A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 8, 17:45 UTC
[Bugfix] Fix the issue where tool calling does not work when using fast detokenization with dsv32 (#33964)
Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
Chauncey committed
Feb 7, 2026 at 01:23 UTC
7bec4351305f7b99ff6385b778760d301e29a6f0
1 file changed
+12
vllm/tool_parsers/deepseekv32_tool_parser.py
+12
@@ -110,6 +110,18 @@ class DeepSeekV32ToolParser(ToolParser):
110
"""Generate a unique tool call ID."""
111
return f"call_{uuid.uuid4().hex[:24]}"
112
113
+ def adjust_request(self, request):
114
+ request = super().adjust_request(request)
115
+ if request.tools and request.tool_choice != "none":
116
+ # Ensure tool call tokens
117
+ # (<｜DSML｜function_calls>, </｜DSML｜function_calls>)
118
+ # are not skippedduring decoding.
119
+ # Even though they are not marked as special tokens,
120
+ # setting skip_special_tokens=False ensures proper handling in
121
+ # transformers 5.x where decoding behavior may have changed.
122
+ request.skip_special_tokens = False
123
+ return request
124
+
125
def _reset_streaming_state(self):
126
"""Reset all streaming state."""
127
self.current_tool_index = 0