diff --git a/vllm/v1/attention/backends/mla/rocm_aiter_mla.py b/vllm/v1/attention/backends/mla/rocm_aiter_mla.py index e6a64ee85f..977a42c6fe 100644 --- a/vllm/v1/attention/backends/mla/rocm_aiter_mla.py +++ b/vllm/v1/attention/backends/mla/rocm_aiter_mla.py @@ -196,6 +196,7 @@ class AiterMLAMetadataBuilder(MLACommonMetadataBuilder[AiterMLAMetadata]): kv_cache_dtype_str = getattr(vllm_config.cache_config, "cache_dtype", "auto") if kv_cache_dtype_str in ("fp8", "fp8_e4m3", "fp8_e5m2"): kv_cache_dtype_str = "fp8" + q_dtype = dtypes.fp8 else: kv_cache_dtype_str = "bf16" kv_dtype = dtypes.d_dtypes.get(kv_cache_dtype_str, dtypes.bf16)