fix(moe): fix RoutedExpertsCapturer assertion failure with DP>1 and MK path (#37879)

This commit is contained in:
Le Yang
2026-04-12 10:28:17 -04:00
committed by GitHub
parent 3244a2ebf2
commit 21fab0a3db
2 changed files with 103 additions and 5 deletions
@@ -176,11 +176,27 @@ class RoutedExpertsCapturer:
end_loc = topk_ids.shape[0]
token_num_per_dp = topk_ids.shape[0]
else: # multi dp
token_num_per_dp = ctx.dp_metadata.num_tokens_across_dp_cpu[self.dp_rank]
cumsum = torch.cumsum(ctx.dp_metadata.num_tokens_across_dp_cpu, dim=0)
assert cumsum[-1] == topk_ids.shape[0]
end_loc = cumsum[self.dp_rank]
start_loc = end_loc - token_num_per_dp
num_tokens_dp = ctx.dp_metadata.num_tokens_across_dp_cpu
token_num_per_dp = int(num_tokens_dp[self.dp_rank].item())
total = int(num_tokens_dp.sum().item())
n = topk_ids.shape[0]
if n == total:
# Naive dispatch: all DP ranks' tokens concatenated before routing.
cumsum = torch.cumsum(num_tokens_dp, dim=0)
end_loc = int(cumsum[self.dp_rank].item())
start_loc = end_loc - token_num_per_dp
elif n == token_num_per_dp:
# Modular-kernel path: DP combine happens inside quant_method.apply;
# select_experts only sees this rank's tokens.
start_loc = 0
end_loc = token_num_per_dp
else:
raise AssertionError(
"RoutedExpertsCapturer: unexpected topk_ids batch dim "
f"{n} (expected {total} or {token_num_per_dp} "
f"for dp_rank={self.dp_rank})"
)
if layer_id >= self._device_buffer.shape[1]:
return