From 1f6a023c44c2c32e51ca1185fcdb8723c1df8046 Mon Sep 17 00:00:00 2001 From: yuxin00j Date: Wed, 1 Jul 2026 06:10:16 +0000 Subject: [PATCH] Fix timer initialization in StepTimeCallback for LLaMA 3.1 8B training --- .../helm_chart/llama_3_1_8b_cpu_sim.py | 8 +------- 1 file changed, 1 insertion(+), 7 deletions(-) diff --git a/gcsfs/tests/perf/macrobenchmarks/workloads/hf-pytorch-lightning-cpu/helm_chart/llama_3_1_8b_cpu_sim.py b/gcsfs/tests/perf/macrobenchmarks/workloads/hf-pytorch-lightning-cpu/helm_chart/llama_3_1_8b_cpu_sim.py index feb7c5b9..67c1ab25 100644 --- a/gcsfs/tests/perf/macrobenchmarks/workloads/hf-pytorch-lightning-cpu/helm_chart/llama_3_1_8b_cpu_sim.py +++ b/gcsfs/tests/perf/macrobenchmarks/workloads/hf-pytorch-lightning-cpu/helm_chart/llama_3_1_8b_cpu_sim.py @@ -285,13 +285,7 @@ def __init__(self): self.ckpt_time = 0.0 def on_train_start(self, trainer, pl_module): - # Initialize timer at training start to avoid AttributeError when resuming mid-epoch - # (where on_train_epoch_start is skipped). - self.start_time = time.perf_counter() - self.ckpt_time = 0.0 - - def on_train_epoch_start(self, trainer, pl_module): - # Start timer at the beginning of the epoch to capture the first batch's data loading time + # Start timer at the beginning of the training to capture the first batch's data loading time self.start_time = time.perf_counter() self.ckpt_time = 0.0