From 0c054e2e0dbe594591143a1b0c0e0b6e31559302 Mon Sep 17 00:00:00 2001 From: Mandar Deshpande Date: Wed, 19 Feb 2025 14:41:14 -0800 Subject: [PATCH] Fix embedding_dim for fp8 attention Summary: By default we set `embedding_dim` to be 3072, but its not expected when ad_head` is also provided as an input to the benchmark resulting in a value error. Differential Revision: D69881611 --- tritonbench/operators/fp8_attention/operator.py | 1 - 1 file changed, 1 deletion(-) diff --git a/tritonbench/operators/fp8_attention/operator.py b/tritonbench/operators/fp8_attention/operator.py index 1be39980..dfc5a3a7 100644 --- a/tritonbench/operators/fp8_attention/operator.py +++ b/tritonbench/operators/fp8_attention/operator.py @@ -41,7 +41,6 @@ def parse_op_args(args: List[str]): parser.add_argument( "--embedding-dim", type=int, - default=3072, help="specify embedding dim, embedding dim = n_heads * head_dim", ) parser.add_argument("--n-heads", type=int, default=48, help="Number of heads")