From 998714b21b413c78db8eb7af7f384dc90c0b10dc Mon Sep 17 00:00:00 2001 From: Xin Yang <105740670+xyang16@users.noreply.github.com> Date: Mon, 18 May 2026 01:32:46 -0700 Subject: [PATCH] [Perf] Add do_not_specialize in fused FP8 RoPE kernel (#42849) Signed-off-by: Xin Yang --- .../attention/ops/deepseek_v4_ops/fused_inv_rope_fp8_quant.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/vllm/v1/attention/ops/deepseek_v4_ops/fused_inv_rope_fp8_quant.py b/vllm/v1/attention/ops/deepseek_v4_ops/fused_inv_rope_fp8_quant.py index 68d33f1aa1..c650eabd00 100644 --- a/vllm/v1/attention/ops/deepseek_v4_ops/fused_inv_rope_fp8_quant.py +++ b/vllm/v1/attention/ops/deepseek_v4_ops/fused_inv_rope_fp8_quant.py @@ -14,7 +14,7 @@ from vllm.triton_utils import tl, triton from vllm.utils.torch_utils import direct_register_custom_op -@triton.jit +@triton.jit(do_not_specialize=["num_tokens"]) def _fused_inv_rope_fp8_quant_per_head( o_ptr, positions_ptr,