diff --git a/tester/input_generation/generation_rules.py b/tester/input_generation/generation_rules.py index cfb48978..551b6cd9 100644 --- a/tester/input_generation/generation_rules.py +++ b/tester/input_generation/generation_rules.py @@ -2957,7 +2957,10 @@ def initialize_from_x(input_binding): for index, item in enumerate(candidate): if isinstance(item, numbers.Integral): if item == 0: - state["maxvalue"] //= shape[index] + # 0 表示抄 x 同位置维度；index 越界属于无效配置， + # 这里不能崩，要让 Paddle 去报 InvalidArgument。 + if index < len(shape): + state["maxvalue"] //= shape[index] elif item != -1: state["maxvalue"] //= int(item) elif rule.is_tensor_config(item): diff --git a/tester/paddle_to_torch/mapping.json b/tester/paddle_to_torch/mapping.json index f9aef13b..7bee6271 100644 --- a/tester/paddle_to_torch/mapping.json +++ b/tester/paddle_to_torch/mapping.json @@ -3955,6 +3955,12 @@ "paddle.Tensor.reshape": { "Rule": "ReshapeRule" }, + "paddle.reshape_": { + "Rule": "ReshapeInplaceRule" + }, + "paddle.Tensor.reshape_": { + "Rule": "ReshapeInplaceRule" + }, "paddle.reverse": { "Rule": "ReverseRule" }, @@ -5151,7 +5157,7 @@ "Rule": "CopsPutAlongAxisRule" }, "paddle._C_ops.reshape_": { - "Rule": "CopsReshapeRule" + "Rule": "ReshapeInplaceRule" }, "paddle._C_ops.scale_": { "Rule": "CopsScaleRule" diff --git a/tester/paddle_to_torch/rules.py b/tester/paddle_to_torch/rules.py index 5c5f6adc..d6c59f6f 100644 --- a/tester/paddle_to_torch/rules.py +++ b/tester/paddle_to_torch/rules.py @@ -6162,6 +6162,76 @@ def apply(self, paddle_api: str) -> ConvertResult: ) +# paddle/phi/infermeta/unary.cc ValidateShape 的逐条复刻，reshape 系列共用。 +# 归一化后 shape 里不再有 0 或 -1，torch.reshape 只需处理确定形状。 +# 三处必须在此拦下的语义（torch 的报错文本不在 classify_runtime_error 白名单里， +# 直接交给 torch 会把无效配置误报成 torch_error）： +# - shape 里的 0 表示抄输入同位置维度，index 必须落在输入维度内，zero-size 输入例外； +# - zero-size 输入下 -1 按非零维乘积之比推导，不是 in_size // capacity（那样恒为 0）； +# - -1 只允许一个，其余负数一律非法。 +_RESHAPE_VALIDATE_SHAPE = """ +if isinstance(shape, torch.Tensor): + shape = shape.tolist() +else: + shape = list(shape) +in_size = x.numel() +in_dims = list(x.shape) +out_shape = [0] * len(shape) +capacity = 1 +shape_zero_cnt = 0 +unk_dim_idx = -1 +for i, s in enumerate(shape): + if s == -1: + if unk_dim_idx != -1: + raise ValueError( + "(InvalidArgument) Only one dimension value of 'shape' in " + "ReshapeOp can be -1" + ) + unk_dim_idx = i + out_shape[i] = -1 + elif s == 0: + shape_zero_cnt += 1 + if i < len(in_dims): + out_shape[i] = 0 if in_size == 0 else in_dims[i] + elif in_size != 0: + raise ValueError( + "(InvalidArgument) If The index of 0 in `shape` >= the input " + "tensor X's dimensions, It can only be Zero-Sized Tensor" + ) + capacity *= out_shape[i] + elif s < 0: + raise ValueError( + "(InvalidArgument) Each dimension value of 'shape' in ReshapeOp " + "must not be negative except one unknown dimension" + ) + else: + out_shape[i] = s + capacity *= s +if capacity == 0 and unk_dim_idx != -1: + in_zero_cnt = 0 + in_pdt = 1 + for d in in_dims: + if d == 0: + in_zero_cnt += 1 + else: + in_pdt *= d + shape_pdt = 1 + for s in shape: + if s != 0 and s != -1: + shape_pdt *= s + if shape_zero_cnt == in_zero_cnt and in_pdt % shape_pdt == 0: + out_shape[unk_dim_idx] = in_pdt // shape_pdt + else: + raise ValueError( + "(InvalidArgument) can not reshape, because the unspecified " + "dimension can be any number and is ambiguous" + ) +elif unk_dim_idx != -1: + out_shape[unk_dim_idx] = in_size // capacity +shape = out_shape +""" + + class ReshapeRule(BaseRule): PADDLE_APIS = ( "paddle.reshape", @@ -6169,29 +6239,51 @@ class ReshapeRule(BaseRule): ) def apply(self, paddle_api: str) -> ConvertResult: - preprocess = """ -if isinstance(shape, torch.Tensor): - shape = shape.tolist() -elif isinstance(shape, tuple): - shape = list(shape) -elements = x.numel() -for i, s in enumerate(shape): - if s == 0 and elements != 0: - shape[i] = x.shape[i] - elements = elements // x.shape[i] - elif s != -1 and s != 0: - elements = elements // s -for i, s in enumerate(shape): - if s == -1: - shape[i] = elements -""" core = """ result = torch.reshape(x, shape) """ return self.build_result( paddle_api, kind=ConversionKind.DIRECT, - preprocess=preprocess, + preprocess=_RESHAPE_VALIDATE_SHAPE, + core=core, + ) + + +class ReshapeInplaceRule(BaseRule): + PADDLE_APIS = ( + "paddle.reshape_", + "paddle.Tensor.reshape_", + "paddle._C_ops.reshape_", + ) + + def apply(self, paddle_api: str) -> ConvertResult: + # torch 没有 reshape_ kernel，只能组合。Paddle 侧限制 reshape_ 的三道门禁里， + # 只有门 1 属于单 API 语义，必须在这里对齐： + # 门 1 paddle/fluid/eager/utils.cc CheckInplace —— 需要梯度的叶子 Tensor 直接抛 + # InvalidArgument；但 python/paddle/tensor/manipulation.py 在目标 shape 与 + # 原 shape 相同时先短路返回 x，不进 C++，所以 shape 未变时不该报错。 + # 门 2 inplace version 校验取决于上游算子的反向依赖，单 API 参考实现里没有上游。 + # 门 3 静态图退化为 reshape，与动态图单 API 比对无关。 + core = """ +if list(x.shape) == shape: + # 门 1 的短路口：Paddle 在 Python 层直接返回 x，不触发 CheckInplace + result = x +else: + if x.requires_grad and x.is_leaf: + raise RuntimeError( + "Leaf Var () that doesn't stop gradient can't use inplace strategy." + ) + # set_ 不能包 no_grad：包了以后 x 保留旧 grad_fn 而 shape 已变，反向会校验失败。 + # 不包时 set_ 把 grad_fn 置为 NotImplemented，而 harness 只对 x 自身求 grad + # （outputs 与 inputs 同对象），autograd 直接返回 grad_output，与 Paddle 一致。 + x.set_(torch.reshape(x, shape)) + result = x +""" + return self.build_result( + paddle_api, + kind=ConversionKind.COMPOSITE, + preprocess=_RESHAPE_VALIDATE_SHAPE, core=core, ) @@ -8542,38 +8634,6 @@ def _infer_broadcast_shape(inp, idx, dim): ) -class CopsReshapeRule(BaseRule): - PADDLE_APIS = ("paddle._C_ops.reshape_",) - - """paddle._C_ops.reshape_(x, shape) → torch.reshape (in-place via set_)""" - - def apply(self, paddle_api: str) -> ConvertResult: - core = """ -x = x -shape = shape -if isinstance(shape, torch.Tensor): - shape = shape.tolist() -elif isinstance(shape, tuple): - shape = list(shape) -elements = x.numel() -for i, s in enumerate(shape): - if s == 0 and elements != 0: - shape[i] = x.shape[i] - elements = elements // x.shape[i] - elif s != -1 and s != 0: - elements = elements // s -for i, s in enumerate(shape): - if s == -1: - shape[i] = elements -result = torch.reshape(x, shape) -""" - return self.build_result( - paddle_api, - kind=ConversionKind.DIRECT, - core=core, - ) - - class CopsScaleRule(BaseRule): PADDLE_APIS = ("paddle._C_ops.scale_", "paddle.Tensor.scale_")