[Bug] Include DSA variant in exact-bucket graph admission (#38992)
This commit is contained in:
@@ -705,6 +705,11 @@ class DecodeCudaGraphRunner(BaseCudaGraphRunner):
|
|||||||
cuda_graph_bs,
|
cuda_graph_bs,
|
||||||
stream_idx=get_current_stream_idx() if self.enable_pdmux else None,
|
stream_idx=get_current_stream_idx() if self.enable_pdmux else None,
|
||||||
variant_label=self._resolve_lora_variant(forward_batch),
|
variant_label=self._resolve_lora_variant(forward_batch),
|
||||||
|
dsa_variant=(
|
||||||
|
self._resolve_dsa_variant(forward_batch)
|
||||||
|
if self.disable_padding
|
||||||
|
else None
|
||||||
|
),
|
||||||
)
|
)
|
||||||
|
|
||||||
is_bs_supported = (
|
is_bs_supported = (
|
||||||
|
|||||||
Reference in New Issue
Block a user