From c1e7065076898c30bbd13eb5aed839bc024c8a7f Mon Sep 17 00:00:00 2001
From: dylanyunlon <dylanyunlon@users.noreply.github.com>
Date: Thu, 13 Aug 2026 09:08:47 +0000
Subject: [PATCH] =?UTF-8?q?fix:=20remove=20expandable=5Fsegments=20?=
 =?UTF-8?q?=E2=80=94=20CoreX=20CUDACachingAllocator=E4=B8=8D=E6=94=AF?=
 =?UTF-8?q?=E6=8C=81?=
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

BI-V100 CoreX PyTorch的CUDACachingAllocator.cpp:545没有实现
expandable segment特性，导致模型加载阶段(VocabParallelEmbedding)
直接INTERNAL ASSERT FAILED崩溃。

替换为max_split_size_mb:512减少内存碎片化。
---
 computility-run.yaml | 12 +-----------
 1 file changed, 1 insertion(+), 11 deletions(-)

diff --git a/computility-run.yaml b/computility-run.yaml
index 38cf2789..962a337b 100644
--- a/computility-run.yaml
+++ b/computility-run.yaml
@@ -48,14 +48,4 @@ env:
     - name: BI100_MOE_COREX_TOPK_SOFTMAX
       value: '1'
     - name: PYTORCH_CUDA_ALLOC_CONF
-      value: expandable_segments:True
-    - name: LD_PRELOAD
-      value: /workspace/qwen3_6_scripts/cccl_preload_allocator.so
-    - name: CCCL_ALLOC_BIN_GROWTH
-      value: '8'
-    - name: CCCL_ALLOC_MIN_BIN
-      value: '3'
-    - name: CCCL_ALLOC_MAX_BIN
-      value: '13'
-    - name: CCCL_ALLOC_MAX_CACHED_MB
-      value: '4096'
+      value: max_split_size_mb:512