fix and test EPLB balancedness calculation (#51813)

Signed-off-by: jdebache <jdebache@nvidia.com>
This commit is contained in:
Julien Debache
2026-08-12 20:27:22 -07:00
committed by GitHub
parent 61826c1c6f
commit 6e502b62cb
2 changed files with 31 additions and 3 deletions
+11 -2
View File
@@ -61,6 +61,14 @@ from .rebalance_execute import (
logger = init_logger(__name__)
def _compute_eplb_load_stats(
num_tokens_per_rank: torch.Tensor,
) -> tuple[torch.Tensor, torch.Tensor]:
avg_tokens = num_tokens_per_rank.mean(dim=1).sum()
max_tokens = num_tokens_per_rank.max(dim=1).values.sum()
return avg_tokens, max_tokens
@dataclass
class EplbStats:
"""
@@ -584,8 +592,9 @@ class EplbState:
# Compute balancedness ratio:
# for each layer:
# (mean load across ranks) / (max load across ranks)
avg_tokens_tensor = num_tokens_per_rank.mean(dim=0).sum(dim=0)
max_tokens_tensor = num_tokens_per_rank.max(dim=0).values.sum(dim=0)
avg_tokens_tensor, max_tokens_tensor = _compute_eplb_load_stats(
num_tokens_per_rank
)
# Just to make type checker happy
tokens_tensors: list[float] = torch.stack(