From feabba825afd465ce032108ea29eba90f7e49ee1 Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Wed, 25 Feb 2026 16:40:50 +0000 Subject: [PATCH] Sync UNSLOTH_FORCE_FLOAT32 in forced float32 trainer path (cherry picked from commit d27c104ec831e8e00571206945cafbc31308fcf1) --- unsloth/models/rl.py | 1 + 1 file changed, 1 insertion(+) diff --git a/unsloth/models/rl.py b/unsloth/models/rl.py index 2f36253c35..dd73c836fc 100755 --- a/unsloth/models/rl.py +++ b/unsloth/models/rl.py @@ -733,6 +733,7 @@ def _patch_trl_rl_trainers(trainer_file = "grpo_trainer"): " force_float32 = True\n" "if force_float32:\n" " # Forced float32 training\n" + " os.environ['UNSLOTH_FORCE_FLOAT32'] = '1'\n" " args.fp16 = False\n" " args.bf16 = False\n" " os.environ['ACCELERATE_MIXED_PRECISION'] = 'no'\n"