minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

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

Browse Source 
Co-authored-by: Mick Qian <mickqian@users.noreply.github.com>
This commit is contained in:
Mick
2026-09-11 22:45:21 +08:00
committed by GitHub
co-authored by Mick Qian
parent 593c7a900d
commit dd67a42634
3 changed files with 5 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/multimodal_gen/runtime/managers/memory_managers/component_residency_strategies.py
+1 -1
View File
@@ -310,7 +310,7 @@ class LayerwiseOffloadStrategy(ComponentResidencyStrategy):
if advise_cold is not None: if advise_cold is not None:
paged_out += int(advise_cold(room_bytes=room_bytes) or 0) paged_out += int(advise_cold(room_bytes=room_bytes) or 0)
if paged_out: if paged_out:
logger.info( logger.debug(
"Layerwise offload: paged out the first %.1f GiB of %s so the " "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.", "next request's stream fits the %.1f GiB the cache can give it.",
paged_out / 1024**3, paged_out / 1024**3,
python/sglang/multimodal_gen/runtime/models/dits/minimax_h3_adaln_cache.py
+3 -3
View File
@@ -306,7 +306,7 @@ class MiniMaxH3AdalnHostTier:
def _skip_for_pressure(self, group_key) -> None: def _skip_for_pressure(self, group_key) -> None:
if self.stats is not None: if self.stats is not None:
self.stats.host_pressure_skips += 1 self.stats.host_pressure_skips += 1
logger.info( logger.debug(
"MiniMax H3 AdaLN host tier: group of %d plan(s) skipped under " "MiniMax H3 AdaLN host tier: group of %d plan(s) skipped under "
"memory pressure; it will recompute on its next occurrence", "memory pressure; it will recompute on its next occurrence",
len(group_key), len(group_key),
@@ -651,7 +651,7 @@ class MiniMaxH3AdalnCache(nn.Module):
self._slots.update(pending_slots) self._slots.update(pending_slots)
self.rebuilds += 1 self.rebuilds += 1
self.stats.built_plans += len(missing) self.stats.built_plans += len(missing)
logger.info( logger.debug(
"MiniMax H3 AdaLN: rebuilt %d plan(s), %d/%d resident, pass #%d", "MiniMax H3 AdaLN: rebuilt %d plan(s), %d/%d resident, pass #%d",
len(missing), len(missing),
len(self._slots), len(self._slots),
@@ -727,7 +727,7 @@ class MiniMaxH3AdalnCache(nn.Module):
self.plan_lengths[assignments[key]] = timesteps.numel() self.plan_lengths[assignments[key]] = timesteps.numel()
self._slots.update(assignments) self._slots.update(assignments)
self.stats.host_hit_plans += len(missing) self.stats.host_hit_plans += len(missing)
logger.info( logger.debug(
"MiniMax H3 AdaLN: %d plan(s) from the host cache, %d/%d resident", "MiniMax H3 AdaLN: %d plan(s) from the host cache, %d/%d resident",
len(missing), len(missing),
len(self._slots), len(self._slots),
python/sglang/multimodal_gen/runtime/pipelines_core/stages/model_specific_stages/helios_denoising.py
+1 -1
View File
@@ -128,7 +128,7 @@ class HeliosChunkedDenoisingStage(PipelineStage):
return return
if want: if want:
if mount_helios_gated_residual(self.transformer): if mount_helios_gated_residual(self.transformer):
logger.info( logger.debug(
"Mounted Helios per-token gated residual for quality=%s", quality "Mounted Helios per-token gated residual for quality=%s", quality
) )
else: else: