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 14, 05:30 UTC
[ROCm][AITER][Bugfix] Switch AITER to use PIECEWISE_AND_FULL compilation (#25104)
Signed-off-by: Rohan138 <rohanpotdar138@gmail.com>
Rohan Potdar committed
Sep 18, 2025 at 12:46 UTC
bbdc0f2366997536207abc212fcdae7a1b688159
1 file changed
+1
-1
vllm/v1/attention/backends/rocm_aiter_fa.py
+1
-1
@@ -232,7 +232,7 @@ class AiterFlashAttentionMetadata:
232
233
class AiterFlashAttentionMetadataBuilder(
234
AttentionMetadataBuilder[AiterFlashAttentionMetadata]):
235
- cudagraph_support = AttentionCGSupport.ALWAYS
235
+ cudagraph_support = AttentionCGSupport.UNIFORM_SINGLE_TOKEN_DECODE
236
237
def __init__(self, kv_cache_spec: AttentionSpec, layer_names: list[str],
238
vllm_config: VllmConfig, device: torch.device):