fix: remove expandable_segments — CoreX CUDACachingAllocator不支持
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。
This commit is contained in:
@@ -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
|
||||
|
||||
Reference in New Issue
Block a user