Merge branch 'master' into gg/flash-attn
This commit is contained in:
Georgi Gerganov
commit
02a645e7b7
94 changed files with 12943 additions and 4640 deletions
151
ggml-metal.m
151
ggml-metal.m
|
|
@ -61,7 +61,11 @@ enum ggml_metal_kernel_type {
|
|||
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS,
|
||||
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS,
|
||||
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS,
|
||||
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S,
|
||||
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S,
|
||||
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S,
|
||||
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL,
|
||||
GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS,
|
||||
GGML_METAL_KERNEL_TYPE_GET_ROWS_I32,
|
||||
GGML_METAL_KERNEL_TYPE_RMS_NORM,
|
||||
GGML_METAL_KERNEL_TYPE_GROUP_NORM,
|
||||
|
|
@ -84,7 +88,11 @@ enum ggml_metal_kernel_type {
|
|||
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32,
|
||||
//GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32,
|
||||
|
|
@ -103,7 +111,11 @@ enum ggml_metal_kernel_type {
|
|||
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32,
|
||||
|
|
@ -119,7 +131,11 @@ enum ggml_metal_kernel_type {
|
|||
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32,
|
||||
|
|
@ -135,7 +151,11 @@ enum ggml_metal_kernel_type {
|
|||
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32,
|
||||
GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32,
|
||||
GGML_METAL_KERNEL_TYPE_ROPE_F32,
|
||||
GGML_METAL_KERNEL_TYPE_ROPE_F16,
|
||||
GGML_METAL_KERNEL_TYPE_ALIBI_F32,
|
||||
|
|
@ -283,6 +303,14 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
|
|||
return NULL;
|
||||
}
|
||||
} else {
|
||||
#if GGML_METAL_EMBED_LIBRARY
|
||||
GGML_METAL_LOG_INFO("%s: using embedded metal library\n", __func__);
|
||||
|
||||
extern const char ggml_metallib_start[];
|
||||
extern const char ggml_metallib_end[];
|
||||
|
||||
NSString * src = [[NSString alloc] initWithBytes:ggml_metallib_start length:(ggml_metallib_end-ggml_metallib_start) encoding:NSUTF8StringEncoding];
|
||||
#else
|
||||
GGML_METAL_LOG_INFO("%s: default.metallib not found, loading from source\n", __func__);
|
||||
|
||||
NSString * sourcePath;
|
||||
|
|
@ -305,6 +333,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
|
|||
GGML_METAL_LOG_ERROR("%s: error: %s\n", __func__, [[error description] UTF8String]);
|
||||
return NULL;
|
||||
}
|
||||
#endif
|
||||
|
||||
@autoreleasepool {
|
||||
// dictionary of preprocessor macros
|
||||
|
|
@ -447,7 +476,11 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
|
|||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS, get_rows_iq2_xxs, true);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS, get_rows_iq2_xs, true);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS, get_rows_iq3_xxs, true);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S, get_rows_iq3_s, true);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S, get_rows_iq2_s, true);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S, get_rows_iq1_s, true);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL, get_rows_iq4_nl, true);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS, get_rows_iq4_xs, true);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GET_ROWS_I32, get_rows_i32, true);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM, rms_norm, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM, group_norm, ctx->support_simdgroup_reduction);
|
||||
|
|
@ -470,7 +503,11 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
|
|||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XXS_F32, mul_mv_iq2_xxs_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_XS_F32, mul_mv_iq2_xs_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32, mul_mv_iq3_xxs_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32, mul_mv_iq3_s_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32, mul_mv_iq2_s_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32, mul_mv_iq1_s_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32, mul_mv_iq4_nl_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32, mul_mv_iq4_xs_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F32_F32, mul_mv_id_f32_f32, ctx->support_simdgroup_reduction);
|
||||
//GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F16, mul_mv_id_f16_f16, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_F16_F32, mul_mv_id_f16_f32, ctx->support_simdgroup_reduction);
|
||||
|
|
@ -489,7 +526,11 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
|
|||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XXS_F32, mul_mv_id_iq2_xxs_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_XS_F32, mul_mv_id_iq2_xs_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32, mul_mv_id_iq3_xxs_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32, mul_mv_id_iq3_s_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32, mul_mv_id_iq2_s_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32, mul_mv_id_iq1_s_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32, mul_mv_id_iq4_nl_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32, mul_mv_id_iq4_xs_f32, ctx->support_simdgroup_reduction);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F32_F32, mul_mm_f32_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_F16_F32, mul_mm_f16_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_Q4_0_F32, mul_mm_q4_0_f32, ctx->support_simdgroup_mm);
|
||||
|
|
@ -505,7 +546,11 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
|
|||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32, mul_mm_iq2_xxs_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32, mul_mm_iq2_xs_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32, mul_mm_iq3_xxs_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32, mul_mm_iq3_s_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32, mul_mm_iq2_s_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32, mul_mm_iq1_s_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32, mul_mm_iq4_nl_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32, mul_mm_iq4_xs_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F32_F32, mul_mm_id_f32_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_F16_F32, mul_mm_id_f16_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_Q4_0_F32, mul_mm_id_q4_0_f32, ctx->support_simdgroup_mm);
|
||||
|
|
@ -521,7 +566,11 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
|
|||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32, mul_mm_id_iq2_xxs_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32, mul_mm_id_iq2_xs_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32, mul_mm_id_iq3_xxs_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32, mul_mm_id_iq3_s_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32, mul_mm_id_iq2_s_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32, mul_mm_id_iq1_s_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32, mul_mm_id_iq4_nl_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32, mul_mm_id_iq4_xs_f32, ctx->support_simdgroup_mm);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F32, rope_f32, true);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ROPE_F16, rope_f16, true);
|
||||
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_ALIBI_F32, alibi_f32, true);
|
||||
|
|
@ -1346,7 +1395,11 @@ static bool ggml_metal_graph_compute(
|
|||
case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XXS_F32].pipeline; break;
|
||||
case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_XS_F32 ].pipeline; break;
|
||||
case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_XXS_F32].pipeline; break;
|
||||
case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ3_S_F32 ].pipeline; break;
|
||||
case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ2_S_F32 ].pipeline; break;
|
||||
case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ1_S_F32 ].pipeline; break;
|
||||
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_NL_F32 ].pipeline; break;
|
||||
case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_IQ4_XS_F32 ].pipeline; break;
|
||||
default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
|
||||
}
|
||||
|
||||
|
|
@ -1481,12 +1534,36 @@ static bool ggml_metal_graph_compute(
|
|||
nth1 = 16;
|
||||
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_XXS_F32].pipeline;
|
||||
} break;
|
||||
case GGML_TYPE_IQ3_S:
|
||||
{
|
||||
nth0 = 4;
|
||||
nth1 = 16;
|
||||
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ3_S_F32].pipeline;
|
||||
} break;
|
||||
case GGML_TYPE_IQ2_S:
|
||||
{
|
||||
nth0 = 4;
|
||||
nth1 = 16;
|
||||
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ2_S_F32].pipeline;
|
||||
} break;
|
||||
case GGML_TYPE_IQ1_S:
|
||||
{
|
||||
nth0 = 4;
|
||||
nth1 = 16;
|
||||
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ1_S_F32].pipeline;
|
||||
} break;
|
||||
case GGML_TYPE_IQ4_NL:
|
||||
{
|
||||
nth0 = 4;
|
||||
nth1 = 16;
|
||||
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_NL_F32].pipeline;
|
||||
} break;
|
||||
case GGML_TYPE_IQ4_XS:
|
||||
{
|
||||
nth0 = 4;
|
||||
nth1 = 16;
|
||||
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_IQ4_XS_F32].pipeline;
|
||||
} break;
|
||||
default:
|
||||
{
|
||||
GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
|
||||
|
|
@ -1519,9 +1596,9 @@ static bool ggml_metal_graph_compute(
|
|||
[encoder setBytes:&r2 length:sizeof(r2) atIndex:17];
|
||||
[encoder setBytes:&r3 length:sizeof(r3) atIndex:18];
|
||||
|
||||
if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
|
||||
src0t == GGML_TYPE_Q5_0 || src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 ||
|
||||
src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_IQ1_S) { // || src0t == GGML_TYPE_Q4_K) {
|
||||
if (src0t == GGML_TYPE_Q4_0 || src0t == GGML_TYPE_Q4_1 ||
|
||||
src0t == GGML_TYPE_Q5_0 || src0t == GGML_TYPE_Q5_1 || src0t == GGML_TYPE_Q8_0 ||
|
||||
src0t == GGML_TYPE_Q2_K || src0t == GGML_TYPE_IQ1_S || src0t == GGML_TYPE_IQ2_S) {
|
||||
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
|
||||
}
|
||||
else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
|
||||
|
|
@ -1529,11 +1606,16 @@ static bool ggml_metal_graph_compute(
|
|||
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
|
||||
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
|
||||
}
|
||||
else if (src0t == GGML_TYPE_IQ3_XXS) {
|
||||
const int mem_size = 256*4+128;
|
||||
else if (src0t == GGML_TYPE_IQ3_XXS || src0t == GGML_TYPE_IQ3_S) {
|
||||
const int mem_size = src0t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
|
||||
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
|
||||
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
|
||||
}
|
||||
else if (src0t == GGML_TYPE_IQ4_NL || src0t == GGML_TYPE_IQ4_XS) {
|
||||
const int mem_size = 32*sizeof(float);
|
||||
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
|
||||
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
|
||||
}
|
||||
else if (src0t == GGML_TYPE_Q4_K) {
|
||||
[encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
|
||||
}
|
||||
|
|
@ -1627,7 +1709,11 @@ static bool ggml_metal_graph_compute(
|
|||
case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XXS_F32].pipeline; break;
|
||||
case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_XS_F32 ].pipeline; break;
|
||||
case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_XXS_F32].pipeline; break;
|
||||
case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ3_S_F32 ].pipeline; break;
|
||||
case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ2_S_F32 ].pipeline; break;
|
||||
case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ1_S_F32 ].pipeline; break;
|
||||
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_NL_F32 ].pipeline; break;
|
||||
case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MM_ID_IQ4_XS_F32 ].pipeline; break;
|
||||
default: GGML_ASSERT(false && "MUL_MAT_ID not implemented");
|
||||
}
|
||||
|
||||
|
|
@ -1765,12 +1851,36 @@ static bool ggml_metal_graph_compute(
|
|||
nth1 = 16;
|
||||
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_XXS_F32].pipeline;
|
||||
} break;
|
||||
case GGML_TYPE_IQ3_S:
|
||||
{
|
||||
nth0 = 4;
|
||||
nth1 = 16;
|
||||
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ3_S_F32].pipeline;
|
||||
} break;
|
||||
case GGML_TYPE_IQ2_S:
|
||||
{
|
||||
nth0 = 4;
|
||||
nth1 = 16;
|
||||
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ2_S_F32].pipeline;
|
||||
} break;
|
||||
case GGML_TYPE_IQ1_S:
|
||||
{
|
||||
nth0 = 4;
|
||||
nth1 = 16;
|
||||
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ1_S_F32].pipeline;
|
||||
} break;
|
||||
case GGML_TYPE_IQ4_NL:
|
||||
{
|
||||
nth0 = 4;
|
||||
nth1 = 16;
|
||||
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_NL_F32].pipeline;
|
||||
} break;
|
||||
case GGML_TYPE_IQ4_XS:
|
||||
{
|
||||
nth0 = 4;
|
||||
nth1 = 16;
|
||||
pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_MUL_MV_ID_IQ4_XS_F32].pipeline;
|
||||
} break;
|
||||
default:
|
||||
{
|
||||
GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src2t);
|
||||
|
|
@ -1819,9 +1929,9 @@ static bool ggml_metal_graph_compute(
|
|||
[encoder setBuffer:id_src_cur offset:offs_src_cur atIndex:23 + j];
|
||||
}
|
||||
|
||||
if (src2t == GGML_TYPE_Q4_0 || src2t == GGML_TYPE_Q4_1 ||
|
||||
src2t == GGML_TYPE_Q5_0 || src2t == GGML_TYPE_Q5_1 || src2t == GGML_TYPE_Q8_0 ||
|
||||
src2t == GGML_TYPE_Q2_K || src2t == GGML_TYPE_IQ1_S) { // || src2t == GGML_TYPE_Q4_K) {
|
||||
if (src2t == GGML_TYPE_Q4_0 || src2t == GGML_TYPE_Q4_1 ||
|
||||
src2t == GGML_TYPE_Q5_0 || src2t == GGML_TYPE_Q5_1 || src2t == GGML_TYPE_Q8_0 ||
|
||||
src2t == GGML_TYPE_Q2_K || src2t == GGML_TYPE_IQ1_S || src2t == GGML_TYPE_IQ2_S) {
|
||||
[encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
|
||||
}
|
||||
else if (src2t == GGML_TYPE_IQ2_XXS || src2t == GGML_TYPE_IQ2_XS) {
|
||||
|
|
@ -1829,11 +1939,16 @@ static bool ggml_metal_graph_compute(
|
|||
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
|
||||
[encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
|
||||
}
|
||||
else if (src2t == GGML_TYPE_IQ3_XXS) {
|
||||
const int mem_size = 256*4+128;
|
||||
else if (src2t == GGML_TYPE_IQ3_XXS || src2t == GGML_TYPE_IQ3_S) {
|
||||
const int mem_size = src2t == GGML_TYPE_IQ3_XXS ? 256*4+128 : 512*4;
|
||||
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
|
||||
[encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
|
||||
}
|
||||
else if (src2t == GGML_TYPE_IQ4_NL || src2t == GGML_TYPE_IQ4_XS) {
|
||||
const int mem_size = 32*sizeof(float);
|
||||
[encoder setThreadgroupMemoryLength:mem_size atIndex:0];
|
||||
[encoder dispatchThreadgroups:MTLSizeMake((ne21 + 3)/4, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
|
||||
}
|
||||
else if (src2t == GGML_TYPE_Q4_K) {
|
||||
[encoder dispatchThreadgroups:MTLSizeMake((ne21 + 3)/4, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
|
||||
}
|
||||
|
|
@ -1875,7 +1990,11 @@ static bool ggml_metal_graph_compute(
|
|||
case GGML_TYPE_IQ2_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XXS].pipeline; break;
|
||||
case GGML_TYPE_IQ2_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_XS ].pipeline; break;
|
||||
case GGML_TYPE_IQ3_XXS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_XXS].pipeline; break;
|
||||
case GGML_TYPE_IQ3_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ3_S ].pipeline; break;
|
||||
case GGML_TYPE_IQ2_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ2_S ].pipeline; break;
|
||||
case GGML_TYPE_IQ1_S: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ1_S ].pipeline; break;
|
||||
case GGML_TYPE_IQ4_NL: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_NL ].pipeline; break;
|
||||
case GGML_TYPE_IQ4_XS: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_IQ4_XS ].pipeline; break;
|
||||
case GGML_TYPE_I32: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_GET_ROWS_I32 ].pipeline; break;
|
||||
default: GGML_ASSERT(false && "not implemented");
|
||||
}
|
||||
|
|
@ -2211,8 +2330,8 @@ static bool ggml_metal_graph_compute(
|
|||
id<MTLComputePipelineState> pipeline = nil;
|
||||
|
||||
switch (order) {
|
||||
case GGML_SORT_ASC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC].pipeline; break;
|
||||
case GGML_SORT_DESC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC].pipeline; break;
|
||||
case GGML_SORT_ORDER_ASC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_ASC].pipeline; break;
|
||||
case GGML_SORT_ORDER_DESC: pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_ARGSORT_F32_I32_DESC].pipeline; break;
|
||||
default: GGML_ASSERT(false);
|
||||
};
|
||||
|
||||
|
|
@ -2776,6 +2895,11 @@ void ggml_backend_metal_log_set_callback(ggml_log_callback log_callback, void *
|
|||
ggml_metal_log_user_data = user_data;
|
||||
}
|
||||
|
||||
static ggml_guid_t ggml_backend_metal_guid(void) {
|
||||
static ggml_guid guid = { 0x81, 0xa1, 0x8b, 0x1e, 0x71, 0xec, 0x79, 0xed, 0x2b, 0x85, 0xdc, 0x8a, 0x61, 0x98, 0x30, 0xe6 };
|
||||
return &guid;
|
||||
}
|
||||
|
||||
ggml_backend_t ggml_backend_metal_init(void) {
|
||||
struct ggml_metal_context * ctx = ggml_metal_init(GGML_DEFAULT_N_THREADS);
|
||||
|
||||
|
|
@ -2786,6 +2910,7 @@ ggml_backend_t ggml_backend_metal_init(void) {
|
|||
ggml_backend_t metal_backend = malloc(sizeof(struct ggml_backend));
|
||||
|
||||
*metal_backend = (struct ggml_backend) {
|
||||
/* .guid = */ ggml_backend_metal_guid(),
|
||||
/* .interface = */ ggml_backend_metal_i,
|
||||
/* .context = */ ctx,
|
||||
};
|
||||
|
|
@ -2794,7 +2919,7 @@ ggml_backend_t ggml_backend_metal_init(void) {
|
|||
}
|
||||
|
||||
bool ggml_backend_is_metal(ggml_backend_t backend) {
|
||||
return backend && backend->iface.get_name == ggml_backend_metal_name;
|
||||
return backend != NULL && ggml_guid_matches(backend->guid, ggml_backend_metal_guid());
|
||||
}
|
||||
|
||||
void ggml_backend_metal_set_n_cb(ggml_backend_t backend, int n_cb) {
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue