[hotfix] fix grad accumulation plus clipping for gemini (#5002)

This commit is contained in:
Baizhou Zhang
2023-11-02 17:59:10 +08:00
committed by GitHub
parent dc003c304c
commit d99b2c961a
4 changed files with 13 additions and 3 deletions

View File

@@ -637,6 +637,7 @@ class Chunk:
# grad chunk is initialized, just reallocate cuda global chunk
self.grad_chunk.cuda_shard = None
self.grad_chunk.is_gathered = True
self.grad_chunk.l2_norm = None
alloc_storage(self.grad_chunk.cuda_global_chunk)
return self.grad_chunk