feat: add moe auxloss and metrics

This commit is contained in:
2026-08-05 18:12:28 +08:00
parent 602b5ce216
commit 9b7e6c205f
9 changed files with 583 additions and 3 deletions
+10
View File
@@ -17,10 +17,15 @@ from astrai.parallel import only_on_rank
from astrai.parallel.setup import get_current_device
from astrai.serialization import Checkpoint
from astrai.trainer.metric_util import (
ctx_get_dead_expert_fraction,
ctx_get_grad_norm,
ctx_get_grad_snr,
ctx_get_load_imbalance_max,
ctx_get_load_imbalance_mean,
ctx_get_loss,
ctx_get_lr,
ctx_get_moe_aux_loss,
ctx_get_router_entropy,
ctx_get_val_loss,
)
from astrai.trainer.train_context import TrainContext
@@ -257,6 +262,11 @@ class MetricCallback(TrainCallback):
"val_loss": ctx_get_val_loss,
"grad_norm": ctx_get_grad_norm,
"grad_snr": ctx_get_grad_snr,
"moe_aux_loss": ctx_get_moe_aux_loss,
"router_entropy": ctx_get_router_entropy,
"dead_expert_fraction": ctx_get_dead_expert_fraction,
"load_imbalance_mean": ctx_get_load_imbalance_mean,
"load_imbalance_max": ctx_get_load_imbalance_max,
}
def _metrics(self, context: TrainContext, names):