From e37b4d283baeb72e4a908d91831a38f795e2acbc Mon Sep 17 00:00:00 2001
From: Claude <claude@project6.dev>
Date: Sat, 8 Aug 2026 07:36:03 +0000
Subject: [PATCH] =?UTF-8?q?env(yaml):=20CCCL=20buddy=5Fallocator=20pattern?=
 =?UTF-8?q?=20=E2=80=94=20PYTORCH=5FCUDA=5FALLOC=5FCONF=20+=20OMP=5FNUM=5F?=
 =?UTF-8?q?THREADS?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

CCCL buddy_allocator.cu teaches: control memory block fragmentation
at the allocator level. Sub168 OOM trace shows 'max_split_size_mb'
suggestion. Adding PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:512
prevents PyTorch memory fragmentation that caused Sub168's final OOM.

OMP_NUM_THREADS=1 matches Sub168 docker log:
  'Reducing Torch parallelism from 64 threads to 1'

CCCL device_reduce policy_selector pattern: hardware-adaptive params
through environment, not code changes. computility-run.yaml env vars
are the serving-safe equivalent of CCCL policy_selector.
---
 computility-run.yaml | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/computility-run.yaml b/computility-run.yaml
index 0be4999a..02adf99e 100644
--- a/computility-run.yaml
+++ b/computility-run.yaml
@@ -51,3 +51,7 @@ env:
       value: /tmp/vllm-request-metrics.jsonl
     - name: VLLM_CACHE_BLOCK_SIZE
       value: '16'
+    - name: PYTORCH_CUDA_ALLOC_CONF
+      value: max_split_size_mb:512
+    - name: OMP_NUM_THREADS
+      value: '1'