Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

ggml : add ggml_top_k (#17365)

Browse Source 
* ggml : add ggml_top_k

* cont : add ggml_argsort_top_k

* metal : add top_k support

* ggml : cleanup

* tests : add virtual err() function for test_case

* ggml : add comments
This commit is contained in:
Georgi Gerganov
2025-11-25 15:31:43 +02:00
committed by GitHub
parent 05872ac885
commit 583cb83416
15 changed files with 516 additions and 85 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-metal/ggml-metal-ops.cpp
+143 -17
View File
@@ -406,6 +406,10 @@ static int ggml_metal_op_encode_impl(ggml_metal_op_t ctx, int idx) {
{
n_fuse = ggml_metal_op_argsort(ctx, idx);
} break;
case GGML_OP_TOP_K:
{
n_fuse = ggml_metal_op_top_k(ctx, idx);
} break;
case GGML_OP_LEAKY_RELU:
{
n_fuse = ggml_metal_op_leaky_relu(ctx, idx);
@@ -3678,14 +3682,19 @@ int ggml_metal_op_argsort(ggml_metal_op_t ctx, int idx) {
}
ggml_metal_kargs_argsort args = {
/*.ne00 =*/ ne00,
/*.ne01 =*/ ne01,
/*.ne02 =*/ ne02,
/*.ne03 =*/ ne03,
/*.nb00 =*/ nb00,
/*.nb01 =*/ nb01,
/*.nb02 =*/ nb02,
/*.nb03 =*/ nb03,
/*.ne00 =*/ ne00,
/*.ne01 =*/ ne01,
/*.ne02 =*/ ne02,
/*.ne03 =*/ ne03,
/*.nb00 =*/ nb00,
/*.nb01 =*/ nb01,
/*.nb02 =*/ nb02,
/*.nb03 =*/ nb03,
/*.ne0 =*/ ne0,
/*.ne1 =*/ ne1,
/*.ne2 =*/ ne2,
/*.ne3 =*/ ne3,
/*.top_k =*/ nth,
};
ggml_metal_encoder_set_pipeline(enc, pipeline);
@@ -3705,15 +3714,20 @@ int ggml_metal_op_argsort(ggml_metal_op_t ctx, int idx) {
ggml_metal_op_concurrency_reset(ctx);
ggml_metal_kargs_argsort_merge args_merge = {
.ne00 = ne00,
.ne01 = ne01,
.ne02 = ne02,
.ne03 = ne03,
.nb00 = nb00,
.nb01 = nb01,
.nb02 = nb02,
.nb03 = nb03,
.len = len,
/*.ne00 =*/ ne00,
/*.ne01 =*/ ne01,
/*.ne02 =*/ ne02,
/*.ne03 =*/ ne03,
/*.nb00 =*/ nb00,
/*.nb01 =*/ nb01,
/*.nb02 =*/ nb02,
/*.nb03 =*/ nb03,
/*.ne0 =*/ ne0,
/*.ne1 =*/ ne1,
/*.ne2 =*/ ne2,
/*.ne3 =*/ ne3,
/*.top_k =*/ ne00,
/*.len =*/ len,
};
// merges per row
@@ -3737,6 +3751,118 @@ int ggml_metal_op_argsort(ggml_metal_op_t ctx, int idx) {
return 1;
}
int ggml_metal_op_top_k(ggml_metal_op_t ctx, int idx) {
ggml_tensor * op = ctx->node(idx);
ggml_metal_library_t lib = ctx->lib;
ggml_metal_encoder_t enc = ctx->enc;
GGML_ASSERT(ggml_is_contiguous_rows(op->src[0]));
GGML_TENSOR_LOCALS( int32_t, ne0, op->src[0], ne);
GGML_TENSOR_LOCALS(uint64_t, nb0, op->src[0], nb);
GGML_TENSOR_LOCALS( int32_t, ne, op, ne);
GGML_TENSOR_LOCALS(uint64_t, nb, op, nb);
ggml_metal_pipeline_t pipeline = ggml_metal_library_get_pipeline_top_k(lib, op);
// bitonic sort requires the number of elements to be power of 2
int nth = 1;
while (nth < ne00 && 2*nth <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline)) {
nth *= 2;
}
// blocks per row
const int npr = (ne00 + nth - 1)/nth;
const size_t smem = GGML_PAD(nth*sizeof(int32_t), 16);
ggml_metal_buffer_id bid_src0 = ggml_metal_get_buffer_id(op->src[0]);
ggml_metal_buffer_id bid_dst = ggml_metal_get_buffer_id(op);
ggml_metal_buffer_id bid_tmp = bid_dst;
bid_tmp.offs += sizeof(int32_t)*ggml_nelements(op->src[0]);
if ((int) ceil(std::log(npr) / std::log(2)) % 2 == 1) {
std::swap(bid_dst, bid_tmp);
}
const int top_k = ne0;
ggml_metal_kargs_argsort args = {
/*.ne00 =*/ ne00,
/*.ne01 =*/ ne01,
/*.ne02 =*/ ne02,
/*.ne03 =*/ ne03,
/*.nb00 =*/ nb00,
/*.nb01 =*/ nb01,
/*.nb02 =*/ nb02,
/*.nb03 =*/ nb03,
/*.ne0 =*/ ne0,
/*.ne1 =*/ ne1,
/*.ne2 =*/ ne2,
/*.ne3 =*/ ne3,
/*.top_k =*/ std::min(nth, top_k), // for each block, keep just the top_k indices
};
if (npr > 1) {
args.ne0 = (npr - 1)*args.top_k + std::min(ne00 - (npr - 1)*nth, args.top_k);
}
ggml_metal_encoder_set_pipeline(enc, pipeline);
ggml_metal_encoder_set_bytes (enc, &args, sizeof(args), 0);
ggml_metal_encoder_set_buffer (enc, bid_src0, 1);
ggml_metal_encoder_set_buffer (enc, bid_dst, 2);
ggml_metal_encoder_set_threadgroup_memory_size(enc, smem, 0);
ggml_metal_encoder_dispatch_threadgroups(enc, npr*ne01, ne02, ne03, nth, 1, 1);
ggml_metal_pipeline_t pipeline_merge = ggml_metal_library_get_pipeline_top_k_merge(lib, op);
int len = args.top_k;
while (len < args.ne0) {
ggml_metal_op_concurrency_reset(ctx);
// merges per row
const int nm = (args.ne0 + 2*len - 1) / (2*len);
const int nth = std::min(512, std::min(len, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline_merge)));
ggml_metal_kargs_argsort_merge args_merge = {
/*.ne00 =*/ ne00,
/*.ne01 =*/ ne01,
/*.ne02 =*/ ne02,
/*.ne03 =*/ ne03,
/*.nb00 =*/ nb00,
/*.nb01 =*/ nb01,
/*.nb02 =*/ nb02,
/*.nb03 =*/ nb03,
/*.ne0 =*/ args.ne0,
/*.ne1 =*/ ne1,
/*.ne2 =*/ ne2,
/*.ne3 =*/ ne3,
/*.top_k =*/ nm == 1 ? top_k : args.ne0, // the final merge outputs top_k elements
/*.len =*/ len,
};
ggml_metal_encoder_set_pipeline(enc, pipeline_merge);
ggml_metal_encoder_set_bytes (enc, &args_merge, sizeof(args_merge), 0);
ggml_metal_encoder_set_buffer (enc, bid_src0, 1);
ggml_metal_encoder_set_buffer (enc, bid_dst, 2);
ggml_metal_encoder_set_buffer (enc, bid_tmp, 3);
ggml_metal_encoder_dispatch_threadgroups(enc, nm*ne01, ne02, ne03, nth, 1, 1);
std::swap(bid_dst, bid_tmp);
len <<= 1;
}
return 1;
}
int ggml_metal_op_leaky_relu(ggml_metal_op_t ctx, int idx) {
ggml_tensor * op = ctx->node(idx);