[None][feat] tool template and parser for dsv4

tensorrt_llm/commands/serve.py

@@ -821,7 +821,7 @@ def serve(
                 f"Cannot auto-detect tool parser for model '{model}'. "
                 f"Supported model types for auto-detection: qwen2, qwen3, "
                 f"qwen3_moe, qwen3_5, qwen3_5_moe, qwen3_next, deepseek_v3, "
-                f"deepseek_v32, kimi_k2, kimi_k25, glm4. "
+                f"deepseek_v32, deepseek_v4, kimi_k2, kimi_k25, glm4. "
                 f"Please specify a parser explicitly: "
                 f"{list(ToolParserFactory.parsers.keys())}",
                 param_hint="--tool_parser")
@@ -835,7 +835,7 @@ def serve(
                 f"Cannot auto-detect reasoning parser for model '{model}'. "
                 f"Supported model types for auto-detection: qwen3, qwen3_moe, "
                 f"qwen3_5, qwen3_5_moe, qwen3_next, deepseek_v3 (R1 only), "
-                f"deepseek_v32 (R1 only), nemotron_h. "
+                f"deepseek_v32 (R1 only), deepseek_v4, nemotron_h. "
                 f"Please specify a parser explicitly: "
                 f"{list(ReasoningParserFactory.keys())}",
                 param_hint="--reasoning_parser")

tensorrt_llm/llmapi/reasoning_parser.py

@@ -80,6 +80,19 @@ def finish(self) -> ReasoningParserResult:
         return ReasoningParserResult()
 
 
+class IdentityReasoningParser(BaseReasoningParser):
+    """Reasoning parser that treats all model output as visible content."""
+
+    reasoning_start = "<think>"
+    reasoning_end = "</think>"
+
+    def parse(self, text: str) -> ReasoningParserResult:
+        return ReasoningParserResult(content=text)
+
+    def parse_delta(self, delta_text: str) -> ReasoningParserResult:
+        return ReasoningParserResult(content=delta_text)
+
+
 @register_reasoning_parser("deepseek-r1", reasoning_at_start=True)
 @register_reasoning_parser("qwen3")
 @register_reasoning_parser("minimax_m2", reasoning_at_start=True)
@@ -173,6 +186,42 @@ def parse_delta(self, delta_text: str) -> ReasoningParserResult:
             "Unreachable code reached in `DeepSeekR1Parser.parse_delta`")
 
 
+@register_reasoning_parser("deepseek_v4")
+class DeepSeekV4ReasoningParser(BaseReasoningParser):
+    """DeepSeek-V4 parser selected by thinking-mode chat template kwargs."""
+
+    reasoning_start = "<think>"
+    reasoning_end = "</think>"
+
+    def __init__(
+        self,
+        *,
+        chat_template_kwargs: Optional[dict[str, Any]] = None,
+    ) -> None:
+        super().__init__(chat_template_kwargs=chat_template_kwargs)
+        chat_template_kwargs = chat_template_kwargs or {}
+        thinking = bool(
+            chat_template_kwargs.get("thinking", False)
+            or chat_template_kwargs.get("enable_thinking", False))
+        if thinking:
+            self._parser = DeepSeekR1Parser(
+                reasoning_at_start=True,
+                chat_template_kwargs=chat_template_kwargs,
+            )
+        else:
+            self._parser = IdentityReasoningParser(
+                chat_template_kwargs=chat_template_kwargs)
+
+    def parse(self, text: str) -> ReasoningParserResult:
+        return self._parser.parse(text)
+
+    def parse_delta(self, delta_text: str) -> ReasoningParserResult:
+        return self._parser.parse_delta(delta_text)
+
+    def finish(self) -> ReasoningParserResult:
+        return self._parser.finish()
+
+
 MODEL_TYPE_TO_REASONING_PARSER: dict[str, str] = {
     "qwen3": "qwen3",
     "qwen3_moe": "qwen3",
@@ -181,6 +230,7 @@ def parse_delta(self, delta_text: str) -> ReasoningParserResult:
     "qwen3_next": "qwen3",
     "deepseek_v3": "deepseek-r1",
     "deepseek_v32": "deepseek-r1",
+    "deepseek_v4": "deepseek_v4",
     "nemotron_h": "nano-v3",
 }
 

tensorrt_llm/serve/tool_parser/deepseekv32_parser.py

@@ -133,7 +133,7 @@ def detect_and_parse(self, text: str, tools: List[Tool]) -> StreamingParseResult
         try:
             # Extract content between function_calls tags
             function_calls_match = re.search(
-                r"<｜DSML｜function_calls>(.*?)</｜DSML｜function_calls>",
+                re.escape(self.bot_token) + r"(.*?)" + re.escape(self.eot_token),
                 text,
                 re.DOTALL,
             )

tensorrt_llm/serve/tool_parser/deepseekv4_parser.py

@@ -0,0 +1,25 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .deepseekv32_parser import DeepSeekV32Parser
+
+
+class DeepSeekV4Parser(DeepSeekV32Parser):
+    """Tool parser for the DeepSeek V4 DSML tool call format."""
+
+    def __init__(self) -> None:
+        super().__init__()
+        self.bot_token = "<｜DSML｜tool_calls>"  # nosec B105
+        self.eot_token = "</｜DSML｜tool_calls>"  # nosec B105

tensorrt_llm/serve/tool_parser/tool_parser_factory.py

@@ -4,6 +4,7 @@
 
 from .base_tool_parser import BaseToolParser
 from .deepseekv3_parser import DeepSeekV3Parser
+from .deepseekv4_parser import DeepSeekV4Parser
 from .deepseekv31_parser import DeepSeekV31Parser
 from .deepseekv32_parser import DeepSeekV32Parser
 from .glm4_parser import Glm4ToolParser
@@ -21,6 +22,7 @@
     "qwen3_next": "qwen3",
     "deepseek_v3": "deepseek_v3",
     "deepseek_v32": "deepseek_v32",
+    "deepseek_v4": "deepseek_v4",
     "kimi_k2": "kimi_k2",
     "kimi_k25": "kimi_k2",
     "glm4": "glm4",
@@ -48,6 +50,7 @@ class ToolParserFactory:
         "deepseek_v3": DeepSeekV3Parser,
         "deepseek_v31": DeepSeekV31Parser,
         "deepseek_v32": DeepSeekV32Parser,
+        "deepseek_v4": DeepSeekV4Parser,
         "glm4": Glm4ToolParser,
         "minimax_m2": MiniMaxM2ToolParser,
     }