Move mem_fraction_static adjustment for multimodal models to server_args.py & Fix session control & Other cleanups (#7748)
This commit is contained in:
@@ -54,7 +54,7 @@ class SessionReqNode:
|
||||
prefix += " -- " + self.childs[0].req.rid
|
||||
ret = self.childs[0]._str_helper(prefix)
|
||||
for child in self.childs[1:]:
|
||||
prefix = " " * len(origin_prefix) + r" \- " + child.req.rid
|
||||
prefix = " " * len(origin_prefix) + " \- " + child.req.rid
|
||||
ret += child._str_helper(prefix)
|
||||
return ret
|
||||
|
||||
@@ -106,14 +106,22 @@ class Session:
|
||||
last_req.origin_input_ids
|
||||
+ last_req.output_ids[: last_req.sampling_params.max_new_tokens]
|
||||
)
|
||||
|
||||
if session_params.drop_previous_output:
|
||||
input_ids = last_req.origin_input_ids[:]
|
||||
|
||||
if session_params.offset and session_params.offset != 0:
|
||||
input_ids = input_ids[: session_params.offset] + req.input_ids
|
||||
else:
|
||||
input_ids += req.input_ids
|
||||
|
||||
input_ids_unpadded = (
|
||||
last_req.origin_input_ids_unpadded
|
||||
+ last_req.output_ids[: last_req.sampling_params.max_new_tokens]
|
||||
)
|
||||
if session_params.drop_previous_output:
|
||||
input_ids_unpadded = last_req.origin_input_ids_unpadded[:]
|
||||
|
||||
if session_params.offset and session_params.offset != 0:
|
||||
input_ids_unpadded = (
|
||||
input_ids_unpadded[: session_params.offset] + req.input_ids
|
||||
@@ -138,10 +146,11 @@ class Session:
|
||||
token_ids_logprob=req.token_ids_logprob,
|
||||
)
|
||||
if last_req is not None:
|
||||
new_req.multimodal_inputs = last_req.mm_inputs
|
||||
new_req.multimodal_inputs = last_req.multimodal_inputs
|
||||
new_req.tokenizer = tokenizer
|
||||
|
||||
if abort:
|
||||
new_req.to_abort = True
|
||||
new_req.set_finish_with_abort("Invalid request session id")
|
||||
else:
|
||||
new_req_node = SessionReqNode(new_req, last_req_node)
|
||||
self.req_nodes[req.rid] = new_req_node
|
||||
|
||||
Reference in New Issue
Block a user