happyz
/
llama.cpp
mirror of https://github.com/ggerganov/llama.cpp.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

metal : extend mul_mv_ext to BF16, Q2_K, Q3_K (#20250) 

Enable mul_mv_ext small-batch kernels (BS 2-8) for BF16, Q2_K,
and Q3_K quantization types. These types previously fell through
to the slower single-row mul_mv path.

BF16 uses the float4 dequantize path (like F16). Q2_K and Q3_K
use the float4x4 K-quant path (like Q4_K/Q5_K/Q6_K).

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
This commit is contained in:
Paul Flynn 2026-03-09 10:48:12 -04:00 committed by GitHub
parent 96cfc4992c
commit e22cd0aa15
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
2 changed files with 20 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

3
ggml/src/ggml-metal/ggml-metal-ops.cpp
View File

@ -1963,6 +1963,7 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
(
op->src[0]->type == GGML_TYPE_F32 || // TODO: helper function
op->src[0]->type == GGML_TYPE_F16 ||
op->src[0]->type == GGML_TYPE_BF16 ||
op->src[0]->type == GGML_TYPE_Q4_0 ||
op->src[0]->type == GGML_TYPE_Q4_1 ||
op->src[0]->type == GGML_TYPE_Q5_0 ||
@ -1977,6 +1978,8 @@ int ggml_metal_op_mul_mat(ggml_metal_op_t ctx, int idx) {
op->src[0]->type == GGML_TYPE_Q4_K ||
op->src[0]->type == GGML_TYPE_Q5_K ||
op->src[0]->type == GGML_TYPE_Q6_K ||
op->src[0]->type == GGML_TYPE_Q2_K ||
op->src[0]->type == GGML_TYPE_Q3_K ||
false) && (ne11 >= 4 && ne11 <= 8)
)
)

17
ggml/src/ggml-metal/ggml-metal.metal
View File

@ -3481,6 +3481,13 @@ template [[host_name("kernel_mul_mv_ext_f16_f32_r1_3")]] kernel mul_mv_ext_q4
template [[host_name("kernel_mul_mv_ext_f16_f32_r1_4")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, half4, 4, dequantize_f16_t4>;
template [[host_name("kernel_mul_mv_ext_f16_f32_r1_5")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, half4, 4, dequantize_f16_t4>;
#if defined(GGML_METAL_HAS_BF16)
template [[host_name("kernel_mul_mv_ext_bf16_f32_r1_2")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, bfloat4, 4, dequantize_bf16_t4>;
template [[host_name("kernel_mul_mv_ext_bf16_f32_r1_3")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, bfloat4, 4, dequantize_bf16_t4>;
template [[host_name("kernel_mul_mv_ext_bf16_f32_r1_4")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, bfloat4, 4, dequantize_bf16_t4>;
template [[host_name("kernel_mul_mv_ext_bf16_f32_r1_5")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<5, bfloat4, 4, dequantize_bf16_t4>;
#endif
template [[host_name("kernel_mul_mv_ext_q4_0_f32_r1_2")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<2, block_q4_0, 32, dequantize_q4_0_t4>;
template [[host_name("kernel_mul_mv_ext_q4_0_f32_r1_3")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<3, block_q4_0, 32, dequantize_q4_0_t4>;
template [[host_name("kernel_mul_mv_ext_q4_0_f32_r1_4")]] kernel mul_mv_ext_q4_f32_t kernel_mul_mv_ext_q4_f32_disp<4, block_q4_0, 32, dequantize_q4_0_t4>;
@ -3531,6 +3538,16 @@ template [[host_name("kernel_mul_mv_ext_q6_K_f32_r1_3")]] kernel mul_mv_ext_q4x4
template [[host_name("kernel_mul_mv_ext_q6_K_f32_r1_4")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<4, block_q6_K, 256, dequantize_q6_K>;
template [[host_name("kernel_mul_mv_ext_q6_K_f32_r1_5")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<5, block_q6_K, 256, dequantize_q6_K>;
template [[host_name("kernel_mul_mv_ext_q2_K_f32_r1_2")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<2, block_q2_K, 256, dequantize_q2_K>;
template [[host_name("kernel_mul_mv_ext_q2_K_f32_r1_3")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<3, block_q2_K, 256, dequantize_q2_K>;
template [[host_name("kernel_mul_mv_ext_q2_K_f32_r1_4")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<4, block_q2_K, 256, dequantize_q2_K>;
template [[host_name("kernel_mul_mv_ext_q2_K_f32_r1_5")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<5, block_q2_K, 256, dequantize_q2_K>;
template [[host_name("kernel_mul_mv_ext_q3_K_f32_r1_2")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<2, block_q3_K, 256, dequantize_q3_K>;
template [[host_name("kernel_mul_mv_ext_q3_K_f32_r1_3")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<3, block_q3_K, 256, dequantize_q3_K>;
template [[host_name("kernel_mul_mv_ext_q3_K_f32_r1_4")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<4, block_q3_K, 256, dequantize_q3_K>;
template [[host_name("kernel_mul_mv_ext_q3_K_f32_r1_5")]] kernel mul_mv_ext_q4x4_f32_t kernel_mul_mv_ext_q4x4_f32_disp<5, block_q3_K, 256, dequantize_q3_K>;
template<typename T0, typename T1, short NR0, typename args_t>
void kernel_mul_mv_t_t_impl(
args_t args,