biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[CI] Revert PRs 34818 and 33600 (#34979)

Browse Source
This commit is contained in:
Lucas Wilkinson
2026-02-20 16:25:50 -05:00
committed by GitHub
parent f24b2de3d3
commit aaefc58ee0
16 changed files with 249 additions and 301 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
vllm/model_executor/layers/attention/mla_attention.py
View File

@@ -407,24 +407,17 @@ class MLAAttention(nn.Module, AttentionLayerBase):
)
# Attributes for forward_impl method
self._vllm_config = get_current_vllm_config()
self._chunked_prefill_workspace_size: int | None = None
self.chunked_prefill_workspace_size = (
MLACommonMetadataBuilder.determine_chunked_prefill_workspace_size(
get_current_vllm_config()
)
)
self._decode_concat_quant_fp8_op = _DecodeConcatQuantFP8(
static=True,
group_shape=GroupShape.PER_TENSOR,
compile_native=True,
)
@property
def chunked_prefill_workspace_size(self) -> int:
if self._chunked_prefill_workspace_size is None:
self._chunked_prefill_workspace_size = (
MLACommonMetadataBuilder.determine_chunked_prefill_workspace_size(
self._vllm_config
)
)
return self._chunked_prefill_workspace_size
def forward(
self,
q: torch.Tensor,