feat(v4): cherry-pick #8670 onto release/deepseekv4

lib/llm/src/preprocessor.rs

@@ -1260,13 +1260,15 @@ impl OpenAIPreprocessor {
             }
             Some("deepseek_r1") | Some("deepseek_v4") | Some("deepseek-v4")
             | Some("deepseekv4") => {
-                if let Some(args) = chat_template_args {
-                    if let Some(thinking) = args.get("thinking") {
-                        return thinking == &serde_json::Value::Bool(false);
-                    }
-                    if let Some(mode) = args.get("thinking_mode").and_then(|v| v.as_str()) {
-                        return mode == "chat";
-                    }
+                if let Some(enabled) =
+                    crate::preprocessor::prompt::thinking_bool_from_args(chat_template_args)
+                {
+                    return !enabled;
+                }
+                if let Some(args) = chat_template_args
+                    && let Some(mode) = args.get("thinking_mode").and_then(|v| v.as_str())
+                {
+                    return mode == "chat";
                 }
                 false
             }
@@ -1875,6 +1877,18 @@ mod tests {
                 true,
                 "deepseekv4 (joined alias) + thinking_mode=chat → disabled",
             ),
+            (
+                Some("deepseek_v4"),
+                Some(&enable_thinking_false),
+                true,
+                "deepseek_v4 + enable_thinking=false → disabled (vLLM alias)",
+            ),
+            (
+                Some("deepseek_v4"),
+                Some(&enable_thinking_true),
+                false,
+                "deepseek_v4 + enable_thinking=true → enabled (vLLM alias)",
+            ),
         ];
 
         for (parser, args, expected, desc) in cases {

lib/llm/src/preprocessor/prompt.rs

@@ -31,6 +31,26 @@ mod template;
 
 pub use template::{ChatTemplate, ContextMixins};
 
+/// Shared helper: extract a boolean thinking toggle from `chat_template_args`.
+///
+/// Reads the two equivalent keys (`thinking`, `enable_thinking` — vLLM's
+/// canonical kwarg) in order and returns the first bool value found, or `None`
+/// if neither key is present (or neither carries a bool). Used by the V4
+/// formatter's `resolve_thinking_mode` and by the reasoning-parser gate in
+/// `OpenAIPreprocessor::is_reasoning_disabled_by_request` so both paths agree
+/// on the signal interpretation.
+pub(crate) fn thinking_bool_from_args(
+    args: Option<&HashMap<String, serde_json::Value>>,
+) -> Option<bool> {
+    let args = args?;
+    for key in ["thinking", "enable_thinking"] {
+        if let Some(v) = args.get(key).and_then(|x| x.as_bool()) {
+            return Some(v);
+        }
+    }
+    None
+}
+
 #[derive(Debug)]
 pub enum TokenInput {
     Single(Vec<u32>),