[Attention] MLA decode optimizations (#12528)

Signed-off-by: Lucas Wilkinson <lwilkinson@neuralmagic.com>
Signed-off-by: simon-mo <xmo@berkeley.edu>
Co-authored-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>
Co-authored-by: simon-mo <simon.mo@hey.com>
Co-authored-by: Michael Goin <mgoin64@gmail.com>
Co-authored-by: Zhuohan Li <zhuohan123@gmail.com>
Co-authored-by: Tyler Michael Smith <tysmith@redhat.com>
Co-authored-by: Alexander Matveev <59768536+alexm-neuralmagic@users.noreply.github.com>
Co-authored-by: simon-mo <xmo@berkeley.edu>
This commit is contained in:
Lucas Wilkinson
2025-01-30 23:49:37 -08:00
committed by GitHub
co-authored by Woosuk Kwon simon-mo Michael Goin Zhuohan Li Tyler Michael Smith Alexander Matveev simon-mo
parent a1fc18c030
commit cabaf4eff3
31 changed files with 2266 additions and 32 deletions
+4
View File
@@ -289,7 +289,9 @@ class CommonAttentionState(AttentionState):
@contextmanager
def graph_capture(self, max_batch_size: int):
self._is_graph_capturing = True
self._graph_slot_mapping = torch.full((max_batch_size, ),
PAD_SLOT_ID,
dtype=torch.long,
@@ -299,7 +301,9 @@ class CommonAttentionState(AttentionState):
device=self.runner.device)
self._graph_block_tables = torch.from_numpy(
self.runner.graph_block_tables).to(device=self.runner.device)
yield
self._is_graph_capturing = False
del self._graph_slot_mapping
del self._graph_seq_lens