[Fix] Snapshot req.prefix_indices when the prefix cache is disabled (#34644)
This commit is contained in:
@@ -519,7 +519,7 @@ class SWARadixCache(KVCacheEventMixin, BasePrefixCache):
|
||||
]
|
||||
|
||||
# `req.prefix_indices` will be used in `PrefillAdder::add_chunked_req` later
|
||||
req.prefix_indices = kv_indices
|
||||
req.prefix_indices = kv_indices.to(dtype=torch.int64, copy=True)
|
||||
return
|
||||
|
||||
token_ids = req.get_fill_ids()
|
||||
|
||||
@@ -749,7 +749,7 @@ class UnifiedRadixCache(BasePrefixCache):
|
||||
kv_indices = self.req_to_token_pool.req_to_token[
|
||||
req.req_pool_idx, : len(token_ids)
|
||||
]
|
||||
req.prefix_indices = kv_indices
|
||||
req.prefix_indices = kv_indices.to(dtype=torch.int64, copy=True)
|
||||
return
|
||||
|
||||
kv_indices_orig = self.req_to_token_pool.req_to_token[
|
||||
|
||||
Reference in New Issue
Block a user