AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Re-apply unroll=1 and 112 channels for gfx950 (#1706)

Browse Source 
* Reapply "[SRC] Enable unroll=1 for gfx950 (#1602)" (#1667)
This reverts commit a6972c0d09.

* Reapply "[GRAPH] Increase default nChannels to 112 for gfx950 (#1596)" (#1620)
This reverts commit 1a2eca1756.

[ROCm/rccl commit: 12517a957e]
This commit is contained in:
Nilesh M Negi
2025-05-28 14:58:10 -05:00
committed by GitHub
parent 22120c6303
commit 19ed482121
9 changed files with 123 additions and 41 deletions
Download Patch File Download Diff File Expand all files Collapse all files
projects/rccl/src/include/nccl_common.h
+4 -3
View File
@@ -74,9 +74,10 @@ typedef enum {
#define NCCL_ALGO_PROTO_IGNORE -1.0
#define NCCL_NUM_UNROLLS 2 // 2/4
#define NCCL_UNROLL_2 0
#define NCCL_UNROLL_4 1
#define NCCL_NUM_UNROLLS 3 // 1/2/4
#define NCCL_UNROLL_1 0
#define NCCL_UNROLL_2 1
#define NCCL_UNROLL_4 2
#define NCCL_NUM_FLOATS 6 // half/float/double/rccl_bfloat16/rccl_float8/rccl_bfloat8
#endif