From dd54005c7474091e2e751c27364b5234a50a5639 Mon Sep 17 00:00:00 2001 From: Daniel Han Date: Tue, 18 Mar 2025 02:05:30 -0700 Subject: [PATCH] Update rl_replacements.py --- unsloth/models/rl_replacements.py | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-) diff --git a/unsloth/models/rl_replacements.py b/unsloth/models/rl_replacements.py index 4071ef835a..841da92da4 100644 --- a/unsloth/models/rl_replacements.py +++ b/unsloth/models/rl_replacements.py @@ -212,7 +212,8 @@ def grpo_trainer__get_per_token_logps(function_name, function): # Otherwise, calculate normally: if not hasattr(self, '_autocast_dtype'): self._autocast_dtype = torch.float16 if os.environ.get('ACCELERATE_MIXED_PRECISION', 'fp16') == 'fp16' else torch.bfloat16 - if os.environ.get('UNSLOTH_FORCE_FLOAT32', '0') == '1': self._autocast_dtype = torch.float32 + if os.environ.get('UNSLOTH_FORCE_FLOAT32', '0') == '1': self._autocast_dtype = torch.float16 + print(self._autocast_dtype) with torch.amp.autocast(device_type = 'cuda', dtype = self._autocast_dtype): # We add 1 to `logits_to_keep` because the last logits of the sequence is later excluded logits = model(input_ids=input_ids, attention_mask=attention_mask, logits_to_keep=logits_to_keep + 1).logits