fix: use force_reasoning parser for nemotron_nano (#6288)

Signed-off-by: Neal Vaidya <nealv@nvidia.com>

docs/pages/agents/tool-calling.md

@@ -39,6 +39,7 @@ Parser to Model Mapping
 | llama3_json | meta-llama/Llama-3.1-*, meta-llama/Llama-3.2-* |
 | harmony | openai/gpt-oss-* |
 | nemotron_deci | nvidia/nemotron-* |
+| nemotron_nano | nvidia/NVIDIA-Nemotron-3-Nano-* |
 | phi4 | Phi-4-* |
 | deepseek_v3 | deepseek-ai/DeepSeek-V3, deepseek-ai/DeepSeek-R1, deepseek-ai/DeepSeek-R1-0528 |
 | deepseek_v3_1 | deepseek-ai/DeepSeek-V3.1 |

lib/llm/src/preprocessor.rs

@@ -949,6 +949,7 @@ impl OpenAIPreprocessor {
 
     /// Check if reasoning parsing should be disabled based on per-request parameters.
     /// For kimi_k25: disabled when chat_template_args contains "thinking": false.
+    /// For nemotron_nano: disabled when chat_template_args contains "enable_thinking": false.
     fn is_reasoning_disabled_by_request(
         reasoning_parser: Option<&str>,
         chat_template_args: Option<&std::collections::HashMap<String, serde_json::Value>>,
@@ -962,6 +963,14 @@ impl OpenAIPreprocessor {
                 }
                 false
             }
+            Some("nemotron_nano") => {
+                if let Some(args) = chat_template_args
+                    && let Some(enable_thinking) = args.get("enable_thinking")
+                {
+                    return enable_thinking == &serde_json::Value::Bool(false);
+                }
+                false
+            }
             _ => false,
         }
     }
@@ -1370,6 +1379,19 @@ mod tests {
             m.insert("thinking".to_string(), serde_json::Value::Bool(false));
             m
         };
+        let enable_thinking_true = {
+            let mut m = std::collections::HashMap::new();
+            m.insert("enable_thinking".to_string(), serde_json::Value::Bool(true));
+            m
+        };
+        let enable_thinking_false = {
+            let mut m = std::collections::HashMap::new();
+            m.insert(
+                "enable_thinking".to_string(),
+                serde_json::Value::Bool(false),
+            );
+            m
+        };
         let empty_args = std::collections::HashMap::new();
 
         // (parser, args, expected_disabled, description)
@@ -1416,6 +1438,31 @@ mod tests {
                 false,
                 "no parser → never disabled",
             ),
+            // nemotron_nano uses "enable_thinking" key
+            (
+                Some("nemotron_nano"),
+                Some(&enable_thinking_false),
+                true,
+                "nemotron_nano + enable_thinking=false → disabled",
+            ),
+            (
+                Some("nemotron_nano"),
+                Some(&enable_thinking_true),
+                false,
+                "nemotron_nano + enable_thinking=true → enabled",
+            ),
+            (
+                Some("nemotron_nano"),
+                None,
+                false,
+                "nemotron_nano + no args → enabled",
+            ),
+            (
+                Some("nemotron_nano"),
+                Some(&empty_args),
+                false,
+                "nemotron_nano + empty args → enabled",
+            ),
         ];
 
         for (parser, args, expected, desc) in cases {

lib/parsers/src/reasoning/mod.rs

@@ -30,7 +30,7 @@ fn get_reasoning_parser_map() -> &'static HashMap<&'static str, ReasoningParserT
         map.insert("step3", ReasoningParserType::Step3);
         map.insert("mistral", ReasoningParserType::Mistral);
         map.insert("granite", ReasoningParserType::Granite);
-        map.insert("nemotron_nano", ReasoningParserType::NemotronDeci); // nemotron nano is <think>...</think>
+        map.insert("nemotron_nano", ReasoningParserType::DeepseekR1); // nemotron nano is ...</think>
         map.insert("glm45", ReasoningParserType::NemotronDeci); // GLM-4.5/5 is <think>...</think>, no force_reasoning
         map.insert(
             "minimax_append_think",