tocmo0nlord/axolotl
1
0
Fork 0
Code Issues Pull Requests Actions 3 Packages Projects Releases Wiki Activity

uniform routing default

Browse Source
This commit is contained in:
Dan Saunders
2025-09-25 15:47:23 -04:00
parent e003a05177
commit 824a641cee
2 changed files with 15 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
scripts/benchmarks/deepseek_v3_moe_sweep.py
View File

@@ -50,9 +50,9 @@ def parse_args() -> argparse.Namespace:
help="Override GROUP_SIZE_M for every configuration",
)
parser.add_argument(
"--uniform-routing",
"--no-uniform-routing",
action="store_true",
help="Force uniform routing for every configuration",
help="Disable uniform routing for every configuration",
)
parser.add_argument(
"--include-mixtral-long",
@@ -79,7 +79,7 @@ def make_namespace(
"warmup": args.warmup,
"iters": args.iters,
"seed": args.seed,
"uniform_routing": args.uniform_routing,
"uniform_routing": not args.no_uniform_routing,
}
)
if args.group_size is not None:
@@ -196,8 +196,9 @@ def main() -> None: # pragma: no cover - utility script
)
rows = []
uniform_flag = not args.no_uniform_routing
print(
f"Running sweep on device={args.device} dtype={args.dtype} backends={BACKENDS} uniform_routing={args.uniform_routing}"
f"Running sweep on device={args.device} dtype={args.dtype} backends={BACKENDS} uniform_routing={uniform_flag}"
)
print(
f"{'model':>10} {'batch':>5} {'seq':>5} {'hidden':>7} {'experts':>7} {'topk':>4} {'groups':>6} {'backend':>8}"