support chunked prefill and fix minor bug

2026-04-14 02:47:35 +08:00
parent 9e8507ef41
commit 8d63a98c03
8 changed files with 65 additions and 53 deletions
@@ -88,7 +88,6 @@ class BlockManager:
            if block.ref_count == 0:
                self._deallocate_block(block_id)
        seq.num_cached_tokens = 0
-        seq.prefilled = False
        seq.block_table.clear()

    def can_append(self, seq: Sequence) -> bool: