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

[Doc] Clarify SWA HybridSWAPoolConfigurator comments on all-SWA vs hybrid semantics (#22443)

Browse Source
This commit is contained in:
Liangsheng Yin
2026-04-09 03:02:16 -07:00
committed by GitHub
parent 8a67fb20ea
commit 9fed58805f
1 changed files with 12 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/model_executor/pool_configurator.py
+12 -6
View File
@@ -216,9 +216,15 @@ class HybridSWAPoolConfigurator(MemoryPoolConfigurator):
* kv_size * kv_size
) )
# Bytes per max_total_num_token. # Bytes per token of max_total_num_tokens.
# For hybrid (full_layers > 0): full_tokens * _cell_size = total memory for both pools. #
# For all-SWA (full_layers == 0): swa_tokens * _cell_size = total SWA memory. # Hybrid (full_layers > 0): max_total = full_tokens, so cell_size accounts
# for both pools: F*nf + r*S*ns (where swa_tokens = full_tokens * r).
#
# All-SWA (full_layers == 0): max_total = swa_tokens directly. The ratio
# is meaningless here -- there is no full pool to relate to, and every
# token beyond the sliding window can be evicted. So cell_size = S*ns,
# with no ratio factor applied.
if self._full_layers_num == 0: if self._full_layers_num == 0:
self._cell_size = self._swa_per_token * self._swa_layers_num self._cell_size = self._swa_per_token * self._swa_layers_num
else: else:
@@ -238,7 +244,8 @@ class HybridSWAPoolConfigurator(MemoryPoolConfigurator):
return (x // page_size) * page_size return (x // page_size) * page_size
if self._full_layers_num == 0: if self._full_layers_num == 0:
# All layers are SWA — no full pool needed # All-SWA: no full pool, max_total = actual SWA pool size.
# Ratio is not applied -- see __init__ comment.
swa_tokens = align_page_size(max_total_num_tokens) swa_tokens = align_page_size(max_total_num_tokens)
logger.info( logger.info(
f"Use sliding window memory pool (all SWA). " f"Use sliding window memory pool (all SWA). "
@@ -250,8 +257,7 @@ class HybridSWAPoolConfigurator(MemoryPoolConfigurator):
swa_max_total_num_tokens=swa_tokens, swa_max_total_num_tokens=swa_tokens,
) )
# full_tokens = max_total_num_tokens (page aligned) # Hybrid: full_tokens = max_total_num_tokens, swa_tokens = full_tokens * ratio
# swa_tokens = full_tokens * ratio (page aligned)
full_tokens = align_page_size(max_total_num_tokens) full_tokens = align_page_size(max_total_num_tokens)
swa_tokens = align_page_size(int(full_tokens * self._swa_full_tokens_ratio)) swa_tokens = align_page_size(int(full_tokens * self._swa_full_tokens_ratio))