Build Rust extensions on demand in source checkouts (#34994)

rust/sglang-server/Cargo.toml

@@ -7,7 +7,7 @@ license.workspace = true
 
 # Consumed by python/setup.py: registers this crate as a PyO3 extension module.
 [package.metadata.sglang]
-python-module = "sglang.srt.server._core"
+python-module = "sglang.srt.rust_extensions._server"
 # Always build optimized, even for an editable install.
 debug = false
 

rust/sglang-server/pyproject.toml

@@ -13,4 +13,4 @@ classifiers = [
 dynamic = ["version"]
 
 [tool.maturin]
-module-name = "_core"
+module-name = "_server"

rust/sglang-server/src/api_server/openai/chat.rs

@@ -836,6 +836,7 @@ pub(super) fn chat_logprobs(extras: Option<&ChunkExtras>) -> ChatChoiceLogprobs
             bytes: Some(token.as_bytes().to_vec()),
             token,
             logprob,
+            token_id: u32::try_from(token_id).ok(),
             top_logprobs,
         });
     }
@@ -1020,6 +1021,7 @@ mod tests {
         let logprobs = chat_logprobs(Some(&extras));
         let token = &logprobs.content.unwrap()[0];
         assert_eq!(token.token, "x");
+        assert_eq!(token.token_id, Some(7));
         assert_eq!(token.top_logprobs.len(), 2);
         assert_eq!(token.top_logprobs[1].token, "y");
     }

rust/sglang-server/src/lib.rs

@@ -292,7 +292,7 @@ static LOG_GUARD: std::sync::OnceLock<tracing_appender::non_blocking::WorkerGuar
     std::sync::OnceLock::new();
 
 #[pymodule]
-fn _core(m: &Bound<'_, PyModule>) -> PyResult<()> {
+fn _server(m: &Bound<'_, PyModule>) -> PyResult<()> {
     // Initialize tracing once; ignore if already set by the host process.
     // Non-blocking writer: emitting threads (axum workers, egress, detok) only
     // enqueue; a dedicated thread does the stdout formatting-flush + syscall.