EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[Bugfix][kvcache] revert multiple kv cache groups (#923)

Browse Source 
Revert multiple kv cache groups related changes as this feature is
reverted in vllm https://github.com/vllm-project/vllm/pull/18459

---------

Signed-off-by: MengqingCao <cmq0113@163.com>
This commit is contained in:
Mengqing Cao
2025-05-22 15:15:33 +08:00
committed by GitHub
parent b4d6672d01
commit 7aa4f85f10
3 changed files with 34 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
vllm_ascend/attention/mla_v1.py
View File

@@ -16,7 +16,6 @@ from vllm.model_executor.layers.rotary_embedding import RotaryEmbedding
from vllm_ascend.attention.attention_v1 import AscendAttentionState
from vllm_ascend.ops.attention import vanilla_chunked_prefill_mla
from vllm_ascend.utils import vllm_version_is
from vllm_ascend.worker.model_runner_v1 import NPUModelRunner
if TYPE_CHECKING:
@@ -239,12 +238,8 @@ class AscendMLAMetadataBuilder:
# function. We should avoid GPU -> CPU sync as much as possible because
# it blocks on all previous kernels.
device = self.runner.device
if vllm_version_is("0.8.5") or vllm_version_is("0.8.5.post1"):
block_table = (self.runner.input_batch.block_table.
get_device_tensor()[:num_reqs])
else:
block_table = (self.runner.input_batch.block_table[0].
get_device_tensor()[:num_reqs])
block_table = (
self.runner.input_batch.block_table.get_device_tensor()[:num_reqs])
slot_mapping = self.runner.slot_mapping_cpu[:num_actual_tokens].to(
device, non_blocking=True)
input_positions = self.runner.positions_cpu[:num_actual_tokens].to(