zfflxx

error when finetune gemma2-2b when Running Evaluation

14

File "torch/_dynamo/variables/constant.py", line 176, in call_method return ConstantVariable.create(op(self.value, add_target)) torch._dynamo.exc.InternalTorchDynamoError: 'bool' object is not subscriptable from user code: File "unsloth/kernels/flex_attention.py", line 73, in slow_attention_softcapping A += causal_mask[:q_len, :q_len] Set TORCH_LOGS="+dynamo"...

why not pass attention_mask to compute logits in grpo_accumulated_loss?

1

I check the code of unsloth, I found this is not the same as the code of trl. So I wonder why, are they equivalent? here is the code of...

Gradient reliability with sample-by-sample vs batch processing

2

I'm examining the training implementation in src/art/unsloth/service.py and have a question about the gradient computation approach. Currently, the code processes samples individually: for offset in range(0, packed_tensors["tokens"].shape[0]): # Process single...