iq1_s") || (tname == "iq1_m") || (tname == "iq2_xxs") || (tname == "iq2_xs") || (tname == "iq2_s"))
|
||||
load_vec_quant = "8";
|
||||
else if ((tname == "q5_0") || (tname == "q8_0") || (tname == "q2_k") || (tname == "q4_k") || (tname == "q5_k") || (tname == "iq3_xxs") || (tname == "iq3_s") || (tname == "iq4_nl") || (tname == "mxfp4"))
|
||||
else if ((tname == "q5_0") || (tname == "q8_0") || (tname == "q2_k") || (tname == "q4_k") || (tname == "q5_k") || (tname == "iq3_xxs") || (tname == "iq3_s") || (tname == "iq4_xs") || (tname == "iq4_nl") || (tname == "mxfp4"))
|
||||
load_vec_quant = "4";
|
||||
|
||||
if (tname == "bf16") {
|
||||
|
||||
Reference in New Issue
Block a user