test(sglang): gate VL mm_agg_router hit-rate check on wide-pad SGLang builds

tests/serve/multimodal_profiles/sglang.py

@@ -10,6 +10,22 @@
     make_image_payload,
     make_image_payload_cached_tokens,
 )
+from tests.utils.payloads import CachedTokensChatPayload
+
+
+class SGLangMmRoutingPayload(CachedTokensChatPayload):
+    def bind_environment(self, env: dict[str, str]) -> None:
+        """Keep router observations while relaxing the floor on wide-pad builds."""
+        # Defer the optional backend import until this test runs.
+        from dynamo.sglang._compat import sglang_has_wide_mm_pad_values
+
+        super().bind_environment(env)
+        if sglang_has_wide_mm_pad_values():
+            # Wide-pad SGLang builds route media by text prefix only (#17653),
+            # so exact media overlap is required only on scalar-pad builds.
+            # The router must still report hit-rate observations.
+            self.min_avg_kv_hit_rate = 0.0
+
 
 # pd = multimodal prefill+decode on one GPU, no separate encode worker.
 SGLANG_TOPOLOGY_SCRIPTS: dict[str, str] = {
@@ -62,6 +78,7 @@
                             ["green"],
                             require_rust_processor_init=True,
                             min_avg_kv_hit_rate=0.9,
+                            payload_cls=SGLangMmRoutingPayload,
                         )
                     )
                 ],
@@ -94,6 +111,7 @@
                             ["green"],
                             require_rust_processor_init=True,
                             min_avg_kv_hit_rate=0.9,
+                            payload_cls=SGLangMmRoutingPayload,
                         )
                     )
                 ],
@@ -116,6 +134,7 @@
                             ["green"],
                             require_rust_processor_init=True,
                             min_avg_kv_hit_rate=0.9,
+                            payload_cls=SGLangMmRoutingPayload,
                         )
                     )
                 ],

tests/serve/test_sglang.py

@@ -28,6 +28,7 @@
 from tests.serve.multimodal_profiles.sglang import (
     SGLANG_MULTIMODAL_PROFILES,
     SGLANG_TOPOLOGY_SCRIPTS,
+    SGLangMmRoutingPayload,
 )
 from tests.utils.constants import DefaultPort, DynamoPortRange
 from tests.utils.engine_process import EngineConfig
@@ -52,7 +53,6 @@
     router_selection_chat_payload_default,
 )
 from tests.utils.payloads import (
-    CachedTokensChatPayload,
     ChatPayload,
     HttpErrorPayload,
     ImageGenerationPayload,
@@ -85,18 +85,6 @@ class SGLangConfig(EngineConfig):
     stragglers: list[str] = field(default_factory=lambda: ["SGLANG:EngineCore"])
 
 
-class SGLangVideoRoutingPayload(CachedTokensChatPayload):
-    def bind_environment(self, env: dict[str, str]) -> None:
-        # Defer the optional backend import until this test runs.
-        from dynamo.sglang._compat import sglang_has_wide_mm_pad_values
-
-        super().bind_environment(env)
-        if sglang_has_wide_mm_pad_values():
-            # Wide IDs use text-prefix routing, not exact media overlap.
-            # Keep response-content and worker cached-token checks enabled.
-            self.min_avg_kv_hit_rate = 0.0
-
-
 sglang_dir = os.environ.get("SGLANG_DIR") or os.path.join(
     WORKSPACE_DIR, "examples/backends/sglang"
 )
@@ -807,7 +795,7 @@ def bind_environment(self, env: dict[str, str]) -> None:
         timeout=450,
         frontend_port=DefaultPort.FRONTEND.value,
         request_payloads=[
-            SGLangVideoRoutingPayload(
+            SGLangMmRoutingPayload(
                 body={
                     "messages": [
                         {

tests/utils/multimodal.py

@@ -102,6 +102,7 @@ def make_image_payload_cached_tokens(
     min_routing_total_blocks: int = 0,
     min_avg_kv_hit_rate: float = 0.0,
     prompt_filler_repeats: int = 0,
+    payload_cls: type[CachedTokensChatPayload] = CachedTokensChatPayload,
 ) -> CachedTokensChatPayload:
     """Image payload that asserts MM-aware KV cache reuse on repeats.
 
@@ -113,11 +114,13 @@ def make_image_payload_cached_tokens(
     ``prompt_filler_repeats`` prepends N copies of a topic-aligned filler so
     the routing-token sequence spans multiple blocks on specs that use a
     coarse routing-block size (e.g. qwen3_5 family at block_size=544).
+    ``payload_cls`` lets a backend substitute a subclass that adjusts the
+    checks at run time (e.g. SGLang's wide-pad gate).
     """
     prompt = _MULTIMODAL_COLOR_PROMPT
     if prompt_filler_repeats > 0:
         prompt = (_COLOR_PROMPT_FILLER * prompt_filler_repeats).strip() + " " + prompt
-    return CachedTokensChatPayload(
+    return payload_cls(
         body={
             "messages": [
                 {

tests/utils/payloads.py

@@ -1474,6 +1474,9 @@ def __init__(
         # block-count, etc.), move into a RouterMetricsAssertion
         # mixin/subclass.
         self.min_avg_kv_hit_rate = min_avg_kv_hit_rate
+        # Require new router_kv_hit_rate observations on R2+. A subclass may
+        # drop the floor at run time and keep this check.
+        self.require_kv_hit_rate_observations = min_avg_kv_hit_rate > 0
         self._metrics_baseline: Optional[tuple[float, float]] = None
 
     def validate(self, response: Any, content: str) -> None:
@@ -1528,7 +1531,7 @@ def validate(self, response: Any, content: str) -> None:
         if (
             self._metrics_baseline is None
             and self._request_count == 1
-            and self.min_avg_kv_hit_rate > 0
+            and self.require_kv_hit_rate_observations
         ):
             self._metrics_baseline = self._scrape_router_kv_hit_rate()
 
@@ -1564,8 +1567,8 @@ def _scrape_router_kv_hit_rate(self) -> Optional[tuple[float, float]]:
 
     def final_validation(self) -> None:
         """Assert cached_tokens >= min_cached_tokens on at least one repeat
-        (only when min_cached_tokens > 0), and (if set) router_kv_hit_rate
-        post-R1 mean >= min_avg_kv_hit_rate.
+        (only when min_cached_tokens > 0), and (if set) new router_kv_hit_rate
+        observations with a post-R1 mean >= min_avg_kv_hit_rate.
         """
         # Only assert cached tokens when a positive threshold is set; a caller
         # validating purely via the router metric passes min_cached_tokens=0.
@@ -1581,11 +1584,11 @@ def final_validation(self) -> None:
                 "✓ Final validation PASSED: cached_tokens found in repeated requests"
             )
 
-        if self.min_avg_kv_hit_rate <= 0:
+        if not self.require_kv_hit_rate_observations:
             return
         if self._metrics_baseline is None:
             raise AssertionError(
-                "min_avg_kv_hit_rate set but no metrics baseline captured "
+                "router_kv_hit_rate check set but no metrics baseline captured "
                 "(R1 validate() didn't run or /metrics was unreachable)."
             )
         after = self._scrape_router_kv_hit_rate()
@@ -1604,6 +1607,12 @@ def final_validation(self) -> None:
                 f"MM-routing likely not engaging on repeat requests."
             )
         avg = d_sum / d_count
+        if self.min_avg_kv_hit_rate <= 0:
+            logger.info(
+                f"✓ router_kv_hit_rate: {d_count} new observations over R2+ "
+                f"(mean {avg:.3f}, no floor)"
+            )
+            return
         if avg < self.min_avg_kv_hit_rate:
             raise AssertionError(
                 f"router_kv_hit_rate: mean over R2+ ({avg:.3f}) below required "