Skip to content

Commit fc18425

Browse files
ggerganovggerganov
authored
ggml : add SSM Metal kernels (ggml-org#8546)
* ggml : add ggml_ssm_conv metal impl * ggml : add ssm_scan metal impl ggml-ci
1 parent 879275a commit fc18425

File tree

4 files changed

+303
-2
lines changed

4 files changed

+303
-2
lines changed

‎ggml/src/ggml-metal.m‎

Lines changed: 122 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -82,6 +82,8 @@
8282
GGML_METAL_KERNEL_TYPE_RMS_NORM,
8383
GGML_METAL_KERNEL_TYPE_GROUP_NORM,
8484
GGML_METAL_KERNEL_TYPE_NORM,
85+
GGML_METAL_KERNEL_TYPE_SSM_CONV_F32,
86+
GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32,
8587
GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32,
8688
GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16,
8789
GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32,
@@ -542,6 +544,8 @@ static void ggml_metal_log(enum ggml_log_level level, const char * format, ...){
542544
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_RMS_NORM, rms_norm, ctx->support_simdgroup_reduction);
543545
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_GROUP_NORM, group_norm, ctx->support_simdgroup_reduction);
544546
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_NORM, norm, true);
547+
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_CONV_F32, ssm_conv_f32, true);
548+
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32, ssm_scan_f32, true);
545549
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F32_F32, mul_mv_f32_f32, ctx->support_simdgroup_reduction);
546550
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F16, mul_mv_f16_f16, ctx->support_simdgroup_reduction);
547551
GGML_METAL_ADD_KERNEL(GGML_METAL_KERNEL_TYPE_MUL_MV_F16_F32, mul_mv_f16_f32, ctx->support_simdgroup_reduction);
@@ -803,6 +807,9 @@ static bool ggml_metal_supports_op(const struct ggml_backend_metal_context * ctx
803807
return false;
804808
}
805809
return ctx->support_simdgroup_mm; // TODO: over-restricted for vec-kernels
810+
case GGML_OP_SSM_CONV:
811+
case GGML_OP_SSM_SCAN:
812+
return true;
806813
case GGML_OP_MUL_MAT:
807814
case GGML_OP_MUL_MAT_ID:
808815
return ctx->support_simdgroup_reduction &&
@@ -1538,6 +1545,121 @@ static enum ggml_status ggml_metal_graph_compute(
15381545
[encoder dispatchThreadgroups:MTLSizeMake(ne00, ne01, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
15391546
}
15401547
} break;
1548+
case GGML_OP_SSM_CONV:
1549+
{
1550+
GGML_ASSERT(src0t == GGML_TYPE_F32);
1551+
GGML_ASSERT(src1t == GGML_TYPE_F32);
1552+
1553+
GGML_ASSERT(ggml_is_contiguous(src0));
1554+
GGML_ASSERT(ggml_is_contiguous(src1));
1555+
1556+
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_CONV_F32].pipeline;
1557+
1558+
[encoder setComputePipelineState:pipeline];
1559+
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
1560+
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
1561+
[encoder setBuffer:id_dst offset:offs_dst atIndex:2];
1562+
[encoder setBytes:&ne00 length:sizeof(ne00) atIndex:3];
1563+
[encoder setBytes:&ne01 length:sizeof(ne01) atIndex:4];
1564+
[encoder setBytes:&ne02 length:sizeof(ne02) atIndex:5];
1565+
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:6];
1566+
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:7];
1567+
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:8];
1568+
[encoder setBytes:&ne10 length:sizeof(ne10) atIndex:9];
1569+
[encoder setBytes:&ne11 length:sizeof(ne11) atIndex:10];
1570+
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:11];
1571+
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:12];
1572+
[encoder setBytes:&ne0 length:sizeof(ne0) atIndex:13];
1573+
[encoder setBytes:&ne1 length:sizeof(ne1) atIndex:14];
1574+
[encoder setBytes:&ne2 length:sizeof(ne2) atIndex:15];
1575+
[encoder setBytes:&nb0 length:sizeof(nb0) atIndex:16];
1576+
[encoder setBytes:&nb1 length:sizeof(nb1) atIndex:17];
1577+
[encoder setBytes:&nb2 length:sizeof(nb2) atIndex:18];
1578+
1579+
[encoder dispatchThreadgroups:MTLSizeMake(ne01, ne1, ne02) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
1580+
} break;
1581+
case GGML_OP_SSM_SCAN:
1582+
{
1583+
struct ggml_tensor * src3 = gf->nodes[i]->src[3];
1584+
struct ggml_tensor * src4 = gf->nodes[i]->src[4];
1585+
struct ggml_tensor * src5 = gf->nodes[i]->src[5];
1586+
1587+
GGML_ASSERT(src3);
1588+
GGML_ASSERT(src4);
1589+
GGML_ASSERT(src5);
1590+
1591+
size_t offs_src3 = 0;
1592+
size_t offs_src4 = 0;
1593+
size_t offs_src5 = 0;
1594+
1595+
id<MTLBuffer> id_src3 = src3 ? ggml_metal_get_buffer(src3, &offs_src3) : nil;
1596+
id<MTLBuffer> id_src4 = src4 ? ggml_metal_get_buffer(src4, &offs_src4) : nil;
1597+
id<MTLBuffer> id_src5 = src5 ? ggml_metal_get_buffer(src5, &offs_src5) : nil;
1598+
1599+
const int64_t ne30 = src3->ne[0]; GGML_UNUSED(ne30);
1600+
const int64_t ne31 = src3->ne[1]; GGML_UNUSED(ne31);
1601+
1602+
const uint64_t nb30 = src3->nb[0];
1603+
const uint64_t nb31 = src3->nb[1];
1604+
1605+
const int64_t ne40 = src4->ne[0]; GGML_UNUSED(ne40);
1606+
const int64_t ne41 = src4->ne[1]; GGML_UNUSED(ne41);
1607+
const int64_t ne42 = src4->ne[2]; GGML_UNUSED(ne42);
1608+
1609+
const uint64_t nb40 = src4->nb[0];
1610+
const uint64_t nb41 = src4->nb[1];
1611+
const uint64_t nb42 = src4->nb[2];
1612+
1613+
const int64_t ne50 = src5->ne[0]; GGML_UNUSED(ne50);
1614+
const int64_t ne51 = src5->ne[1]; GGML_UNUSED(ne51);
1615+
const int64_t ne52 = src5->ne[2]; GGML_UNUSED(ne52);
1616+
1617+
const uint64_t nb50 = src5->nb[0];
1618+
const uint64_t nb51 = src5->nb[1];
1619+
const uint64_t nb52 = src5->nb[2];
1620+
1621+
const int64_t d_state = ne00;
1622+
const int64_t d_inner = ne01;
1623+
const int64_t n_seq_tokens = ne11;
1624+
const int64_t n_seqs = ne02;
1625+
1626+
id<MTLComputePipelineState> pipeline = ctx->kernels[GGML_METAL_KERNEL_TYPE_SSM_SCAN_F32].pipeline;
1627+
1628+
[encoder setComputePipelineState:pipeline];
1629+
[encoder setBuffer:id_src0 offset:offs_src0 atIndex:0];
1630+
[encoder setBuffer:id_src1 offset:offs_src1 atIndex:1];
1631+
[encoder setBuffer:id_src2 offset:offs_src2 atIndex:2];
1632+
[encoder setBuffer:id_src3 offset:offs_src3 atIndex:3];
1633+
[encoder setBuffer:id_src4 offset:offs_src4 atIndex:4];
1634+
[encoder setBuffer:id_src5 offset:offs_src5 atIndex:5];
1635+
[encoder setBuffer:id_dst offset:offs_dst atIndex:6];
1636+
1637+
[encoder setBytes:&d_state length:sizeof(d_state) atIndex:7];
1638+
[encoder setBytes:&d_inner length:sizeof(d_inner) atIndex:8];
1639+
[encoder setBytes:&n_seq_tokens length:sizeof(n_seq_tokens) atIndex:9];
1640+
[encoder setBytes:&n_seqs length:sizeof(n_seqs) atIndex:10];
1641+
1642+
[encoder setBytes:&nb00 length:sizeof(nb00) atIndex:11];
1643+
[encoder setBytes:&nb01 length:sizeof(nb01) atIndex:12];
1644+
[encoder setBytes:&nb02 length:sizeof(nb02) atIndex:13];
1645+
[encoder setBytes:&nb10 length:sizeof(nb10) atIndex:14];
1646+
[encoder setBytes:&nb11 length:sizeof(nb11) atIndex:15];
1647+
[encoder setBytes:&nb12 length:sizeof(nb12) atIndex:16];
1648+
[encoder setBytes:&nb13 length:sizeof(nb13) atIndex:17];
1649+
[encoder setBytes:&nb20 length:sizeof(nb20) atIndex:18];
1650+
[encoder setBytes:&nb21 length:sizeof(nb21) atIndex:19];
1651+
[encoder setBytes:&nb22 length:sizeof(nb22) atIndex:20];
1652+
[encoder setBytes:&nb30 length:sizeof(nb30) atIndex:21];
1653+
[encoder setBytes:&nb31 length:sizeof(nb31) atIndex:22];
1654+
[encoder setBytes:&nb40 length:sizeof(nb40) atIndex:23];
1655+
[encoder setBytes:&nb41 length:sizeof(nb41) atIndex:24];
1656+
[encoder setBytes:&nb42 length:sizeof(nb42) atIndex:25];
1657+
[encoder setBytes:&nb50 length:sizeof(nb50) atIndex:26];
1658+
[encoder setBytes:&nb51 length:sizeof(nb51) atIndex:27];
1659+
[encoder setBytes:&nb52 length:sizeof(nb52) atIndex:28];
1660+
1661+
[encoder dispatchThreadgroups:MTLSizeMake(d_inner, n_seqs, 1) threadsPerThreadgroup:MTLSizeMake(1, 1, 1)];
1662+
} break;
15411663
case GGML_OP_MUL_MAT:
15421664
{
15431665
GGML_ASSERT(ne00 == ne10);

‎ggml/src/ggml-metal.metal‎

Lines changed: 121 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -667,6 +667,127 @@ kernel void kernel_diag_mask_inf_8(
667667
}
668668
}
669669

670+
// ref: ggml.c:ggml_compute_forward_ssm_conv_f32
671+
// TODO: optimize
672+
kernel void kernel_ssm_conv_f32(
673+
device const void * src0,
674+
device const void * src1,
675+
device float * dst,
676+
constant int64_t & ne00,
677+
constant int64_t & ne01,
678+
constant int64_t & ne02,
679+
constant uint64_t & nb00,
680+
constant uint64_t & nb01,
681+
constant uint64_t & nb02,
682+
constant int64_t & ne10,
683+
constant int64_t & ne11,
684+
constant uint64_t & nb10,
685+
constant uint64_t & nb11,
686+
constant int64_t & ne0,
687+
constant int64_t & ne1,
688+
constant int64_t & ne2,
689+
constant uint64_t & nb0,
690+
constant uint64_t & nb1,
691+
constant uint64_t & nb2,
692+
uint3 tgpig[[threadgroup_position_in_grid]],
693+
uint3 tpitg[[thread_position_in_threadgroup]],
694+
uint3 ntg[[threads_per_threadgroup]]) {
695+
const int64_t ir = tgpig.x;
696+
const int64_t i2 = tgpig.y;
697+
const int64_t i3 = tgpig.z;
698+
699+
const int64_t nc = ne10;
700+
const int64_t ncs = ne00;
701+
const int64_t nr = ne01;
702+
const int64_t n_t = ne1;
703+
const int64_t n_s = ne2;
704+
705+
device const float * s = (device const float *) ((device const char *) src0 + ir*nb01 + i2*nb00 + i3*nb02);
706+
device const float * c = (device const float *) ((device const char *) src1 + ir*nb11);
707+
device float * x = (device float *) ((device char *) dst + ir*nb0 + i2*nb1 + i3*nb2);
708+
709+
float sumf = 0.0f;
710+
711+
for (int64_t i0 = 0; i0 < nc; ++i0) {
712+
sumf += s[i0] * c[i0];
713+
}
714+
715+
x[0] = sumf;
716+
}
717+
718+
// ref: ggml.c:ggml_compute_forward_ssm_scan_f32
719+
// TODO: optimize
720+
kernel void kernel_ssm_scan_f32(
721+
device const void * src0,
722+
device const void * src1,
723+
device const void * src2,
724+
device const void * src3,
725+
device const void * src4,
726+
device const void * src5,
727+
device float * dst,
728+
constant int64_t & d_state,
729+
constant int64_t & d_inner,
730+
constant int64_t & n_seq_tokens,
731+
constant int64_t & n_seqs,
732+
constant uint64_t & nb00,
733+
constant uint64_t & nb01,
734+
constant uint64_t & nb02,
735+
constant uint64_t & nb10,
736+
constant uint64_t & nb11,
737+
constant uint64_t & nb12,
738+
constant uint64_t & nb13,
739+
constant uint64_t & nb20,
740+
constant uint64_t & nb21,
741+
constant uint64_t & nb22,
742+
constant uint64_t & nb30,
743+
constant uint64_t & nb31,
744+
constant uint64_t & nb40,
745+
constant uint64_t & nb41,
746+
constant uint64_t & nb42,
747+
constant uint64_t & nb50,
748+
constant uint64_t & nb51,
749+
constant uint64_t & nb52,
750+
uint3 tgpig[[threadgroup_position_in_grid]],
751+
uint3 tpitg[[thread_position_in_threadgroup]],
752+
uint3 ntg[[threads_per_threadgroup]]) {
753+
const int64_t ir = tgpig.x;
754+
const int64_t i3 = tgpig.y;
755+
756+
const int64_t nc = d_state;
757+
const int64_t nr = d_inner;
758+
const int64_t n_t = n_seq_tokens;
759+
const int64_t n_s = n_seqs;
760+
761+
for (int64_t i2 = 0; i2 < n_t; ++i2) {
762+
device const float * s0 = (device const float *) ((device const char *) src0 + ir*nb01 + i3*nb02);
763+
device const float * x = (device const float *) ((device const char *) src1 + ir*nb10 + i2*nb11 + i3*nb12);
764+
device const float * dt = (device const float *) ((device const char *) src2 + ir*nb20 + i2*nb21 + i3*nb22);
765+
device const float * A = (device const float *) ((device const char *) src3 + ir*nb31);
766+
device const float * B = (device const float *) ((device const char *) src4 + i2*nb41 + i3*nb42);
767+
device const float * C = (device const float *) ((device const char *) src5 + i2*nb51 + i3*nb52);
768+
device float * y = (device float *) ((device char *) dst + ir*nb10 + i2*nb11 + i3*nb12); // TODO: do not use src1 strides
769+
device float * s = (device float *) ((device char *) dst + ir*nb01 + i3*nb02 + nb13);
770+
771+
if (i2 > 0) {
772+
s0 = s;
773+
}
774+
775+
// i1 == 0
776+
float dt_soft_plus = dt[0] <= 20.0f ? log(1.0f + exp(dt[0])) : dt[0];
777+
float x_dt = x[0] * dt_soft_plus;
778+
float sumf = 0.0f;
779+
780+
for (int64_t i0 = 0; i0 < nc; ++i0) {
781+
int64_t i = i0;
782+
float state = (s0[i] * exp(dt_soft_plus * A[i])) + (B[i0] * x_dt);
783+
sumf += state * C[i0];
784+
s[i] = state;
785+
}
786+
787+
y[0] = sumf;
788+
}
789+
}
790+
670791
kernel void kernel_norm(
671792
device const void * src0,
672793
device float * dst,

‎ggml/src/ggml.c‎

Lines changed: 2 additions & 2 deletions
Original file line numberDiff line numberDiff line change
@@ -15898,8 +15898,8 @@ static void ggml_compute_forward_ssm_scan_f32(
1589815898
const float * A = (const float *) ((const char *) src3->data + ir0*(src3->nb[1])); // {d_state, d_inner}
1589915899
const float * B = (const float *) ((const char *) src4->data + i2*(src4->nb[1]) + i3*(src4->nb[2])); // {d_state, n_t, n_s}
1590015900
const float * C = (const float *) ((const char *) src5->data + i2*(src5->nb[1]) + i3*(src5->nb[2])); // {d_state, n_t, n_s}
15901-
float * y = (float *) ((char *) dst->data + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s}
15902-
float * s = (float *) ((char *) dst->data + ir0*(src0->nb[1]) + i3*(src0->nb[2]) + src1->nb[3]); // {d_state, d_inner, n_s}
15901+
float * y = ( float *) (( char *) dst->data + ir0*(src1->nb[0]) + i2*(src1->nb[1]) + i3*(src1->nb[2])); // {d_inner, n_t, n_s}
15902+
float * s = ( float *) (( char *) dst->data + ir0*(src0->nb[1]) + i3*(src0->nb[2]) + src1->nb[3]); // {d_state, d_inner, n_s}
1590315903

1590415904
// use the output as the source for the next token-wise iterations
1590515905
if (i2 > 0) { s0 = s; }

‎tests/test-backend-ops.cpp‎

Lines changed: 58 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -949,6 +949,58 @@ struct test_rms_norm : public test_case {
949949
}
950950
};
951951

952+
// GGML_OP_SSM_CONV
953+
struct test_ssm_conv : public test_case {
954+
const ggml_type type;
955+
const std::array<int64_t, 4> ne_a;
956+
const std::array<int64_t, 4> ne_b;
957+
958+
std::string vars() override {
959+
return VARS_TO_STR3(type, ne_a, ne_b);
960+
}
961+
962+
test_ssm_conv(ggml_type type = GGML_TYPE_F32,
963+
std::array<int64_t, 4> ne_a = {10, 10, 10, 1},
964+
std::array<int64_t, 4> ne_b = {3, 3, 1, 1})
965+
: type(type), ne_a(ne_a), ne_b(ne_b) {}
966+
967+
ggml_tensor * build_graph(ggml_context * ctx) override {
968+
ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne_a.data());
969+
ggml_tensor * b = ggml_new_tensor(ctx, type, 4, ne_b.data());
970+
ggml_tensor * out = ggml_ssm_conv(ctx, a, b);
971+
return out;
972+
}
973+
};
974+
975+
// GGML_OP_SSM_SCAN
976+
struct test_ssm_scan : public test_case {
977+
const ggml_type type;
978+
979+
const int64_t d_state;
980+
const int64_t d_inner;
981+
const int64_t n_seq_tokens;
982+
const int64_t n_seqs;
983+
984+
std::string vars() override {
985+
return VARS_TO_STR5(type, d_state, d_inner, n_seq_tokens, n_seqs);
986+
}
987+
988+
test_ssm_scan(ggml_type type = GGML_TYPE_F32,
989+
int64_t d_state = 32, int64_t d_inner = 32, int64_t n_seq_tokens = 32, int64_t n_seqs = 32)
990+
: type(type), d_state(d_state), d_inner(d_inner), n_seq_tokens(n_seq_tokens), n_seqs(n_seqs) {}
991+
992+
ggml_tensor * build_graph(ggml_context * ctx) override {
993+
ggml_tensor * s = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ d_state, d_inner, n_seqs, 1 }.data());
994+
ggml_tensor * x = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ d_inner, n_seq_tokens, n_seqs, 1 }.data());
995+
ggml_tensor * dt = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ d_inner, n_seq_tokens, n_seqs, 1 }.data());
996+
ggml_tensor * A = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ d_state, d_inner, 1 , 1 }.data());
997+
ggml_tensor * B = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ d_state, n_seq_tokens, n_seqs, 1 }.data());
998+
ggml_tensor * C = ggml_new_tensor(ctx, type, 4, std::vector<int64_t>{ d_state, n_seq_tokens, n_seqs, 1 }.data());
999+
ggml_tensor * out = ggml_ssm_scan(ctx, s, x, dt, A, B, C);
1000+
return out;
1001+
}
1002+
};
1003+
9521004
// GGML_OP_MUL_MAT
9531005
struct test_mul_mat : public test_case {
9541006
const ggml_type type_a;
@@ -2240,6 +2292,12 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
22402292
test_cases.emplace_back(new test_rms_norm(GGML_TYPE_F32, {64, 10, 10, 10}, eps));
22412293
}
22422294

2295+
test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {4, 1536, 1, 1}, {4, 1536, 1, 1}));
2296+
test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {8, 1536, 1, 1}, {4, 1536, 1, 1}));
2297+
test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {4, 1536, 4, 1}, {4, 1536, 1, 1}));
2298+
2299+
test_cases.emplace_back(new test_ssm_scan(GGML_TYPE_F32, 16, 1024, 32, 4));
2300+
22432301
#if 1
22442302
for (ggml_type type_a : base_types) {
22452303
for (ggml_type type_b : {GGML_TYPE_F32, GGML_TYPE_F16}) {

0 commit comments

Comments
 (0)