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

model_runner: extract post-memory-pool wiring into _init_post_memory (#31601)

Browse Source
This commit is contained in:
Sam Shleifer
2026-07-19 00:57:54 +08:00
committed by GitHub
parent d86ae51fcf
commit b3a0185cab
1 changed files with 5 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/srt/model_executor/model_runner.py
+5
View File
@@ -745,6 +745,11 @@ class ModelRunner:
# Keep a reference so the shared byte buffer is not GC'd. # Keep a reference so the shared byte buffer is not GC'd.
self._unified_memory_pool = result.unified_memory_pool self._unified_memory_pool = result.unified_memory_pool
self._init_post_memory_pool_components()
def _init_post_memory_pool_components(self):
"""Post-pool component wiring, split out of alloc_memory_pool so forks
that build bespoke memory pools can reuse it after allocating them."""
# Must be called AFTER init_memory_pool so the pool object exists for # Must be called AFTER init_memory_pool so the pool object exists for
# canary to monkey-patch, and BEFORE init_decode_cuda_graph so warmup # canary to monkey-patch, and BEFORE init_decode_cuda_graph so warmup
# forwards captured into the graph see the patched pool methods. # forwards captured into the graph see the patched pool methods.