[responsesAPI] support input output messages for non harmony models (#29549)

Signed-off-by: Andrew Xia <axia@fb.com>
Co-authored-by: Andrew Xia <axia@fb.com>
This commit is contained in:
Andrew Xia
2025-12-02 23:59:23 -08:00
committed by GitHub
co-authored by Andrew Xia
parent bbfb55c29e
commit 3a7751485b
4 changed files with 64 additions and 11 deletions
+22
View File
@@ -23,6 +23,7 @@ from vllm.entrypoints.openai.parser.responses_parser import (
)
from vllm.entrypoints.openai.protocol import (
ResponseInputOutputItem,
ResponseRawMessageAndToken,
ResponsesRequest,
)
from vllm.entrypoints.responses_utils import construct_tool_dicts
@@ -148,6 +149,8 @@ def _create_json_parse_error_messages(
class SimpleContext(ConversationContext):
"""This is a context that cannot handle MCP tool calls"""
def __init__(self):
self.last_output = None
self.num_prompt_tokens = 0
@@ -158,6 +161,9 @@ class SimpleContext(ConversationContext):
# not implemented yet for SimpleContext
self.all_turn_metrics = []
self.input_messages: list[ResponseRawMessageAndToken] = []
self.output_messages: list[ResponseRawMessageAndToken] = []
def append_output(self, output) -> None:
self.last_output = output
if not isinstance(output, RequestOutput):
@@ -166,6 +172,22 @@ class SimpleContext(ConversationContext):
self.num_cached_tokens = output.num_cached_tokens or 0
self.num_output_tokens += len(output.outputs[0].token_ids or [])
if len(self.input_messages) == 0:
output_prompt = output.prompt or ""
output_prompt_token_ids = output.prompt_token_ids or []
self.input_messages.append(
ResponseRawMessageAndToken(
message=output_prompt,
tokens=output_prompt_token_ids,
)
)
self.output_messages.append(
ResponseRawMessageAndToken(
message=output.outputs[0].text,
tokens=output.outputs[0].token_ids,
)
)
def append_tool_output(self, output) -> None:
raise NotImplementedError("Should not be called.")