Complete server warmup before scripted runtime scripts start (#27445)

python/sglang/test/scripted_runtime/http_server.py

@@ -151,12 +151,13 @@ class ScriptedHttpServer:
         # _reset_engine_state, which POSTs to this server's own HTTP port, so
         # block until the port is bound and routing before any script can run.
         #
-        # Wait for *any* HTTP response, not status 200: in scripted mode the
-        # scheduler is driven by the script, so normal warmup never completes
-        # and /health stays 503 (server_status == Starting) for the whole run.
-        # A 503 still proves the socket is bound and routes are registered,
-        # which is all the control POSTs need.
-        url = f"{self._base_url}/health"
+        # Poll /model_info rather than /health: once the server reports Up,
+        # /health runs the generation-based health check (a real probe request
+        # through the scheduler), which cannot complete while the scheduler
+        # waits for scripts between RunScript commands. Any /model_info
+        # response proves the socket is bound and routes are registered, which
+        # is all the control POSTs need.
+        url = f"{self._base_url}/model_info"
         deadline = time.monotonic() + HTTP_READY_TIMEOUT_S
         while time.monotonic() < deadline:
             if not self._server_process.is_alive():

python/sglang/test/scripted_runtime/scheduler_hook.py

@@ -2,6 +2,7 @@ from __future__ import annotations
 
 import logging
 import sys
+import time
 import traceback
 from dataclasses import dataclass
 from pathlib import Path
@@ -36,6 +37,9 @@ if TYPE_CHECKING:
 logger = logging.getLogger(__name__)
 
 RESET_DRAIN_MAX_STEPS: int = 200
+# Below the test-side LISTENER_ACCEPT_TIMEOUT_S so a stuck warmup surfaces as
+# this specific error instead of a generic handshake timeout.
+WARMUP_DRIVE_TIMEOUT_S: float = 120.0
 
 
 @dataclass(frozen=True, slots=True)
@@ -47,6 +51,47 @@ class ScriptedBatchRecord:
     chunked_rid: Optional[str]
 
 
+def _drive_engine_through_warmup(ctx: ScriptedContext) -> Generator:
+    """Run the engine until the server warmup request has been received and
+    fully processed, so scripts never observe foreign warmup traffic."""
+    scheduler = ctx.scheduler
+    server_args = scheduler.server_args
+    if server_args.skip_server_warmup:
+        logger.info("scripted_runtime: skip_server_warmup set, not driving warmup")
+        return
+
+    logger.info("scripted_runtime: driving engine until server warmup completes")
+    start_time = time.monotonic()
+
+    # is_fully_idle() can transiently report idle while a PP microbatch result
+    # is still in flight, so require it to hold for two full microbatch
+    # rotations after the warmup request was observed on the recv socket.
+    quiesce_iters = 2 * (server_args.pp_size + server_args.pp_async_batch_depth)
+    proxy = ctx._tokenizer_recv_proxy
+    deadline = start_time + WARMUP_DRIVE_TIMEOUT_S
+
+    idle_streak = 0
+    while idle_streak < quiesce_iters:
+        if time.monotonic() >= deadline:
+            raise RuntimeError(
+                "scripted_runtime: server warmup did not complete within "
+                f"{WARMUP_DRIVE_TIMEOUT_S}s "
+                f"(work_reqs_seen={proxy.work_reqs_seen}, "
+                f"idle_streak={idle_streak})"
+            )
+        yield
+        if proxy.work_reqs_seen > 0 and scheduler.is_fully_idle():
+            idle_streak += 1
+        else:
+            idle_streak = 0
+
+    logger.info(
+        "scripted_runtime: server warmup drained in %.1fs (work_reqs_seen=%d)",
+        time.monotonic() - start_time,
+        proxy.work_reqs_seen,
+    )
+
+
 def _reset_engine_state(ctx: ScriptedContext) -> Generator:
     scheduler = ctx.scheduler
 
@@ -106,6 +151,7 @@ class ScriptedSchedulerHook:
         ctx_zmq = zmq.Context()
         socket = get_zmq_socket(ctx_zmq, zmq.PAIR, endpoint, bind=False)
         try:
+            yield from _drive_engine_through_warmup(self._context)
             socket.send_pyobj(HookReady())
             while True:
                 msg = socket.recv_pyobj()

python/sglang/test/scripted_runtime/tokenizer_recv_proxy.py

@@ -6,12 +6,27 @@ from typing import Any, Callable
 
 import zmq
 
+from sglang.srt.managers.io_struct import (
+    BatchTokenizedEmbeddingReqInput,
+    BatchTokenizedGenerateReqInput,
+    TokenizedEmbeddingReqInput,
+    TokenizedGenerateReqInput,
+)
+
+_WORK_REQ_TYPES = (
+    TokenizedGenerateReqInput,
+    TokenizedEmbeddingReqInput,
+    BatchTokenizedGenerateReqInput,
+    BatchTokenizedEmbeddingReqInput,
+)
+
 
 class ScriptedTokenizerRecvProxy:
 
     def __init__(self, *, underlying: zmq.Socket) -> None:
         self._underlying = underlying
         self._buffer: deque = deque()
+        self.work_reqs_seen: int = 0
 
     def recv_pyobj(self, flags: int = 0) -> Any:
         self._drain_underlying()
@@ -52,4 +67,6 @@ class ScriptedTokenizerRecvProxy:
                 req = self._underlying.recv_pyobj(zmq.NOBLOCK)
             except zmq.ZMQError:
                 break
+            if isinstance(req, _WORK_REQ_TYPES):
+                self.work_reqs_seen += 1
             self._buffer.append(req)