fix: enable corex topk_softmax kernel — verified correct on real BI-V100
verify_topk_softmax.py results: IDs match: True (0/32 mismatches) Max weight diff: 0.00000003 Speedup: 2.54x vs PyTorch (0.025ms vs 0.064ms) Previous disable was based on speculation, not measurement.
This commit is contained in:
@@ -46,6 +46,6 @@ env:
|
|||||||
- name: BI100_GDN_RESTORE_MODE
|
- name: BI100_GDN_RESTORE_MODE
|
||||||
value: hybrid64
|
value: hybrid64
|
||||||
- name: BI100_MOE_COREX_TOPK_SOFTMAX
|
- name: BI100_MOE_COREX_TOPK_SOFTMAX
|
||||||
value: '0'
|
value: '1'
|
||||||
- name: PYTORCH_CUDA_ALLOC_CONF
|
- name: PYTORCH_CUDA_ALLOC_CONF
|
||||||
value: expandable_segments:True
|
value: expandable_segments:True
|
||||||
|
|||||||
Reference in New Issue
Block a user