[INVESTIGATE] sglang + DSv4: kwarg workaround breaks live — needs diagnosis

.github/workflows/copyright-check.ps1

@@ -84,7 +84,7 @@ $global:copyright_results = @{
 
 $ignored_files = @('.clang-format', '.gitattributes', '.gitignore', '.gitkeep', '.patch', 'Cargo.lock', 'LICENSE', 'uv.lock', 'rust-toolchain.toml', 'codespell.txt', 'exclusions.txt')
 write-debug "<copyright-check> ignored_files = ['$($ignored_files -join "','")']."
-$ignored_paths = @('.github', '.mypy_cache', '.pytest_cache', 'lib/llm/tests/data/sample-models', 'lib/llm/tests/data/deepseek-v3.2')
+$ignored_paths = @('.github', '.mypy_cache', '.pytest_cache', 'lib/llm/tests/data/sample-models', 'lib/llm/tests/data/deepseek-v3.2', 'lib/llm/tests/data/deepseek-v4')
 write-debug "<copyright-check> ignored_paths = ['$($ignored_paths -join "','")']."
 $ignored_types = @('.bat', '.gif', '.ico', '.ipynb', '.jpg', '.jpeg', '.patch', '.png', '.pyc', '.pyi', '.rst', '.zip', '.md', '.json')
 write-debug "<copyright-check> ignored_types = ['$($ignored_types -join "', '")']."

components/src/dynamo/frontend/sglang_prepost.py

@@ -174,7 +174,7 @@ def build_tool_call_guided_decoding(
         )
         constraint = parser.get_structure_constraint(
             tool_choice,
-            parallel_tool_calls=parallel_tool_calls,
+            # parallel_tool_calls=parallel_tool_calls,
         )
 
     if isinstance(constraint, tuple) and len(constraint) == 2:

components/src/dynamo/sglang/request_handlers/llm/decode_handler.py

@@ -308,7 +308,7 @@ async def generate(
                 **input_param,
                 sampling_params=sampling_params,
                 stream=True,
-                return_routed_experts=return_routed_experts,
+                # return_routed_experts=return_routed_experts,
                 bootstrap_host=bootstrap_info["bootstrap_host"],
                 bootstrap_port=bootstrap_info["bootstrap_port"],
                 bootstrap_room=bootstrap_info["bootstrap_room"],
@@ -346,7 +346,7 @@ async def generate(
                 video_data=video_data,
                 sampling_params=sampling_params,
                 stream=True,
-                return_routed_experts=return_routed_experts,
+                # return_routed_experts=return_routed_experts,
                 external_trace_header=trace_header,
                 rid=trace_id,
                 data_parallel_rank=dp_rank,

lib/llm/src/preprocessor.rs

@@ -1227,8 +1227,9 @@ impl OpenAIPreprocessor {
     /// For kimi_k25: disabled when chat_template_args contains "thinking": false.
     /// For nemotron_nano: disabled when chat_template_args contains "enable_thinking": false
     ///   or "force_nonempty_content": true.
-    /// For deepseek_r1: disabled when chat_template_args contains "thinking": false
-    ///   or "thinking_mode": "chat".
+    /// For deepseek_r1 / deepseek_v4: disabled when chat_template_args contains
+    ///   "thinking": false or "thinking_mode": "chat" — matches the V4 formatter's
+    ///   `resolve_thinking_mode` convention, so the parser and the prompt stay in sync.
     fn is_reasoning_disabled_by_request(
         reasoning_parser: Option<&str>,
         chat_template_args: Option<&std::collections::HashMap<String, serde_json::Value>>,
@@ -1257,7 +1258,8 @@ impl OpenAIPreprocessor {
                 }
                 false
             }
-            Some("deepseek_r1") => {
+            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);
@@ -1829,6 +1831,50 @@ mod tests {
                 false,
                 "nemotron_nano + empty args → enabled",
             ),
+            // deepseek_v4 — same convention as deepseek_r1; verify all three aliases
+            // (deepseek_v4 / deepseek-v4 / deepseekv4) plus both signal keys.
+            (
+                Some("deepseek_v4"),
+                Some(&thinking_false),
+                true,
+                "deepseek_v4 + thinking=false → disabled",
+            ),
+            (
+                Some("deepseek_v4"),
+                Some(&thinking_true),
+                false,
+                "deepseek_v4 + thinking=true → enabled",
+            ),
+            (
+                Some("deepseek_v4"),
+                Some(&thinking_mode_chat),
+                true,
+                "deepseek_v4 + thinking_mode=chat → disabled",
+            ),
+            (
+                Some("deepseek_v4"),
+                Some(&thinking_mode_thinking),
+                false,
+                "deepseek_v4 + thinking_mode=thinking → enabled",
+            ),
+            (
+                Some("deepseek_v4"),
+                None,
+                false,
+                "deepseek_v4 + no args → enabled",
+            ),
+            (
+                Some("deepseek-v4"),
+                Some(&thinking_false),
+                true,
+                "deepseek-v4 (hyphen alias) + thinking=false → disabled",
+            ),
+            (
+                Some("deepseekv4"),
+                Some(&thinking_mode_chat),
+                true,
+                "deepseekv4 (joined alias) + thinking_mode=chat → disabled",
+            ),
         ];
 
         for (parser, args, expected, desc) in cases {

lib/llm/src/preprocessor/prompt.rs

@@ -26,6 +26,7 @@ use std::sync::Arc;
 use crate::preprocessor::media::MediaDecoder;
 
 pub mod deepseek_v32;
+pub mod deepseek_v4;
 mod template;
 
 pub use template::{ChatTemplate, ContextMixins};