This commit is contained in:
Dan Saunders
2025-09-22 22:48:11 -04:00
parent 94cbc6d42d
commit 5b97633faa
2 changed files with 11 additions and 2 deletions

View File

@@ -142,7 +142,11 @@ def benchmark_deepseek_v3(args: argparse.Namespace) -> dict:
raise SystemExit("CUDA requested but not available")
baseline_module = build_module(args)
original_moe = DeepseekV3MoE.moe
original_moe = getattr(
DeepseekV3MoE,
"_axolotl_triton_original_moe",
DeepseekV3MoE.moe,
)
baseline_module.moe = MethodType(original_moe, baseline_module)
state_dict = baseline_module.state_dict()