PTQ-Level Efficiency, QAT-Level Accuracy | 训练后量化的效率,量化感知训练的精度
Tensor reshape bugfix (#364)
WeiweiZhang13ec66bd9e8fd
diff --git a/auto_round/data_type/utils.py b/auto_round/data_type/utils.pyindex 4d774c88b..95538b7e7 100644--- a/auto_round/data_type/utils.py+++ b/auto_round/data_type/utils.py@@ -67,7 +67,11 @@ def revert_tensor_by_pad(data: torch.Tensor, orig_shape: tuple, pad_len: int):if pad_len == 0:return data.reshape(orig_shape)else:- data_new = data.reshape(data.shape[0], -1)+ if len(orig_shape) > 2:+ tmp_shape = torch.prod(torch.tensor(orig_shape[:-1])).item()+ else:+ tmp_shape = orig_shape[0]+ data_new = data.reshape(tmp_shape, -1)data_new = data_new[:, :-pad_len]data_new = data_new.reshape(orig_shape)return data_new@@ -150,3 +154,4 @@ def floor_ste(x: torch.Tensor):torch.Tensor"""return (x.floor() - x).detach() + x+diff --git a/test/test_autoround.py b/test/test_autoround.pyindex 5e831aca4..2b842505f 100644--- a/test/test_autoround.py+++ b/test/test_autoround.py@@ -350,8 +350,24 @@ class TestAutoRound(unittest.TestCase):inputs = tokenizer(text, return_tensors="pt").to(model.device)res = tokenizer.decode(model.generate(**inputs, max_new_tokens=1)[0])+ def test_tensor_reshape(self):+ model_name = "facebook/opt-125m"+ model = AutoModelForCausalLM.from_pretrained(model_name, torch_dtype=torch.float32, trust_remote_code=True, device_map='auto')+ bits, group_size, sym = 4, 100, False+ autoround = AutoRound(+ self.model,+ self.tokenizer,+ bits=bits,+ group_size=group_size,+ sym=sym,+ iters=2,+ seqlen=2,+ dataset=self.llm_dataloader,+ )+ autoround.quantize()if __name__ == "__main__":unittest.main()+