diff --git a/docs/design/attention_backends.md b/docs/design/attention_backends.md index 65e93657e7..d10c23038a 100644 --- a/docs/design/attention_backends.md +++ b/docs/design/attention_backends.md @@ -106,6 +106,7 @@ Priority is **1 = highest** (tried first). | 2 | `FLASH_ATTN` | | 3 | `TRITON_ATTN` | | 4 | `FLEX_ATTENTION` | +| 5 | `TURBOQUANT` | **Ampere/Hopper (SM 8.x-9.x):** @@ -115,6 +116,7 @@ Priority is **1 = highest** (tried first). | 2 | `FLASHINFER` | | 3 | `TRITON_ATTN` | | 4 | `FLEX_ATTENTION` | +| 5 | `TURBOQUANT` | ### MLA Attention (DeepSeek-style) diff --git a/vllm/platforms/cuda.py b/vllm/platforms/cuda.py index bbbd9af0cf..d79d319182 100644 --- a/vllm/platforms/cuda.py +++ b/vllm/platforms/cuda.py @@ -131,6 +131,7 @@ def _get_backend_priorities( AttentionBackendEnum.FLASH_ATTN, AttentionBackendEnum.TRITON_ATTN, AttentionBackendEnum.FLEX_ATTENTION, + AttentionBackendEnum.TURBOQUANT, ] else: return [ @@ -138,6 +139,7 @@ def _get_backend_priorities( AttentionBackendEnum.FLASHINFER, AttentionBackendEnum.TRITON_ATTN, AttentionBackendEnum.FLEX_ATTENTION, + AttentionBackendEnum.TURBOQUANT, ] @@ -255,11 +257,6 @@ class CudaPlatformBase(Platform): valid_backends_priorities = [] invalid_reasons: dict[AttentionBackendEnum, tuple[int, list[str]]] = {} - # TurboQuant KV cache: route directly to TQ backend - kv_cache_dtype = attn_selector_config.kv_cache_dtype - if kv_cache_dtype is not None and kv_cache_dtype.startswith("turboquant_"): - return [(AttentionBackendEnum.TURBOQUANT, 0)], {} - backend_priorities = _get_backend_priorities( attn_selector_config.use_mla, device_capability,