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

HIP: Patch failed testcase in WMMA-MMQ kernels for RDNA 4 (#17502)

Browse Source 
* patch failed test case MUL_MAT(type_a=q4_0,type_b=f32,m=576,n=512,k=576,bs=[1,1],nr=[1,1],per=[0,1,2,3],k_v=0,o=1) for enabling WMMA on RDNA4

* Quick clean up on mma.cuh to add ggml_cuda_memcpy_1 back in for half2 and bfloat162
This commit is contained in:
Jiacheng (Jason) Chen
2025-11-26 11:18:48 +01:00
committed by GitHub
parent eeb5605de2
commit 3e18dba9fd
2 changed files with 21 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-cuda/mmq.cuh
+1 -1
View File
@@ -3701,7 +3701,7 @@ static size_t mmq_get_nbytes_shared(const int mmq_x, const int mmq_y, const int
const tile_x_sizes txs = mmq_get_dp4a_tile_x_sizes(type, mmq_y);
const int mmq_tile_x_k = mmq_get_mma_tile_x_k(type);
const size_t nbs_ids = mmq_x*sizeof(int);
const size_t nbs_x = (turing_mma_available(cc) || amd_mfma_available(cc)) ? mmq_y*mmq_tile_x_k*sizeof(int) : txs.qs*sizeof(int) + txs.dm*sizeof(half2) + txs.sc*sizeof(int);
const size_t nbs_x = (turing_mma_available(cc) || amd_mfma_available(cc) || amd_wmma_available(cc)) ? mmq_y*mmq_tile_x_k*sizeof(int) : txs.qs*sizeof(int) + txs.dm*sizeof(half2) + txs.sc*sizeof(int);
const size_t nbs_y = mmq_x*sizeof(block_q8_1_mmq);
return nbs_ids + nbs_x + GGML_PAD(nbs_y, nwarps*warp_size*sizeof(int));
}