[HiCache] feat: truncate mamba prefetch length to available host KV size (#26945)
This commit is contained in:
@@ -1735,6 +1735,16 @@ class HiMambaRadixCache(MambaRadixCache):
|
||||
prefetch_length,
|
||||
self.evict_host,
|
||||
)
|
||||
if host_indices is None:
|
||||
# truncate the prefetch length to the page-aligned available host size
|
||||
available_size = self.cache_controller.mem_pool_host.available_size()
|
||||
prefetch_length = available_size - (available_size % self.page_size)
|
||||
if prefetch_length < self.prefetch_threshold:
|
||||
self._release_host_node(last_host_node, release_mamba=False)
|
||||
return
|
||||
new_input_tokens = new_input_tokens[:prefetch_length]
|
||||
host_indices = self.cache_controller.mem_pool_host.alloc(prefetch_length)
|
||||
|
||||
if host_indices is None:
|
||||
self._release_host_node(last_host_node, release_mamba=False)
|
||||
return
|
||||
|
||||
Reference in New Issue
Block a user