refactor(qwen35): split TP response contract into tp_executor/responses.rs

pegainfer-qwen35/src/cublas_thread.rs

@@ -0,0 +1,40 @@
+//! Bind the calling thread to a model's CUDA device/context and cuBLAS
+//! handle; the guard destroys the handle when the binding ends.
+
+use anyhow::Result;
+
+use crate::weights::Qwen35Model;
+
+pub(crate) struct CublasThreadGuard;
+
+impl Drop for CublasThreadGuard {
+    fn drop(&mut self) {
+        unsafe {
+            crate::ffi::cublas_destroy();
+        }
+    }
+}
+
+/// Bind this thread to `model`'s device and context, then initialize cuBLAS.
+/// `role` names the thread kind in failure messages ("scheduler", "TP worker");
+/// every thread that runs model work must hold the returned guard.
+pub(crate) fn bind_model_thread(model: &Qwen35Model, role: &str) -> Result<CublasThreadGuard> {
+    let ctx = model.device_ctx();
+    unsafe {
+        let err = crate::ffi::cuda_set_device(ctx.device_ordinal as i32);
+        if err != 0 {
+            return Err(anyhow::anyhow!(
+                "Failed to set CUDA device {} on Qwen3.5 {role} thread: cudaError={}",
+                ctx.device_ordinal,
+                err
+            ));
+        }
+    }
+    ctx.ctx.bind_to_thread().map_err(|e| {
+        anyhow::anyhow!("Failed to bind CUDA context to Qwen3.5 {role} thread: {e}")
+    })?;
+    unsafe {
+        crate::ffi::cublas_init();
+    }
+    Ok(CublasThreadGuard)
+}

pegainfer-qwen35/src/lib.rs

@@ -9,6 +9,7 @@
 mod batch_decode;
 pub(crate) mod batch_decode_graph;
 pub(crate) mod config;
+mod cublas_thread;
 mod decode_buffers;
 mod executor;
 mod ffi;

pegainfer-qwen35/src/scheduler/mod.rs

@@ -379,20 +379,26 @@ pub(crate) fn start_with_capacity_and_policy(
 
     let join_handle = thread::Builder::new()
         .name("scheduler-qwen35".into())
-        .spawn(move || match bind_model_thread(backend.model()) {
-            Ok(_guard) => {
-                let _ = startup_tx.send(Ok(()));
-                scheduler_loop(
-                    SchedulerBackend::Single(backend),
-                    submit_rx,
-                    seed,
-                    max_prefill_tokens,
-                    scheduler_policy,
-                    load_tx,
-                );
-            }
-            Err(err) => {
-                let _ = startup_tx.send(Err(err));
+        .spawn(move || {
+            match crate::cublas_thread::bind_model_thread(backend.model(), "scheduler") {
+                Ok(_guard) => {
+                    if let Err(err) = backend.model().tune_decode_gemm_algos() {
+                        let _ = startup_tx.send(Err(err));
+                        return;
+                    }
+                    let _ = startup_tx.send(Ok(()));
+                    scheduler_loop(
+                        SchedulerBackend::Single(backend),
+                        submit_rx,
+                        seed,
+                        max_prefill_tokens,
+                        scheduler_policy,
+                        load_tx,
+                    );
+                }
+                Err(err) => {
+                    let _ = startup_tx.send(Err(err));
+                }
             }
         })
         .expect("failed to spawn Qwen3.5 scheduler thread");
@@ -491,38 +497,6 @@ fn servable_len(max_context: usize, max_pages: usize, page_size: usize) -> u32 {
         .unwrap_or(u32::MAX)
 }
 
-struct CublasThreadGuard;
-
-impl Drop for CublasThreadGuard {
-    fn drop(&mut self) {
-        unsafe {
-            crate::ffi::cublas_destroy();
-        }
-    }
-}
-
-fn bind_model_thread(model: &Qwen35Model) -> Result<CublasThreadGuard> {
-    let ctx = model.device_ctx();
-    unsafe {
-        let err = crate::ffi::cuda_set_device(ctx.device_ordinal as i32);
-        if err != 0 {
-            return Err(anyhow::anyhow!(
-                "Failed to set CUDA device {} on Qwen3.5 scheduler thread: cudaError={}",
-                ctx.device_ordinal,
-                err
-            ));
-        }
-    }
-    ctx.ctx.bind_to_thread().map_err(|e| {
-        anyhow::anyhow!("Failed to bind CUDA context to Qwen3.5 scheduler thread: {e}")
-    })?;
-    unsafe {
-        crate::ffi::cublas_init();
-    }
-    model.tune_decode_gemm_algos()?;
-    Ok(CublasThreadGuard)
-}
-
 // ── Main loop ───────────────────────────────────────────────────────────
 
 fn publish_load(