[sgl-router] Add SGLang-compatible DeepSeek V4.1 Flash rendering (#40532)

experimental/sgl-router/Cargo.lock

@@ -918,9 +918,9 @@ dependencies = [
 
 [[package]]
 name = "dynamo-renderer"
-version = "5.1.2"
+version = "5.2.0"
 source = "registry+https://github.com/rust-lang/crates.io-index"
-checksum = "3ae2eaa139651c535aeaad8856c5546709608931ccd4d24d3a529f6c5233c4b6"
+checksum = "d4609cdbd65f8bd15c74f4531c1560d04496e8737dfb6b65b45487644b878963"
 dependencies = [
  "anyhow",
  "chrono",

experimental/sgl-router/Cargo.toml

@@ -27,7 +27,7 @@ dynamo-tokenizers = "=1.8.1"
 # Renders the model's chat template (HF Jinja, or Dynamo's built-in encoder for
 # template-less models like DeepSeek-V4) so cache-aware routing hashes the same
 # tokens the engine caches.
-dynamo-renderer = "=5.1.2"
+dynamo-renderer = "=5.2.0"
 # `OAIChatLikeRequest` speaks minijinja values.
 minijinja = "2.24"
 

experimental/sgl-router/README.md

@@ -204,7 +204,7 @@ settings are not inferred from the router's environment. Disabling forwarding pr
 engine behavior but does not establish parity for local routing hashes.
 
 Also set `--disable-input-ids-forwarding` for array-only templates: Dynamo may wrap
-string content into arrays differently from the worker. Dynamo 5.1.2 does not expose
+string content into arrays differently from the worker. The pinned Dynamo renderer does not expose
 its conversion flag, so the router cannot automatically block these templates.
 Detailed content-format parity coverage follows in #39133.
 
@@ -221,6 +221,11 @@ official/preview effort profile is detected from the checkpoint's
 `config.json`, as in SGLang. Reference prompts live in `tests/fixtures/deepseek/`
 and are regenerated by `tests/scripts/generate_deepseek_parity.py`.
 
+V4.1 Flash uses Dynamo's separate V4.1 encoder with SGLang's numeric reasoning
+budgets, tool payloads, and `<｜System｜>` markers. Developer messages and media
+are left to the worker (the pinned encoder renders them differently), and a
+non-default `SGLANG_DSV41_REASONING_EFFORT` needs the forwarding precautions above.
+
 ## Kimi-K3
 
 Kimi-K3 renders through dynamo-render's native XTML formatter with SGLang's

experimental/sgl-router/src/tokenizer/chat_formatter.rs

@@ -40,7 +40,7 @@ pub struct ChatFormatter {
     defaults: ChatTemplateKwargs,
     /// Stripped from a separately tokenized continuation prefix, as SGLang does.
     bos_token: Option<String>,
-    deepseek_v4: Option<super::deepseek::V4Profile>,
+    deepseek: Option<super::deepseek::Encoder>,
     is_kimi_k3: bool,
 }
 
@@ -49,7 +49,14 @@ impl ChatFormatter {
     pub fn load(model_id: &str, tokenizer_path: &str) -> Result<Option<Self>> {
         let files = super::adapter::ModelFiles::open(tokenizer_path);
         let config = files.json("config.json")?.unwrap_or_default();
-        let model_type = config["model_type"].as_str().map(str::to_owned);
+        let mut model_type = config["model_type"].as_str().map(str::to_owned);
+        // SGLang also recognizes V4.1 checkpoints that retain a V4 model_type.
+        if config["architectures"][0]
+            .as_str()
+            .is_some_and(|a| a.contains("DeepseekV41"))
+        {
+            model_type = Some("deepseek_v41".into());
+        }
         if let Some(kimi) = Self::kimi_native(model_type.as_deref(), model_id) {
             return Ok(Some(kimi));
         }
@@ -59,8 +66,7 @@ impl ChatFormatter {
             Some(t) if t.starts_with("deepseek_v4") => {
                 let mut formatter = Self::deepseek_native(model_type.as_deref(), model_id);
                 if let Some(formatter) = &mut formatter {
-                    formatter.deepseek_v4 =
+                    formatter.configure_deepseek(&files, &config)?;
-                        Some(super::deepseek::V4Profile::load(&files, &config)?);
                 }
                 return Ok(formatter);
             }
@@ -73,13 +79,22 @@ impl ChatFormatter {
         let mut formatter = Self::from_tokenizer_config(cfg, jinja.as_deref())?
             .or_else(|| Self::deepseek_native(model_type.as_deref(), model_id));
         if let Some(formatter) = &mut formatter {
-            if formatter.deepseek_v4.is_some() {
+            formatter.configure_deepseek(&files, &config)?;
-                formatter.deepseek_v4 = Some(super::deepseek::V4Profile::load(&files, &config)?);
-            }
         }
         Ok(formatter)
     }
 
+    fn configure_deepseek(
+        &mut self,
+        files: &super::adapter::ModelFiles,
+        config: &JsonValue,
+    ) -> Result<()> {
+        if let Some(super::deepseek::Encoder::V4(profile)) = &mut self.deepseek {
+            *profile = super::deepseek::V4Profile::load(files, config)?;
+        }
+        Ok(())
+    }
+
     /// HF Jinja template from `tokenizer_config.json`, overridden by a sibling
     /// `chat_template.jinja` when present (transformers' precedence); `Ok(None)`
     /// when the model ships neither.
@@ -146,7 +161,7 @@ impl ChatFormatter {
             formatter,
             defaults,
             bos_token,
-            deepseek_v4: None,
+            deepseek: None,
             is_kimi_k3: false,
         }))
     }
@@ -161,23 +176,18 @@ impl ChatFormatter {
             formatter,
             defaults: HashMap::new(),
             bos_token: Some("[BOS]".into()),
-            deepseek_v4: None,
+            deepseek: None,
             is_kimi_k3: true,
         })
     }
 
-    /// Native V4 and V3.2 formatters. Other V4-family encoders must be
+    /// Native DeepSeek encoders; V4.1 uses its own wire format.
-    /// supported explicitly; a V4.1 checkpoint cannot use the V4 wire format.
     pub fn deepseek_native(model_type: Option<&str>, model_id: &str) -> Option<Self> {
         let name = model_name(model_id);
         let model_type = model_type.map(str::to_lowercase);
-        if model_type.is_none() && (name.contains("v4.1") || name.contains("v41")) {
+        if model_type.as_deref().is_some_and(|t| {
-            return None;
+            t.starts_with("deepseek_v4") && !matches!(t, "deepseek_v4" | "deepseek_v41")
-        }
+        }) {
-        if model_type
-            .as_deref()
-            .is_some_and(|t| t.starts_with("deepseek_v4") && t != "deepseek_v4")
-        {
             return None;
         }
         let PromptFormatter::OAI(formatter) = deepseek_formatter_for(&model_type, &name)?;
@@ -190,6 +200,11 @@ impl ChatFormatter {
             .map_or(version.split(['-', '_', '.']).next() == Some("v4"), |t| {
                 t == "deepseek_v4"
             });
+        let is_deepseek_v41 = model_type
+            .as_deref()
+            .map_or(name.contains("v4.1") || name.contains("v41"), |t| {
+                t == "deepseek_v41"
+            });
         // Engine defaults: chat mode (`SGLANG_DEFAULT_THINKING=false`) and no
         // reasoning-effort preamble; dynamo-render defaults to thinking at high effort.
         let defaults = HashMap::from([
@@ -200,7 +215,11 @@ impl ChatFormatter {
             formatter,
             defaults,
             bos_token: Some("<｜begin▁of▁sentence｜>".into()),
-            deepseek_v4: is_deepseek_v4.then(Default::default),
+            deepseek: if is_deepseek_v41 {
+                Some(super::deepseek::Encoder::V41)
+            } else {
+                is_deepseek_v4.then(|| super::deepseek::Encoder::V4(Default::default()))
+            },
             is_kimi_k3: false,
         })
     }
@@ -280,8 +299,8 @@ impl ChatFormatter {
         if self.is_kimi_k3 {
             super::kimi::normalize(request, &mut messages, &mut kwargs)?;
         }
-        if self.deepseek_v4.is_some() {
+        if let Some(encoder) = self.deepseek {
-            super::deepseek::normalize_messages(&mut messages)?;
+            encoder.normalize(&mut messages)?;
         }
         let mut prefix = String::new();
         if let Some(last) = messages.last_mut().filter(|m| m["role"] == "assistant") {
@@ -294,17 +313,25 @@ impl ChatFormatter {
                 }
             }
         }
-        if self.deepseek_v4.is_some() {
+        if self.deepseek.is_some() {
             if let Some(task) = request.get("task").filter(|v| !v.is_null()).cloned() {
+                // V4.1 also treats a mid-conversation system turn as the task target.
+                let v41 = matches!(self.deepseek, Some(super::deepseek::Encoder::V41));
                 let message = messages
                     .iter_mut()
+                    .enumerate()
                     .rev()
-                    .find(|m| matches!(m["role"].as_str(), Some("user" | "developer")))
+                    .find(|(i, m)| match m["role"].as_str() {
+                        Some("user" | "developer") => true,
+                        Some("system") => v41 && *i > 0,
+                        _ => false,
+                    })
+                    .map(|(_, m)| m)
                     .context("task requires a user or developer message")?;
                 message["task"] = task;
             }
         }
-        if let Some(profile) = self.deepseek_v4 {
+        if let Some(profile) = self.deepseek {
             return Ok((
                 RenderedPrompt::text(profile.render(request, messages, &kwargs)?),
                 prefix,
@@ -657,7 +684,8 @@ mod tests {
         assert!(ChatFormatter::deepseek_native(None, "deepseek-ai/DeepSeek-V4-Flash").is_some());
         assert!(ChatFormatter::deepseek_native(None, "deepseek-v4-tiny").is_some());
         assert!(ChatFormatter::deepseek_native(Some("deepseek_v4"), "alias").is_some());
-        assert!(ChatFormatter::deepseek_native(Some("deepseek_v41"), "alias").is_none());
+        assert!(ChatFormatter::deepseek_native(Some("deepseek_v41"), "alias").is_some());
+        assert!(ChatFormatter::deepseek_native(None, "deepseek-ai/DeepSeek-V4.1-Flash").is_some());
         assert!(ChatFormatter::deepseek_native(Some("deepseek_v32"), "DeepSeek-V3.2").is_some());
         assert!(ChatFormatter::deepseek_native(Some("inkling_mm_model"), "inkling").is_none());
         assert!(ChatFormatter::deepseek_native(Some("llama"), "deepseek-v4").is_none());

experimental/sgl-router/src/tokenizer/deepseek.rs

@@ -4,11 +4,47 @@
 //! SGLang's native DeepSeek serving semantics around Dynamo's V4 encoder.
 
 use anyhow::{bail, ensure, Context, Result};
-use dynamo_renderer::deepseek::v4::{self, ReasoningEffort, ThinkingMode};
+use dynamo_renderer::deepseek::{
+    v4::{self, ReasoningEffort, ThinkingMode},
+    v41,
+};
 use serde_json::{json, Map, Value};
 
 use super::{adapter::ModelFiles, chat_formatter::ChatTemplateKwargs};
 
+#[derive(Clone, Copy)]
+pub(super) enum Encoder {
+    V4(V4Profile),
+    V41,
+}
+
+impl Encoder {
+    pub fn normalize(self, messages: &mut [Value]) -> Result<()> {
+        normalize_messages(messages, matches!(self, Self::V4(_)))
+    }
+
+    pub fn render(
+        self,
+        request: &Value,
+        mut messages: Vec<Value>,
+        kwargs: &ChatTemplateKwargs,
+    ) -> Result<String> {
+        // SGLang drops a later user's task when merging it into an existing
+        // user/tool-result turn. Dynamo otherwise preserves that field.
+        for index in 1..messages.len() {
+            if messages[index]["role"] == "user"
+                && matches!(messages[index - 1]["role"].as_str(), Some("user" | "tool"))
+            {
+                messages[index].as_object_mut().unwrap().remove("task");
+            }
+        }
+        match self {
+            Self::V4(profile) => profile.render(request, messages, kwargs),
+            Self::V41 => render_v41(request, messages, kwargs),
+        }
+    }
+}
+
 #[derive(Clone, Copy, Debug, Default, PartialEq, Eq)]
 pub(super) enum V4Profile {
     #[default]
@@ -77,15 +113,6 @@ impl V4Profile {
         if let Some(tools) = request["tools"].as_array().filter(|t| !t.is_empty()) {
             messages[0]["tools"] = normalize_tools(tools)?.into();
         }
-        // SGLang drops a later user's task when merging it into an existing
-        // user/tool-result turn. Dynamo otherwise preserves that field.
-        for index in 1..messages.len() {
-            if messages[index]["role"] == "user"
-                && matches!(messages[index - 1]["role"].as_str(), Some("user" | "tool"))
-            {
-                messages[index].as_object_mut().unwrap().remove("task");
-            }
-        }
         let effort = match (
             self,
             request_effort(request).as_ref().and_then(Value::as_str),
@@ -145,9 +172,11 @@ pub(super) fn request_effort(request: &Value) -> Option<Value> {
 /// Before continuation extraction, SGLang flattens V4 parts with spaces and
 /// parses assistant arguments as JSON objects. Dynamo expects those arguments
 /// serialized, but must not apply its permissive malformed-JSON fallback here.
-pub(super) fn normalize_messages(messages: &mut [Value]) -> Result<()> {
+fn normalize_messages(messages: &mut [Value], flatten_content: bool) -> Result<()> {
     for message in messages {
-        if let Some(parts) = message["content"].as_array() {
+        // V4.1 keeps parts lists (its encoder joins them), so a parts-list
+        // final assistant turn is not a continuation prefix there, as in SGLang.
+        if let Some(parts) = message["content"].as_array().filter(|_| flatten_content) {
             message["content"] = parts
                 .iter()
                 .filter(|p| matches!(p["type"].as_str(), Some("text" | "input_text")))
@@ -213,6 +242,90 @@ fn normalize_tools(tools: &[Value]) -> Result<Vec<Value>> {
         .collect()
 }
 
+/// Use Dynamo's low-level encoder so SGLang, rather than Dynamo's OpenAI
+/// defaults, controls tool selection and the numeric reasoning budget.
+fn render_v41(
+    request: &Value,
+    mut messages: Vec<Value>,
+    kwargs: &ChatTemplateKwargs,
+) -> Result<String> {
+    ensure!(
+        !messages.is_empty(),
+        "DeepSeek requires messages after continuation extraction"
+    );
+    let thinking = kwargs
+        .get("thinking")
+        .is_some_and(|v| minijinja::Value::from_serialize(v).is_true());
+    // Dynamo maps developer to system, unlike this SGLang encoder. Leave
+    // that unsupported shape to the worker instead of forwarding wrong IDs.
+    ensure!(
+        !messages.iter().any(|m| m["role"] == "developer"),
+        "V4.1 developer messages require engine-side rendering"
+    );
+    if let Some(tools) = request["tools"].as_array().filter(|t| !t.is_empty()) {
+        if messages[0]["role"] != "system" {
+            messages.insert(0, json!({"role":"system", "content":""}));
+        }
+        // dsv41_tool_payload: only supplied fields, in the OpenAI field order.
+        messages[0]["tools"] = tools
+            .iter()
+            .map(|tool| {
+                let f = &tool["function"];
+                let mut function: Map<String, Value> = [
+                    "name",
+                    "description",
+                    "parameters",
+                    "strict",
+                    "defer_loading",
+                ]
+                .into_iter()
+                .filter_map(|key| {
+                    f.get(key)
+                        .filter(|v| !v.is_null())
+                        .map(|v| (key.into(), v.clone()))
+                })
+                .collect();
+                if !function.contains_key("defer_loading") {
+                    if let Some(v) = tool.get("defer_loading").filter(|v| !v.is_null()) {
+                        function.insert("defer_loading".into(), v.clone());
+                    }
+                }
+                json!({"type":"function", "function":function})
+            })
+            .collect();
+    }
+    // chat_encoding.parse_dsv41_reasoning_effort; unsupported values use the
+    // engine default (`high`), matching SGLANG_DSV41_REASONING_EFFORT unset.
+    let budget = match request_effort(request) {
+        Some(Value::String(s)) => match s.as_str() {
+            "low" => Some(25),
+            "high" => Some(50),
+            "xhigh" => Some(75),
+            "max" => Some(100),
+            _ => None,
+        },
+        Some(Value::Number(n)) => match (n.as_i64(), n.as_f64()) {
+            (Some(i), _) => u8::try_from(i).ok().filter(|i| (1..=100).contains(i)),
+            (None, Some(f)) => (0.0..=0.99)
+                .contains(&f)
+                .then(|| (f * 100.0).round_ties_even().max(1.0) as u8),
+            _ => None,
+        },
+        _ => None,
+    }
+    .unwrap_or(50);
+    v41::encode_messages(
+        &messages,
+        if thinking {
+            ThinkingMode::Thinking
+        } else {
+            ThinkingMode::Chat
+        },
+        true,
+        budget,
+    )
+}
+
 #[cfg(test)]
 mod tests {
     use super::*;

experimental/sgl-router/src/tokenizer/mod.rs

@@ -85,7 +85,7 @@ impl TokenizerRegistry {
             tracing::warn!(model = %m.id,
                 "router-generated input_ids forwarding enabled: requires matching worker model \
                  files and template defaults; native DeepSeek assumes SGLANG_DEFAULT_THINKING=false \
-                 and no SGLANG_DSV4_REASONING_EFFORT preamble; worker parser overrides \
+                 and default SGLANG_DSV4_REASONING_EFFORT / SGLANG_DSV41_REASONING_EFFORT; worker parser overrides \
                  (including --tool-call-parser deepseekv32), content-format detection, and \
                  conversation-template stop strings are not replicated. Use \
                  --disable-input-ids-forwarding for array-only templates or when these assumptions do not hold");
@@ -357,7 +357,8 @@ mod tests {
             .render(&request)
             .unwrap()
             .contains("<|open|>message"));
-        assert!(resolve("deepseek_v41").is_none());
+        assert_eq!(resolve("deepseek_v41").unwrap().render(&serde_json::json!({"messages":[{"role":"system","content":"S"},{"role":"user","content":"hi"}]})).unwrap(),
+            "<｜begin▁of▁sentence｜><｜System｜>S<｜User｜>hi<｜Assistant｜></think>");
         assert_eq!(
             resolve("deepseek_v4").unwrap().render(&request).unwrap(),
             "<｜begin▁of▁sentence｜><｜User｜>hi<｜Assistant｜></think>"

experimental/sgl-router/tests/component/tokenizer/deepseek.rs

@@ -77,3 +77,22 @@ fn v4_matches_sglang_serving() {
         "deepseek_v4",
     );
 }
+
+#[test]
+fn v41_matches_sglang_serving() {
+    check_fixture(
+        include_str!("../../fixtures/deepseek/v41.json"),
+        "deepseek_v41",
+    );
+}
+
+#[test]
+fn v41_leaves_unsupported_shapes_to_the_worker() {
+    let formatter = ChatFormatter::deepseek_native(Some("deepseek_v41"), "alias").unwrap();
+    for request in [
+        json!({"messages":[{"role":"developer","content":"rule"}]}),
+        json!({"messages":[{"role":"user","content":[{"type":"image_url","image_url":{"url":"https://example.com/a.png"}}]}]}),
+    ] {
+        assert!(formatter.render(&request).is_err(), "{request}");
+    }
+}

experimental/sgl-router/tests/fixtures/deepseek/v41.json

@@ -0,0 +1,27 @@
+{"model": "deepseek-ai/DeepSeek-V4.1-Flash", "revision": "dba1be0a40aa45a94ad051997016db3960a90277", "cases": [
+{"name":"user","profile":"v41","request":{"model":"m","messages":[{"role":"user","content":"Hi"}]},"prompt":"<｜begin▁of▁sentence｜><｜User｜>Hi<｜Assistant｜></think>","token_count":5,"token_sha256":"95aeed952d0173ec8849b13faec66eb61ec4412cd9c41dae0c405257559365a6"},
+{"name":"system","profile":"v41","request":{"model":"m","messages":[{"role":"system","content":"Be terse"},{"role":"user","content":"Hi"}]},"prompt":"<｜begin▁of▁sentence｜><｜System｜>Be terse<｜User｜>Hi<｜Assistant｜></think>","token_count":9,"token_sha256":"01befc17a5a1030ed5de3c3d7924514de09fc4fb9724e396586458bdcc396c1d"},
+{"name":"multi_turn","profile":"v41","request":{"model":"m","messages":[{"role":"user","content":"Hi"},{"role":"assistant","content":"Answer","reasoning_content":"Prior reasoning"},{"role":"user","content":"Again"}]},"prompt":"<｜begin▁of▁sentence｜><｜User｜>Hi<｜Assistant｜></think>Answer<｜end▁of▁sentence｜><｜User｜>Again<｜Assistant｜></think>","token_count":11,"token_sha256":"5ab68a07b5a4b1fa1811d68fe156ece6b5284af15cac9b76933c28d9909e4ee2"},
+{"name":"parts","profile":"v41","request":{"model":"m","messages":[{"role":"user","content":[{"type":"text","text":"Hello"},{"type":"text","text":"world"}]}]},"prompt":"<｜begin▁of▁sentence｜><｜User｜>Hello\n\nworld<｜Assistant｜></think>","token_count":7,"token_sha256":"1e631e7763d21541ad375b0ff431453be6515189b7588b30a285dec4da726ae7"},
+{"name":"continuation_parts","profile":"v41","request":{"model":"m","messages":[{"role":"user","content":"Hi"},{"role":"assistant","content":[{"type":"text","text":"Hello"},{"type":"text","text":"world"}]}],"continue_final_message":true},"prompt":"<｜begin▁of▁sentence｜><｜User｜>Hi<｜Assistant｜></think>Hello\n\nworld<｜end▁of▁sentence｜>","token_count":9,"token_sha256":"bbbd4e92e321aa42394288a7575735657c094ab48ea6aea8f609b37f84d70bd5"},
+{"name":"continuation_bos","profile":"v41","request":{"model":"m","messages":[{"role":"user","content":"Hi"},{"role":"assistant","content":"<｜begin▁of▁sentence｜>abc"}],"continue_final_message":true},"prompt":"<｜begin▁of▁sentence｜><｜User｜>Hi<｜Assistant｜></think><｜begin▁of▁sentence｜>abc","token_count":6,"token_sha256":"defef25e49dca2301b583fb67cabe122b8aa1bdd889727ea1e8b083be7149c14"},
+{"name":"tools","profile":"v41","request":{"model":"m","messages":[{"role":"user","content":"Hi"}],"tools":[{"type":"function","function":{"name":"foo","parameters":{"type":"object","properties":{}}}},{"type":"function","function":{"name":"bar","description":"第二个","strict":true}}]},"prompt":"<｜begin▁of▁sentence｜><｜System｜>\n\n## Tools\n\nYou have access to a set of tools to help answer the user's question. You can invoke tools by writing a \"<｜DSML｜ calls>\" block like the following:\n\n<｜DSML｜ calls>\n<｜DSML｜ invoke name=\"$TOOL_NAME\">\n<｜DSML｜ parameter name=\"$PARAMETER_NAME\" string=\"true|false\">$PARAMETER_VALUE</｜DSML｜ parameter>\n...\n</｜DSML｜ invoke>\n<｜DSML｜ invoke name=\"$TOOL_NAME2\">\n...\n</｜DSML｜ invoke>\n</｜DSML｜ calls>\n\nString parameters should be specified as is and set `string=\"true\"`. For all other types (numbers, booleans, arrays, objects), pass the value in JSON format and set `string=\"false\"`.\n\nIf thinking_mode is enabled (triggered by <think>), you MUST output your complete reasoning inside <think>...</think> BEFORE any tool calls or final response.\n\nOtherwise, output directly after </think> with tool calls or final response.\n\n### Available Tool Schemas\n\n{\"name\": \"foo\", \"parameters\": {\"type\": \"object\", \"properties\": {}}}\n{\"name\": \"bar\", \"description\": \"第二个\", \"strict\": true}\n\nYou MUST strictly follow the above defined tool name and parameter schemas to invoke tool calls.\n<｜User｜>Hi<｜Assistant｜></think>","token_count":256,"token_sha256":"89eaf9e9642f765c5f244da76ce14d680a6c22bf97f6a859d1946d528649d458"},
+{"name":"tools_none","profile":"v41","request":{"model":"m","messages":[{"role":"user","content":"Hi"}],"tools":[{"type":"function","function":{"name":"foo","parameters":{"type":"object","properties":{}}}},{"type":"function","function":{"name":"bar","description":"第二个","strict":true}}],"tool_choice":"none"},"prompt":"<｜begin▁of▁sentence｜><｜System｜>\n\n## Tools\n\nYou have access to a set of tools to help answer the user's question. You can invoke tools by writing a \"<｜DSML｜ calls>\" block like the following:\n\n<｜DSML｜ calls>\n<｜DSML｜ invoke name=\"$TOOL_NAME\">\n<｜DSML｜ parameter name=\"$PARAMETER_NAME\" string=\"true|false\">$PARAMETER_VALUE</｜DSML｜ parameter>\n...\n</｜DSML｜ invoke>\n<｜DSML｜ invoke name=\"$TOOL_NAME2\">\n...\n</｜DSML｜ invoke>\n</｜DSML｜ calls>\n\nString parameters should be specified as is and set `string=\"true\"`. For all other types (numbers, booleans, arrays, objects), pass the value in JSON format and set `string=\"false\"`.\n\nIf thinking_mode is enabled (triggered by <think>), you MUST output your complete reasoning inside <think>...</think> BEFORE any tool calls or final response.\n\nOtherwise, output directly after </think> with tool calls or final response.\n\n### Available Tool Schemas\n\n{\"name\": \"foo\", \"parameters\": {\"type\": \"object\", \"properties\": {}}}\n{\"name\": \"bar\", \"description\": \"第二个\", \"strict\": true}\n\nYou MUST strictly follow the above defined tool name and parameter schemas to invoke tool calls.\n<｜User｜>Hi<｜Assistant｜></think>","token_count":256,"token_sha256":"89eaf9e9642f765c5f244da76ce14d680a6c22bf97f6a859d1946d528649d458"},
+{"name":"tools_named","profile":"v41","request":{"model":"m","messages":[{"role":"user","content":"Hi"}],"tools":[{"type":"function","function":{"name":"foo","parameters":{"type":"object","properties":{}}}},{"type":"function","function":{"name":"bar","description":"第二个","strict":true}}],"tool_choice":{"type":"function","function":{"name":"bar"}}},"prompt":"<｜begin▁of▁sentence｜><｜System｜>\n\n## Tools\n\nYou have access to a set of tools to help answer the user's question. You can invoke tools by writing a \"<｜DSML｜ calls>\" block like the following:\n\n<｜DSML｜ calls>\n<｜DSML｜ invoke name=\"$TOOL_NAME\">\n<｜DSML｜ parameter name=\"$PARAMETER_NAME\" string=\"true|false\">$PARAMETER_VALUE</｜DSML｜ parameter>\n...\n</｜DSML｜ invoke>\n<｜DSML｜ invoke name=\"$TOOL_NAME2\">\n...\n</｜DSML｜ invoke>\n</｜DSML｜ calls>\n\nString parameters should be specified as is and set `string=\"true\"`. For all other types (numbers, booleans, arrays, objects), pass the value in JSON format and set `string=\"false\"`.\n\nIf thinking_mode is enabled (triggered by <think>), you MUST output your complete reasoning inside <think>...</think> BEFORE any tool calls or final response.\n\nOtherwise, output directly after </think> with tool calls or final response.\n\n### Available Tool Schemas\n\n{\"name\": \"foo\", \"parameters\": {\"type\": \"object\", \"properties\": {}}}\n{\"name\": \"bar\", \"description\": \"第二个\", \"strict\": true}\n\nYou MUST strictly follow the above defined tool name and parameter schemas to invoke tool calls.\n<｜User｜>Hi<｜Assistant｜></think>","token_count":256,"token_sha256":"89eaf9e9642f765c5f244da76ce14d680a6c22bf97f6a859d1946d528649d458"},
+{"name":"message_tools","profile":"v41","request":{"model":"m","messages":[{"role":"system","content":"SYS","tools":[{"type":"function","function":{"name":"foo","parameters":{"type":"object","properties":{}}}},{"type":"function","function":{"name":"bar","description":"第二个","strict":true}}]},{"role":"user","content":"Hi"}]},"prompt":"<｜begin▁of▁sentence｜><｜System｜>SYS\n\n## Tools\n\nYou have access to a set of tools to help answer the user's question. You can invoke tools by writing a \"<｜DSML｜ calls>\" block like the following:\n\n<｜DSML｜ calls>\n<｜DSML｜ invoke name=\"$TOOL_NAME\">\n<｜DSML｜ parameter name=\"$PARAMETER_NAME\" string=\"true|false\">$PARAMETER_VALUE</｜DSML｜ parameter>\n...\n</｜DSML｜ invoke>\n<｜DSML｜ invoke name=\"$TOOL_NAME2\">\n...\n</｜DSML｜ invoke>\n</｜DSML｜ calls>\n\nString parameters should be specified as is and set `string=\"true\"`. For all other types (numbers, booleans, arrays, objects), pass the value in JSON format and set `string=\"false\"`.\n\nIf thinking_mode is enabled (triggered by <think>), you MUST output your complete reasoning inside <think>...</think> BEFORE any tool calls or final response.\n\nOtherwise, output directly after </think> with tool calls or final response.\n\n### Available Tool Schemas\n\n{\"description\": null, \"name\": \"foo\", \"parameters\": {\"type\": \"object\", \"properties\": {}}, \"strict\": false}\n{\"description\": \"第二个\", \"name\": \"bar\", \"parameters\": null, \"strict\": true}\n\nYou MUST strictly follow the above defined tool name and parameter schemas to invoke tool calls.\n<｜User｜>Hi<｜Assistant｜></think>","token_count":272,"token_sha256":"d6ebecca3f9e9a498e18f1a5140ff8d2e833951d64329ff970ff8ef6157f0a0c"},
+{"name":"empty_message_tools","profile":"v41","request":{"model":"m","messages":[{"role":"system","content":"SYS","tools":[]},{"role":"user","content":"Hi"}]},"prompt":"<｜begin▁of▁sentence｜><｜System｜>SYS<｜User｜>Hi<｜Assistant｜></think>","token_count":8,"token_sha256":"0c718f5b123ecfacb6fc0d0091dfc63b1b7959c9122b2365fe823b270b6c3d35"},
+{"name":"tool_results","profile":"v41","request":{"model":"m","messages":[{"role":"user","content":"Hi"},{"role":"assistant","content":null,"tool_calls":[{"id":"a","type":"function","function":{"name":"foo","arguments":"{\"x\":\"雪\", \"y\":[1, true]}"}},{"id":"b","type":"function","function":{"name":"bar","arguments":{"z":2}}}],"reasoning_content":"Call both"},{"role":"tool","content":[{"type":"text","text":"Hello"},{"type":"text","text":"world"}],"tool_call_id":"b"},{"role":"tool","content":null,"tool_call_id":"a"}],"tools":[{"type":"function","function":{"name":"foo","parameters":{"type":"object","properties":{}}}},{"type":"function","function":{"name":"bar","description":"第二个","strict":true}}],"chat_template_kwargs":{"thinking":true}},"prompt":"<｜begin▁of▁sentence｜><｜System｜>Reasoning Effort: 50 (range 1-100, the higher the value, the more thorough the reasoning)\n\n\n\n## Tools\n\nYou have access to a set of tools to help answer the user's question. You can invoke tools by writing a \"<｜DSML｜ calls>\" block like the following:\n\n<｜DSML｜ calls>\n<｜DSML｜ invoke name=\"$TOOL_NAME\">\n<｜DSML｜ parameter name=\"$PARAMETER_NAME\" string=\"true|false\">$PARAMETER_VALUE</｜DSML｜ parameter>\n...\n</｜DSML｜ invoke>\n<｜DSML｜ invoke name=\"$TOOL_NAME2\">\n...\n</｜DSML｜ invoke>\n</｜DSML｜ calls>\n\nString parameters should be specified as is and set `string=\"true\"`. For all other types (numbers, booleans, arrays, objects), pass the value in JSON format and set `string=\"false\"`.\n\nIf thinking_mode is enabled (triggered by <think>), you MUST output your complete reasoning inside <think>...</think> BEFORE any tool calls or final response.\n\nOtherwise, output directly after </think> with tool calls or final response.\n\n### Available Tool Schemas\n\n{\"name\": \"foo\", \"parameters\": {\"type\": \"object\", \"properties\": {}}}\n{\"name\": \"bar\", \"description\": \"第二个\", \"strict\": true}\n\nYou MUST strictly follow the above defined tool name and parameter schemas to invoke tool calls.\n<｜User｜>Hi<｜Assistant｜><think>Call both</think>\n\n<｜DSML｜ calls>\n<｜DSML｜ invoke name=\"foo\">\n<｜DSML｜ parameter name=\"x\" string=\"true\">雪</｜DSML｜ parameter>\n<｜DSML｜ parameter name=\"y\" string=\"false\">[1, true]</｜DSML｜ parameter>\n</｜DSML｜ invoke>\n<｜DSML｜ invoke name=\"bar\">\n<｜DSML｜ parameter name=\"z\" string=\"false\">2</｜DSML｜ parameter>\n</｜DSML｜ invoke>\n</｜DSML｜ calls><｜end▁of▁sentence｜><｜User｜><tool_result></tool_result>\n\n<tool_result>Hello\n\nworld</tool_result><｜Assistant｜><think>","token_count":388,"token_sha256":"f9fb80a884084408d01041c452a9ece38a9c3a959a5fd3ed0289db3f6611c203"},
+{"name":"kwargs_effort","profile":"v41","request":{"model":"m","messages":[{"role":"user","content":"Hi"}],"chat_template_kwargs":{"thinking":true,"reasoning_effort":"max"}},"prompt":"<｜begin▁of▁sentence｜><｜System｜>Reasoning Effort: 100 (range 1-100, the higher the value, the more thorough the reasoning)\n\n<｜User｜>Hi<｜Assistant｜><think>","token_count":31,"token_sha256":"a29f18c6e65c09656d098c43e548224b67ced6a674f65186b7e132fb6d997d75"},
+{"name":"effort_conflict","profile":"v41","request":{"model":"m","messages":[{"role":"user","content":"Hi"}],"reasoning_effort":"high","chat_template_kwargs":{"reasoning_effort":"low"}},"prompt":"<｜begin▁of▁sentence｜><｜System｜>Reasoning Effort: 25 (range 1-100, the higher the value, the more thorough the reasoning)\n\n<｜User｜>Hi<｜Assistant｜><think>","token_count":31,"token_sha256":"80aa41f1cbf4d847db08af7c5f4dde508ea75aed6e2cb6b94f3a0a806e275d46"},
+{"name":"drop_thinking_ignored","profile":"v41","request":{"model":"m","messages":[{"role":"user","content":"Hi"},{"role":"assistant","content":"Answer","reasoning_content":"Prior reasoning"},{"role":"user","content":"Again"}],"chat_template_kwargs":{"thinking":true,"drop_thinking":false}},"prompt":"<｜begin▁of▁sentence｜><｜System｜>Reasoning Effort: 50 (range 1-100, the higher the value, the more thorough the reasoning)\n\n<｜User｜>Hi<｜Assistant｜></think>Answer<｜end▁of▁sentence｜><｜User｜>Again<｜Assistant｜><think>","token_count":37,"token_sha256":"f6cb0ab0ad4a220297622920684237a208098dd2b2cd368599bc49ebba7994f9"},
+{"name":"thinking_false","profile":"v41","request":{"model":"m","messages":[{"role":"user","content":"Hi"}],"reasoning_effort":"high","chat_template_kwargs":{"thinking":false}},"prompt":"<｜begin▁of▁sentence｜><｜User｜>Hi<｜Assistant｜></think>","token_count":5,"token_sha256":"95aeed952d0173ec8849b13faec66eb61ec4412cd9c41dae0c405257559365a6"},
+{"name":"thinking_none","profile":"v41","request":{"model":"m","messages":[{"role":"user","content":"Hi"}],"reasoning_effort":"none","chat_template_kwargs":{"thinking":true}},"prompt":"<｜begin▁of▁sentence｜><｜System｜>Reasoning Effort: 50 (range 1-100, the higher the value, the more thorough the reasoning)\n\n<｜User｜>Hi<｜Assistant｜><think>","token_count":31,"token_sha256":"4add8297343cd581f7ded6ecba6c06c5f70f402802db9a021d469efdc2138a15"},
+{"name":"consecutive_task","profile":"v41","request":{"model":"m","messages":[{"role":"user","content":"Hi"},{"role":"user","content":"Again"}],"task":"domain"},"prompt":"<｜begin▁of▁sentence｜><｜User｜>Hi\n\nAgain<｜Assistant｜></think>","token_count":7,"token_sha256":"d05e248d8a06f2c06df20b7a71af0d4dfae1cd91ce9f614719b5842674042d6b"},
+{"name":"effort_high","profile":"v41","request":{"model":"m","messages":[{"role":"user","content":"Hi"}],"reasoning_effort":"high"},"prompt":"<｜begin▁of▁sentence｜><｜System｜>Reasoning Effort: 50 (range 1-100, the higher the value, the more thorough the reasoning)\n\n<｜User｜>Hi<｜Assistant｜><think>","token_count":31,"token_sha256":"4add8297343cd581f7ded6ecba6c06c5f70f402802db9a021d469efdc2138a15"},
+{"name":"effort_max","profile":"v41","request":{"model":"m","messages":[{"role":"user","content":"Hi"}],"reasoning_effort":"max"},"prompt":"<｜begin▁of▁sentence｜><｜System｜>Reasoning Effort: 100 (range 1-100, the higher the value, the more thorough the reasoning)\n\n<｜User｜>Hi<｜Assistant｜><think>","token_count":31,"token_sha256":"a29f18c6e65c09656d098c43e548224b67ced6a674f65186b7e132fb6d997d75"},
+{"name":"task_action","profile":"v41","request":{"model":"m","messages":[{"role":"user","content":"Hi"}],"task":"action"},"prompt":"<｜begin▁of▁sentence｜><｜User｜>Hi<｜Assistant｜></think><｜action｜>","token_count":6,"token_sha256":"925bcaae248cc41efabd3e5d8f660f17513e43174f1d34aba51a1399aba76e8e"},
+{"name":"effort_low","profile":"v41","request":{"model":"m","messages":[{"role":"user","content":"Hi"}],"reasoning_effort":"low"},"prompt":"<｜begin▁of▁sentence｜><｜System｜>Reasoning Effort: 25 (range 1-100, the higher the value, the more thorough the reasoning)\n\n<｜User｜>Hi<｜Assistant｜><think>","token_count":31,"token_sha256":"80aa41f1cbf4d847db08af7c5f4dde508ea75aed6e2cb6b94f3a0a806e275d46"},
+{"name":"effort_xhigh","profile":"v41","request":{"model":"m","messages":[{"role":"user","content":"Hi"}],"reasoning_effort":"xhigh"},"prompt":"<｜begin▁of▁sentence｜><｜System｜>Reasoning Effort: 75 (range 1-100, the higher the value, the more thorough the reasoning)\n\n<｜User｜>Hi<｜Assistant｜><think>","token_count":31,"token_sha256":"54a0e744ad23d9a75775f4a71851980f51a778a1011d73de2f370b1d5499a5b7"},
+{"name":"effort_0.75","profile":"v41","request":{"model":"m","messages":[{"role":"user","content":"Hi"}],"reasoning_effort":0.75},"prompt":"<｜begin▁of▁sentence｜><｜System｜>Reasoning Effort: 75 (range 1-100, the higher the value, the more thorough the reasoning)\n\n<｜User｜>Hi<｜Assistant｜><think>","token_count":31,"token_sha256":"54a0e744ad23d9a75775f4a71851980f51a778a1011d73de2f370b1d5499a5b7"},
+{"name":"kwargs_budget","request":{"model":"m","messages":[{"role":"user","content":"Hi"}],"chat_template_kwargs":{"thinking":true,"reasoning_effort":80}},"prompt":"<｜begin▁of▁sentence｜><｜System｜>Reasoning Effort: 80 (range 1-100, the higher the value, the more thorough the reasoning)\n\n<｜User｜>Hi<｜Assistant｜><think>","token_count":31,"token_sha256":"32a4462bbf34bce35dd5c163574188ebd71896ac46d58966e0a6c2481b1b4081"}
+]}

experimental/sgl-router/tests/scripts/generate_deepseek_parity.py

@@ -25,6 +25,10 @@ from sglang.srt.entrypoints.openai.serving_chat import OpenAIServingChat
 
 ROOT = Path(__file__).resolve().parents[1] / "fixtures/deepseek"
 MODELS = {
+    "v41": (
+        "deepseek-ai/DeepSeek-V4.1-Flash",
+        "dba1be0a40aa45a94ad051997016db3960a90277",
+    ),
     "v4": ("deepseek-ai/DeepSeek-V4-Flash", "60d8d70770c6776ff598c94bb586a859a38244f1"),
 }
 
@@ -45,12 +49,18 @@ class RecordingTokenizer:
 def main():
     ROOT.mkdir(exist_ok=True)
     for family, (model, revision) in MODELS.items():
-        path = snapshot_download(model, revision=revision, local_files_only=True)
+        path = snapshot_download(
+            model,
+            revision=revision,
+            local_files_only=True,
+            allow_patterns=["config.json", "tokenizer*.json"],
+        )
         tok = RecordingTokenizer(
             AutoTokenizer.from_pretrained(path, local_files_only=True)
         )
         server = object.__new__(OpenAIServingChat)
         server.chat_encoding_spec = "ds" + family
+        server._dsv41_default_reasoning_effort = "high"
         server.tokenizer_manager = SimpleNamespace(tokenizer=tok)
         server.template_manager = SimpleNamespace(
             jinja_template_content_format="string"