From 5eec9da50260edfe4fc9fd9d68cc889a6d901e50 Mon Sep 17 00:00:00 2001 From: Chen Cui Date: Tue, 19 May 2026 16:16:06 -0700 Subject: [PATCH] =?UTF-8?q?Revert=20"[recipe]=20feat:=20enable=20THD=20pac?= =?UTF-8?q?king=20by=20default=20for=20Qwen3.5-VL=20finetune=20=E2=80=A6"?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit This reverts commit ae0ae7ef516f15fcd5b544c85ef3276ef895bf40. --- src/megatron/bridge/recipes/qwen_vl/qwen35_vl.py | 5 +---- 1 file changed, 1 insertion(+), 4 deletions(-) diff --git a/src/megatron/bridge/recipes/qwen_vl/qwen35_vl.py b/src/megatron/bridge/recipes/qwen_vl/qwen35_vl.py index 8c9ef45d8c..4bfa6dcc14 100644 --- a/src/megatron/bridge/recipes/qwen_vl/qwen35_vl.py +++ b/src/megatron/bridge/recipes/qwen_vl/qwen35_vl.py @@ -125,10 +125,7 @@ def _qwen35_vl_apply_common( # Dataset configuration cfg.dataset.seq_length = 4096 cfg.dataset.hf_processor_path = hf_path - # Enable THD packing by default for SFT/PEFT: the step function builds - # cu_seqlens so attention skips padding FLOPs on short samples (especially - # when PP/EP forces micro-batches to pad to seq_length). - cfg.dataset.pack_sequences_in_batch = True + cfg.dataset.pack_sequences_in_batch = False # DDP settings cfg.ddp.overlap_grad_reduce = False