[V1][Attention] Split triton_attn in triton-only and rocm specific backends (#24648)

Signed-off-by: Burkhard Ringlein <ngl@zurich.ibm.com>
2025-09-22 17:20:28 +02:00
parent c10101a3eb
commit 175811e3b5
5 changed files with 482 additions and 123 deletions
--- a/vllm/engine/arg_utils.py
+++ b/vllm/engine/arg_utils.py
@@ -1494,6 +1494,7 @@ class EngineArgs:
            "FLEX_ATTENTION",
            "TREE_ATTN",
            "XFORMERS_VLLM_V1",
+            "ROCM_ATTN_VLLM_V1",
        ]
        if (envs.is_set("VLLM_ATTENTION_BACKEND")
                and envs.VLLM_ATTENTION_BACKEND not in V1_BACKENDS):