PR #37385: [Bugfix] Fix GLM-4 tool parser double serialization in streaming

• Repository: vllm-project/vllm
• Author: karanb192
• State: open | merged: False
• Link: https://github.com/vllm-project/vllm/pull/37385

Description (problem / solution / changelog)

Summary

Fix double serialization bug in GLM-4 tool parser streaming where prev_tool_call_arr stored arguments as a parsed dict instead of a JSON string, causing malformed final delta chunks in streaming tool call responses.

Root Cause

In glm4_moe_tool_parser.py, the streaming finalization code stored arguments as a dict:

args_dict = json.loads(full_args_str)
self.prev_tool_call_arr[self.current_tool_id] = {
    "arguments": args_dict,  # stored as dict -- BUG
}

The serving layer (chat_completion/serving.py) then called json.dumps(args) on this dict, which could produce different JSON formatting than streamed_args_for_tool, causing the replace() logic to fail and emit a spurious cumulative final delta.

Fix

Store arguments as the raw JSON string (matching other tool parsers like Qwen3Coder, DeepSeek-V3.2, Seed-OSS):

json.loads(full_args_str)  # validate only
self.prev_tool_call_arr[self.current_tool_id] = {
    "arguments": full_args_str,  # stored as string
}

The serving layer already handles both str and dict types at lines 1143-1146:

if isinstance(args, str):
    expected_call = args  # use directly, no double serialization
else:
    expected_call = json.dumps(args, ensure_ascii=False)

Test Plan

pytest tests/tool_parsers/test_glm4_moe_tool_parser.py -v
pytest tests/tool_parsers/test_glm47_moe_tool_parser.py -v

Fixes #36833

Changed files

• tests/tool_parsers/test_glm47_moe_tool_parser.py (modified, +4/-1)
• tests/tool_parsers/test_glm4_moe_tool_parser.py (modified, +17/-11)
• vllm/tool_parsers/glm4_moe_tool_parser.py (modified, +3/-3)