[diffusion] UX: quiet request-path cache diagnostics (#38783)

Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>

python/sglang/multimodal_gen/runtime/managers/memory_managers/component_residency_strategies.py

@@ -310,7 +310,7 @@ class LayerwiseOffloadStrategy(ComponentResidencyStrategy):
                 if advise_cold is not None:
                     paged_out += int(advise_cold(room_bytes=room_bytes) or 0)
             if paged_out:
-                logger.info(
+                logger.debug(
                     "Layerwise offload: paged out the first %.1f GiB of %s so the "
                     "next request's stream fits the %.1f GiB the cache can give it.",
                     paged_out / 1024**3,

python/sglang/multimodal_gen/runtime/models/dits/minimax_h3_adaln_cache.py

@@ -306,7 +306,7 @@ class MiniMaxH3AdalnHostTier:
     def _skip_for_pressure(self, group_key) -> None:
         if self.stats is not None:
             self.stats.host_pressure_skips += 1
-        logger.info(
+        logger.debug(
             "MiniMax H3 AdaLN host tier: group of %d plan(s) skipped under "
             "memory pressure; it will recompute on its next occurrence",
             len(group_key),
@@ -651,7 +651,7 @@ class MiniMaxH3AdalnCache(nn.Module):
         self._slots.update(pending_slots)
         self.rebuilds += 1
         self.stats.built_plans += len(missing)
-        logger.info(
+        logger.debug(
             "MiniMax H3 AdaLN: rebuilt %d plan(s), %d/%d resident, pass #%d",
             len(missing),
             len(self._slots),
@@ -727,7 +727,7 @@ class MiniMaxH3AdalnCache(nn.Module):
             self.plan_lengths[assignments[key]] = timesteps.numel()
         self._slots.update(assignments)
         self.stats.host_hit_plans += len(missing)
-        logger.info(
+        logger.debug(
             "MiniMax H3 AdaLN: %d plan(s) from the host cache, %d/%d resident",
             len(missing),
             len(self._slots),

python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/helios_denoising.py

@@ -128,7 +128,7 @@ class HeliosChunkedDenoisingStage(PipelineStage):
             return
         if want:
             if mount_helios_gated_residual(self.transformer):
-                logger.info(
+                logger.debug(
                     "Mounted Helios per-token gated residual for quality=%s", quality
                 )
         else: