[feat] add new aten-op-plugin-AtomGit

已合并

[feat] add new aten #5718

culechan创建于 14 天前从culechan/op-plugin： repeatinter 合入到 Ascend/op-plugin： master

[feat] add new aten #5718

已合并

culechan创建于 14 天前

从culechan/op-plugin repeatinter合入到Ascend/op-plugin master

共 4 个文件变更+95-0

Mop_plugin/config/op_plugin_functions.yaml+6-0

	@@ -4294,6 +4294,9 @@ official:
4294	acl_op: v2.1	4294	acl_op: v2.1
4295	op_api: v2.1	4295	op_api: v2.1
4296		4296
		4297	+ - func: repeat_interleave.Tensor(Tensor repeats, *, SymInt? output_size=None) -> Tensor
		4298	+ op_api: [v2.2, newest]
		4299	+
4297	- func: replication_pad1d(Tensor self, SymInt[2] padding) -> Tensor	4300	- func: replication_pad1d(Tensor self, SymInt[2] padding) -> Tensor
4298	acl_op: [v2.1, newest]	4301	acl_op: [v2.1, newest]
4299	op_api: [v2.1, newest]	4302	op_api: [v2.1, newest]
	@@ -7700,6 +7703,9 @@ symint:
7700	acl_op: [v2.2, newest]	7703	acl_op: [v2.2, newest]
7701	op_api: [v2.2, newest]	7704	op_api: [v2.2, newest]
7702		7705
		7706	+ - func: repeat_interleave.Tensor(Tensor repeats, *, SymInt? output_size=None) -> Tensor
		7707	+ op_api: [v2.2, newest]
		7708	+
7703	- func: repeat_interleave.self_int(Tensor self, SymInt repeats, int? dim=None, *, SymInt? output_size=None) -> Tensor	7709	- func: repeat_interleave.self_int(Tensor self, SymInt repeats, int? dim=None, *, SymInt? output_size=None) -> Tensor
7704	acl_op: [v2.2, newest]	7710	acl_op: [v2.2, newest]
7705	op_api: [v2.2, newest]	7711	op_api: [v2.2, newest]

Mop_plugin/ops/opapi/RepeatInterLeaveKernelNpuOpApi.cpp+27-0

	@@ -215,5 +215,32 @@ at::Tensor repeat_interleave_symint(
215	}	215	}
216	return result;	216	return result;
217	}	217	}
		218	+
		219	+
		220	+at::Tensor repeat_interleave_symint(
		221	+ const at::Tensor& repeats,
		222	+ c10::optional<c10::SymInt> output_size)
		223	+{
		224	+ TORCH_CHECK(
		225	+ repeats.dim() == 1, "repeat_interleave only accept 1D vector as repeat");
		226	+ TORCH_CHECK(
		227	+ repeats.scalar_type() == at::kLong \|\| repeats.scalar_type() == at::kInt,
		228	+ "repeats has to be Long or Int tensor");
		229	+ if (repeats.size(0) == 0) {
		230	+ return at::empty_like(repeats, LEGACY_CONTIGUOUS_MEMORY_FORMAT);
		231	+ }
		232	+ at::Tensor cumsum = repeats.cumsum(0);
		233	+ int64_t total = 0;
		234	+ if (output_size.has_value()) {
		235	+ total = output_size.value().expect_int();
		236	+ } else {
		atomgit-bot14 天前 🟡 Medium Priority 变更行 226-228：当 `output_size.has_value()` 为 true 时，`total` 被直接设为 `output_size.value().expect_int()`，但未校验该值是否与 `repeats` 各元素之和一致。对比同文件中其他重载（如行 200-208），它们通过 `apply_result_tensor` 中的 `TORCH_CHECK(output_size_expected == output_size, ...)` 进行此项校验。新增重载完全缺失此校验链。触发条件：`output_size` 与 `repeats.sum()` 不一致。例如 `repeat_interleave(tensor([2,3]), output_size=10)` —— 实际 repeats 和为 5，但 output_size=10。失败模式：创建的 result tensor 大小为 `output_size`（10），但 `aclnnRepeatInterleaveTensor` 可能按实际 repeats 写入数据（5 个元素），导致 result 后半部分未初始化 / 越界写入，产生数据错乱。建议：在 `output_size.has_value()` 分支内增加校验：用 `cumsum[-1].item<int64_t>()` 得到实际 repeats 总和，与 `output_size` 值做 `TORCH_CHECK` 比对，确保二者一致。或将 cumsum 计算移到 if 之前共享，并在 output_size 分支中校验。 culechan 12 天前评论： 12 天前评论：
		237	+ total = cumsum[-1].item<int64_t>();
		atomgit-bot14 天前 🟡 Medium Priority 变更行 224、229：当 `repeats` 是空的一维张量（shape `[0]`）时，`repeats.cumsum(0)` 返回空张量，紧接着 `cumsum[-1]` 会访问越界索引，导致运行时崩溃。该函数无任何对 `repeats.numel() == 0` 的前置保护。其他重载通过 `check_tensor_repeats` 间接约束了形状（需要与 self 张量维度匹配），但本重载无 self 参数，完全依赖 repeats 自身形状。触发条件：`repeat_interleave_symint(torch.tensor([]), c10::nullopt)`。失败模式：`cumsum[-1]` 抛出 `IndexError` 或段错误。建议：在函数开头增加 `TORCH_CHECK(repeats.numel() > 0, ...)` 或 `TORCH_CHECK(repeats.dim() == 1 && repeats.numel() > 0, ...)` 校验 repeats 非空且为一维张量，与其他重载中 `check_tensor_repeats` 的防护水平对齐。 culechan 12 天前评论： 12 天前评论：
		238	+ TORCH_CHECK((repeats >= 0).all().to(at::kBool).item<bool>(), "repeats cannot be negative.");
		239	+ }
		atomgit-bot14 天前 🟡 Medium Priority 变更行 226-231：新增的 `repeat_interleave_symint` 重载中，`TORCH_CHECK((repeats >= 0).all().item<uint8_t>, ...)` 仅位于 `else` 分支（即 `output_size` 无值时的路径）。当调用者同时传入 `output_size` 和负数 `repeats` 时，负数校验被完全跳过，`repeats` 中的负值不会被拦截。对比同一文件中其他所有 `repeat_interleave_symint` 重载（行 97、161、197），它们的 repeats 非负校验均在 if-else 之外独立执行，无论 `output_size` 是否有值都会生效。此新增重载与此模式不一致，构成逻辑缺陷。触发条件：调用 `repeat_interleave.Tensor(repeats_tensor_with_negative, output_size=some_val)`。失败模式：负数 repeats 未被拒绝，`aclnnRepeatInterleaveTensor` 可能产生未定义行为或静默错误结果。建议：将 `TORCH_CHECK` 移到 if-else 之外，使其无论 `output_size` 是否有值都执行。参照行 161 的模式：在参数校验阶段（即 if-else 之前）单独执行 repeats 的非负检查。
		240	+ at::Tensor result = at::empty({total}, repeats.options());
		241	+ EXEC_NPU_CMD(aclnnRepeatInterleaveTensor, repeats, total, result);
		242	+ return result;
		243	+}
		244	+
218	#endif	245	#endif
219	}	246	}

Mtest/core_tests/torch_npu_OpApi_schema_all.json+3-0

	@@ -4823,6 +4823,9 @@
4823	"func: repeat_interleave.self_Tensor(Tensor self, Tensor repeats, int? dim=None, *, SymInt? output_size=None) -> Tensor": {	4823	"func: repeat_interleave.self_Tensor(Tensor self, Tensor repeats, int? dim=None, *, SymInt? output_size=None) -> Tensor": {
4824	"version": ["v2.3", "newest"]	4824	"version": ["v2.3", "newest"]
4825	},	4825	},
		4826	+ "func: repeat_interleave.Tensor(Tensor repeats, *, SymInt? output_size=None) -> Tensor": {
		4827	+ "version": ["v2.2", "newest"]
		4828	+ },
4826	"func: repeat_interleave.self_int(Tensor self, SymInt repeats, int? dim=None, *, SymInt? output_size=None) -> Tensor": {	4829	"func: repeat_interleave.self_int(Tensor self, SymInt repeats, int? dim=None, *, SymInt? output_size=None) -> Tensor": {
4827	"version": ["v2.3", "newest"]	4830	"version": ["v2.3", "newest"]
4828	},	4831	},

Mtest/test_base_ops/test_repeat_interleave.py+59-0

	@@ -124,6 +124,65 @@ class TestRepeatInterleave(TestCase):
124	npu_output = self.npu_op_exec_without_dim(cpu_input1.npu(), input2)	124	npu_output = self.npu_op_exec_without_dim(cpu_input1.npu(), input2)
125	self.assertRtolEqual(cpu_output, npu_output)	125	self.assertRtolEqual(cpu_output, npu_output)
126		126
		127	+ def cpu_op_exec_repeats_only(self, repeats):
		128	+ output = torch.repeat_interleave(repeats)
		129	+ output = output.numpy()
		130	+ return output
		131	+
		132	+ def npu_op_exec_repeats_only(self, repeats):
		133	+ output = torch.repeat_interleave(repeats)
		134	+ output = output.cpu()
		135	+ output = output.numpy()
		136	+ return output
		137	+
		138	+ def cpu_op_exec_repeats_output_size(self, repeats, output_size):
		139	+ output = torch.repeat_interleave(repeats, output_size=output_size)
		140	+ output = output.numpy()
		141	+ return output
		142	+
		143	+ def npu_op_exec_repeats_output_size(self, repeats, output_size):
		144	+ output = torch.repeat_interleave(repeats, output_size=output_size)
		145	+ output = output.cpu()
		146	+ output = output.numpy()
		147	+ return output
		148	+
		149	+ def test_repeat_interleave_repeats_tensor_int64(self):
		150	+ repeats = torch.tensor([2, 3, 1, 4], dtype=torch.int64)
		151	+ cpu_output = self.cpu_op_exec_repeats_only(repeats)
		152	+ npu_output = self.npu_op_exec_repeats_only(repeats.npu())
		153	+ self.assertRtolEqual(cpu_output, npu_output)
		154	+
		155	+ def test_repeat_interleave_repeats_tensor_int32(self):
		156	+ repeats = torch.tensor([2, 3, 1, 4], dtype=torch.int32)
		157	+ cpu_output = self.cpu_op_exec_repeats_only(repeats)
		158	+ npu_output = self.npu_op_exec_repeats_only(repeats.npu())
		159	+ self.assertRtolEqual(cpu_output, npu_output)
		160	+
		161	+ def test_repeat_interleave_repeats_tensor_with_output_size(self):
		162	+ repeats = torch.tensor([2, 3, 1, 4], dtype=torch.int64)
		163	+ output_size = 10
		164	+ cpu_output = self.cpu_op_exec_repeats_output_size(repeats, output_size)
		165	+ npu_output = self.npu_op_exec_repeats_output_size(repeats.npu(), output_size)
		166	+ self.assertRtolEqual(cpu_output, npu_output)
		167	+
		168	+ def test_repeat_interleave_repeats_tensor_large(self):
		169	+ repeats = torch.randint(1, 5, (100,), dtype=torch.int64)
		170	+ cpu_output = self.cpu_op_exec_repeats_only(repeats)
		171	+ npu_output = self.npu_op_exec_repeats_only(repeats.npu())
		172	+ self.assertRtolEqual(cpu_output, npu_output)
		173	+
		174	+ def test_repeat_interleave_repeats_tensor_single_element(self):
		175	+ repeats = torch.tensor([5], dtype=torch.int64)
		176	+ cpu_output = self.cpu_op_exec_repeats_only(repeats)
		177	+ npu_output = self.npu_op_exec_repeats_only(repeats.npu())
		178	+ self.assertRtolEqual(cpu_output, npu_output)
		179	+
		180	+ def test_repeat_interleave_repeats_tensor_empty(self):
		181	+ repeats = torch.tensor([], dtype=torch.int64)
		182	+ cpu_output = self.cpu_op_exec_repeats_only(repeats)
		183	+ npu_output = self.npu_op_exec_repeats_only(repeats.npu())
		184	+ self.assertRtolEqual(cpu_output, npu_output)
		185	+
127		186
128	if __name__ == '__main__':	187	if __name__ == '__main__':
129	run_tests()	188	run_tests()