Superminaren/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-11 04:56:56 +02:00
Code Issues Packages Projects Releases Wiki Activity

Adding CDNA, RDNA2 and RDNA4

Browse Source
This commit is contained in:
ravel7524
2026-09-01 13:04:01 +02:00
committed by Carl Philipp Klemm
parent a606a0d849
commit a76a341dcf
1 changed files with 10 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/mmq.cuh
+10 -3
View File
@@ -1472,6 +1472,13 @@ static void launch_mul_mat_q(ggml_backend_cuda_context & ctx, const mmq_args & a
ntx_fd);
}
static bool mmq_use_routed_moe_ncols_picker(const int cc) {
return GGML_CUDA_CC_IS_CDNA(cc) ||
GGML_CUDA_CC_IS_RDNA2(cc) ||
GGML_CUDA_CC_IS_RDNA3(cc) ||
GGML_CUDA_CC_IS_RDNA4(cc);
}
template <ggml_type type, bool fallback>
void mul_mat_q_switch_J(ggml_backend_cuda_context & ctx, const mmq_args & args, cudaStream_t stream) {
const int id = ggml_cuda_get_device();
@@ -1479,10 +1486,10 @@ void mul_mat_q_switch_J(ggml_backend_cuda_context & ctx, const mmq_args & args,
const size_t smpbo = ggml_cuda_info().devices[id].smpbo;
int64_t ncols_picker = args.ncols_max;
if (args.expert_bounds != nullptr && GGML_CUDA_CC_IS_RDNA3(cc) && args.nchannels_x > 0) {
if (args.expert_bounds != nullptr && mmq_use_routed_moe_ncols_picker(cc) && args.nchannels_x > 0) {
// In routed MoE, ncols_max is the worst-case per-expert width. Size the
// MMQ N-tile from the typical routed width while it is below the RDNA3
// max tile width. The launch grid still uses args.ncols_max.
// MMQ N-tile from the typical routed width while it is below the current
// architecture's max tile width. The launch grid still uses args.ncols_max.
const int J_max = ggml_cuda_mmq_get_J_max(type, fallback, cc, 128);
const int64_t ncols_typical = (args.ncols_dst + args.nchannels_x - 1) / args.nchannels_x;
if (ncols_typical >= 1 && ncols_typical < J_max && ncols_typical < ncols_picker) {