A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 3, 11:45 UTC
[Perf] Add do_not_specialize in fused FP8 RoPE kernel (#42849)
Signed-off-by: Xin Yang <xyangx@amazon.com>
Xin Yang committed
May 18, 2026 at 01:32 UTC
998714b21b413c78db8eb7af7f384dc90c0b10dc
1 file changed
+1
-1
vllm/v1/attention/ops/deepseek_v4_ops/fused_inv_rope_fp8_quant.py
+1
-1
@@ -14,7 +14,7 @@ from vllm.triton_utils import tl, triton
14
from vllm.utils.torch_utils import direct_register_custom_op
15
16
17
-@triton.jit
17
+@triton.jit(do_not_specialize=["num_tokens"])
18
def _fused_inv_rope_fp8_quant_per_head(
19
o_ptr,
20
positions_ptr,