fix(fa3): translate page table to SWA loc in EAGLE3 topk>1 spec metadata (#24617)
This commit is contained in:
@@ -2483,10 +2483,20 @@ class FlashAttentionBackend(AttentionBackend):
|
|||||||
else metadata_swa.page_table
|
else metadata_swa.page_table
|
||||||
)
|
)
|
||||||
|
|
||||||
|
page_table_a = metadata.page_table
|
||||||
|
page_table_b = metadata_expand.page_table
|
||||||
|
if self.use_sliding_window_kv_pool:
|
||||||
|
page_table_a = self.token_to_kv_pool.translate_loc_from_full_to_swa(
|
||||||
|
page_table_a
|
||||||
|
)
|
||||||
|
page_table_b = self.token_to_kv_pool.translate_loc_from_full_to_swa(
|
||||||
|
page_table_b
|
||||||
|
)
|
||||||
|
|
||||||
prepare_swa_spec_page_table_triton(
|
prepare_swa_spec_page_table_triton(
|
||||||
page_table,
|
page_table,
|
||||||
metadata.page_table,
|
page_table_a,
|
||||||
metadata_expand.page_table,
|
page_table_b,
|
||||||
metadata.cache_seqlens_int32,
|
metadata.cache_seqlens_int32,
|
||||||
metadata_expand.cache_seqlens_int32,
|
metadata_expand.cache_seqlens_int32,
|
||||||
self.speculative_num_draft_tokens,
|
self.speculative_num_draft_tokens,
|
||||||
|
|||||||
Reference in New Issue
Block a user