DeepSeekR1Parser ignores a per-request enable_thinking=False, returning empty content

System Info

• TensorRT-LLM: 1.3.0rc23 (built from f20ea652); behaviour re-checked against main @ fbda11f5
• GPU: NVIDIA GeForce RTX 5090 (sm_120), driver 595.71.05
• PyTorch 2.11.0+cu130, CUDA 13.0, Python 3.12.13
• Ubuntu 25.10, kernel 6.17.0-41-generic
• Backend: PyTorch; single GPU, no TP/PP
• Model where observed: Qwen3.6-35B-A3B-NVFP4 with FP8 KV cache (where model-specific)

Who can help?

No response

Reproduction

DeepSeekR1Parser is registered with reasoning_at_start=True for qwen3_5, minimax_m2 and
minimax_m2_append_think. When a request passes chat_template_kwargs: {"enable_thinking": false},
the rendered template injects no <think> and the model emits no </think>, so parse() partitions
on an absent tag and returns the whole reply as reasoning_content with an empty content.

A conventional OpenAI client reads message.content and therefore receives nothing for a request
that succeeded. Two clients on one server — one wanting thinking, one not — cannot share it.

Expected behavior

With enable_thinking=False, the reply is returned in message.content.

actual behavior

parse() partitions on an absent , so the whole reply lands in reasoning_content and content is empty. An OpenAI-style client reading message.content receives nothing.

additional notes

NemotronV3ReasoningParser (a DeepSeekR1Parser subclass) already implements the rule:

reasoning_at_start = chat_template_kwargs.get("enable_thinking", reasoning_at_start)

and test_nano_v3_reasoning_parser already covers {"enable_thinking": False}. The base class does
not, so the keys that register it directly do not benefit.

Apply the same rule in DeepSeekR1Parser.__init__, so every reasoning_at_start=True registration
honours a per-request enable_thinking=False. The value is already threaded into the constructor by
ReasoningParserFactory.create_reasoning_parser().

Per CONTRIBUTING.md we are raising this Issue Request before submitting the patch.

Questions for maintainers

1. Would you prefer the rule hoisted to the base class (making the Nemotron subclass's assignment
   redundant), or duplicated per subclass?
2. Should the existing test_nano_v3_reasoning_parser parametrisation be extended to a base-class
   registration such as qwen3_5?

Start with DeepSeekR1Parser.init and ReasoningParserFactory.create_reasoning_parser(), then read the existing test_nano_v3_reasoning_parser coverage for enable_thinking=False. Confirm the base parser honors the per-request setting for a direct reasoning_at_start=True registration, and run the relevant reasoning-parser tests to verify that the response appears in content.