[sgl-router] Rename chat encoder to chat formatter (#39459)

Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>

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

@@ -34,8 +34,9 @@ use std::sync::Arc;
 pub struct RequestTokens {
     /// The prompt token ids.
     pub ids: Vec<u32>,
-    /// Whether the token ids are safe to forward as engine `input_ids`.
-    pub engine_equivalent: bool,
+    /// Whether the IDs came from rendered chat messages.
+    /// Forwarding also requires the request safety guard.
+    pub rendered_from_chat: bool,
 }
 
 /// External indexer answer prepared by the async ingress path for the
@@ -45,19 +46,18 @@ pub struct ExternalPrefixSignal {
     pub query_blocks: usize,
 }
 
-/// Tokenizes a request for routing. Chat-encoder tokens are engine-equivalent;
-/// raw prompt tokens are used only for routing.
+/// Tokenizes requests for routing, preferring chat rendering over raw text.
 pub fn request_tokens_for(
     tokenizers: &TokenizerRegistry,
     model_id: &ModelId,
     value: &serde_json::Value,
 ) -> Option<RequestTokens> {
-    if tokenizers.has_chat_encoder(&model_id.0) {
+    if tokenizers.has_chat_formatter(&model_id.0) {
         if let Some(messages) = value.get("messages").filter(|m| m.is_array()) {
             if let Some(ids) = tokenizers.encode_chat(&model_id.0, messages) {
                 return Some(RequestTokens {
                     ids,
-                    engine_equivalent: true,
+                    rendered_from_chat: true,
                 });
             }
         }
@@ -66,7 +66,7 @@ pub fn request_tokens_for(
     let ids = tokenize_text(tokenizers, model_id, &text)?;
     Some(RequestTokens {
         ids,
-        engine_equivalent: false,
+        rendered_from_chat: false,
     })
 }
 
@@ -511,11 +511,11 @@ pub trait Policy: Send + Sync + std::fmt::Debug {
 
     /// Whether this policy's routing decision needs request tokens (i.e.
     /// it routes by prompt prefix). Ingress tokenization itself is no longer
-    /// gated on this — that is a model property (`has_chat_encoder`) decided at
+    /// gated on this — that is a model property (`has_chat_formatter`) decided at
     /// ingress via [`request_tokens_for`]. This flag is the EXTRA gate that
     /// keeps the cache-aware policy's RAW-prompt routing path alive: a
-    /// cache-aware model with no chat encoder still wants its `/v1/completions`
-    /// /`text` prompt tokenized for tree matching, which `has_chat_encoder`
+    /// cache-aware model with no chat formatter still wants its `/v1/completions`
+    /// /`text` prompt tokenized for tree matching, which `has_chat_formatter`
     /// alone would not trigger. Default `false` for load-only and sticky
     /// routes; only the cache-aware policy overrides it.
     fn needs_request_tokens(&self) -> bool {

experimental/sgl-router/src/server/metrics.rs

@@ -735,7 +735,7 @@ impl MetricsRegistry {
 
     /// Bump `sgl_router_ingress_tokenize_errors_total{model_id}`.
     ///
-    /// Count encoder failures only for chats eligible for `input_ids`
+    /// Count formatter failures only for chats eligible for `input_ids`
     /// forwarding. Requests excluded by the guard are expected fallbacks.
     /// Pairs with the per-model WARN log in `encode_chat`.
     pub fn record_ingress_tokenize_error(&self, model_id: &str) {
@@ -1191,7 +1191,7 @@ impl MetricsRegistry {
 
         // ingress_tokenize_errors_total
         out.push_str(
-            "# HELP sgl_router_ingress_tokenize_errors_total Plain text chat requests on a chat-encoder model whose ingress rendering or tokenization failed, silently falling back to engine-side tokenization (the input_ids offload was defeated).\n",
+            "# HELP sgl_router_ingress_tokenize_errors_total Plain text chat requests on a chat-formatter model whose ingress rendering or tokenization failed, silently falling back to engine-side tokenization (the input_ids offload was defeated).\n",
         );
         out.push_str("# TYPE sgl_router_ingress_tokenize_errors_total counter\n");
         let guard = self.ingress_tokenize_errors_total.lock();

experimental/sgl-router/src/server/routes/chat.rs

@@ -504,16 +504,16 @@ pub async fn chat_completions(
 
     // Tokenize once at ingress whenever it can pay off — decoupled from the
     // routing policy, because forwarding `input_ids` is a property of the
-    // MODEL (does it have a chat encoder so the router can produce
+    // MODEL (does it have a chat formatter so the router can produce
     // engine-equivalent tokens?), not of how we pick the worker. Two gates:
     //
-    //   * `has_chat_encoder` → a chat request on this model yields
+    //   * `has_chat_formatter` → a chat request on this model yields
     //     engine-equivalent ids we can forward as `input_ids` so the engine
     //     skips re-tokenizing. This enables the offload for EVERY policy —
     //     sticky and round-robin included — not just cache-aware.
     //   * `needs_request_tokens()` → the cache-aware policy ALSO wants the
     //     raw-prompt path tokenized for tree matching even on a model with no
-    //     chat encoder (`/v1/completions` / `text`), which the first gate
+    //     chat formatter (`/v1/completions` / `text`), which the first gate
     //     alone wouldn't trigger.
     //
     //   * Bucket routing also needs the prompt token count.
@@ -524,7 +524,7 @@ pub async fn chat_completions(
     // tokenization and the outgoing-body injection below (and PD bootstrap
     // injection). `parse_probe` already validated the object shape.
     let want_tokens = should_tokenize_request(
-        ctx.tokenizers.has_chat_encoder(&model_str),
+        ctx.tokenizers.has_chat_formatter(&model_str),
         policy.needs_request_tokens(),
         ctx.bucket_selector.is_enabled(),
     );
@@ -798,8 +798,8 @@ pub async fn chat_completions(
 
     // Forward the router-computed tokens to the engine as `input_ids` so it
     // skips re-tokenizing the same prompt — but only when they are
-    // engine-equivalent (chat-encoder path) AND the request contains nothing
-    // the router's encoder didn't replicate (see `input_ids_safe_to_forward`).
+    // engine-equivalent (chat-formatter path) AND the request contains nothing
+    // the router's formatter didn't replicate (see `input_ids_safe_to_forward`).
     // Otherwise omit them and the engine tokenizes from `messages` as usual —
     // a transparent, always-correct fallback (`messages` are always retained
     // in the forwarded body). `forward_input_ids` is `Some` only when
@@ -807,19 +807,19 @@ pub async fn chat_completions(
     // predicate always has a parsed body to inspect.
     let forward_input_ids: Option<&[u32]> = match (request_tokens.as_ref(), request_value.as_ref())
     {
-        (Some(t), Some(v)) if t.engine_equivalent && input_ids_safe_to_forward(v) => {
+        (Some(t), Some(v)) if t.rendered_from_chat && input_ids_safe_to_forward(v) => {
             Some(t.ids.as_slice())
         }
         _ => None,
     };
 
-    // Surface a broken offload: when the encoder SHOULD have produced
+    // Surface a broken offload: when the formatter SHOULD have produced
     // engine-equivalent ids but didn't, the chat request silently fell back to
     // engine-side tokenization. Count only that case (see
     // `ingress_tokenize_offload_failed`); successful forwards and expected
     // omissions are not problems.
     if ingress_tokenize_offload_failed(
-        ctx.tokenizers.has_chat_encoder(&model_str),
+        ctx.tokenizers.has_chat_formatter(&model_str),
         request_value.as_ref(),
         request_tokens.as_ref(),
     ) {
@@ -1198,11 +1198,11 @@ fn parse_optional_positive_f64_header(
 }
 
 fn should_tokenize_request(
-    has_chat_encoder: bool,
+    has_chat_formatter: bool,
     policy_needs_request_tokens: bool,
     bucket_enabled: bool,
 ) -> bool {
-    has_chat_encoder || policy_needs_request_tokens || bucket_enabled
+    has_chat_formatter || policy_needs_request_tokens || bucket_enabled
 }
 
 /// Estimate prefill-token count from the raw request body for use as
@@ -1339,28 +1339,28 @@ fn build_outgoing_body(
 /// uses it verbatim and ignores everything that would otherwise steer its
 /// `messages`-side tokenization (only stop tokens / tool-call constraint are
 /// still taken from `messages`). So any request field that changes that
-/// tokenization but which the router's chat encoder does not replicate makes
+/// tokenization but which the router's chat formatter does not replicate makes
 /// the forwarded ids wrong. This predicate is conservative by construction —
 /// any such signal returns `false` and the engine tokenizes from `messages`
 /// (always correct).
 ///
 /// Replicated-and-safe: plain text `messages` with a string `content`.
 /// Not replicated → omit:
-///   * `tools` / `functions` — the encoder doesn't render tool schemas.
+///   * `tools` / `functions` — the formatter doesn't render tool schemas.
 ///   * non-string or missing `content` (arrays, `null`): the engine normalizes
-///     these before rendering; the router's encoder renders them verbatim.
+///     these before rendering; the router's formatter renders them verbatim.
 ///   * `chat_template` — an OpenAI-compatible per-request template override
 ///     (e.g. vLLM); the router renders with the model's default template, so a
 ///     custom one would diverge. (SGLang ignores it today, but block it so the
 ///     offload stays correct across engines / future versions.)
 ///   * `chat_template_kwargs` (carries `enable_thinking`/`thinking`),
 ///     `reasoning` / `reasoning_effort`, `task` — thinking/mode toggles the
-///     encoder renders in the engine's default mode only.
+///     formatter renders in the engine's default mode only.
 ///   * `continue_final_message: true`, or a trailing `assistant` message — the
-///     engine rewrites/strips the final assistant turn; the encoder renders it
+///     engine rewrites/strips the final assistant turn; the formatter renders it
 ///     verbatim.
 ///
-/// NOTE: the router's chat encoder renders in the engine's default
+/// NOTE: the router's chat formatter renders in the engine's default
 /// (non-thinking) mode. Current sglang derives thinking from the request
 /// (`chat_template_kwargs`), which this guard already omits, so a plain request
 /// the router rendered matches the engine. The only way to diverge is an engine
@@ -1376,7 +1376,7 @@ fn input_ids_safe_to_forward(value: &serde_json::Value) -> bool {
         return false;
     }
     // Fields that steer the engine's template tokenization but which the
-    // router's encoder does not thread through.
+    // router's formatter does not thread through.
     for key in [
         "chat_template",
         "chat_template_kwargs",
@@ -1400,15 +1400,15 @@ fn input_ids_safe_to_forward(value: &serde_json::Value) -> bool {
 
 /// Whether to increment `sgl_router_ingress_tokenize_errors_total`.
 ///
-/// Count chats with a configured encoder that pass the forwarding guard
-/// but lack engine-equivalent tokens. Excluded requests are expected fallbacks,
+/// Count chats with a configured formatter that pass the forwarding guard
+/// but lack chat-rendered tokens. Excluded requests are expected fallbacks,
 /// even when rendering fails.
 fn ingress_tokenize_offload_failed(
-    has_chat_encoder: bool,
+    has_chat_formatter: bool,
     request_value: Option<&serde_json::Value>,
     request_tokens: Option<&RequestTokens>,
 ) -> bool {
-    if !has_chat_encoder {
+    if !has_chat_formatter {
         return false;
     }
     let chat_request = request_value.is_some_and(|v| {
@@ -1417,7 +1417,7 @@ fn ingress_tokenize_offload_failed(
     if !chat_request {
         return false;
     }
-    !request_tokens.is_some_and(|t| t.engine_equivalent)
+    !request_tokens.is_some_and(|t| t.rendered_from_chat)
 }
 
 /// Whether the final chat message has `role: "assistant"` (a prefix /
@@ -1433,7 +1433,7 @@ fn last_message_is_assistant(value: &serde_json::Value) -> bool {
 }
 
 /// Whether the request carries tool / function definitions. The router's chat
-/// encoder renders only `messages`, so its `input_ids` would omit the tool
+/// formatter renders only `messages`, so its `input_ids` would omit the tool
 /// schemas the engine's template injects into the prompt — the caller must let
 /// the engine tokenize these itself.
 fn request_has_tools(value: &serde_json::Value) -> bool {
@@ -1448,7 +1448,7 @@ fn request_has_tools(value: &serde_json::Value) -> bool {
 }
 
 /// Detect non-string or missing content, which requires engine tokenization:
-/// the engine normalizes arrays and nulls differently from the router's encoder.
+/// the engine normalizes arrays and nulls differently from the router's formatter.
 fn request_has_non_text_content(value: &serde_json::Value) -> bool {
     value
         .get("messages")
@@ -1747,7 +1747,7 @@ mod tests {
     }
 
     /// Tool / function requests are detected so the caller omits `input_ids`
-    /// (the router's encoder doesn't render tools).
+    /// (the router's formatter doesn't render tools).
     #[test]
     fn request_has_tools_detects_tools_and_functions() {
         assert!(request_has_tools(
@@ -1792,7 +1792,7 @@ mod tests {
     }
 
     /// Every field the engine honors on the `messages` path but which the
-    /// router's encoder does not replicate must block forwarding — otherwise
+    /// router's formatter does not replicate must block forwarding — otherwise
     /// the engine uses the router's ids verbatim and silently runs a different
     /// prompt than the request asked for.
     #[test]
@@ -1848,14 +1848,14 @@ mod tests {
         assert!(parsed.get("input_ids").is_none());
     }
 
-    /// A chat request on a chat-encoder model that yields engine-equivalent
+    /// A chat request on a chat-formatter model that yields engine-equivalent
     /// ids (encode succeeded) is NOT a failure — the offload worked.
     #[test]
-    fn offload_failed_false_when_tokens_engine_equivalent() {
+    fn offload_failed_false_when_tokens_rendered_from_chat() {
         let value = serde_json::json!({"messages":[{"role":"user","content":"hi"}]});
         let tokens = RequestTokens {
             ids: vec![1, 2, 3],
-            engine_equivalent: true,
+            rendered_from_chat: true,
         };
         assert!(!ingress_tokenize_offload_failed(
             true,
@@ -1874,22 +1874,22 @@ mod tests {
         assert!(!ingress_tokenize_offload_failed(true, Some(&value), None));
     }
 
-    /// Missing tokens count as a failure for an eligible chat with an encoder.
+    /// Missing tokens count as a failure for an eligible chat with a formatter.
     #[test]
-    fn offload_failed_true_when_chat_encoder_request_has_no_tokens() {
+    fn offload_failed_true_when_chat_formatter_request_has_no_tokens() {
         let value = serde_json::json!({"messages":[{"role":"user","content":"hi"}]});
         assert!(ingress_tokenize_offload_failed(true, Some(&value), None));
     }
 
-    /// Encode produced ids but NOT via the chat encoder (raw fallback,
-    /// `engine_equivalent = false`) on a chat-encoder model + chat request →
+    /// Encode produced ids but NOT via the chat formatter (raw fallback,
+    /// `rendered_from_chat = false`) on a chat-formatter model + chat request →
     /// the chat-encode render/encode failed and fell through to the raw path.
     #[test]
-    fn offload_failed_true_when_tokens_not_engine_equivalent() {
+    fn offload_failed_true_when_tokens_not_rendered_from_chat() {
         let value = serde_json::json!({"messages":[{"role":"user","content":"hi"}]});
         let tokens = RequestTokens {
             ids: vec![1, 2, 3],
-            engine_equivalent: false,
+            rendered_from_chat: false,
         };
         assert!(ingress_tokenize_offload_failed(
             true,
@@ -1898,15 +1898,15 @@ mod tests {
         ));
     }
 
-    /// Non-chat-encoder models never expected the offload → not a failure even
+    /// Non-chat-formatter models never expected the offload → not a failure even
     /// with no tokens.
     #[test]
-    fn offload_failed_false_without_chat_encoder() {
+    fn offload_failed_false_without_chat_formatter() {
         let value = serde_json::json!({"messages":[{"role":"user","content":"hi"}]});
         assert!(!ingress_tokenize_offload_failed(false, Some(&value), None));
     }
 
-    /// A non-chat (no `messages`) request on a chat-encoder model — e.g.
+    /// A non-chat (no `messages`) request on a chat-formatter model — e.g.
     /// `/v1/completions` `prompt` — never expected the chat-encode offload, so
     /// the absence of engine-equivalent ids is not a failure.
     #[test]

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

@@ -15,7 +15,7 @@ use std::sync::Arc;
 /// How to turn a chat request's `messages` into the prompt the engine tokenizes
 /// and caches. Cache-aware routing renders this before hashing so its query
 /// tokens match the engine's stored blocks.
-pub enum ChatEncoder {
+pub enum ChatFormatter {
     /// HuggingFace Jinja chat template from `tokenizer_config.json` (most
     /// models). Boxed: it holds a minijinja `Environment`, far larger than the
     /// other variants.
@@ -24,26 +24,26 @@ pub enum ChatEncoder {
     DeepSeekV4,
 }
 
-impl ChatEncoder {
+impl ChatFormatter {
     /// Render `messages` into the engine-equivalent prompt text.
     fn render(&self, messages: &serde_json::Value) -> Result<String> {
         match self {
-            ChatEncoder::Jinja(t) => t.render(messages),
-            ChatEncoder::DeepSeekV4 => Ok(dsv4::render_messages(messages)),
+            ChatFormatter::Jinja(t) => t.render(messages),
+            ChatFormatter::DeepSeekV4 => Ok(dsv4::render_messages(messages)),
         }
     }
 }
 
-/// A model's chat encoder plus its fallback-logging state.
-struct ChatEncoderEntry {
-    encoder: ChatEncoder,
+/// A model's chat formatter plus its fallback-logging state.
+struct ChatFormatterEntry {
+    formatter: ChatFormatter,
     fallback_warned: AtomicBool,
 }
 
-impl ChatEncoderEntry {
-    fn new(encoder: ChatEncoder) -> Self {
+impl ChatFormatterEntry {
+    fn new(formatter: ChatFormatter) -> Self {
         Self {
-            encoder,
+            formatter,
             fallback_warned: AtomicBool::new(false),
         }
     }
@@ -56,11 +56,11 @@ impl ChatEncoderEntry {
     fn log_fallback(&self, model_id: &str, cause: &str) {
         if !self.fallback_warned.swap(true, Ordering::Relaxed) {
             tracing::warn!(model = %model_id, %cause,
-                "chat-encoder failed; falling back to raw prompt-text hashing \
+                "chat-formatter failed; falling back to raw prompt-text hashing \
                  (cache-aware overlap degrades for this model; further failures log at debug)");
         } else {
             tracing::debug!(model = %model_id, %cause,
-                "chat-encoder failed; falling back to raw prompt-text hashing");
+                "chat-formatter failed; falling back to raw prompt-text hashing");
         }
     }
 }
@@ -68,12 +68,12 @@ impl ChatEncoderEntry {
 #[derive(Default)]
 pub struct TokenizerRegistry {
     inner: DashMap<String, Arc<Tokenizer>>,
-    /// Per-model chat encoder, present only when the model's prompt format is
-    /// known (a `tokenizer_config.json` chat template, or a built-in encoder
+    /// Per-model chat formatter, present only when the model's prompt format is
+    /// known (a `tokenizer_config.json` chat template, or a built-in formatter
     /// like DeepSeek-V4's). Cache-aware routing uses it to tokenize chat
     /// requests the way the engine does; models without one fall back to raw
     /// prompt-text tokenization.
-    encoders: DashMap<String, Arc<ChatEncoderEntry>>,
+    formatters: DashMap<String, Arc<ChatFormatterEntry>>,
 }
 
 impl std::fmt::Debug for TokenizerRegistry {
@@ -90,28 +90,32 @@ impl TokenizerRegistry {
         let m = &cfg.model;
         let t = adapter::load(&m.tokenizer_path)?;
         me.inner.insert(m.id.clone(), t);
-        // Resolve the chat encoder, best-effort: a Jinja template from
-        // tokenizer_config.json, else a built-in encoder for a recognized model
+        // Resolve the chat formatter, best-effort: a Jinja template from
+        // tokenizer_config.json, else a built-in formatter for a recognized model
         // (DeepSeek-V4), else none (chat traffic routes via raw text). Every
         // path logs its outcome — whether chat-aware routing is live for this
         // model is the single most useful signal for diagnosing "cache-aware
         // routing degraded to overlap=0 on chat traffic", so it must never be
         // silent.
-        if let Some(encoder) = me.resolve_chat_encoder(&m.id, &m.tokenizer_path) {
-            me.encoders
-                .insert(m.id.clone(), Arc::new(ChatEncoderEntry::new(encoder)));
+        if let Some(formatter) = me.resolve_chat_formatter(&m.id, &m.tokenizer_path) {
+            me.formatters
+                .insert(m.id.clone(), Arc::new(ChatFormatterEntry::new(formatter)));
         }
         Ok(me)
     }
 
-    /// Pick the chat encoder for a model, logging the outcome on every branch.
-    fn resolve_chat_encoder(&self, model_id: &str, tokenizer_path: &str) -> Option<ChatEncoder> {
+    /// Pick the chat formatter for a model, logging the outcome on every branch.
+    fn resolve_chat_formatter(
+        &self,
+        model_id: &str,
+        tokenizer_path: &str,
+    ) -> Option<ChatFormatter> {
         match adapter::load_tokenizer_config(tokenizer_path) {
             Ok(Some(cfg_json)) => match ChatTemplate::from_tokenizer_config(&cfg_json) {
                 Ok(Some(tmpl)) => {
                     tracing::info!(model = %model_id,
                         "chat-template routing enabled; chat requests route by templated tokens");
-                    return Some(ChatEncoder::Jinja(Box::new(tmpl)));
+                    return Some(ChatFormatter::Jinja(Box::new(tmpl)));
                 }
                 Ok(None) => {} // no template — fall through to built-in detection
                 Err(e) => tracing::warn!(model = %model_id, error = %e,
@@ -123,11 +127,11 @@ impl TokenizerRegistry {
         }
         if is_deepseek_v4(model_id) {
             tracing::info!(model = %model_id,
-                "DeepSeek-V4 routing enabled; chat requests route via the built-in V4 encoder");
-            return Some(ChatEncoder::DeepSeekV4);
+                "DeepSeek-V4 routing enabled; chat requests route via the built-in V4 formatter");
+            return Some(ChatFormatter::DeepSeekV4);
         }
         tracing::info!(model = %model_id,
-            "no chat template or built-in encoder; chat traffic routes via raw prompt text");
+            "no chat template or built-in formatter; chat traffic routes via raw prompt text");
         None
     }
 
@@ -135,24 +139,24 @@ impl TokenizerRegistry {
         self.inner.get(model_id).map(|r| Arc::clone(&*r))
     }
 
-    /// Whether this model has a chat encoder (and thus the chat-aware
+    /// Whether this model has a chat formatter (and thus the chat-aware
     /// tokenization path is available for it).
-    pub fn has_chat_encoder(&self, model_id: &str) -> bool {
-        self.encoders.contains_key(model_id)
+    pub fn has_chat_formatter(&self, model_id: &str) -> bool {
+        self.formatters.contains_key(model_id)
     }
 
-    /// Render `messages` through the model's chat encoder, then tokenize the
+    /// Render `messages` through the model's chat formatter, then tokenize the
     /// result the same way the engine does (`add_special_tokens = false`, so the
-    /// encoder's literal `bos_token`/role markers carry the specials). Returns
+    /// formatter's literal `bos_token`/role markers carry the specials). Returns
     /// `None` — caller falls back to raw routing — when the model has no
-    /// encoder, no tokenizer, or rendering/encoding fails or yields no tokens.
+    /// formatter, no tokenizer, or rendering/encoding fails or yields no tokens.
     pub fn encode_chat(&self, model_id: &str, messages: &serde_json::Value) -> Option<Vec<u32>> {
         // Clone the Arc and drop the DashMap guard before the CPU-bound
         // render+encode (mirrors `get`), so no shard read-lock is held across it.
-        let entry = Arc::clone(&*self.encoders.get(model_id)?);
+        let entry = Arc::clone(&*self.formatters.get(model_id)?);
         let tokenizer = self.get(model_id)?;
         let rendered = entry
-            .encoder
+            .formatter
             .render(messages)
             .inspect_err(|e| {
                 // `{e:#}` prints the full anyhow chain, so the underlying
@@ -178,18 +182,18 @@ impl TokenizerRegistry {
         self.inner.iter().map(|kv| kv.key().clone()).collect()
     }
 
-    /// Attach a chat encoder to an already-loaded model. Lets policy tests in
+    /// Attach a chat formatter to an already-loaded model. Lets policy tests in
     /// other modules exercise the chat-aware routing path without a co-located
     /// fixture.
     #[cfg(test)]
-    pub(crate) fn attach_chat_encoder_for_test(&self, model_id: &str, encoder: ChatEncoder) {
-        self.encoders.insert(
+    pub(crate) fn attach_chat_formatter_for_test(&self, model_id: &str, formatter: ChatFormatter) {
+        self.formatters.insert(
             model_id.to_string(),
-            Arc::new(ChatEncoderEntry::new(encoder)),
+            Arc::new(ChatFormatterEntry::new(formatter)),
         );
     }
 
-    /// Convenience: attach a Jinja chat encoder built from an inline
+    /// Convenience: attach a Jinja chat formatter built from an inline
     /// `tokenizer_config.json` value.
     #[cfg(test)]
     pub(crate) fn attach_chat_template_for_test(
@@ -200,12 +204,12 @@ impl TokenizerRegistry {
         let template = ChatTemplate::from_tokenizer_config(tokenizer_config)
             .expect("valid test chat template")
             .expect("test tokenizer_config has a chat_template");
-        self.attach_chat_encoder_for_test(model_id, ChatEncoder::Jinja(Box::new(template)));
+        self.attach_chat_formatter_for_test(model_id, ChatFormatter::Jinja(Box::new(template)));
     }
 }
 
 /// Whether `model_id` denotes a DeepSeek-V4 model, which the engine encodes via
-/// the built-in [`dsv4`] encoder rather than a Jinja template. Heuristic on the
+/// the built-in [`dsv4`] formatter rather than a Jinja template. Heuristic on the
 /// served model id (the router has no model architecture from `/server_info`);
 /// scoped to "deepseek" + "v4" so it doesn't claim V3-family models, whose
 /// encoding differs.
@@ -420,7 +424,7 @@ mod tests {
             "bos_token": "<s>",
         });
         reg.attach_chat_template_for_test("tiny", &cfg);
-        assert!(reg.has_chat_encoder("tiny"));
+        assert!(reg.has_chat_formatter("tiny"));
 
         let messages = serde_json::json!([{"role":"user","content":"hi"}]);
         let chat_ids = reg.encode_chat("tiny", &messages).expect("encode_chat");
@@ -435,10 +439,10 @@ mod tests {
 
         // encode_chat is exactly tokenize(render(messages)).
         let rendered = reg
-            .encoders
+            .formatters
             .get("tiny")
             .unwrap()
-            .encoder
+            .formatter
             .render(&messages)
             .unwrap();
         assert_eq!(chat_ids, adapter::encode(&tok, &rendered).unwrap());
@@ -451,7 +455,7 @@ mod tests {
             "tiny".into(),
             adapter::load("tests/fixtures/tiny_tokenizer.json").unwrap(),
         );
-        assert!(!reg.has_chat_encoder("tiny"));
+        assert!(!reg.has_chat_formatter("tiny"));
         let messages = serde_json::json!([{"role":"user","content":"hi"}]);
         assert!(reg.encode_chat("tiny", &messages).is_none());
     }
@@ -473,7 +477,7 @@ mod tests {
                 "bos_token": "<s>",
             }),
         );
-        assert!(reg.has_chat_encoder("tiny"));
+        assert!(reg.has_chat_formatter("tiny"));
         let messages = serde_json::json!([{"role":"user","content":"hi"}]);
         assert!(
             reg.encode_chat("tiny", &messages).is_none(),

experimental/sgl-router/tests/proxy/cache_aware_input_ids.rs

@@ -6,7 +6,7 @@
 //! skips re-tokenizing the same prompt). Asserts the gating contract through
 //! the real chat handler + a MockWorker backend:
 //!
-//! * A plain text chat request on the engine-equivalent chat-encoder path →
+//! * A plain text chat request on the engine-equivalent chat-formatter path →
 //!   the forwarded body carries `input_ids` AND retains `messages`.
 //! * A request carrying `tools` → `input_ids` omitted (the router's encoder
 //!   doesn't render tool schemas, so its ids would diverge from the engine).
@@ -35,8 +35,8 @@ fn build_ctx(url: String) -> Arc<AppContext> {
     let cfg = config();
     let tokenizers = Arc::new(TokenizerRegistry::load_from_config(&cfg).unwrap());
     assert!(
-        tokenizers.has_chat_encoder(MODEL),
-        "deepseek-v4 model id must auto-attach the built-in chat encoder"
+        tokenizers.has_chat_formatter(MODEL),
+        "deepseek-v4 model id must auto-attach the built-in chat formatter"
     );
     let registry = Arc::new(WorkerRegistry::default());
     let _ = registry.add(WorkerSpec {

experimental/sgl-router/tests/proxy/common/cache_aware_fixture.rs

@@ -4,7 +4,7 @@
 //! Shared router config for the cache-aware proxy tests.
 //!
 //! The model id contains `deepseek-v4` so the tokenizer registry auto-attaches the
-//! built-in V4 chat encoder — the engine-equivalent path — with no template fixture.
+//! built-in V4 chat formatter — the engine-equivalent path — with no template fixture.
 
 use sgl_router::config::{
     ActiveLoadConfig, CacheAwareConfig, Config, DiscoveryBackend, ModelConfig, ObservabilityConfig,

experimental/sgl-router/tests/proxy/roundrobin_input_ids.rs

@@ -2,10 +2,10 @@
 // SPDX-License-Identifier: Apache-2.0
 
 //! `input_ids` forwarding is policy-independent: a load-only **round-robin**
-//! policy on a chat-encoder model still forwards `input_ids` to the engine
+//! policy on a chat-formatter model still forwards `input_ids` to the engine
 //! (the engine-tokenization offload), even though it picks workers round-robin
 //! and ignores the tokens for routing. Tokenization is gated on the model's
-//! chat encoder at ingress, not on the policy.
+//! chat formatter at ingress, not on the policy.
 
 use axum::body::Body;
 use axum::http::{Request, StatusCode};
@@ -66,7 +66,7 @@ fn build_ctx(url: String) -> Arc<AppContext> {
     // The handler tokenizes via the AppContext's registry (which carries the V4
     // encoder); the RoundRobin policy itself needs no tokenizer.
     let tokenizers = Arc::new(TokenizerRegistry::load_from_config(&cfg).unwrap());
-    assert!(tokenizers.has_chat_encoder(MODEL));
+    assert!(tokenizers.has_chat_formatter(MODEL));
     let registry = Arc::new(WorkerRegistry::default());
     let _ = registry.add(WorkerSpec {
         id: WorkerId(url.clone()),
@@ -103,7 +103,7 @@ fn captured(mock: &MockWorker) -> Value {
 }
 
 /// A round-robin (load-only) policy still forwards `input_ids` on a
-/// chat-encoder model — the offload is decoupled from routing.
+/// chat-formatter model — the offload is decoupled from routing.
 #[tokio::test]
 async fn round_robin_plain_chat_forwards_input_ids() {
     let mock = MockWorker::start(vec![]).await;
@@ -122,7 +122,7 @@ async fn round_robin_plain_chat_forwards_input_ids() {
     let ids = body.get("input_ids").and_then(|v| v.as_array());
     assert!(
         ids.is_some_and(|a| !a.is_empty()),
-        "round-robin must forward input_ids on a chat-encoder model; got {body}"
+        "round-robin must forward input_ids on a chat-formatter model; got {body}"
     );
     assert!(
         body.get("messages").is_some(),
@@ -154,7 +154,7 @@ async fn round_robin_tool_request_omits_input_ids() {
     );
 }
 
-/// A successful plain-chat forward on a chat-encoder model must NOT emit
+/// A successful plain-chat forward on a chat-formatter model must NOT emit
 /// `sgl_router_ingress_tokenize_errors_total` — that counter fires only when the
 /// offload was expected but the encoder failed. A tool request on the same model
 /// is an *expected* omission (its ids are still engine-equivalent; the

experimental/sgl-router/tests/proxy/sticky_input_ids.rs

@@ -3,8 +3,8 @@
 
 //! Tokenize-once at ingress under the STICKY policy. The engine-tokenization
 //! offload (`input_ids` forwarding) is a property of the MODEL — does it have a
-//! chat encoder? — not of the routing policy, so a sticky-routed request on a
-//! chat-encoder model must forward `input_ids` exactly like cache-aware does,
+//! chat formatter? — not of the routing policy, so a sticky-routed request on a
+//! chat-formatter model must forward `input_ids` exactly like cache-aware does,
 //! while still pinning sessions O(1) by header.
 //!
 //! Asserts through the real chat handler + `MockWorker` backends:
@@ -17,7 +17,7 @@
 //!   routing is unchanged by the added tokenization).
 //!
 //! The model id contains `deepseek-v4` so the tokenizer registry auto-attaches
-//! the built-in V4 chat encoder — the engine-equivalent path — without a
+//! the built-in V4 chat formatter — the engine-equivalent path — without a
 //! template fixture.
 
 use axum::body::Body;
@@ -82,14 +82,14 @@ fn config() -> Config {
 
 /// Build an `AppContext` running the sticky policy over the given workers.
 /// The tokenizer registry is loaded from config (real tiny tokenizer + the
-/// auto-attached V4 chat encoder) so the ingress can tokenize — the sticky
+/// auto-attached V4 chat formatter) so the ingress can tokenize — the sticky
 /// policy itself holds no tokenizer.
 fn build_ctx(worker_urls: &[String]) -> Arc<AppContext> {
     let cfg = config();
     let tokenizers = Arc::new(TokenizerRegistry::load_from_config(&cfg).unwrap());
     assert!(
-        tokenizers.has_chat_encoder(MODEL),
-        "deepseek-v4 model id must auto-attach the built-in chat encoder"
+        tokenizers.has_chat_formatter(MODEL),
+        "deepseek-v4 model id must auto-attach the built-in chat formatter"
     );
     let registry = Arc::new(WorkerRegistry::default());
     for (i, url) in worker_urls.iter().enumerate() {