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 15, 11:00 UTC
[Bugfix][ROCm] Fix AITER MLA V1 (#17880)
Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
vllmellm committed
May 9, 2025 at 16:38 UTC
217db4baa6466f89d893386ea1cefee176dc45d6
1 file changed
+1
-3
vllm/v1/attention/backends/mla/rocm_aiter_mla.py
+1
-3
@@ -95,8 +95,7 @@ class AiterMLAMetadataBuilder(MLACommonMetadataBuilder[AiterMLAMetadata]):
95
paged_kv_last_page_len,
96
)
97
98
- def _build_decode(self, input_positions: torch.Tensor,
99
- block_table: torch.Tensor,
98
+ def _build_decode(self, block_table: torch.Tensor,
99
seq_lens: torch.Tensor) -> AiterMLADecodeMetadata:
100
101
(
@@ -106,7 +105,6 @@ class AiterMLAMetadataBuilder(MLACommonMetadataBuilder[AiterMLAMetadata]):
105
) = self._get_paged_kv_tensors(block_table, seq_lens)
106
107
attn_metadata = AiterMLADecodeMetadata(
109
- input_positions=input_positions,
108
block_table=block_table,
109
seq_lens=seq_lens,
110
paged_kv_indptr=paged_kv_indptr,