[Bugfix] Give each data-parallel engine its own global RNG streams (#59788)
Signed-off-by: aoshen02 <aoshen@inferact.ai> Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5.5
parent
877ddcc456
commit
d09ff77637
@@ -530,8 +530,12 @@ def kv_cache_dtype_str_to_dtype(
|
||||
return STR_DTYPE_TO_TORCH_DTYPE[kv_cache_dtype]
|
||||
|
||||
|
||||
def set_random_seed(seed: int | None) -> None:
|
||||
def set_random_seed(seed: int | None, data_parallel_index: int = 0) -> None:
|
||||
if seed is not None:
|
||||
if data_parallel_index:
|
||||
# DP engines share the seed but must not sample unseeded requests alike.
|
||||
ss = np.random.SeedSequence((seed, data_parallel_index))
|
||||
seed = int(ss.generate_state(1)[0])
|
||||
random.seed(seed)
|
||||
np.random.seed(seed)
|
||||
torch.manual_seed(seed)
|
||||
|
||||
Reference in New Issue
Block a user