support fa3 rope3d (#3622)

2025-10-05 16:48:03 +08:00 · 2025-08-27 11:31:29 +08:00
parent 85afa72763
commit ad319a87cc
4 changed files with 17 additions and 7 deletions
--- a/fastdeploy/model_executor/layers/attention/flash_attn_backend.py
+++ b/fastdeploy/model_executor/layers/attention/flash_attn_backend.py
@@ -311,6 +311,7 @@ class FlashAttentionBackend(AttentionBackend):
                metadata.kv_token_num_cpu[0].item(),
                self.max_seq_len,
                getattr(layer, "cache_quant_type_str", "none"),
+                self.rope_3d,
            )

            res_encoder = self.flash_attn_func(