Skip to content

Commit

Permalink
fix : Converge the Ring-LL region at TUNER_MAX_RANKS
Browse files Browse the repository at this point in the history
Currently the Ring-LL region for AG0x0 and RS0x0 starts with
the point {0, 16} and extends till {TUNER_MAX_SIZE, TUNER_MAX_RANKS}.
But the polygon was not completely closed without a point on the y-axis.
Added a points {0, TUNER_MAX_RANKS}. This ensures the polygon is closed
correctly.

Signed-off-by: Arun Karthik <akkart@amazon.com>
(cherry picked from commit fbae0f9)
  • Loading branch information
arunkarthik-akkart committed Dec 5, 2024
1 parent 8231c44 commit 1779045
Showing 1 changed file with 4 additions and 4 deletions.
8 changes: 4 additions & 4 deletions src/tuner/nccl_ofi_regions.c
Original file line number Diff line number Diff line change
Expand Up @@ -356,8 +356,8 @@ static ncclResult_t region_init_internal_p5en(nccl_ofi_tuner_region_context_t *r
const nccl_ofi_tuner_region_t regions[] = {
{.algorithm = NCCL_ALGO_RING,
.protocol = NCCL_PROTO_LL,
.num_vertices = 6,
.vertices = {{0, 16}, {131072, 16}, {262144, 32}, {8388608, 256}, {33554432, 1024}, extended_ring_ll}},
.num_vertices = 7,
.vertices = {{0, 16}, {131072, 16}, {262144, 32}, {8388608, 256}, {33554432, 1024}, extended_ring_ll, {0, TUNER_MAX_RANKS}}},
{.algorithm = NCCL_ALGO_RING,
.protocol = NCCL_PROTO_LL128,
.num_vertices = 10,
Expand Down Expand Up @@ -399,8 +399,8 @@ static ncclResult_t region_init_internal_p5en(nccl_ofi_tuner_region_context_t *r
const nccl_ofi_tuner_region_t regions[] = {
{.algorithm = NCCL_ALGO_RING,
.protocol = NCCL_PROTO_LL,
.num_vertices = 6,
.vertices = {{0, 16}, {131072, 16}, {262144, 32}, {8388608, 256}, {33554432, 1024}, extended_ring_ll}},
.num_vertices = 7,
.vertices = {{0, 16}, {131072, 16}, {262144, 32}, {8388608, 256}, {33554432, 1024}, extended_ring_ll, {0, TUNER_MAX_RANKS}}},
{.algorithm = NCCL_ALGO_RING,
.protocol = NCCL_PROTO_LL128,
.num_vertices = 10,
Expand Down

0 comments on commit 1779045

Please sign in to comment.