正在为 MindSpore rotary_position_embedding 算子设计分布式实现方案，计划报告已输出完毕，包含 5 个文件清单、25 条 UT 和 3 组 ST 测试用例。下一步是确认方案后开始动手实现。 (disable recaps in /config)

正在为 MindSpore RPE（旋转位置编码）实现 HyperParallel 分布式算子。计划文档已确认，下一步开始编写代码实现。 (disable recaps in /config)

1-npu-dense-lightning-indexer-softmax-l-cheerful-yeti.md
1-torch-mindspore-virtual-stonebraker.md
3-commit-1-3-dense-lightning-indexer-gr-snoopy-toast.md
ISSUE.md
PR_description.md
ci310_install_troubleshooting.md
claude_workflow.md
dfunction.md
distributed_op_migration_audit_report.md
distributed_op_migration_guide.md
dsa_distributed_op_analysis.md
ep_deleted_code.md
etp.md
expert_parallel.md
hyper-parallel-ep-commit-f0fa78f686caa8f-prancy-donut.md
images
issues
lightning_indexer_implementation_plan.md
memoized-seeking-honey.md
mhc

# Copyright 2026 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");

File created successfully at: /root/workspace/hyper-parallel/docs/rotary_position_embedding_dist_op_plan.md (file state is current in your context — no need to Read it back)

activation_with_axis_ops.yaml
argmax_with_value_ops.yaml
argsort.yaml
chunk_view_ops.yaml
concat_ops.yaml
cumsum_ops.yaml
element_wise_ops.yaml
element_wise_ops_with_shape.yaml
embedding_ops.yaml
expand_dims_ops.yaml
flash_attention_score.yaml
gather_ops.yaml
histc_ext.yaml
inplace_scatter_value_ops.yaml
lightning_indexer_ops.yaml
matmul_ops.yaml
multinomial_ops.yaml
norm_ops.yaml
npu_dense_lightning_indexer_grad_kl_loss_ops.yaml
npu_dense_lightning_indexer_softmax_lse_ops.yaml

# Copyright 2026 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
"""MindSpore distributed ST for npu_sparse_flash_attention.

Pipeline: ops.lightning_indexer(q_idx, k_idx, w) → sparse_indices
          → ops.sparse_flash_attention(q, k, v, sparse_indices, ...).

The two ops use different head_dim projections of the same sequence:
  - lightning_indexer : q_idx/k_idx head_dim = D_IDX = 128 (kernel constraint)
  - sparse_flash_attention : q/k head_dim = D_DIM = 512 (MLA attention_mode=2 constraint)

For BSND+CP the lightning_indexer distributed op slices k_idx to the causal window
``k_idx[:, :S1_local*(split_id+1), :, :]``; the sparse_flash_attention distributed op
applies the same slice to k/v/key_rope, so sparse_indices remain valid.  This mirrors
MindFormers adjust_bsnd_input logic.

Distributed sharding patterns verified against MindFormers dsa_attention.py shard():
  - BSND: q/si sharded on B (dp) and/or S1 (cp); k/v sharded on B (dp) only.
  - TND:  q/si sharded on T1 = dp*cp combined; k/v sharded on T2 = dp only.
"""
import numpy as np
import mindspore as ms
import mindspore.communication.management as D
from mindspore import Tensor, ops
from mindspore.ops import sparse_flash_attention
from hyper_parallel import init_device_mesh
from hyper_parallel.core.dtensor.dtensor import DTensor, distribute_tensor
from hyper_parallel.core.dtensor.placement_types import Shard, Replicate, Partial


def setup_module():
    """Initialize MindSpore context and communication."""
    ms.context.set_context(mode=ms.context.PYNATIVE_MODE, device_target="Ascend")
    D.init()


np.random.seed(42)
ms.set_seed(42)
ms.set_deterministic(True)

# sparse_flash_attention head dims (attention_mode=2 MLA kernel constraints).
N2 = 1
D_DIM = 512       # qk_head_dim for sparse_flash_attention (must be 512)
DR_DIM = 64       # rope head_dim
SCALE_VALUE = 0.135234

# lightning_indexer head dim (kernel constraint: must be 128).
D_IDX = 128

# BSND layout — replicated/DP tests.
B_BSND = 4
S1_BSND = 4
S2_BSND = 1024
N1_BSND = 8

# Number of top-k blocks for lightning_indexer (non-CP BSND).
SPARSE_COUNT_BSND = 8

# BSND layout — CP/dp+cp tests (realistic self-attention scenario).
# S1_BSND_CP=S2_BSND_CP are the GLOBAL sequence lengths (Q and K come from the same
# sequence).  CP=2 splits Q on S1: each rank processes S1_local=S1_BSND_CP//2 query tokens.
# K is Replicate (full length on each rank); the distributed op truncates it to the
# causal window per rank: rank r sees k[:, :S1_local*(r+1), :, :].
# S1=S2 (self-attention) ensures causal truncation is lossless for rank 0.
# SPARSE_COUNT_CP <= S1_local = S1_BSND_CP // 2.
B_BSND_CP = 2
N1_BSND_CP = 32
S1_BSND_CP = 1024
S2_BSND_CP = 1024
SPARSE_COUNT_CP = 128

# TND layout: T1=T2=1024, N1=8.
S_TND = 1024
N1_TND = 8
BATCH_DIM = 128
SPARSE_COUNT_TND = 8

_INPUT_NAMES = ("query", "key", "value", "query_rope", "key_rope")
_FWD_OUTPUT_NAMES = ("attention_out", "softmax_max", "softmax_sum")


def _make_tnd_seq_lens():
    """Build TND actual_seq_lengths tensors (prefix sums, int32, shape (128,))."""
    step = S_TND // BATCH_DIM
    arr = (np.arange(BATCH_DIM, dtype=np.int32) + 1) * step
    t = Tensor(arr, ms.int32)
    return t, Tensor(arr.copy(), ms.int32)

100	
101	
102	def _make_tnd_local_seq_lens():
103	    """Build LOCAL TND actual_seq_lengths for DP test (half batch per rank)."""
104	    local_batch = BATCH_DIM // 2
105	    local_s = S_TND // 2
106	    step = local_s // local_batch
107	    arr = (np.arange(local_batch, dtype=np.int32) + 1) * step
108	    t = Tensor(arr, ms.int32)
109	    return t, Tensor(arr.copy(), ms.int32)
110	
111	
112	
113	def _np_to_bf16(arr):
114	    """Convert float32 numpy array to MindSpore bfloat16 Tensor."""
115	    return Tensor(arr, ms.bfloat16)
116	
117	
118	def _to_f32np(t):
119	    """Convert a MindSpore tensor (any dtype) to float32 numpy array."""
120	    return t.astype(ms.float32).asnumpy()
121	
122	
123	def _generate_inputs(layout='BSND', s1=None, s2=None, n1=None, b=None):
124	    """Generate random float32 numpy arrays for BSND or TND layout.
125	
126	    Returns two groups of tensors:
127	      - SFA group (D_DIM=512): q, k, v, q_rope, k_rope
128	      - Indexer group (D_IDX=128): q_idx, k_idx, w
129	        q_idx/k_idx share the same B/S1/S2 dims as q/k but with D_IDX head_dim.
130	        w has shape (B, S1, N1) for BSND or (T1, N1) for TND.
131	
132	    Args:
133	        layout: 'BSND' or 'TND'.
134	        s1: Query sequence length (BSND only; defaults to S1_BSND).
135	        s2: Key/value sequence length (BSND only; defaults to S2_BSND).
136	        n1: Number of query heads (BSND only; defaults to N1_BSND).
137	        b: Batch size (BSND only; defaults to B_BSND).
138	
139	    Returns:
140	        tuple: (q_np, k_np, v_np, q_idx_np, k_idx_np, w_np, q_rope_np, k_rope_np)
141	    """
142	    if layout == 'BSND':
143	        s1 = s1 if s1 is not None else S1_BSND
144	        s2 = s2 if s2 is not None else S2_BSND
145	        n1 = n1 if n1 is not None else N1_BSND
146	        b = b if b is not None else B_BSND
147	        # SFA tensors (D_DIM=512)
148	        q_np = np.random.randn(b, s1, n1, D_DIM).astype(np.float32)
149	        k_np = np.random.randn(b, s2, N2, D_DIM).astype(np.float32)
150	        v_np = k_np.copy()
151	        q_rope_np = np.random.randn(b, s1, n1, DR_DIM).astype(np.float32)
152	        k_rope_np = np.random.randn(b, s2, N2, DR_DIM).astype(np.float32)
153	        # Indexer tensors (D_IDX=128)
154	        q_idx_np = np.random.randn(b, s1, n1, D_IDX).astype(np.float32)
155	        k_idx_np = np.random.randn(b, s2, N2, D_IDX).astype(np.float32)
156	        w_np = np.random.randn(b, s1, n1).astype(np.float32)
157	    else:  # TND
158	        # SFA tensors (D_DIM=512)
159	        q_np = np.random.randn(S_TND, N1_TND, D_DIM).astype(np.float32)
160	        k_np = np.random.randn(S_TND, N2, D_DIM).astype(np.float32)
161	        v_np = k_np.copy()
162	        q_rope_np = np.random.randn(S_TND, N1_TND, DR_DIM).astype(np.float32)
163	        k_rope_np = np.random.randn(S_TND, N2, DR_DIM).astype(np.float32)
164	        # Indexer tensors (D_IDX=128)
165	        q_idx_np = np.random.randn(S_TND, N1_TND, D_IDX).astype(np.float32)
166	        k_idx_np = np.random.randn(S_TND, N2, D_IDX).astype(np.float32)
167	        w_np = np.random.randn(S_TND, N1_TND).astype(np.float32)
168	    return q_np, k_np, v_np, q_idx_np, k_idx_np, w_np, q_rope_np, k_rope_np
169	
170	
171	def _call_sfa(q, k, v, si, q_rope, k_rope,
172	              actual_seq_lengths_query=None, actual_seq_lengths_kv=None,
173	              layout='BSND'):
174	    """Call sparse_flash_attention with an explicit layout string.
175	
176	    For BSND, actual_seq_lengths_query and actual_seq_lengths_kv should be
177	    None (kernel infers from tensor shape); for TND they must be passed.
178	    Omitted parameters use kernel defaults: sparse_mode=3 (rightDownCausal),
179	    sparse_block_size=1, pre_tokens/next_tokens=INT64_MAX.
180	
181	    Args:
182	        q: Query tensor (D_DIM=512).
183	        k: Key tensor (D_DIM=512).
184	        v: Value tensor (D_DIM=512).
185	        si: Sparse indices tensor (int32), generated by lightning_indexer.
186	        q_rope: Query rope tensor.
187	        k_rope: Key rope tensor.
188	        actual_seq_lengths_query: Cumulative query lengths (TND only).
189	        actual_seq_lengths_kv: Cumulative kv lengths (TND only).
190	        layout: Layout string ('BSND' or 'TND').
191	
192	    Returns:
193	        tuple: (attention_out, softmax_max, softmax_sum)
194	    """
195	    return sparse_flash_attention(
196	        q, k, v, si, SCALE_VALUE,
197	        actual_seq_lengths_query=actual_seq_lengths_query,
198	        actual_seq_lengths_kv=actual_seq_lengths_kv,
199	        query_rope=q_rope, key_rope=k_rope,
200	        layout_query=layout, layout_kv=layout,
201	        attention_mode=2,
202	        return_softmax_lse=True,
203	    )
204	
205	
206	def _get_sparse_indices_bsnd(q_idx, k_idx, w, sparse_count):
207	    """Run lightning_indexer in BSND mode and return sparse_indices.
208	
209	    Uses q_idx/k_idx with D_IDX=128 (lightning_indexer kernel constraint).
210	    sparse_mode defaults to 3 (rightDownCausal); for BSND+CP the CP impl
211	    truncates k_idx to the causal window before calling the kernel.
212	
213	    Args:
214	        q_idx: Query index tensor (B, S1, N1, D_IDX=128).
215	        k_idx: Key index tensor (B, S2, N2, D_IDX=128).
216	        w: Weights tensor (B, S1, N1).
217	        sparse_count: Number of top-k blocks to select.
218	
219	    Returns:

# Copyright 2026 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
"""parallel run case"""
import os
import signal
import multiprocessing as mp
from typing import Union


class TorchCase:
    """torch case messages"""

    def __init__(self, file_name: str, case_name: str, master_port: int, num_proc: int = 1):
        self.file_name = file_name
        self.case_name = case_name
        self.master_port = master_port
        self.num_proc = num_proc


class MindSporeCase:
    """mindspore case messages"""

    def __init__(self, file_name: str, case_name: str, master_port: int, worker_num: int = 1, local_worker_num: int = 1,
                 glog_v: int = 3):
        self.glog_v = glog_v
        self.file_name = file_name
        self.case_name = case_name
        self.master_port = master_port
        self.num_proc = worker_num
        self.local_worker_num = local_worker_num


def run_case(visible_devices, case: Union[TorchCase, MindSporeCase]):
    """
    run case in child process
    """
    # become the leader of a new process group so that os.killpg on timeout
    # kills torchrun/msrun worker sub-processes as well as this wrapper
    os.setsid()
    # set visible devices for current case
    os.environ['ASCEND_RT_VISIBLE_DEVICES'] = ','.join(map(str, visible_devices))
    if isinstance(case, TorchCase):
        # pylint: disable=C0415
        from tests.torch.utils import torchrun_case
        torchrun_case(case.file_name, case.case_name, case.master_port, case.num_proc)
    elif isinstance(case, MindSporeCase):
        # pylint: disable=C0415
        from tests.mindspore.st.utils import msrun_case
        msrun_case(case.glog_v, case.file_name, case.case_name, case.master_port, case.num_proc, case.local_worker_num)


def parallel_run(cases: Union[list[TorchCase], list[MindSporeCase]], global_num_proc: int = 8):
    """
    parallel run cases

    Args:
        cases (list[Case]): list of case messages to be run parallel
        global_num_proc (int, optional): number of total num of process. Defaults to 8.
    """
    # assign devices
    sum_num_proc = 0
    assignments = []

    for case in cases:
        num = case.num_proc
        devices = list(range(sum_num_proc, sum_num_proc + num))
        assignments.append(devices)
        sum_num_proc += num

24:import yaml
234:        self._env_yaml_dir: Optional[str] = os.environ.get("HYPER_PARALLEL_OPS_YAML_DIR")
236:        # The following attributes are initialized in _setup_yaml_dir()
237:        self.work_dir = ""  # Initialized in _setup_yaml_dir()
238:        self.yaml_dir = ""  # Initialized in _setup_yaml_dir()
242:        self.layout_infer_ops = self.safe_load_yaml_from_dir()
282:        self._setup_yaml_dir(self._env_yaml_dir)
285:    def _setup_yaml_dir(self, env_yaml_dir: Optional[str]):
287:        Feature: Configure yaml_dir/work_dir for OpDispatcher
289:                     If env_yaml_dir is an absolute path, use it directly; otherwise treat it
290:                     as a path relative to the project work_dir. If env_yaml_dir is not set,
291:                     fall back to the default 'shard/ops/yaml' under work_dir.
292:        Expectation: self.yaml_dir and self.work_dir are set to valid values used later by
293:                     safe_load_yaml_from_dir(); no functional behavior is changed.
295:        if env_yaml_dir:
296:            if os.path.isabs(env_yaml_dir):
297:                self.yaml_dir = env_yaml_dir
303:                self.yaml_dir = env_yaml_dir
305:            self.yaml_dir = "shard/ops/yaml"
794:    def safe_load_yaml_from_dir(self) -> dict:

19:import glob
20:import importlib
87:    global _no_skip_ops
97:    global _no_skip_ops
108:    global _dtensor_dispatch
119:    global _dtensor_dispatch
338:            module = importlib.import_module(module_name)
346:            module = importlib.import_module(module_name)
351:                module = importlib.import_module(module_file)
736:        global_shape = input_tensor.shape
743:        extra_args.append(global_shape)
744:        cache_key_values = [str(layout_id), str(begin), str(end), str(global_shape)]
806:        for yaml_file_path in glob.glob(os.path.join(yaml_path, '*.yaml')):
845:                global_shape=first_arg.shape,
914:            True when the op is whitelisted or DTensor dispatch is globally disabled.

330	                     provided via env, fall back to importing 'distributed_op_file' directly.
331	        Expectation: The distributed op class is imported and instantiated successfully,
332	                     or the original import error is raised; no functional behavior is changed.
333	        """
334	        class_name = config["distributed_op_class"]
335	
336	        if "distributed_op_module" in config:
337	            module_name = config["distributed_op_module"]
338	            module = importlib.import_module(module_name)
339	            op_class = getattr(module, class_name)
340	            _ = op_class(op_name)
341	            return
342	
343	        module_file = config["distributed_op_file"]
344	        try:
345	            module_name = "hyper_parallel.core.shard.ops." + module_file
346	            module = importlib.import_module(module_name)
347	            op_class = getattr(module, class_name)
348	            _ = op_class(op_name)
349	        except (ModuleNotFoundError, ImportError):
350	            if self._env_python_path:
351	                module = importlib.import_module(module_file)
352	                op_class = getattr(module, class_name)
353	                _ = op_class(op_name)
354	            else:
355	                raise
356	
357	    @staticmethod
358	    def _process_args_and_kwargs(
359	        args, kwargs
360	    ) -> tuple[list, list, list, dict, list]:
361	        """_process_args_and_kwargs"""
362	        input_layouts = []
363	        extra_args = []
364	        input_args = []
365	        input_kwargs = kwargs.copy()
366	        cache_key_values = []
367	
368	        for arg in args:
369	            if arg is None:

794	    def safe_load_yaml_from_dir(self) -> dict:
795	        """
796	        Load yaml dictionary from directory.
797	
798	        Returns:
799	            dict: Merged dictionary of all operator configurations loaded from YAML files.
800	        """
801	        yaml_dict = {}
802	        yaml_path = os.path.join(self.work_dir, self.yaml_dir) if self.work_dir else self.yaml_dir
803	        if not os.path.isdir(yaml_path):
804	            raise ValueError(f"Invalid yaml directory path: {yaml_path}")
805	
806	        for yaml_file_path in glob.glob(os.path.join(yaml_path, '*.yaml')):
807	            with open(yaml_file_path, 'r', encoding="utf-8") as f:
808	                yaml_data = yaml.safe_load(f)
809	
810	            yaml_data = OpDispatcher._merge_default(yaml_data)
811	            for name, data in yaml_data.items():
812	                if name in yaml_dict:
813	                    raise ValueError(f"Duplicate yaml object with name '{name}'.")
814	                yaml_dict[name] = data
815	
816	        return yaml_dict
817	
818	    def _dispatch_random_op(self, op_name: str, op_call: callable, args, kwargs):
819	        """Handle dispatch for random ops that operate on DTensors."""
820	        first_arg = next(
821	            (x for x in chain(args, kwargs.values()) if isinstance(x, DTensor)),
822	            None,
823	        )
824	        # Fall back to the default op if no DTensor is found.
825	        if first_arg is None:
826	            return op_call(*args, **kwargs)
827	
828	        local_args = [arg.to_local() if isinstance(arg, DTensor) else arg for arg in args]
829	        local_kwargs = {k: v.to_local() if isinstance(v, DTensor) else v for k, v in kwargs.items()}
830	        first_local_arg = first_arg.to_local()
831	
832	        if self._rng_tracker is None and is_rng_supported_mesh():
833	            self._rng_tracker = OffsetBasedRNGTracker()
834	
835	        maybe_user_generator = local_kwargs.pop("generator", None)
836	        if (
837	            self._rng_tracker is not None
838	            and not first_local_arg.is_meta
839	            and self._rng_tracker.distribute_region_enabled
840	        ):
841	            # pylint: disable=W0212
842	            with self._rng_tracker._distribute_region(
843	                device_mesh=first_arg.device_mesh,

1011:_OP_DISPATCHER = OpDispatcher()

ugrep: error: error at position 19
(?m)OpDispatcher\(\)
                   \___empty (sub)expression

/root/workspace/hyper-parallel/hyper_parallel/core/shard/_op_dispatch.py:1011:_OP_DISPATCHER = OpDispatcher()

import copy as cp
import inspect
import warnings
from typing import Any, Callable, Optional, Sequence, Set, Tuple, Union
import numpy as np
from hyper_parallel.core.dtensor.device_mesh import _mesh_resources
from hyper_parallel.core.dtensor.layout import Layout, DeviceMesh, _get_slice_tensor_by_layout
from hyper_parallel.core.dtensor.placement_types import Placement, Replicate
from hyper_parallel.platform import get_platform
from hyper_parallel.platform.platform import PlatformType
from hyper_parallel.core.utils import compute_local_shape_and_global_offset

62:        from hyper_parallel.core.shard._op_dispatch import disable_dtensor_dispatch, add_no_skip_ops
69:        from hyper_parallel.core.shard._op_dispatch import enable_dtensor_dispatch, remove_no_skip_ops

# Copyright 2025 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================

(Bash completed with no output)

op: <hyper_parallel.core.shard.ops.parallel_npu_sparse_flash_attention.SparseFlashAttentionDistributedOp object at 0x7c391d80ea70>
type: <class 'hyper_parallel.core.shard.ops.parallel_npu_sparse_flash_attention.SparseFlashAttentionDistributedOp'>

# Copyright 2025 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================
"""Interfaces for parallel-related functionality"""

(Bash completed with no output)

op_dispatch loaded!

/root/workspace/hyper-parallel/hyper_parallel/platform/torch/dtensor.py:76:        from hyper_parallel.core.shard._op_dispatch import _OP_DISPATCHER

/root/workspace/hyper-parallel/hyper_parallel/core/dtensor/dtensor.py:62:        from hyper_parallel.core.shard._op_dispatch import disable_dtensor_dispatch, add_no_skip_ops
/root/workspace/hyper-parallel/hyper_parallel/core/dtensor/dtensor.py:69:        from hyper_parallel.core.shard._op_dispatch import enable_dtensor_dispatch, remove_no_skip_ops
/root/workspace/hyper-parallel/hyper_parallel/core/shard/dfunction.py:20:from hyper_parallel.core.shard._op_dispatch import _OP_DISPATCHER
/root/workspace/hyper-parallel/hyper_parallel/platform/mindspore/dtensor.py:129:        from hyper_parallel.core.shard._op_dispatch import _OP_DISPATCHER
/root/workspace/hyper-parallel/hyper_parallel/platform/torch/dtensor.py:76:        from hyper_parallel.core.shard._op_dispatch import _OP_DISPATCHER

(Bash completed with no output)

(Bash completed with no output)

# Copyright 2025 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ============================================================================

/root/workspace/hyper-parallel/hyper_parallel/__init__.py:30:from hyper_parallel.core.shard.dfunction import DFunction

/root/workspace/hyper-parallel/hyper_parallel/__init__.py:30:from hyper_parallel.core.shard.dfunction import DFunction
/root/workspace/hyper-parallel/hyper_parallel/platform/mindspore/custom_ops/custom_op_impl.py:21:from hyper_parallel.core.shard.dfunction import DFunction

File created successfully at: /root/workspace/hyper-parallel/hyper_parallel/core/shard/ops/parallel_rotary_position_embedding.py (file state is current in your context — no need to Read it back)

File created successfully at: /root/workspace/hyper-parallel/hyper_parallel/core/shard/ops/yaml/rotary_position_embedding_ops.yaml (file state is current in your context — no need to Read it back)

File created successfully at: /root/workspace/hyper-parallel/tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py (file state is current in your context — no need to Read it back)

Conversation compacted

1	# Copyright 2026 Huawei Technologies Co., Ltd
2	#
3	# Licensed under the Apache License, Version 2.0 (the "License");
4	# you may not use this file except in compliance with the License.
5	# You may obtain a copy of the License at
6	#
7	# http://www.apache.org/licenses/LICENSE-2.0
8	#
9	# Unless required by applicable law or agreed to in writing, software
10	# distributed under the License is distributed on an "AS IS" BASIS,
11	# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12	# See the License for the specific language governing permissions and
13	# limitations under the License.
14	# ============================================================================
15	"""Test runner for npu_sparse_flash_attention distributed ST (MindSpore)."""
16	from pathlib import Path
17	
18	from tests.common.mark_utils import arg_mark
19	from tests.common.parallel_case import parallel_run, MindSporeCase
20	
21	IMPL_FILE = str(
22	    Path(__file__).resolve().parent / "npu_sparse_flash_attention_shard_in_python.py"
23	)
24	
25	
26	@arg_mark(plat_marks=["platform_ascend910b"], level_mark="level1", card_mark="allcards", essential_mark="essential")
27	def test_parallel_sfa_ms_group1():
28	    """
29	    Feature: parallel run case in npu_sparse_flash_attention_shard_in_python (BSND+TND replicated
30	             and DP, 8 cards)
31	    Description:
32	        1. test_sfa_bsnd_replicated — BSND all replicated, fwd+bwd vs standalone
33	        2. test_sfa_bsnd_dp — BSND B-dim data parallel, fwd+bwd vs standalone
34	        3. test_sfa_tnd_replicated — TND all replicated, fwd+bwd vs standalone
35	        4. test_sfa_tnd_dp — TND T1-dim data parallel (q AND k Shard(0)); local seq_lens
36	           per rank; q_split==k_split so no CP seq_len adjustment.
37	    Expectation: Run success.
38	    """
39	    parallel_run([
40	        MindSporeCase(IMPL_FILE, "test_sfa_bsnd_replicated", 19200, 2, 2, 2),
41	        MindSporeCase(IMPL_FILE, "test_sfa_bsnd_dp", 19201, 2, 2, 2),
42	        MindSporeCase(IMPL_FILE, "test_sfa_tnd_replicated", 19203, 2, 2, 2),
43	        MindSporeCase(IMPL_FILE, "test_sfa_tnd_dp", 19202, 2, 2, 2),
44	    ])
45	
46	
47	@arg_mark(plat_marks=["platform_ascend910b"], level_mark="level1", card_mark="allcards", essential_mark="essential")
48	def test_parallel_sfa_ms_group2():
49	    """
50	    Feature: parallel run case in npu_sparse_flash_attention_shard_in_python (BSND+TND CP, 4 cards)
51	    Description:
52	        1. test_sfa_bsnd_cp — BSND, q/si/q_rope Shard(1) on S1, k/v/k_rope Replicate;
53	           validates TP-as-CP pattern (MindFormers: S1 sharded by cp_tp combined)
54	        2. test_sfa_tnd_cp — TND, q/si/q_rope Shard(0) on T1, k/v/k_rope Replicate;
55	           seq_lens adjusted per rank by _tnd_cp_impl
56	    Expectation: Run success.
57	    """
58	    parallel_run([
59	        MindSporeCase(IMPL_FILE, "test_sfa_bsnd_cp", 19205, 2, 2, 2),
60	        MindSporeCase(IMPL_FILE, "test_sfa_tnd_cp", 19204, 2, 2, 2),
61	    ])
62	
63	
64	@arg_mark(plat_marks=["platform_ascend910b"], level_mark="level0", card_mark="allcards", essential_mark="essential")
65	def test_parallel_sfa_ms_group3():
66	    """
67	    Feature: parallel run case in npu_sparse_flash_attention_shard_in_python (BSND+TND dp+cp
68	             2-D mesh, 8 cards)
69	    Description:
70	        1. test_sfa_bsnd_dp_cp — 4-card 2-D mesh (dp=2, cp=2); B sharded by dp, S1 sharded
71	           by cp; k/v/key_rope replicated on cp. Mirrors MindFormers BSND shard() spec.
72	        2. test_sfa_tnd_dp_cp — 4-card 2-D mesh (dp=2, cp=2); T1 of q/si/q_rope sharded
73	           by dp AND cp (4-way split); T2 of k/v/k_rope sharded by dp only (2-way split).
74	           Mirrors MindFormers TND shard() spec with dp_cp_tp combined sequence split.
75	    Expectation: Run success.
76	    """
77	    parallel_run([
78	        MindSporeCase(IMPL_FILE, "test_sfa_bsnd_dp_cp", 19206, 4, 4, 2),
79	        MindSporeCase(IMPL_FILE, "test_sfa_tnd_dp_cp", 19207, 4, 4, 2),
80	    ])
81	

1	# Copyright 2026 Huawei Technologies Co., Ltd
2	#
3	# Licensed under the Apache License, Version 2.0 (the "License");
4	# you may not use this file except in compliance with the License.
5	# You may obtain a copy of the License at
6	#
7	# http://www.apache.org/licenses/LICENSE-2.0
8	#
9	# Unless required by applicable law or agreed to in writing, software
10	# distributed under the License is distributed on an "AS IS" BASIS,
11	# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12	# See the License for the specific language governing permissions and
13	# limitations under the License.
14	# ============================================================================
15	"""MindSpore distributed ST for npu_sparse_flash_attention.
16	
17	Pipeline: ops.lightning_indexer(q_idx, k_idx, w) → sparse_indices
18	          → ops.sparse_flash_attention(q, k, v, sparse_indices, ...).
19	
20	The two ops use different head_dim projections of the same sequence:
21	  - lightning_indexer : q_idx/k_idx head_dim = D_IDX = 128 (kernel constraint)
22	  - sparse_flash_attention : q/k head_dim = D_DIM = 512 (MLA attention_mode=2 constraint)
23	
24	For BSND+CP the lightning_indexer distributed op slices k_idx to the causal window
25	``k_idx[:, :S1_local*(split_id+1), :, :]``; the sparse_flash_attention distributed op
26	applies the same slice to k/v/key_rope, so sparse_indices remain valid.  This mirrors
27	MindFormers adjust_bsnd_input logic.
28	
29	Distributed sharding patterns verified against MindFormers dsa_attention.py shard():
30	  - BSND: q/si sharded on B (dp) and/or S1 (cp); k/v sharded on B (dp) only.
31	  - TND:  q/si sharded on T1 = dp*cp combined; k/v sharded on T2 = dp only.
32	"""
33	import numpy as np
34	import mindspore as ms
35	import mindspore.communication.management as D
36	from mindspore import Tensor, ops
37	from mindspore.ops import sparse_flash_attention
38	from hyper_parallel import init_device_mesh
39	from hyper_parallel.core.dtensor.dtensor import DTensor, distribute_tensor
40	from hyper_parallel.core.dtensor.placement_types import Shard, Replicate, Partial
41	
42	
43	def setup_module():
44	    """Initialize MindSpore context and communication."""
45	    ms.context.set_context(mode=ms.context.PYNATIVE_MODE, device_target="Ascend")
46	    D.init()
47	
48	
49	np.random.seed(42)
50	ms.set_seed(42)
51	ms.set_deterministic(True)
52	
53	# sparse_flash_attention head dims (attention_mode=2 MLA kernel constraints).
54	N2 = 1
55	D_DIM = 512       # qk_head_dim for sparse_flash_attention (must be 512)
56	DR_DIM = 64       # rope head_dim
57	SCALE_VALUE = 0.135234
58	
59	# lightning_indexer head dim (kernel constraint: must be 128).
60	D_IDX = 128
61	
62	# BSND layout — replicated/DP tests.
63	B_BSND = 4
64	S1_BSND = 4
65	S2_BSND = 1024
66	N1_BSND = 8
67	
68	# Number of top-k blocks for lightning_indexer (non-CP BSND).
69	SPARSE_COUNT_BSND = 8
70	
71	# BSND layout — CP/dp+cp tests (realistic self-attention scenario).
72	# S1_BSND_CP=S2_BSND_CP are the GLOBAL sequence lengths (Q and K come from the same
73	# sequence).  CP=2 splits Q on S1: each rank processes S1_local=S1_BSND_CP//2 query tokens.
74	# K is Replicate (full length on each rank); the distributed op truncates it to the
75	# causal window per rank: rank r sees k[:, :S1_local*(r+1), :, :].
76	# S1=S2 (self-attention) ensures causal truncation is lossless for rank 0.
77	# SPARSE_COUNT_CP <= S1_local = S1_BSND_CP // 2.
78	B_BSND_CP = 2
79	N1_BSND_CP = 32
80	S1_BSND_CP = 1024
81	S2_BSND_CP = 1024
82	SPARSE_COUNT_CP = 128
83	
84	# TND layout: T1=T2=1024, N1=8.
85	S_TND = 1024
86	N1_TND = 8
87	BATCH_DIM = 128
88	SPARSE_COUNT_TND = 8
89	
90	_INPUT_NAMES = ("query", "key", "value", "query_rope", "key_rope")
91	_FWD_OUTPUT_NAMES = ("attention_out", "softmax_max", "softmax_sum")
92	
93	
94	def _make_tnd_seq_lens():
95	    """Build TND actual_seq_lengths tensors (prefix sums, int32, shape (128,))."""
96	    step = S_TND // BATCH_DIM
97	    arr = (np.arange(BATCH_DIM, dtype=np.int32) + 1) * step
98	    t = Tensor(arr, ms.int32)
99	    return t, Tensor(arr.copy(), ms.int32)
100	
101	
102	def _make_tnd_local_seq_lens():
103	    """Build LOCAL TND actual_seq_lengths for DP test (half batch per rank)."""
104	    local_batch = BATCH_DIM // 2
105	    local_s = S_TND // 2
106	    step = local_s // local_batch
107	    arr = (np.arange(local_batch, dtype=np.int32) + 1) * step
108	    t = Tensor(arr, ms.int32)
109	    return t, Tensor(arr.copy(), ms.int32)
110	
111	
112	
113	def _np_to_bf16(arr):
114	    """Convert float32 numpy array to MindSpore bfloat16 Tensor."""
115	    return Tensor(arr, ms.bfloat16)
116	
117	
118	def _to_f32np(t):
119	    """Convert a MindSpore tensor (any dtype) to float32 numpy array."""
120	    return t.astype(ms.float32).asnumpy()
121	
122	
123	def _generate_inputs(layout='BSND', s1=None, s2=None, n1=None, b=None):
124	    """Generate random float32 numpy arrays for BSND or TND layout.
125	
126	    Returns two groups of tensors:
127	      - SFA group (D_DIM=512): q, k, v, q_rope, k_rope
128	      - Indexer group (D_IDX=128): q_idx, k_idx, w
129	        q_idx/k_idx share the same B/S1/S2 dims as q/k but with D_IDX head_dim.
130	        w has shape (B, S1, N1) for BSND or (T1, N1) for TND.
131	
132	    Args:
133	        layout: 'BSND' or 'TND'.
134	        s1: Query sequence length (BSND only; defaults to S1_BSND).
135	        s2: Key/value sequence length (BSND only; defaults to S2_BSND).
136	        n1: Number of query heads (BSND only; defaults to N1_BSND).
137	        b: Batch size (BSND only; defaults to B_BSND).
138	
139	    Returns:
140	        tuple: (q_np, k_np, v_np, q_idx_np, k_idx_np, w_np, q_rope_np, k_rope_np)
141	    """
142	    if layout == 'BSND':
143	        s1 = s1 if s1 is not None else S1_BSND
144	        s2 = s2 if s2 is not None else S2_BSND
145	        n1 = n1 if n1 is not None else N1_BSND
146	        b = b if b is not None else B_BSND
147	        # SFA tensors (D_DIM=512)
148	        q_np = np.random.randn(b, s1, n1, D_DIM).astype(np.float32)
149	        k_np = np.random.randn(b, s2, N2, D_DIM).astype(np.float32)
150	        v_np = k_np.copy()
151	        q_rope_np = np.random.randn(b, s1, n1, DR_DIM).astype(np.float32)
152	        k_rope_np = np.random.randn(b, s2, N2, DR_DIM).astype(np.float32)
153	        # Indexer tensors (D_IDX=128)
154	        q_idx_np = np.random.randn(b, s1, n1, D_IDX).astype(np.float32)
155	        k_idx_np = np.random.randn(b, s2, N2, D_IDX).astype(np.float32)
156	        w_np = np.random.randn(b, s1, n1).astype(np.float32)
157	    else:  # TND
158	        # SFA tensors (D_DIM=512)
159	        q_np = np.random.randn(S_TND, N1_TND, D_DIM).astype(np.float32)
160	        k_np = np.random.randn(S_TND, N2, D_DIM).astype(np.float32)
161	        v_np = k_np.copy()
162	        q_rope_np = np.random.randn(S_TND, N1_TND, DR_DIM).astype(np.float32)
163	        k_rope_np = np.random.randn(S_TND, N2, DR_DIM).astype(np.float32)
164	        # Indexer tensors (D_IDX=128)
165	        q_idx_np = np.random.randn(S_TND, N1_TND, D_IDX).astype(np.float32)
166	        k_idx_np = np.random.randn(S_TND, N2, D_IDX).astype(np.float32)
167	        w_np = np.random.randn(S_TND, N1_TND).astype(np.float32)
168	    return q_np, k_np, v_np, q_idx_np, k_idx_np, w_np, q_rope_np, k_rope_np
169	
170	
171	def _call_sfa(q, k, v, si, q_rope, k_rope,
172	              actual_seq_lengths_query=None, actual_seq_lengths_kv=None,
173	              layout='BSND'):
174	    """Call sparse_flash_attention with an explicit layout string.
175	
176	    For BSND, actual_seq_lengths_query and actual_seq_lengths_kv should be
177	    None (kernel infers from tensor shape); for TND they must be passed.
178	    Omitted parameters use kernel defaults: sparse_mode=3 (rightDownCausal),
179	    sparse_block_size=1, pre_tokens/next_tokens=INT64_MAX.
180	
181	    Args:
182	        q: Query tensor (D_DIM=512).
183	        k: Key tensor (D_DIM=512).
184	        v: Value tensor (D_DIM=512).
185	        si: Sparse indices tensor (int32), generated by lightning_indexer.
186	        q_rope: Query rope tensor.
187	        k_rope: Key rope tensor.
188	        actual_seq_lengths_query: Cumulative query lengths (TND only).
189	        actual_seq_lengths_kv: Cumulative kv lengths (TND only).
190	        layout: Layout string ('BSND' or 'TND').
191	
192	    Returns:
193	        tuple: (attention_out, softmax_max, softmax_sum)
194	    """
195	    return sparse_flash_attention(
196	        q, k, v, si, SCALE_VALUE,
197	        actual_seq_lengths_query=actual_seq_lengths_query,
198	        actual_seq_lengths_kv=actual_seq_lengths_kv,
199	        query_rope=q_rope, key_rope=k_rope,
200	        layout_query=layout, layout_kv=layout,
201	        attention_mode=2,
202	        return_softmax_lse=True,
203	    )
204	
205	
206	def _get_sparse_indices_bsnd(q_idx, k_idx, w, sparse_count):
207	    """Run lightning_indexer in BSND mode and return sparse_indices.
208	
209	    Uses q_idx/k_idx with D_IDX=128 (lightning_indexer kernel constraint).
210	    sparse_mode defaults to 3 (rightDownCausal); for BSND+CP the CP impl
211	    truncates k_idx to the causal window before calling the kernel.
212	
213	    Args:
214	        q_idx: Query index tensor (B, S1, N1, D_IDX=128).
215	        k_idx: Key index tensor (B, S2, N2, D_IDX=128).
216	        w: Weights tensor (B, S1, N1).
217	        sparse_count: Number of top-k blocks to select.
218	
219	    Returns:
220	        sparse_indices tensor (int32), shape (B, S1, N2, sparse_count).
221	    """
222	    out = ops.lightning_indexer(q_idx, k_idx, w, sparse_count=sparse_count, return_value=True)
223	    return out[0]
224	
225	
226	def _get_sparse_indices_tnd(q_idx, k_idx, w, qlen, klen, sparse_count):
227	    """Run lightning_indexer in TND mode and return sparse_indices.
228	
229	    Uses q_idx/k_idx with D_IDX=128 (lightning_indexer kernel constraint).
230	
231	    Args:
232	        q_idx: Query index tensor (T1, N1, D_IDX=128).
233	        k_idx: Key index tensor (T2, N2, D_IDX=128).
234	        w: Weights tensor (T1, N1).
235	        qlen: Cumulative query sequence lengths (int32 Tensor).
236	        klen: Cumulative key sequence lengths (int32 Tensor).
237	        sparse_count: Number of top-k blocks to select.
238	
239	    Returns:
240	        sparse_indices tensor (int32), shape (T1, N2, sparse_count).
241	    """
242	    out = ops.lightning_indexer(
243	        q_idx, k_idx, w,
244	        actual_seq_lengths_query=qlen,
245	        actual_seq_lengths_key=klen,
246	        layout_query='TND', layout_key='TND',
247	        sparse_count=sparse_count,
248	        return_value=True,
249	    )
250	    return out[0]
251	
252	
253	def _run_standalone(q_np, k_np, v_np, q_idx_np, k_idx_np, w_np, q_rope_np, k_rope_np,
254	                    actual_seq_q=None, actual_seq_kv=None, layout='BSND', sparse_count=None):
255	    """Run standalone lightning_indexer + SFA forward and backward.
256	
257	    lightning_indexer(q_idx, k_idx, w) → si → sparse_flash_attention(q, k, v, si, ...).
258	
259	    Args:
260	        q_np: SFA query numpy array (D_DIM=512).
261	        k_np: SFA key numpy array (D_DIM=512).
262	        v_np: SFA value numpy array (D_DIM=512).
263	        q_idx_np: Indexer query numpy array (D_IDX=128).
264	        k_idx_np: Indexer key numpy array (D_IDX=128).
265	        w_np: Indexer weights numpy array.
266	        q_rope_np: Query rope numpy array.
267	        k_rope_np: Key rope numpy array.
268	        actual_seq_q: Sequence lengths for query.
269	        actual_seq_kv: Sequence lengths for kv.
270	        layout: 'BSND' or 'TND'.
271	        sparse_count: Number of top-k blocks.
272	
273	    Returns:
274	        tuple: (fwd_outs, grad_outs) — each a tuple of float32 ndarrays.
275	    """
276	    q = _np_to_bf16(q_np)
277	    k = _np_to_bf16(k_np)
278	    v = _np_to_bf16(v_np)
279	    q_idx = _np_to_bf16(q_idx_np)
280	    k_idx = _np_to_bf16(k_idx_np)
281	    w = _np_to_bf16(w_np)
282	    q_rope = _np_to_bf16(q_rope_np)
283	    k_rope = _np_to_bf16(k_rope_np)
284	
285	    if layout == 'BSND':
286	        sc = sparse_count or SPARSE_COUNT_BSND
287	        si = _get_sparse_indices_bsnd(q_idx, k_idx, w, sc)
288	    else:
289	        sc = sparse_count or SPARSE_COUNT_TND
290	        si = _get_sparse_indices_tnd(q_idx, k_idx, w, actual_seq_q, actual_seq_kv, sc)
291	
292	    out = _call_sfa(q, k, v, si, q_rope, k_rope, actual_seq_q, actual_seq_kv, layout)
293	    fwd_outs = tuple(_to_f32np(o) for o in out)
294	
295	    grads = ms.grad(_call_sfa, (0, 1, 2, 4, 5))(
296	        q, k, v, si, q_rope, k_rope, actual_seq_q, actual_seq_kv, layout
297	    )
298	    grad_outs = tuple(_to_f32np(g) for g in grads)
299	    return fwd_outs, grad_outs
300	
301	
302	def _get_grad_placements(d_inputs: tuple) -> list:
303	    """Derive the correct DTensor placement for each input's gradient.
304	
305	    Args:
306	        d_inputs: Tuple of DTensor inputs.
307	
308	    Returns:
309	        list of placement tuples, one per input.
310	    """
311	    n_dims = len(d_inputs[0].layout.placements)
312	    dim_has_shard = [
313	        any(isinstance(d_inp.layout.placements[d], Shard) for d_inp in d_inputs)
314	        for d in range(n_dims)
315	    ]
316	    result = []
317	    for d_inp in d_inputs:
318	        mesh_placements = d_inp.layout.placements
319	        grad_placements = tuple(
320	            Partial("sum") if isinstance(p, Replicate) and dim_has_shard[dim_idx] else p
321	            for dim_idx, p in enumerate(mesh_placements)
322	        )
323	        result.append(grad_placements)
324	    return result
325	
326	
327	def _assert_fwd(dist_outs, ref_outs, tag):
328	    """Assert all three forward outputs match reference within tolerance."""
329	    for i, (name, ref) in enumerate(zip(_FWD_OUTPUT_NAMES, ref_outs)):
330	        dist_np = _to_f32np(dist_outs[i].full_tensor())
331	        assert np.allclose(dist_np, ref, atol=1e-3, rtol=1e-3), (
332	            f"{tag} forward {name} mismatch: "
333	            f"max_diff={np.abs(dist_np - ref).max()}"
334	        )
335	
336	
337	_BWD_TOL = {
338	    'key': (1e-1, 1e-1),
339	    'key_rope': (1e-1, 1e-1),
340	    'value': (1e-2, 1e-2),
341	}
342	_BWD_TOL_DEFAULT = (1e-3, 1e-3)
343	
344	
345	def _assert_bwd(raw_grads, d_inputs, ref_grad, tag, tol=None):
346	    """Assert backward gradients match reference within tolerance.
347	
348	    Args:
349	        tol: Optional per-input tolerance dict mapping input name to (atol, rtol).
350	             Pass _BWD_TOL for BSND+CP tests; omit for all other tests (uniform 1e-3).
351	    """
352	    grad_placements = _get_grad_placements(d_inputs)
353	    for name, raw_g, d_inp, ref, g_placements in zip(
354	            _INPUT_NAMES, raw_grads, d_inputs, ref_grad, grad_placements):
355	        dist_g = (
356	            DTensor.from_local(raw_g, d_inp.layout.mesh, g_placements)
357	            .full_tensor()
358	            .astype(ms.float32)
359	            .asnumpy()
360	        )
361	        atol, rtol = tol.get(name, _BWD_TOL_DEFAULT) if tol else _BWD_TOL_DEFAULT
362	        assert np.allclose(dist_g, ref, atol=atol, rtol=rtol), (
363	            f"{tag} backward grad_{name} mismatch: "
364	            f"max_diff={np.abs(dist_g - ref).max()}"
365	        )
366	
367	
368	def test_sfa_bsnd_replicated():
369	    """
370	    Feature: npu_sparse_flash_attention (MindSpore) BSND forward/backward, all replicated.
371	    Description:
372	        - All inputs replicated on 2-device mesh.
373	        - sparse_indices from lightning_indexer(q_idx, k_idx, w) with D_IDX=128.
374	        - SFA called with q/k/v of D_DIM=512 (MLA attention_mode=2).
375	        - bfloat16, attention_mode=2, v=k.clone().
376	    Expectation: Distributed outputs match standalone within tolerance.
377	    """
378	    q_np, k_np, v_np, q_idx_np, k_idx_np, w_np, q_rope_np, k_rope_np = _generate_inputs(layout='BSND')
379	    ref_fwd, ref_grad = _run_standalone(
380	        q_np, k_np, v_np, q_idx_np, k_idx_np, w_np, q_rope_np, k_rope_np,
381	        sparse_count=SPARSE_COUNT_BSND,
382	    )
383	
384	    mesh = init_device_mesh(device_type="npu", mesh_shape=(2,), mesh_dim_names=("dp",))
385	    dq = distribute_tensor(_np_to_bf16(q_np), mesh, (Replicate(),))
386	    dk = distribute_tensor(_np_to_bf16(k_np), mesh, (Replicate(),))
387	    dv = distribute_tensor(_np_to_bf16(v_np), mesh, (Replicate(),))
388	    dq_idx = distribute_tensor(_np_to_bf16(q_idx_np), mesh, (Replicate(),))
389	    dk_idx = distribute_tensor(_np_to_bf16(k_idx_np), mesh, (Replicate(),))
390	    dw = distribute_tensor(_np_to_bf16(w_np), mesh, (Replicate(),))
391	    dq_rope = distribute_tensor(_np_to_bf16(q_rope_np), mesh, (Replicate(),))
392	    dk_rope = distribute_tensor(_np_to_bf16(k_rope_np), mesh, (Replicate(),))
393	
394	    dsi = _get_sparse_indices_bsnd(dq_idx, dk_idx, dw, SPARSE_COUNT_BSND)
395	
396	    out = _call_sfa(dq, dk, dv, dsi, dq_rope, dk_rope)
397	    _assert_fwd(out, ref_fwd, "BSND Replicated")
398	
399	    raw_grads = ms.grad(_call_sfa, (0, 1, 2, 4, 5))(
400	        dq, dk, dv, dsi, dq_rope, dk_rope
401	    )
402	    _assert_bwd(raw_grads, (dq, dk, dv, dq_rope, dk_rope), ref_grad, "BSND Replicated")
403	
404	
405	def test_sfa_bsnd_dp():
406	    """
407	    Feature: npu_sparse_flash_attention (MindSpore) BSND forward/backward with B-dim DP.
408	    Description:
409	        - 2-device dp mesh; all inputs sharded on B (dim 0).
410	        - sparse_indices from distributed lightning_indexer(dq_idx, dk_idx, dw).
411	        - bfloat16, attention_mode=2, v=k.clone().
412	    Expectation: Distributed outputs match standalone within tolerance.
413	    """
414	    q_np, k_np, v_np, q_idx_np, k_idx_np, w_np, q_rope_np, k_rope_np = _generate_inputs(layout='BSND')
415	    ref_fwd, ref_grad = _run_standalone(
416	        q_np, k_np, v_np, q_idx_np, k_idx_np, w_np, q_rope_np, k_rope_np,
417	        sparse_count=SPARSE_COUNT_BSND,
418	    )
419	
420	    mesh = init_device_mesh(device_type="npu", mesh_shape=(2,), mesh_dim_names=("dp",))
421	    dq = distribute_tensor(_np_to_bf16(q_np), mesh, (Shard(0),))
422	    dk = distribute_tensor(_np_to_bf16(k_np), mesh, (Shard(0),))
423	    dv = distribute_tensor(_np_to_bf16(v_np), mesh, (Shard(0),))
424	    dq_idx = distribute_tensor(_np_to_bf16(q_idx_np), mesh, (Shard(0),))
425	    dk_idx = distribute_tensor(_np_to_bf16(k_idx_np), mesh, (Shard(0),))
426	    dw = distribute_tensor(_np_to_bf16(w_np), mesh, (Shard(0),))
427	    dq_rope = distribute_tensor(_np_to_bf16(q_rope_np), mesh, (Shard(0),))
428	    dk_rope = distribute_tensor(_np_to_bf16(k_rope_np), mesh, (Shard(0),))
429	
430	    dsi = _get_sparse_indices_bsnd(dq_idx, dk_idx, dw, SPARSE_COUNT_BSND)
431	
432	    out = _call_sfa(dq, dk, dv, dsi, dq_rope, dk_rope)
433	    _assert_fwd(out, ref_fwd, "BSND DP")
434	
435	    raw_grads = ms.grad(_call_sfa, (0, 1, 2, 4, 5))(
436	        dq, dk, dv, dsi, dq_rope, dk_rope
437	    )
438	    _assert_bwd(raw_grads, (dq, dk, dv, dq_rope, dk_rope), ref_grad, "BSND DP")
439	
440	
441	def test_sfa_bsnd_cp():
442	    """
443	    Feature: npu_sparse_flash_attention (MindSpore) BSND with S1-dim context parallelism.
444	    Description:
445	        - 2-device dp_cp mesh; q/q_idx/w/q_rope sharded on S1 (dim 1);
446	          k/k_idx/v/k_rope replicated.
447	        - lightning_indexer distributed op slices k_idx to causal window
448	          k_idx[:, :S1_local*(split_id+1), :, :] per rank (D_IDX=128).
449	        - sparse_flash_attention distributed op applies the same slice to k/v/key_rope
450	          (D_DIM=512), ensuring sparse_indices remain valid.
451	        - S1=S2 (self-attention) ensures causal truncation is lossless.
452	        - SPARSE_COUNT_CP <= S1_local = S1_BSND_CP // 2.
453	        - bfloat16, attention_mode=2, v=k.clone().
454	    Expectation: Distributed outputs match standalone within tolerance.
455	    """
456	    q_np, k_np, v_np, q_idx_np, k_idx_np, w_np, q_rope_np, k_rope_np = _generate_inputs(
457	        layout='BSND', s1=S1_BSND_CP, s2=S2_BSND_CP, n1=N1_BSND_CP, b=B_BSND_CP
458	    )
459	    ref_fwd, ref_grad = _run_standalone(
460	        q_np, k_np, v_np, q_idx_np, k_idx_np, w_np, q_rope_np, k_rope_np,
461	        sparse_count=SPARSE_COUNT_CP,
462	    )
463	
464	    mesh = init_device_mesh(device_type="npu", mesh_shape=(2,), mesh_dim_names=("dp_cp",))
465	    dq = distribute_tensor(_np_to_bf16(q_np), mesh, (Shard(1),))
466	    dk = distribute_tensor(_np_to_bf16(k_np), mesh, (Replicate(),))
467	    dv = distribute_tensor(_np_to_bf16(v_np), mesh, (Replicate(),))
468	    dq_idx = distribute_tensor(_np_to_bf16(q_idx_np), mesh, (Shard(1),))
469	    dk_idx = distribute_tensor(_np_to_bf16(k_idx_np), mesh, (Replicate(),))
470	    dw = distribute_tensor(_np_to_bf16(w_np), mesh, (Shard(1),))
471	    dq_rope = distribute_tensor(_np_to_bf16(q_rope_np), mesh, (Shard(1),))
472	    dk_rope = distribute_tensor(_np_to_bf16(k_rope_np), mesh, (Replicate(),))
473	
474	    dsi = _get_sparse_indices_bsnd(dq_idx, dk_idx, dw, SPARSE_COUNT_CP)
475	
476	    out = _call_sfa(dq, dk, dv, dsi, dq_rope, dk_rope)
477	    _assert_fwd(out, ref_fwd, "BSND CP")
478	
479	    raw_grads = ms.grad(_call_sfa, (0, 1, 2, 4, 5))(
480	        dq, dk, dv, dsi, dq_rope, dk_rope
481	    )
482	    _assert_bwd(raw_grads, (dq, dk, dv, dq_rope, dk_rope), ref_grad, "BSND CP", tol=_BWD_TOL)
483	
484	
485	def test_sfa_bsnd_dp_cp():
486	    """
487	    Feature: npu_sparse_flash_attention (MindSpore) BSND with 2-D dp+cp mesh.
488	    Description:
489	        - 4-card 2-D mesh (dp=2, cp=2); q/q_idx/w/q_rope sharded on B (dp) and S1 (cp).
490	          k/k_idx/v/k_rope sharded on B (dp), S2 replicated.
491	        - Mirrors MindFormers dsa_attention.py BSND shard():
492	            q/si = layout("dp", "cp_tp", "None", "None")
493	            k/v  = layout("dp", "None", "None", "None")
494	        - lightning_indexer slices k_idx to causal window per (dp, cp) rank.
495	        - sparse_flash_attention applies the same slice to k/v/key_rope.
496	        - S1=S2 (self-attention) ensures causal truncation is lossless.
497	        - Each rank: B_local = B_BSND_CP//dp = 1, S1_local = S1_BSND_CP//cp = S1_BSND_CP//2.
498	        - bfloat16, attention_mode=2, v=k.clone().
499	    Expectation: Distributed outputs match standalone within tolerance.
500	    """
501	    q_np, k_np, v_np, q_idx_np, k_idx_np, w_np, q_rope_np, k_rope_np = _generate_inputs(
502	        layout='BSND', s1=S1_BSND_CP, s2=S2_BSND_CP, n1=N1_BSND_CP, b=B_BSND_CP
503	    )
504	    ref_fwd, ref_grad = _run_standalone(
505	        q_np, k_np, v_np, q_idx_np, k_idx_np, w_np, q_rope_np, k_rope_np,
506	        sparse_count=SPARSE_COUNT_CP,
507	    )
508	
509	    mesh = init_device_mesh(device_type="npu", mesh_shape=(2, 2), mesh_dim_names=("dp", "cp"))
510	    dq = distribute_tensor(_np_to_bf16(q_np), mesh, (Shard(0), Shard(1)))
511	    dk = distribute_tensor(_np_to_bf16(k_np), mesh, (Shard(0), Replicate()))
512	    dv = distribute_tensor(_np_to_bf16(v_np), mesh, (Shard(0), Replicate()))
513	    dq_idx = distribute_tensor(_np_to_bf16(q_idx_np), mesh, (Shard(0), Shard(1)))
514	    dk_idx = distribute_tensor(_np_to_bf16(k_idx_np), mesh, (Shard(0), Replicate()))
515	    dw = distribute_tensor(_np_to_bf16(w_np), mesh, (Shard(0), Shard(1)))
516	    dq_rope = distribute_tensor(_np_to_bf16(q_rope_np), mesh, (Shard(0), Shard(1)))
517	    dk_rope = distribute_tensor(_np_to_bf16(k_rope_np), mesh, (Shard(0), Replicate()))
518	
519	    dsi = _get_sparse_indices_bsnd(dq_idx, dk_idx, dw, SPARSE_COUNT_CP)
520	
521	    out = _call_sfa(dq, dk, dv, dsi, dq_rope, dk_rope)
522	    _assert_fwd(out, ref_fwd, "BSND dp+cp")
523	
524	    raw_grads = ms.grad(_call_sfa, (0, 1, 2, 4, 5))(
525	        dq, dk, dv, dsi, dq_rope, dk_rope
526	    )
527	    _assert_bwd(raw_grads, (dq, dk, dv, dq_rope, dk_rope), ref_grad, "BSND dp+cp", tol=_BWD_TOL)
528	
529	
530	def test_sfa_tnd_replicated():
531	    """
532	    Feature: npu_sparse_flash_attention (MindSpore) TND forward/backward, all replicated.
533	    Description:
534	        - sparse_indices from lightning_indexer(q_idx, k_idx, w) with D_IDX=128.
535	        - SFA called with q/k/v of D_DIM=512.
536	        - actual_seq_lengths as int32 Tensor of shape (128,) with prefix sums.
537	        - bfloat16, attention_mode=2, v=k.clone().
538	    Expectation: Distributed outputs match standalone within tolerance.
539	    """
540	    q_np, k_np, v_np, q_idx_np, k_idx_np, w_np, q_rope_np, k_rope_np = _generate_inputs(layout='TND')
541	    actual_seq_q, actual_seq_kv = _make_tnd_seq_lens()
542	    ref_fwd, ref_grad = _run_standalone(
543	        q_np, k_np, v_np, q_idx_np, k_idx_np, w_np, q_rope_np, k_rope_np,
544	        actual_seq_q=actual_seq_q, actual_seq_kv=actual_seq_kv,
545	        layout='TND', sparse_count=SPARSE_COUNT_TND,
546	    )
547	
548	    mesh = init_device_mesh(device_type="npu", mesh_shape=(2,), mesh_dim_names=("dp",))
549	    dq = distribute_tensor(_np_to_bf16(q_np), mesh, (Replicate(),))
550	    dk = distribute_tensor(_np_to_bf16(k_np), mesh, (Replicate(),))
551	    dv = distribute_tensor(_np_to_bf16(v_np), mesh, (Replicate(),))
552	    dq_idx = distribute_tensor(_np_to_bf16(q_idx_np), mesh, (Replicate(),))
553	    dk_idx = distribute_tensor(_np_to_bf16(k_idx_np), mesh, (Replicate(),))
554	    dw = distribute_tensor(_np_to_bf16(w_np), mesh, (Replicate(),))
555	    dq_rope = distribute_tensor(_np_to_bf16(q_rope_np), mesh, (Replicate(),))
556	    dk_rope = distribute_tensor(_np_to_bf16(k_rope_np), mesh, (Replicate(),))
557	
558	    d_actual_seq_q = distribute_tensor(actual_seq_q, mesh, (Replicate(),))
559	    d_actual_seq_kv = distribute_tensor(actual_seq_kv, mesh, (Replicate(),))
560	
561	    dsi = _get_sparse_indices_tnd(dq_idx, dk_idx, dw, d_actual_seq_q, d_actual_seq_kv, SPARSE_COUNT_TND)
562	
563	    out = _call_sfa(dq, dk, dv, dsi, dq_rope, dk_rope, d_actual_seq_q, d_actual_seq_kv, 'TND')
564	    _assert_fwd(out, ref_fwd, "TND Replicated")
565	
566	    raw_grads = ms.grad(_call_sfa, (0, 1, 2, 4, 5))(
567	        dq, dk, dv, dsi, dq_rope, dk_rope, d_actual_seq_q, d_actual_seq_kv, 'TND'
568	    )
569	    _assert_bwd(raw_grads, (dq, dk, dv, dq_rope, dk_rope), ref_grad, "TND Replicated")
570	
571	
572	def test_sfa_tnd_dp():
573	    """
574	    Feature: npu_sparse_flash_attention (MindSpore) TND forward/backward with T1-dim DP.
575	    Description:
576	        - 2-device dp mesh; all tensors sharded on T1 (dim 0).
577	        - sparse_indices from distributed lightning_indexer (TND DP mode).
578	        - Both q and k sharded on T1 — pure DP, no CP seq_len adjustment.
579	        - bfloat16, attention_mode=2, v=k.clone().
580	    Expectation: Gathered distributed outputs match standalone within tolerance.
581	    """
582	    q_np, k_np, v_np, q_idx_np, k_idx_np, w_np, q_rope_np, k_rope_np = _generate_inputs(layout='TND')
583	    actual_seq_q, actual_seq_kv = _make_tnd_seq_lens()
584	    ref_fwd, ref_grad = _run_standalone(
585	        q_np, k_np, v_np, q_idx_np, k_idx_np, w_np, q_rope_np, k_rope_np,
586	        actual_seq_q=actual_seq_q, actual_seq_kv=actual_seq_kv,
587	        layout='TND', sparse_count=SPARSE_COUNT_TND,
588	    )
589	
590	    mesh = init_device_mesh(device_type="npu", mesh_shape=(2,), mesh_dim_names=("dp",))
591	    dq = distribute_tensor(_np_to_bf16(q_np), mesh, (Shard(0),))
592	    dk = distribute_tensor(_np_to_bf16(k_np), mesh, (Shard(0),))
593	    dv = distribute_tensor(_np_to_bf16(v_np), mesh, (Shard(0),))
594	    dq_idx = distribute_tensor(_np_to_bf16(q_idx_np), mesh, (Shard(0),))
595	    dk_idx = distribute_tensor(_np_to_bf16(k_idx_np), mesh, (Shard(0),))
596	    dw = distribute_tensor(_np_to_bf16(w_np), mesh, (Shard(0),))
597	    dq_rope = distribute_tensor(_np_to_bf16(q_rope_np), mesh, (Shard(0),))
598	    dk_rope = distribute_tensor(_np_to_bf16(k_rope_np), mesh, (Shard(0),))
599	
600	    d_actual_seq_q = distribute_tensor(actual_seq_q, mesh, (Replicate(),))
601	    d_actual_seq_kv = distribute_tensor(actual_seq_kv, mesh, (Replicate(),))
602	
603	    dsi = _get_sparse_indices_tnd(dq_idx, dk_idx, dw, d_actual_seq_q, d_actual_seq_kv, SPARSE_COUNT_TND)
604	
605	    out = _call_sfa(dq, dk, dv, dsi, dq_rope, dk_rope, d_actual_seq_q, d_actual_seq_kv, 'TND')
606	    _assert_fwd(out, ref_fwd, "TND DP")
607	
608	    raw_grads = ms.grad(_call_sfa, (0, 1, 2, 4, 5))(
609	        dq, dk, dv, dsi, dq_rope, dk_rope, d_actual_seq_q, d_actual_seq_kv, 'TND'
610	    )
611	    _assert_bwd(raw_grads, (dq, dk, dv, dq_rope, dk_rope), ref_grad, "TND DP")
612	
613	
614	def test_sfa_tnd_cp():
615	    """
616	    Feature: npu_sparse_flash_attention (MindSpore) TND with context parallelism.
617	    Description:
618	        - 2-device dp_cp mesh; q/q_idx/w/q_rope sharded on T1; k/k_idx/v/k_rope replicated.
619	        - sparse_indices from distributed lightning_indexer (TND CP mode, D_IDX=128).
620	        - actual_seq_lengths adjusted per rank by _tnd_cp_impl in both ops.
621	        - bfloat16, attention_mode=2, v=k.clone().
622	    Expectation: Distributed outputs match standalone within tolerance.
623	    """
624	    q_np, k_np, v_np, q_idx_np, k_idx_np, w_np, q_rope_np, k_rope_np = _generate_inputs(layout='TND')
625	    actual_seq_q, actual_seq_kv = _make_tnd_seq_lens()
626	    ref_fwd, ref_grad = _run_standalone(
627	        q_np, k_np, v_np, q_idx_np, k_idx_np, w_np, q_rope_np, k_rope_np,
628	        actual_seq_q=actual_seq_q, actual_seq_kv=actual_seq_kv,
629	        layout='TND', sparse_count=SPARSE_COUNT_TND,
630	    )
631	
632	    mesh = init_device_mesh(device_type="npu", mesh_shape=(2,), mesh_dim_names=("dp_cp",))
633	    dq = distribute_tensor(_np_to_bf16(q_np), mesh, (Shard(0),))
634	    dk = distribute_tensor(_np_to_bf16(k_np), mesh, (Replicate(),))
635	    dv = distribute_tensor(_np_to_bf16(v_np), mesh, (Replicate(),))
636	    dq_idx = distribute_tensor(_np_to_bf16(q_idx_np), mesh, (Shard(0),))
637	    dk_idx = distribute_tensor(_np_to_bf16(k_idx_np), mesh, (Replicate(),))
638	    dw = distribute_tensor(_np_to_bf16(w_np), mesh, (Shard(0),))
639	    dq_rope = distribute_tensor(_np_to_bf16(q_rope_np), mesh, (Shard(0),))
640	    dk_rope = distribute_tensor(_np_to_bf16(k_rope_np), mesh, (Replicate(),))
641	
642	    d_actual_seq_q = distribute_tensor(actual_seq_q, mesh, (Replicate(),))
643	    d_actual_seq_kv = distribute_tensor(actual_seq_kv, mesh, (Replicate(),))
644	
645	    dsi = _get_sparse_indices_tnd(dq_idx, dk_idx, dw, d_actual_seq_q, d_actual_seq_kv, SPARSE_COUNT_TND)
646	
647	    out = _call_sfa(dq, dk, dv, dsi, dq_rope, dk_rope, d_actual_seq_q, d_actual_seq_kv, 'TND')
648	    _assert_fwd(out, ref_fwd, "TND CP")
649	
650	    raw_grads = ms.grad(_call_sfa, (0, 1, 2, 4, 5))(
651	        dq, dk, dv, dsi, dq_rope, dk_rope, d_actual_seq_q, d_actual_seq_kv, 'TND'
652	    )
653	    _assert_bwd(raw_grads, (dq, dk, dv, dq_rope, dk_rope), ref_grad, "TND CP")
654	
655	
656	def test_sfa_tnd_dp_cp():
657	    """
658	    Feature: npu_sparse_flash_attention (MindSpore) TND with 2-D dp+cp mesh.
659	    Description:
660	        - 4-card 2-D mesh (dp=2, cp=2); T1 of q/q_idx/w/q_rope sharded by BOTH dp and cp.
661	          T2 of k/k_idx/v/k_rope sharded by dp only, cp Replicate.
662	        - Mirrors MindFormers dsa_attention.py TND shard():
663	            q/si = layout("dp_cp_tp", "None", "None")
664	            k/v  = layout("dp", "None", "None")
665	        - sparse_indices from distributed lightning_indexer (TND DP+CP mode, D_IDX=128).
666	        - _tnd_cp_impl adjusts actual_seq_lengths per rank in both ops.
667	        - bfloat16, attention_mode=2, v=k.clone().
668	    Expectation: Distributed outputs match standalone within tolerance.
669	    """
670	    q_np, k_np, v_np, q_idx_np, k_idx_np, w_np, q_rope_np, k_rope_np = _generate_inputs(layout='TND')
671	    actual_seq_q, actual_seq_kv = _make_tnd_seq_lens()
672	    ref_fwd, ref_grad = _run_standalone(
673	        q_np, k_np, v_np, q_idx_np, k_idx_np, w_np, q_rope_np, k_rope_np,
674	        actual_seq_q=actual_seq_q, actual_seq_kv=actual_seq_kv,
675	        layout='TND', sparse_count=SPARSE_COUNT_TND,
676	    )
677	
678	    mesh = init_device_mesh(device_type="npu", mesh_shape=(2, 2), mesh_dim_names=("dp", "cp"))
679	    dq = distribute_tensor(_np_to_bf16(q_np), mesh, (Shard(0), Shard(0)))
680	    dk = distribute_tensor(_np_to_bf16(k_np), mesh, (Shard(0), Replicate()))
681	    dv = distribute_tensor(_np_to_bf16(v_np), mesh, (Shard(0), Replicate()))
682	    dq_idx = distribute_tensor(_np_to_bf16(q_idx_np), mesh, (Shard(0), Shard(0)))
683	    dk_idx = distribute_tensor(_np_to_bf16(k_idx_np), mesh, (Shard(0), Replicate()))
684	    dw = distribute_tensor(_np_to_bf16(w_np), mesh, (Shard(0), Shard(0)))
685	    dq_rope = distribute_tensor(_np_to_bf16(q_rope_np), mesh, (Shard(0), Shard(0)))
686	    dk_rope = distribute_tensor(_np_to_bf16(k_rope_np), mesh, (Shard(0), Replicate()))
687	
688	    d_actual_seq_q = distribute_tensor(actual_seq_q, mesh, (Replicate(),))
689	    d_actual_seq_kv = distribute_tensor(actual_seq_kv, mesh, (Replicate(),))
690	
691	    dsi = _get_sparse_indices_tnd(dq_idx, dk_idx, dw, d_actual_seq_q, d_actual_seq_kv, SPARSE_COUNT_TND)
692	
693	    out = _call_sfa(dq, dk, dv, dsi, dq_rope, dk_rope, d_actual_seq_q, d_actual_seq_kv, 'TND')
694	    _assert_fwd(out, ref_fwd, "TND dp+cp")
695	
696	    raw_grads = ms.grad(_call_sfa, (0, 1, 2, 4, 5))(
697	        dq, dk, dv, dsi, dq_rope, dk_rope, d_actual_seq_q, d_actual_seq_kv, 'TND'
698	    )
699	    _assert_bwd(raw_grads, (dq, dk, dv, dq_rope, dk_rope), ref_grad, "TND dp+cp")
700	

1	# Copyright 2026 Huawei Technologies Co., Ltd
2	#
3	# Licensed under the Apache License, Version 2.0 (the "License");
4	# you may not use this file except in compliance with the License.
5	# You may obtain a copy of the License at
6	#
7	# http://www.apache.org/licenses/LICENSE-2.0
8	#
9	# Unless required by applicable law or agreed to in writing, software
10	# distributed under the License is distributed on an "AS IS" BASIS,
11	# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12	# See the License for the specific language governing permissions and
13	# limitations under the License.
14	# ============================================================================
15	"""parallel run case"""
16	import os
17	import signal
18	import multiprocessing as mp
19	from typing import Union
20	
21	
22	class TorchCase:
23	    """torch case messages"""
24	
25	    def __init__(self, file_name: str, case_name: str, master_port: int, num_proc: int = 1):
26	        self.file_name = file_name
27	        self.case_name = case_name
28	        self.master_port = master_port
29	        self.num_proc = num_proc
30	
31	
32	class MindSporeCase:
33	    """mindspore case messages"""
34	
35	    def __init__(self, file_name: str, case_name: str, master_port: int, worker_num: int = 1, local_worker_num: int = 1,
36	                 glog_v: int = 3):
37	        self.glog_v = glog_v
38	        self.file_name = file_name
39	        self.case_name = case_name
40	        self.master_port = master_port
41	        self.num_proc = worker_num
42	        self.local_worker_num = local_worker_num
43	
44	
45	def run_case(visible_devices, case: Union[TorchCase, MindSporeCase]):
46	    """
47	    run case in child process
48	    """
49	    # become the leader of a new process group so that os.killpg on timeout
50	    # kills torchrun/msrun worker sub-processes as well as this wrapper
51	    os.setsid()
52	    # set visible devices for current case
53	    os.environ['ASCEND_RT_VISIBLE_DEVICES'] = ','.join(map(str, visible_devices))
54	    if isinstance(case, TorchCase):
55	        # pylint: disable=C0415
56	        from tests.torch.utils import torchrun_case
57	        torchrun_case(case.file_name, case.case_name, case.master_port, case.num_proc)
58	    elif isinstance(case, MindSporeCase):
59	        # pylint: disable=C0415
60	        from tests.mindspore.st.utils import msrun_case
61	        msrun_case(case.glog_v, case.file_name, case.case_name, case.master_port, case.num_proc, case.local_worker_num)
62	
63	
64	def parallel_run(cases: Union[list[TorchCase], list[MindSporeCase]], global_num_proc: int = 8):
65	    """
66	    parallel run cases
67	
68	    Args:
69	        cases (list[Case]): list of case messages to be run parallel
70	        global_num_proc (int, optional): number of total num of process. Defaults to 8.
71	    """
72	    # assign devices
73	    sum_num_proc = 0
74	    assignments = []
75	
76	    for case in cases:
77	        num = case.num_proc
78	        devices = list(range(sum_num_proc, sum_num_proc + num))
79	        assignments.append(devices)
80	        sum_num_proc += num
81	    # assert sum num_proc
82	    assert sum_num_proc <= global_num_proc, (f"sum num_proc {sum_num_proc} greater than "
83	                                             f"global_num_proc {global_num_proc}")
84	
85	    # create child process (run_case calls os.setsid to own a process group,
86	    # so os.killpg on timeout kills torchrun/msrun workers too)
87	    processes = []
88	    for _, (case, devices) in enumerate(zip(cases, assignments)):
89	        p = mp.Process(target=run_case, args=(devices, case))
90	        p.start()
91	        processes.append(p)
92	
93	    # wait child process terminates (timeout=900s to prevent infinite hang on distributed deadlock)
94	    timed_out = []
95	    for i, p in enumerate(processes):
96	        p.join(timeout=900)
97	        if p.is_alive():
98	            try:
99	                os.killpg(os.getpgid(p.pid), signal.SIGKILL)
100	            except ProcessLookupError:
101	                pass
102	            p.join()
103	            timed_out.append(cases[i].case_name)
104	
105	    # check results for all cases
106	    if timed_out:
107	        raise AssertionError(f"Cases timed out (possible collective deadlock): {timed_out}")
108	    failed = [cases[i].case_name for i, p in enumerate(processes) if p.exitcode != 0]
109	    assert not failed, f"List cases failed: {failed}"
110	

__pycache__
activation_with_axis_shard_in_python.py
argmax_with_value_shard_in_python.py
argsort_shard_in_python.py
bmm_shard_in_python.py
chunk_view_shard_in_python.py
cumsum_ext_shard_in_python.py
elementwise_shard_in_python.py
expand_dims_shard_in_python.py
expand_dims_view_shard_in_python.py
flash_attention_score_shard_in_python.py
gather_d_shard_in_python.py
gather_nd_shard_in_python.py
histc_ext_shard_in_python.py
inplace_scatter_value_shard_in_python.py
lightning_indexer_shard_in_python.py
masked_fill_scalar_shard_in_python.py
max_dim_shard_in_python.py
ms_flash_attention_score_shard_in_python.py
muls_shard_in_python.py
norm_shard_in_python.py
npu_dense_lightning_indexer_grad_kl_loss_custom_op.py
npu_dense_lightning_indexer_softmax_lse_custom_op.py
npu_mhc_post_custom_op.py
npu_mhc_pre_sinkhorn_custom_op.py
npu_sparse_flash_attention_shard_in_python.py
npu_sparse_lightning_indexer_grad_kl_loss_custom_op.py
one_hot_ext_shard_in_python.py
ones_like_shard_in_python.py
reduce_shard_in_python.py
scatter_update_shard_in_python.py
silu_shard_in_python.py
test_activation_with_axis_shard_in_python.py
test_argmax_with_value_shard_in_python.py
test_argsort_shard_in_python.py
test_bmm_shard_in_python.py
test_chunk_view_shard_in_python.py
test_cumsum_ext_shard_in_python.py
test_elementwise_shard_in_python.py
test_expand_dims_shard_in_python.py
test_expand_dims_view_shard_in_python.py
test_flash_attention_score_shard_in_python.py
test_gather_d_shard_in_python.py
test_gather_nd_shard_in_python.py
test_histc_ext_shard_in_python.py
test_inplace_scatter_value_shard_in_python.py
test_lightning_indexer_shard_in_python.py
test_masked_fill_scalar_shard_in_python.py
test_max_dim_shard_in_python.py
test_ms_flash_attention_score_shard_in_python.py
test_muls_shard_in_python.py
test_norm_shard_in_python.py
test_npu_dense_lightning_indexer_grad_kl_loss_custom_op.py
test_npu_dense_lightning_indexer_softmax_lse_custom_op.py
test_npu_mhc_post_custom_op.py
test_npu_mhc_pre_sinkhorn_custom_op.py
test_npu_sparse_flash_attention_shard_in_python.py
test_npu_sparse_lightning_indexer_grad_kl_loss_custom_op.py
test_one_hot_ext_shard_in_python.py
test_ops_ones_like_shard_in_python.py
test_reduce_shard_in_python.py
test_scatter_update_shard_in_python.py
test_silu_shard_in_python.py
test_transpose_ext_view_shard_in_python.py
transpose_ext_view_shard_in_python.py

File created successfully at: /root/workspace/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py (file state is current in your context — no need to Read it back)

File created successfully at: /root/workspace/hyper-parallel/tests/mindspore/st/shard/ops/test_rotary_position_embedding_shard_in_python.py (file state is current in your context — no need to Read it back)

============================= test session starts ==============================
platform linux -- Python 3.10.20, pytest-9.0.3, pluggy-1.6.0 -- /root/miniconda3/envs/ci3.10/bin/python
cachedir: .pytest_cache
rootdir: /root/workspace/hyper-parallel
collecting ... collected 22 items

tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestNormalizeRpeArgs::test_mode_keyword_arg_2 PASSED [  4%]
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestNormalizeRpeArgs::test_mode_positional_arg_3 PASSED [  9%]
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestNormalizeRpeArgs::test_three_positional_args_mode_defaults_to_zero_1 PASSED [ 13%]
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestYamlRegistration::test_pascal_case_registration_4 PASSED [ 18%]
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestInferLayoutPositive::test_all_replicated_3d_tnd_13 PASSED [ 22%]
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestInferLayoutPositive::test_all_replicated_4d_5 PASSED [ 27%]
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestInferLayoutPositive::test_cp_s_cos_replicated_9 PASSED [ 31%]
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestInferLayoutPositive::test_cp_s_cos_same_shard_8 PASSED [ 36%]
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestInferLayoutPositive::test_dp_b_cos_replicated_6 PASSED [ 40%]
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestInferLayoutPositive::test_dp_cp_2d_mesh_11 PASSED [ 45%]
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestInferLayoutPositive::test_dp_tp_2d_mesh_10 PASSED [ 50%]
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestInferLayoutPositive::test_output_is_independent_deepcopy_15 PASSED [ 54%]
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestInferLayoutPositive::test_tnd_dp_t_14 PASSED [ 59%]
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestInferLayoutPositive::test_tp_cp_2d_mesh_12 PASSED [ 63%]
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestInferLayoutPositive::test_tp_n_cos_broadcast_7 PASSED [ 68%]
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestInferLayoutNegative::test_cos_b_mismatch_raises_19 PASSED [ 72%]
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestInferLayoutNegative::test_cos_d_sharded_raises_17 PASSED [ 77%]
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestInferLayoutNegative::test_cos_n_sharded_x_n_replicated_raises_21 PASSED [ 81%]
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestInferLayoutNegative::test_partial_input_raises_22 PASSED [ 86%]
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestInferLayoutNegative::test_sin_d_sharded_raises_18 PASSED [ 90%]
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestInferLayoutNegative::test_sin_s_mismatch_raises_20 PASSED [ 95%]
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py::TestInferLayoutNegative::test_x_d_sharded_raises_16 PASSED [100%]

============================== 22 passed in 0.12s ==============================

************* Module hyper_parallel.core.shard.ops.parallel_rotary_position_embedding
hyper_parallel/core/shard/ops/parallel_rotary_position_embedding.py:132:4: W0221: Number of parameters was 3 in 'DistributedOp.infer_layout' and is now 2 in overriding 'RotaryPositionEmbeddingDistributedOp.infer_layout' method (arguments-differ)
************* Module tests.ut.core.shard.ops.test_parallel_rotary_position_embedding
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py:24:0: C0413: Import "from hyper_parallel.core.dtensor.dtensor import _build_layout, _LAYOUT_CACHE" should be placed at the top of the module (wrong-import-position)
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py:25:0: C0413: Import "from hyper_parallel.core.dtensor.placement_types import Shard, Replicate" should be placed at the top of the module (wrong-import-position)
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py:26:0: C0413: Import "from hyper_parallel.core.shard.ops.parallel_rotary_position_embedding import RotaryPositionEmbeddingDistributedOp, _normalize_rpe_args" should be placed at the top of the module (wrong-import-position)
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py:30:0: C0413: Import "from hyper_parallel.core.shard.ops.parallel_ops_register import get_distributed_op" should be placed at the top of the module (wrong-import-position)
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py:31:0: C0413: Import "from hyper_parallel.core.dtensor.device_mesh import init_device_mesh, _DEVICE_MESH_MAP" should be placed at the top of the module (wrong-import-position)
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py:32:0: C0413: Import "from hyper_parallel.platform.platform import EXISTING_COMM_GROUPS" should be placed at the top of the module (wrong-import-position)

-----------------------------------
Your code has been rated at 9.83/10

82:    def infer_layout(self, layouts: tuple, extra_args: Optional[tuple] = None) -> Optional[tuple]:

# Copyright 2026 Huawei Technologies Co., Ltd
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at

316:    def infer_layout(self, cache_values: list) -> Tuple[tuple, None]:

from unittest.mock import MagicMock, patch

import numpy as np

os.environ["HYPER_PARALLEL_PLATFORM"] = "torch"

from hyper_parallel.core.dtensor.dtensor import _build_layout, _LAYOUT_CACHE
from hyper_parallel.core.dtensor.placement_types import Shard, Replicate
from hyper_parallel.core.shard.ops.parallel_npu_sparse_flash_attention import (
    SparseFlashAttentionDistributedOp,
    _normalize_sfa_args,
)
from hyper_parallel.core.shard.ops.parallel_ops_register import get_distributed_op
from hyper_parallel.core.dtensor.device_mesh import init_device_mesh, _DEVICE_MESH_MAP
from hyper_parallel.platform.platform import EXISTING_COMM_GROUPS


class TestSparseFlashAttentionDistributedOp(unittest.TestCase):

hyper_parallel/core/shard/ops/parallel_npu_sparse_flash_attention.py:316:4: W0221: Number of parameters was 3 in 'DistributedOp.infer_layout' and is now 2 in overriding 'SparseFlashAttentionDistributedOp.infer_layout' method (arguments-differ)
tests/ut/core/shard/ops/test_parallel_npu_sparse_flash_attention.py:24:0: C0413: Import "from hyper_parallel.core.dtensor.dtensor import _build_layout, _LAYOUT_CACHE" should be placed at the top of the module (wrong-import-position)
tests/ut/core/shard/ops/test_parallel_npu_sparse_flash_attention.py:25:0: C0413: Import "from hyper_parallel.core.dtensor.placement_types import Shard, Replicate" should be placed at the top of the module (wrong-import-position)
tests/ut/core/shard/ops/test_parallel_npu_sparse_flash_attention.py:26:0: C0413: Import "from hyper_parallel.core.shard.ops.parallel_npu_sparse_flash_attention import SparseFlashAttentionDistributedOp, _normalize_sfa_args" should be placed at the top of the module (wrong-import-position)
tests/ut/core/shard/ops/test_parallel_npu_sparse_flash_attention.py:30:0: C0413: Import "from hyper_parallel.core.shard.ops.parallel_ops_register import get_distributed_op" should be placed at the top of the module (wrong-import-position)
tests/ut/core/shard/ops/test_parallel_npu_sparse_flash_attention.py:31:0: C0413: Import "from hyper_parallel.core.dtensor.device_mesh import init_device_mesh, _DEVICE_MESH_MAP" should be placed at the top of the module (wrong-import-position)
tests/ut/core/shard/ops/test_parallel_npu_sparse_flash_attention.py:32:0: C0413: Import "from hyper_parallel.platform.platform import EXISTING_COMM_GROUPS" should be placed at the top of the module (wrong-import-position)
Your code has been rated at 9.82/10

?? .claude/commands/ascend-test.md
?? .claude/skills/ascend-test/
?? .sisyphus/
?? 1-torch-mindspore-virtual-stonebraker.md
?? 1.patch
?? bcba11b5a0b945fc9a1a3ccf0a50728e46746fa5.patch
?? custom_shard_issue.md
?? d_function_issue.md
?? docs/1-npu-dense-lightning-indexer-softmax-l-cheerful-yeti.md
?? docs/1-torch-mindspore-virtual-stonebraker.md
?? docs/3-commit-1-3-dense-lightning-indexer-gr-snoopy-toast.md
?? docs/ISSUE.md
?? docs/PR_description.md
?? docs/ci310_install_troubleshooting.md
?? docs/distributed_op_migration_audit_report.md
?? docs/distributed_op_migration_guide.md
?? docs/dsa_distributed_op_analysis.md
?? docs/ep_deleted_code.md
?? docs/etp.md
?? docs/hyper-parallel-ep-commit-f0fa78f686caa8f-prancy-donut.md
?? docs/issues/
?? docs/lightning_indexer_implementation_plan.md
?? docs/memoized-seeking-honey.md
?? docs/mhc/
?? docs/moe_ep_refactor_design.md
?? docs/np.md
?? docs/npu-sparse-flash-attention-torch-mindsp-wobbly-sunset.md
?? docs/npu_dense_lightning_indexer_grad_kl_loss_dist_op.md
?? docs/npu_dense_lightning_indexer_softmax_lse_dist_op.md
?? docs/npu_dense_lightning_indexer_softmax_lse_issue.md
?? docs/npu_sparse_flash_attention_dist_op.md
?? docs/npu_sparse_lightning_indexer_grad_kl_loss_bsnd_cp_fix.md
?? docs/npu_sparse_lightning_indexer_grad_kl_loss_dist_op.md
?? docs/rfc-moe-parallel.md
?? docs/rfc-operator-parallel.md
?? docs/rotary_position_embedding_dist_op_plan.md
?? docs/sfa_operator_params_comparison.md
?? docs/shimmering-riding-catmull_1.md
?? docs/shimmering-riding-catmull_2.md
?? docs/shimmering-riding-catmull_3.md
?? docs/velvety-shimmying-cake.md
?? email.sh
?? expert_parallel_pr.md
?? hyper_connection.py
?? hyper_parallel/core/shard/ops/parallel_rotary_position_embedding.py
?? hyper_parallel/core/shard/ops/yaml/rotary_position_embedding_ops.yaml
?? install.sh
?? npu_mhc_pre_sinkhorn_custom_op.py
?? pylintrc
?? run_ep_test.sh
?? test_npu.sh
?? test_sfa_torch.py
?? tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py
?? tests/mindspore/st/shard/ops/test_rotary_position_embedding_shard_in_python.py
?? tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py

(Bash completed with no output)

File created successfully at: /root/workspace/hyper-parallel/.claude/skills/ascend-test/ascend-test-work-rope.log (file state is current in your context — no need to Read it back)

A  hyper_parallel/core/shard/ops/parallel_rotary_position_embedding.py
A  hyper_parallel/core/shard/ops/yaml/rotary_position_embedding_ops.yaml
A  tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py
A  tests/mindspore/st/shard/ops/test_rotary_position_embedding_shard_in_python.py
A  tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py
?? .claude/commands/ascend-test.md
?? .claude/skills/ascend-test/
?? .sisyphus/
?? 1-torch-mindspore-virtual-stonebraker.md
?? 1.patch
?? bcba11b5a0b945fc9a1a3ccf0a50728e46746fa5.patch
?? custom_shard_issue.md
?? d_function_issue.md
?? docs/1-npu-dense-lightning-indexer-softmax-l-cheerful-yeti.md
?? docs/1-torch-mindspore-virtual-stonebraker.md
?? docs/3-commit-1-3-dense-lightning-indexer-gr-snoopy-toast.md
?? docs/ISSUE.md
?? docs/PR_description.md
?? docs/ci310_install_troubleshooting.md
?? docs/distributed_op_migration_audit_report.md
?? docs/distributed_op_migration_guide.md
?? docs/dsa_distributed_op_analysis.md
?? docs/ep_deleted_code.md
?? docs/etp.md
?? docs/hyper-parallel-ep-commit-f0fa78f686caa8f-prancy-donut.md
?? docs/issues/
?? docs/lightning_indexer_implementation_plan.md
?? docs/memoized-seeking-honey.md
?? docs/mhc/
?? docs/moe_ep_refactor_design.md
?? docs/np.md
?? docs/npu-sparse-flash-attention-torch-mindsp-wobbly-sunset.md
?? docs/npu_dense_lightning_indexer_grad_kl_loss_dist_op.md
?? docs/npu_dense_lightning_indexer_softmax_lse_dist_op.md
?? docs/npu_dense_lightning_indexer_softmax_lse_issue.md
?? docs/npu_sparse_flash_attention_dist_op.md
?? docs/npu_sparse_lightning_indexer_grad_kl_loss_bsnd_cp_fix.md
?? docs/npu_sparse_lightning_indexer_grad_kl_loss_dist_op.md
?? docs/rfc-moe-parallel.md
?? docs/rfc-operator-parallel.md
?? docs/rotary_position_embedding_dist_op_plan.md
?? docs/sfa_operator_params_comparison.md
?? docs/shimmering-riding-catmull_1.md
?? docs/shimmering-riding-catmull_2.md
?? docs/shimmering-riding-catmull_3.md
?? docs/velvety-shimmying-cake.md
?? email.sh
?? expert_parallel_pr.md
?? hyper_connection.py
?? install.sh
?? npu_mhc_pre_sinkhorn_custom_op.py
?? pylintrc
?? run_ep_test.sh
?? test_npu.sh
?? test_sfa_torch.py

[master_dist_op 244a8d0] feat: add distributed op RotaryPositionEmbedding with UT and MindSpore ST
 5 files changed, 1083 insertions(+)
 create mode 100644 hyper_parallel/core/shard/ops/parallel_rotary_position_embedding.py
 create mode 100644 hyper_parallel/core/shard/ops/yaml/rotary_position_embedding_ops.yaml
 create mode 100644 tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py
 create mode 100644 tests/mindspore/st/shard/ops/test_rotary_position_embedding_shard_in_python.py
 create mode 100644 tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py

master_dist_op

Authorized users only. All activities may be monitored and reported.
To 8.92.7.125:/home/hdd/bare-repo-hyper-parallel/
 * [new branch]      master_dist_op -> master_dist_op

Authorized users only. All activities may be monitored and reported.
Switched to branch 'master_dist_op'
HEAD is now at 244a8d0 feat: add distributed op RotaryPositionEmbedding with UT and MindSpore ST

Authorized users only. All activities may be monitored and reported.
244a8d0 feat: add distributed op RotaryPositionEmbedding with UT and MindSpore ST

<persisted-output>
Output too large (535.3KB). Full output saved to: /root/.claude/projects/-root-workspace-hyper-parallel/d0ef256d-0777-4cfc-b653-4343c4a7ae32/tool-results/bai3axly5.txt

Preview (first 2KB):
Authorized users only. All activities may be monitored and reported.
============================= test session starts ==============================
platform linux -- Python 3.9.25, pytest-8.4.2, pluggy-1.6.0 -- /home/hdd/miniconda3/envs/ci3.9/bin/python3.9
cachedir: .pytest_cache
rootdir: /home/hdd/hyper-parallel
collecting ... collected 4 items

test_rotary_position_embedding_shard_in_python.py::test_parallel_rpe_ms_group1 /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
  setattr(self, word, getattr(machar, word).flat[0])
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
  return self._float_to_str(self.smallest_subnormal)
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for <class 'numpy.float32'> type is zero.
  setattr(self, word, getattr(machar, word).flat[0])
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for <class 'numpy.float32'> type is zero.
  return self._float_to_str(self.smallest_subnormal)
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
  setattr(self, word, getattr(machar, word).flat[0])
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
  return self._float_to_str(self.smallest_subnormal)
...
</persisted-output>

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perlayer.py:50
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perlayer.py:50: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("minmax_update_perlayer")

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
======================= 1 passed, 22 warnings in 16.61s ========================
[INFO] DISTRIBUTED(258196,ffff2a7cf120,python3.9):2026-05-19-10:28:03.155.682 [mindspore/ccsrc/cluster/rpc/core/communicator/tcp_server.cc:313] Start] Event base dispatch success!
[INFO] DISTRIBUTED(258196,ffff29fbf120,python3.9):2026-05-19-10:28:03.155.682 [mindspore/ccsrc/cluster/rpc/core/communicator/tcp_client.cc:357] Start] Event base dispatch success!
../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer.py:54
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer.py:54: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("fake_quant_per_layer")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer_grad.py:81
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer_grad.py:81: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("fake_quant_per_layer_grad")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perchannel.py:50
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perchannel.py:50: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("minmax_update_perchannel")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perlayer.py:50
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perlayer.py:50: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("minmax_update_perlayer")

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
======================= 1 passed, 22 warnings in 16.49s ========================
../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer.py:54
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer.py:54: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("fake_quant_per_layer")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer_grad.py:81
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer_grad.py:81: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("fake_quant_per_layer_grad")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perchannel.py:50
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perchannel.py:50: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("minmax_update_perchannel")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perlayer.py:50
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perlayer.py:50: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("minmax_update_perlayer")

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
======================= 1 passed, 22 warnings in 16.44s ========================
[INFO] DISTRIBUTED(258228,ffff2f7ef120,python3.9):2026-05-19-10:28:03.213.454 [mindspore/ccsrc/cluster/rpc/core/communicator/tcp_client.cc:357] Start] Event base dispatch success!
[INFO] DISTRIBUTED(258228,ffff2ffff120,python3.9):2026-05-19-10:28:03.213.453 [mindspore/ccsrc/cluster/rpc/core/communicator/tcp_server.cc:313] Start] Event base dispatch success!
[INFO] DISTRIBUTED(258240,ffff4555f120,python3.9):2026-05-19-10:28:03.216.401 [mindspore/ccsrc/cluster/rpc/core/communicator/tcp_client.cc:357] Start] Event base dispatch success!
[INFO] DISTRIBUTED(258240,ffff45d6f120,python3.9):2026-05-19-10:28:03.216.401 [mindspore/ccsrc/cluster/rpc/core/communicator/tcp_server.cc:313] Start] Event base dispatch success!
../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer.py:54
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer.py:54: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("fake_quant_per_layer")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer_grad.py:81
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer_grad.py:81: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("fake_quant_per_layer_grad")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perchannel.py:50
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perchannel.py:50: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("minmax_update_perchannel")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perlayer.py:50
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perlayer.py:50: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("minmax_update_perlayer")

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
======================= 1 passed, 22 warnings in 16.64s ========================
[INFO] DISTRIBUTED(258185,ffff26fdf120,python3.9):2026-05-19-10:28:03.228.979 [mindspore/ccsrc/cluster/rpc/core/communicator/tcp_client.cc:357] Start] Event base dispatch success!
[INFO] DISTRIBUTED(258185,ffff277ef120,python3.9):2026-05-19-10:28:03.228.970 [mindspore/ccsrc/cluster/rpc/core/communicator/tcp_server.cc:313] Start] Event base dispatch success!
PASSED

=============================== warnings summary ===============================
../../../../common/mark_utils.py:38
  /home/hdd/hyper-parallel/tests/common/mark_utils.py:38: PytestUnknownMarkWarning: Unknown pytest.mark.platform_ascend910b - is this a typo?  You can register custom marks to avoid this warning - for details, see https://docs.pytest.org/en/stable/how-to/mark.html
    func = getattr(pytest.mark, plat_mark)(func)

../../../../common/mark_utils.py:39
  /home/hdd/hyper-parallel/tests/common/mark_utils.py:39: PytestUnknownMarkWarning: Unknown pytest.mark.level1 - is this a typo?  You can register custom marks to avoid this warning - for details, see https://docs.pytest.org/en/stable/how-to/mark.html
    func = getattr(pytest.mark, level_mark)(func)

../../../../common/mark_utils.py:40
  /home/hdd/hyper-parallel/tests/common/mark_utils.py:40: PytestUnknownMarkWarning: Unknown pytest.mark.allcards - is this a typo?  You can register custom marks to avoid this warning - for details, see https://docs.pytest.org/en/stable/how-to/mark.html
    func = getattr(pytest.mark, card_mark)(func)

../../../../common/mark_utils.py:41
  /home/hdd/hyper-parallel/tests/common/mark_utils.py:41: PytestUnknownMarkWarning: Unknown pytest.mark.essential - is this a typo?  You can register custom marks to avoid this warning - for details, see https://docs.pytest.org/en/stable/how-to/mark.html
    func = getattr(pytest.mark, essential_mark)(func)

tests/mindspore/st/shard/ops/test_rotary_position_embedding_shard_in_python.py::test_parallel_rpe_ms_group1
tests/mindspore/st/shard/ops/test_rotary_position_embedding_shard_in_python.py::test_parallel_rpe_ms_group1
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/torch_npu/utils/collect_env.py:58: UserWarning: Warning: The /usr/local/Ascend/cann-9.0.0 owner does not match the current owner.
    warnings.warn(f"Warning: The {path} owner does not match the current owner.")

tests/mindspore/st/shard/ops/test_rotary_position_embedding_shard_in_python.py::test_parallel_rpe_ms_group1
tests/mindspore/st/shard/ops/test_rotary_position_embedding_shard_in_python.py::test_parallel_rpe_ms_group1
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/torch_npu/utils/collect_env.py:58: UserWarning: Warning: The /usr/local/Ascend/cann-9.0.0/aarch64-linux/ascend_ops_install.info owner does not match the current owner.
    warnings.warn(f"Warning: The {path} owner does not match the current owner.")

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
================== 4 passed, 8 warnings in 163.90s (0:02:43) ===================

完成了 RotaryPositionEmbedding 分布式算子的全部开发：实现文件、YAML 注册、22 个 UT、8 个 MindSpore ST 用例，服务器上 4 个 group 全部通过。下一步可提 PR。 (disable recaps in /config)

The file /root/workspace/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py has been updated successfully. (file state is current in your context — no need to Read it back)

Your code has been rated at 10.00/10

[master_dist_op a7add13] refactor: improve RPE ST tests — multi-mode, _get_grad_placements bwd, tighten tol to 1e-3
 1 file changed, 138 insertions(+), 65 deletions(-)

Authorized users only. All activities may be monitored and reported.
To 8.92.7.125:/home/hdd/bare-repo-hyper-parallel/
   244a8d0..a7add13  master_dist_op -> master_dist_op

Authorized users only. All activities may be monitored and reported.
M	tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py
Already on 'master_dist_op'
HEAD is now at a7add13 refactor: improve RPE ST tests — multi-mode, _get_grad_placements bwd, tighten tol to 1e-3
a7add13 refactor: improve RPE ST tests — multi-mode, _get_grad_placements bwd, tighten tol to 1e-3

Exit code 1

Authorized users only. All activities may be monitored and reported.
============================= test session starts ==============================
platform linux -- Python 3.9.25, pytest-8.4.2, pluggy-1.6.0 -- /home/hdd/miniconda3/envs/ci3.9/bin/python3.9
cachedir: .pytest_cache
rootdir: /home/hdd/hyper-parallel
collecting ... collected 4 items

test_rotary_position_embedding_shard_in_python.py::test_parallel_rpe_ms_group1 /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
  setattr(self, word, getattr(machar, word).flat[0])
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
  return self._float_to_str(self.smallest_subnormal)
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for <class 'numpy.float32'> type is zero.
  setattr(self, word, getattr(machar, word).flat[0])
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for <class 'numpy.float32'> type is zero.
  return self._float_to_str(self.smallest_subnormal)
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
  setattr(self, word, getattr(machar, word).flat[0])
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
  return self._float_to_str(self.smallest_subnormal)
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for <class 'numpy.float32'> type is zero.
  setattr(self, word, getattr(machar, word).flat[0])
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for <class 'numpy.float32'> type is zero.
  return self._float_to_str(self.smallest_subnormal)
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
  setattr(self, word, getattr(machar, word).flat[0])
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
  return self._float_to_str(self.smallest_subnormal)
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for <class 'numpy.float32'> type is zero.
  setattr(self, word, getattr(machar, word).flat[0])
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for <class 'numpy.float32'> type is zero.
  return self._float_to_str(self.smallest_subnormal)
Start scheduler process, log file:./logs//home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python/test_rpe_replicated/scheduler.log. Execute command: pytest -s -v /home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py::test_rpe_replicated
Start worker process with rank id:0, log file:./logs//home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python/test_rpe_replicated/worker_0.log. Environment variable [RANK_ID=0] is exported. Execute command: pytest -s -v /home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py::test_rpe_replicated
Start scheduler process, log file:./logs//home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python/test_rpe_tp_n/scheduler.log. Execute command: pytest -s -v /home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py::test_rpe_tp_n
Start worker process with rank id:0, log file:./logs//home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python/test_rpe_tp_n/worker_0.log. Environment variable [RANK_ID=0] is exported. Execute command: pytest -s -v /home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py::test_rpe_tp_n
Start worker process with rank id:1, log file:./logs//home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python/test_rpe_replicated/worker_1.log. Environment variable [RANK_ID=1] is exported. Execute command: pytest -s -v /home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py::test_rpe_replicated
Start work

... [20009 characters truncated] ...

39:47.786.844 [mindspore/ccsrc/runtime/hardware_abstract/collective/collective_manager.cc:1034] CreateDeviceCommunicator] Start to send/fetch unqiueid for communication group hccl_world_group
[WARNING] ME(541647,fffec3fff120,python3.9):2026-05-19-10:39:47.787.067 [mindspore/ccsrc/runtime/hardware_abstract/collective/collective_manager.cc:1040] CreateDeviceCommunicator] End to send/fetch unqiueid for communication group hccl_world_group
[WARNING] DEVICE(541647,fffec1fbf120,python3.9):2026-05-19-10:39:47.787.287 [mindspore/ccsrc/plugin/ascend/res_manager/collective/ascend_communication_group.cc:276] InitByRootInfoConfig] Start to initialize communicator by HcclCommInitRootInfoConfig for hccl_world_group, hcclBufferSize is 200 MB, hcclDeterministic is 1
[WARNING] ME(541664,ffff94b39c40,python3.9):2026-05-19-10:39:49.259.164 [mindspore/ccsrc/runtime/hardware_abstract/collective/collective_manager.cc:368] CreateCommunicationGroup] Create collective communication group: hccl_world_group [const vector]{0, 1}, async: 0
[WARNING] ME(541664,fffeb1fbf120,python3.9):2026-05-19-10:39:49.259.507 [mindspore/ccsrc/runtime/hardware_abstract/collective/collective_manager.cc:1034] CreateDeviceCommunicator] Start to send/fetch unqiueid for communication group hccl_world_group
[WARNING] ME(541664,fffeb1fbf120,python3.9):2026-05-19-10:39:49.259.829 [mindspore/ccsrc/runtime/hardware_abstract/collective/collective_manager.cc:1040] CreateDeviceCommunicator] End to send/fetch unqiueid for communication group hccl_world_group
[WARNING] DEVICE(541664,fffeb17af120,python3.9):2026-05-19-10:39:49.260.104 [mindspore/ccsrc/plugin/ascend/res_manager/collective/ascend_communication_group.cc:276] InitByRootInfoConfig] Start to initialize communicator by HcclCommInitRootInfoConfig for hccl_world_group, hcclBufferSize is 200 MB, hcclDeterministic is 1
[WARNING] DEVICE(541647,fffec1fbf120,python3.9):2026-05-19-10:39:49.738.341 [mindspore/ccsrc/plugin/ascend/res_manager/collective/ascend_communication_group.cc:291] InitByRootInfoConfig] End to initialize communicator by HcclCommInitRootInfoConfig for hccl_world_group
[WARNING] DEVICE(541664,fffeb17af120,python3.9):2026-05-19-10:39:49.845.688 [mindspore/ccsrc/plugin/ascend/res_manager/collective/ascend_communication_group.cc:291] InitByRootInfoConfig] End to initialize communicator by HcclCommInitRootInfoConfig for hccl_world_group
[ERROR] DEVICE(541647,ffffb659bc40,python3.9):2026-05-19-10:39:50.563.301 [mindspore/ccsrc/plugin/ascend/res_manager/ascend_res_manager.cc:830] BaseCopy] Src ptr:0 or dst ptr:0x25090ae0 is null, please check the address is set correctly.
FAILED

=================================== FAILURES ===================================
________________________________ test_rpe_tp_n _________________________________

    def test_rpe_tp_n():
        """
        Feature: RotaryPositionEmbedding (MindSpore) N-dim tensor parallel, 2-card tp mesh.
        Description:
            - 1-D tp mesh (2 cards); x/cos/sin all Shard(1).
            - cos/sin full shape (B, N, S, D) — same sharding as x.
            - float16, mode=2 (quarter).
            - Forward and all three gradients verified; all grad placements = Shard(1).
        Expectation: Distributed forward and gradients match standalone within tolerance.
        """
        mode = 2
        x_np = np.random.randn(B, N, S, D_DIM).astype(np.float32)
        cos_np = np.random.randn(B, N, S, D_DIM).astype(np.float32)
        sin_np = np.random.randn(B, N, S, D_DIM).astype(np.float32)
>       ref_fwd, ref_grads = _run_standalone(x_np, cos_np, sin_np, mode=mode)

rotary_position_embedding_shard_in_python.py:249: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
rotary_position_embedding_shard_in_python.py:111: in _run_standalone
    return fwd_np, tuple(_to_f32np(g) for g in grads)
rotary_position_embedding_shard_in_python.py:111: in <genexpr>
    return fwd_np, tuple(_to_f32np(g) for g in grads)
rotary_position_embedding_shard_in_python.py:75: in _to_f32np
    return t.astype(ms.float32).asnumpy()
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

self = <[RuntimeError('SyncCopy failed for Tensor(shape=[4, 4, 16, 64], dtype=Float32, value=[...])\n\n----------------------...---------------------------\nmindspore/core/ir/tensor.cc:490 cpu\n') raised in repr()] Tensor object at 0xffff4dd33f70>

    def asnumpy(self):
        """
        Convert tensor to numpy array. Returns self tensor as a NumPy ndarray. This tensor and the returned ndarray
        share the same underlying storage. Changes to self tensor will be reflected in the ndarray.
    
        .. warning::
            Non-backward-compatible change after version 2.9.0: the interface name will be changed to `numpy`.
    
        Returns:
            A numpy ndarray which shares the same underlying storage with the tensor.
    
        Examples:
            >>> from mindspore import Tensor
            >>> import numpy

PASSEDPASSED
PASSED
PASSED
PASSEDPASSED
PASSED
PASSEDPASSEDPASSEDPASSED
PASSEDPASSEDPASSEDPASSED
PASSED
PASSEDPASSEDPASSEDPASSED
PASSEDPASSEDPASSEDPASSED
PASSED
PASSEDPASSEDPASSEDPASSEDPASSEDPASSED
PASSED
PASSED
PASSED

Authorized users only. All activities may be monitored and reported.
self = <[RuntimeError('SyncCopy failed for Tensor(shape=[4, 4, 16, 64], dtype=Float32, value=[...])\n\n----------------------...---------------------------\nmindspore/core/ir/tensor.cc:490 cpu\n') raised in repr()] Tensor object at 0xffff456abb80>
E       (Please search "CANN Common Error Analysis" at https://www.mindspore.cn/en for error code description)
self = <[RuntimeError('SyncCopy failed for Tensor(shape=[4, 4, 16, 64], dtype=Float32, value=[...])\n\n----------------------...---------------------------\nmindspore/core/ir/tensor.cc:490 cpu\n') raised in repr()] Tensor object at 0xffff27fabb80>
E       (Please search "CANN Common Error Analysis" at https://www.mindspore.cn/en for error code description)
FAILED rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp_sp - Runt...
FAILED rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp_sp - Runt...
FAILED rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp_sp - Runt...
FAILED rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp_sp - Runt...
FAILED rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp_sp - Runt...
FAILED rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp_sp - Runt...
FAILED rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp_sp - Runt...
FAILED rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp_sp - Runt...
        failed = [cases[i].case_name for i, p in enumerate(processes) if p.exitcode != 0]
        failed = [cases[i].case_name for i, p in enumerate(processes) if p.exitcode != 0]
        failed = [cases[i].case_name for i, p in enumerate(processes) if p.exitcode != 0]
        failed = [cases[i].case_name for i, p in enumerate(processes) if p.exitcode != 0]
FAILED test_rotary_position_embedding_shard_in_python.py::test_parallel_rpe_ms_group1
FAILED test_rotary_position_embedding_shard_in_python.py::test_parallel_rpe_ms_group2
FAILED test_rotary_position_embedding_shard_in_python.py::test_parallel_rpe_ms_group3
FAILED test_rotary_position_embedding_shard_in_python.py::test_parallel_rpe_ms_group4

1	# RPE 分布式算子 ST 验证工作日志
2	# 算子: RotaryPositionEmbedding
3	# 文件:
4	#   hyper_parallel/core/shard/ops/parallel_rotary_position_embedding.py
5	#   hyper_parallel/core/shard/ops/yaml/rotary_position_embedding_ops.yaml
6	#   tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py
7	#   tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py
8	#   tests/mindspore/st/shard/ops/test_rotary_position_embedding_shard_in_python.py
9	

The file /root/workspace/hyper-parallel/.claude/skills/ascend-test/ascend-test-work-rope.log has been updated successfully. (file state is current in your context — no need to Read it back)

The file /root/workspace/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py has been updated successfully. (file state is current in your context — no need to Read it back)

Your code has been rated at 10.00/10 (previous run: 10.00/10, +0.00)

Exit code 1
The following paths are ignored by one of your .gitignore files:
.claude/skills/ascend-test/ascend-test-work-rope.log
hint: Use -f if you really want to add them.
hint: Turn this message off by running
hint: "git config advice.addIgnoredFile false"

[master_dist_op 574a74b] fix: RPE ST — split fwd/fwd+bwd standalone, restrict mode=2/3 to fwd-only tests
 1 file changed, 50 insertions(+), 29 deletions(-)

Authorized users only. All activities may be monitored and reported.
To 8.92.7.125:/home/hdd/bare-repo-hyper-parallel/
   a7add13..574a74b  master_dist_op -> master_dist_op

Authorized users only. All activities may be monitored and reported.
M	tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py
Already on 'master_dist_op'
HEAD is now at 574a74b fix: RPE ST — split fwd/fwd+bwd standalone, restrict mode=2/3 to fwd-only tests
574a74b fix: RPE ST — split fwd/fwd+bwd standalone, restrict mode=2/3 to fwd-only tests

Exit code 1

Authorized users only. All activities may be monitored and reported.
============================= test session starts ==============================
platform linux -- Python 3.9.25, pytest-8.4.2, pluggy-1.6.0 -- /home/hdd/miniconda3/envs/ci3.9/bin/python3.9
cachedir: .pytest_cache
rootdir: /home/hdd/hyper-parallel
collecting ... collected 4 items

test_rotary_position_embedding_shard_in_python.py::test_parallel_rpe_ms_group1 /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
  setattr(self, word, getattr(machar, word).flat[0])
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
  return self._float_to_str(self.smallest_subnormal)
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for <class 'numpy.float32'> type is zero.
  setattr(self, word, getattr(machar, word).flat[0])
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for <class 'numpy.float32'> type is zero.
  return self._float_to_str(self.smallest_subnormal)
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
  setattr(self, word, getattr(machar, word).flat[0])
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
  return self._float_to_str(self.smallest_subnormal)
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for <class 'numpy.float32'> type is zero.
  setattr(self, word, getattr(machar, word).flat[0])
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for <class 'numpy.float32'> type is zero.
  return self._float_to_str(self.smallest_subnormal)
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
  setattr(self, word, getattr(machar, word).flat[0])
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
  return self._float_to_str(self.smallest_subnormal)
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for <class 'numpy.float32'> type is zero.
  setattr(self, word, getattr(machar, word).flat[0])
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for <class 'numpy.float32'> type is zero.
  return self._float_to_str(self.smallest_subnormal)
Start scheduler process, log file:./logs//home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python/test_rpe_replicated/scheduler.log. Execute command: pytest -s -v /home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py::test_rpe_replicated
Start worker process with rank id:0, log file:./logs//home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python/test_rpe_replicated/worker_0.log. Environment variable [RANK_ID=0] is exported. Execute command: pytest -s -v /home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py::test_rpe_replicated
Start scheduler process, log file:./logs//home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python/test_rpe_tp_n/scheduler.log. Execute command: pytest -s -v /home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py::test_rpe_tp_n
Start worker process with rank id:0, log file:./logs//home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python/test_rpe_tp_n/worker_0.log. Environment variable [RANK_ID=0] is exported. Execute command: pytest -s -v /home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py::test_rpe_tp_n
Start worker process with rank id:1, log file:./logs//home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python/test_rpe_replicated/worker_1.log. Environment variable [RANK_ID=1] is exported. Execute command: pytest -s -v /home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py::test_rpe_replicated
Start work

... [20012 characters truncated] ...

abstract/collective/collective_manager.cc:368] CreateCommunicationGroup] Create collective communication group: hccl_world_group [const vector]{0, 1}, async: 0
[WARNING] ME(744274,fffeb17af120,python3.9):2026-05-19-10:50:33.030.820 [mindspore/ccsrc/runtime/hardware_abstract/collective/collective_manager.cc:1034] CreateDeviceCommunicator] Start to send/fetch unqiueid for communication group hccl_world_group
[WARNING] ME(744274,fffeb17af120,python3.9):2026-05-19-10:50:33.031.059 [mindspore/ccsrc/runtime/hardware_abstract/collective/collective_manager.cc:1040] CreateDeviceCommunicator] End to send/fetch unqiueid for communication group hccl_world_group
[WARNING] DEVICE(744274,fffe8efdf120,python3.9):2026-05-19-10:50:33.031.277 [mindspore/ccsrc/plugin/ascend/res_manager/collective/ascend_communication_group.cc:276] InitByRootInfoConfig] Start to initialize communicator by HcclCommInitRootInfoConfig for hccl_world_group, hcclBufferSize is 200 MB, hcclDeterministic is 1
[WARNING] ME(744278,ffff83151c40,python3.9):2026-05-19-10:50:33.571.874 [mindspore/ccsrc/runtime/hardware_abstract/collective/collective_manager.cc:368] CreateCommunicationGroup] Create collective communication group: hccl_world_group [const vector]{0, 1}, async: 0
[WARNING] ME(744278,fffec51df120,python3.9):2026-05-19-10:50:33.572.184 [mindspore/ccsrc/runtime/hardware_abstract/collective/collective_manager.cc:1034] CreateDeviceCommunicator] Start to send/fetch unqiueid for communication group hccl_world_group
[WARNING] ME(744290,ffffb880fc40,python3.9):2026-05-19-10:50:34.797.486 [mindspore/ccsrc/runtime/hardware_abstract/collective/collective_manager.cc:368] CreateCommunicationGroup] Create collective communication group: hccl_world_group [const vector]{0, 1}, async: 0
[WARNING] ME(744290,fffed67cf120,python3.9):2026-05-19-10:50:34.797.822 [mindspore/ccsrc/runtime/hardware_abstract/collective/collective_manager.cc:1034] CreateDeviceCommunicator] Start to send/fetch unqiueid for communication group hccl_world_group
[WARNING] ME(744290,fffed67cf120,python3.9):2026-05-19-10:50:34.798.149 [mindspore/ccsrc/runtime/hardware_abstract/collective/collective_manager.cc:1040] CreateDeviceCommunicator] End to send/fetch unqiueid for communication group hccl_world_group
[WARNING] DEVICE(744290,fffed5fbf120,python3.9):2026-05-19-10:50:34.798.430 [mindspore/ccsrc/plugin/ascend/res_manager/collective/ascend_communication_group.cc:276] InitByRootInfoConfig] Start to initialize communicator by HcclCommInitRootInfoConfig for hccl_world_group, hcclBufferSize is 200 MB, hcclDeterministic is 1
[WARNING] ME(744262,ffff9c87ec40,python3.9):2026-05-19-10:50:35.057.230 [mindspore/ccsrc/runtime/hardware_abstract/collective/collective_manager.cc:368] CreateCommunicationGroup] Create collective communication group: hccl_world_group [const vector]{0, 1}, async: 0
[WARNING] ME(744262,fffeb9fbf120,python3.9):2026-05-19-10:50:35.064.229 [mindspore/ccsrc/runtime/hardware_abstract/collective/collective_manager.cc:1034] CreateDeviceCommunicator] Start to send/fetch unqiueid for communication group hccl_world_group
[WARNING] ME(744262,fffeb9fbf120,python3.9):2026-05-19-10:50:35.064.420 [mindspore/ccsrc/runtime/hardware_abstract/collective/collective_manager.cc:1040] CreateDeviceCommunicator] End to send/fetch unqiueid for communication group hccl_world_group
[WARNING] DEVICE(744262,fffe977ef120,python3.9):2026-05-19-10:50:35.064.648 [mindspore/ccsrc/plugin/ascend/res_manager/collective/ascend_communication_group.cc:276] InitByRootInfoConfig] Start to initialize communicator by HcclCommInitRootInfoConfig for hccl_world_group, hcclBufferSize is 200 MB, hcclDeterministic is 1
[WARNING] ME(744278,fffec51df120,python3.9):2026-05-19-10:50:35.072.882 [mindspore/ccsrc/runtime/hardware_abstract/collective/collective_manager.cc:1040] CreateDeviceCommunicator] End to send/fetch unqiueid for communication group hccl_world_group
[WARNING] DEVICE(744278,fffec49cf120,python3.9):2026-05-19-10:50:35.073.163 [mindspore/ccsrc/plugin/ascend/res_manager/collective/ascend_communication_group.cc:276] InitByRootInfoConfig] Start to initialize communicator by HcclCommInitRootInfoConfig for hccl_world_group, hcclBufferSize is 200 MB, hcclDeterministic is 1
[WARNING] DEVICE(744274,fffe8efdf120,python3.9):2026-05-19-10:50:35.203.651 [mindspore/ccsrc/plugin/ascend/res_manager/collective/ascend_communication_group.cc:291] InitByRootInfoConfig] End to initialize communicator by HcclCommInitRootInfoConfig for hccl_world_group
[WARNING] DEVICE(744290,fffed5fbf120,python3.9):2026-05-19-10:50:35.382.781 [mindspore/ccsrc/plugin/ascend/res_manager/collective/ascend_communication_group.cc:291] InitByRootInfoConfig] End to initialize communicator by HcclCommInitRootInfoConfig for hccl_world_group
[WARNING] DEVICE(744262,fffe977ef120,python3.9):2026-05-19-10:50:35.547.401 [mindspore/ccsrc/plugin/ascend/res_manager/collective/ascend_communication_group.cc:291] InitByRootInfoConfig] End to initialize communicator by HcclCom

Authorized users only. All activities may be monitored and reported.
Start worker process with rank id:3, log file:./logs//home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python/test_rpe_dp_tp_sp/worker_3.log. Environment variable [RANK_ID=3] is exported. Execute command: pytest -s -v /home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp_sp
Start worker process with rank id:4, log file:./logs//home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python/test_rpe_dp_tp_sp/worker_4.log. Environment variable [RANK_ID=4] is exported. Execute command: pytest -s -v /home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp_sp
Start worker process with rank id:5, log file:./logs//home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python/test_rpe_dp_tp_sp/worker_5.log. Environment variable [RANK_ID=5] is exported. Execute command: pytest -s -v /home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp_sp
collecting ... Start worker process with rank id:6, log file:./logs//home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python/test_rpe_dp_tp_sp/worker_6.log. Environment variable [RANK_ID=6] is exported. Execute command: pytest -s -v /home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp_sp
collecting ... Start worker process with rank id:7, log file:./logs//home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python/test_rpe_dp_tp_sp/worker_7.log. Environment variable [RANK_ID=7] is exported. Execute command: pytest -s -v /home/hdd/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp_sp
rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp_sp [WARNING] ME(802323:281473736625216,MainProcess):2026-05-19-10:55:13.639.28 [mindspore/context.py:1338] For 'context.set_context', the parameter 'device_target' will be deprecated and removed in a future version. Please use the api mindspore.set_device() instead.
rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp_sp [WARNING] ME(802345:281473530047552,MainProcess):2026-05-19-10:55:13.128.628 [mindspore/context.py:1338] For 'context.set_context', the parameter 'device_target' will be deprecated and removed in a future version. Please use the api mindspore.set_device() instead.
rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp_sp [WARNING] DISTRIBUTED(802345,ffffa9c5ac40,python3.9):2026-05-19-10:55:13.130.875 [mindspore/ccsrc/cluster/rpc/tcp/tcp_comm.cc:485] Connect] Connection 20 source: 127.0.0.1:51284, destination: 127.0.0.1:19307
rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp_sp [WARNING] ME(802303:281473561144384,MainProcess):2026-05-19-10:55:13.178.529 [mindspore/context.py:1338] For 'context.set_context', the parameter 'device_target' will be deprecated and removed in a future version. Please use the api mindspore.set_device() instead.
rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp_sp [WARNING] ME(802334:281473449589824,MainProcess):2026-05-19-10:55:13.193.743 [mindspore/context.py:1338] For 'context.set_context', the parameter 'device_target' will be deprecated and removed in a future version. Please use the api mindspore.set_device() instead.
rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp_sp [WARNING] ME(802313:281473449675840,MainProcess):2026-05-19-10:55:13.297.717 [mindspore/context.py:1338] For 'context.set_context', the parameter 'device_target' will be deprecated and removed in a future version. Please use the api mindspore.set_device() instead.
rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp_sp [WARNING] ME(802368:281473808792640,MainProcess):2026-05-19-10:55:13.380.682 [mindspore/context.py:1338] For 'context.set_context', the parameter 'device_target' will be deprecated and removed in a future version. Please use the api mindspore.set_device() instead.
rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp_sp [WARNING] ME(802356:281473436482624,MainProcess):2026-05-19-10:55:13.463.439 [mindspore/context.py:1338] For 'context.set_context', the parameter 'device_target' will be deprecated and removed in a future version. Please use the api mindspore.set_device() instead.
======================= 1 passed, 22 warnings in 16.11s ========================
======================= 1 passed, 22 warnings in 16.07s ========================
======================= 1 passed, 22 warnings in 16.15s ========================
======================= 1 passed, 22 warnings in 16.32s ========================
======================= 1 passed, 22 warnings in 16.24s ========================
======================= 1 passed, 22 warnings in 16.36s ========================
======================= 1 passed, 22 warnings in 16.28s ========================
======================= 1 passed, 22 warnings in 16.20s ========================
_________________________ test_parallel_rpe_ms_group2 __________________________
    def test_parallel_rpe_ms_group2():
            1. test_rpe_dp_tp — 2-D (dp=2, tp=2) mesh; x/cos/sin (Shard(0), Shard(1));
            2. test_rpe_dp_sp — 2-D (dp=2, sp=2) mesh; x (Shard(0), Shard(2)),
            MindSporeCase(IMPL_FILE, "test_rpe_dp_tp", 19303, 4, 4, 2),
            MindSporeCase(IMPL_FILE, "test_rpe_dp_sp", 19304, 4, 4, 2),
        # assert sum num_proc
        assert sum_num_proc <= global_num_proc, (f"sum num_proc {sum_num_proc} greater than "
            raise AssertionError(f"Cases timed out (possible collective deadlock): {timed_out}")
        failed = [cases[i].case_name for i, p in enumerate(processes) if p.exitcode != 0]
>       assert not failed, f"List cases failed: {failed}"
E       AssertionError: List cases failed: ['test_rpe_dp_tp']
../../../../common/parallel_case.py:109: AssertionError
tests/mindspore/st/shard/ops/test_rotary_position_embedding_shard_in_python.py::test_parallel_rpe_ms_group1
tests/mindspore/st/shard/ops/test_rotary_position_embedding_shard_in_python.py::test_parallel_rpe_ms_group1
tests/mindspore/st/shard/ops/test_rotary_position_embedding_shard_in_python.py::test_parallel_rpe_ms_group1
tests/mindspore/st/shard/ops/test_rotary_position_embedding_shard_in_python.py::test_parallel_rpe_ms_group1
FAILED test_rotary_position_embedding_shard_in_python.py::test_parallel_rpe_ms_group2
============= 1 failed, 3 passed, 8 warnings in 157.81s (0:02:37) ==============

Authorized users only. All activities may be monitored and reported.
    @fusion_manager.register("correction_mul_grad")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/correction_mul_grad.py:143
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/correction_mul_grad.py:143: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("correction_mul_grad_reduce")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_learned_scale_quant_perlayer.py:50
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_learned_scale_quant_perlayer.py:50: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("fake_learned_scale_quant_perlayer")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_learned_scale_quant_perlayer_grad.py:92
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_learned_scale_quant_perlayer_grad.py:92: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("fake_learned_scale_quant_perlayer_grad_d")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_learned_scale_quant_perlayer_grad_reduce.py:49
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_learned_scale_quant_perlayer_grad_reduce.py:49: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("fake_learned_scale_quant_perlayer_grad_d_reduce")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_learned_scale_quant_perchannel.py:50
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_learned_scale_quant_perchannel.py:50: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("fake_learned_scale_quant_perchannel")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_learned_scale_quant_perchannel_grad.py:91
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_learned_scale_quant_perchannel_grad.py:91: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("fake_learned_scale_quant_perchannel_grad_d")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_learned_scale_quant_perchannel_grad_reduce.py:48
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_learned_scale_quant_perchannel_grad_reduce.py:48: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("fake_learned_scale_quant_perchannel_grad_d_reduce")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perchannel.py:52
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perchannel.py:52: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("fake_quant_perchannel")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perchannel_grad.py:81
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perchannel_grad.py:81: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("fake_quant_perchannel_grad")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer.py:54
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer.py:54: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("fake_quant_per_layer")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer_grad.py:81
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer_grad.py:81: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("fake_quant_per_layer_grad")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perchannel.py:50
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perchannel.py:50: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("minmax_update_perchannel")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perlayer.py:50
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perlayer.py:50: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("minmax_update_perlayer")

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
=========================== short test summary info ============================
FAILED rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp - Runtime...
======================= 1 failed, 22 warnings in 13.67s ========================
[INFO] DISTRIBUTED(788890,ffff25c6f120,python3.9):2026-05-19-10:54:11.916.441 [mindspore/ccsrc/cluster/rpc/core/communicator/tcp_server.cc:313] Start] Event base dispatch success!
[INFO] DISTRIBUTED(788890,ffff2545f120,python3.9):2026-05-19-10:54:11.916.438 [mindspore/ccsrc/cluster/rpc/core/communicator/tcp_client.cc:357] Start] Event base dispatch success!

Authorized users only. All activities may be monitored and reported.
FAILED

=================================== FAILURES ===================================
________________________________ test_rpe_dp_tp ________________________________

    def test_rpe_dp_tp():
        """
        Feature: RotaryPositionEmbedding (MindSpore) 2-D dp×tp mesh, B by dp, N by tp.
        Description:
            - 4-card 2-D mesh (dp=2, tp=2); x/cos/sin all (Shard(0), Shard(1)).
            - cos/sin full shape (B, N, S, D) — same placements as x.
            - float16, mode=1 (rotate_interleaved).
            - Forward and all three gradients verified; all grad placements = (Shard(0), Shard(1)).
        Expectation: Distributed fwd and all gradients match standalone within tolerance.
        """
        mode = 1
        x_np = np.random.randn(B, N, S, D_DIM).astype(np.float32)
        cos_np = np.random.randn(B, N, S, D_DIM).astype(np.float32)
        sin_np = np.random.randn(B, N, S, D_DIM).astype(np.float32)
>       ref_fwd, ref_grads = _run_standalone(x_np, cos_np, sin_np, mode=mode)

rotary_position_embedding_shard_in_python.py:303: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
rotary_position_embedding_shard_in_python.py:130: in _run_standalone
    fwd_np = _to_f32np(fn(x, cos, sin))
rotary_position_embedding_shard_in_python.py:77: in _to_f32np
    return t.astype(ms.float32).asnumpy()
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 

self = <[RuntimeError('SyncCopy failed for Tensor(shape=[4, 4, 16, 64], dtype=Float32, value=[...])\n\n----------------------...---------------------------\nmindspore/core/ir/tensor.cc:490 cpu\n') raised in repr()] Tensor object at 0xffff285090d0>

    def asnumpy(self):
        """
        Convert tensor to numpy array. Returns self tensor as a NumPy ndarray. This tensor and the returned ndarray
        share the same underlying storage. Changes to self tensor will be reflected in the ndarray.
    
        .. warning::
            Non-backward-compatible change after version 2.9.0: the interface name will be changed to `numpy`.
    
        Returns:
            A numpy ndarray which shares the same underlying storage with the tensor.
    
        Examples:
            >>> from mindspore import Tensor
            >>> import numpy as np
            >>> x = Tensor(np.array([1, 2], dtype=np.float32))
            >>> y = x.asnumpy()
            >>> y[0] = 11
            >>> print(x)
            [11.  2.]
            >>> print(y)
            [11.  2.]
        """
        if self.has_init:
            self.init_data()
>       return TensorPy_.asnumpy(self)
E       RuntimeError: aclnnRotaryPositionEmbeddingGetWorkspaceSize call failed, please check!
E       
E       ----------------------------------------------------
E       - Ascend Error Message:
E       ----------------------------------------------------
E       [PID: 788890] 2026-05-19-10:54:09.521.649 Communication_Error_Bind_IP_Port(EI0019): Failed to enable listening for the host network adapter socket. Reason: The IP address 8.92.7.125 and port 65536 have already been bound.[THREAD:791460]
E               Solution: 1. Check whether this port has been occupied by another process. If yes, you can make adjustment using the environment variable HCCL_IF_BASE_PORT and use sysctl -w net.ipv4.ip_local_reserved_ports=****-**** to adjust the scope of reserved ports. 2. Check whether the service process is started multiple times on a device during this service.
E               TraceBack (most recent call last):
E               The shape of the input x, cos and sin is not supported.[FUNC:TilingSplit][FILE:rope_interleaved_tiling.cpp][LINE:321][THREAD:791775]
E               TilingSplit fail.[FUNC:DoOpTiling][FILE:rope_interleaved_tiling.cpp][LINE:364][THREAD:791775]
E               Tiling failed[THREAD:791775]
E               Tiling Failed.[THREAD:791775]
E               Kernel GetWorkspace failed. opType: 4[THREAD:791775]
E               RotaryPositionEmbedding ADD_TO_LAUNCHER_LIST_AICORE failed.[THREAD:791775]
E       
E       (Please search "CANN Common Error Analysis" at https://www.mindspore.cn/en for error code description)
E       
E       ----------------------------------------------------
E       - C++ Call Stack: (For framework developers)
E       ----------------------------------------------------
E       mindspore/ops/kernel/ascend/aclnn/pyboost_impl/auto_generate/pyboost_ascend_ops_3.cc:3129 operator()

/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/common/tensor.py:1013: RuntimeError
=============================== warnings summary ===============================
../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
    setattr(self, word, getattr(machar, word).flat[0])

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
    return self._float_to_str(self.smallest_subnormal)

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for <class 'numpy.float32'> type is zero.
    setattr(self, word, getattr(machar, word).flat[0])

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for <class 'numpy.float32'> type is zero.
    return self._float_to_str(self.smallest_subnormal)

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/batchnorm_fold2.py:57
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/batchnorm_fold2.py:57: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("batchnorm_fold2")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/batchnorm_fold2_grad.py:56
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/batchnorm_fold2_grad.py:56: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("batchnorm_fold2_grad")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/batchnorm_fold2_grad_reduce.py:48
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/batchnorm_fold2_grad_reduce.py:48: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("batchnorm_fold2_grad_reduce")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/correction_mul.py:51
--
FAILED rotary_position_embedding_shard_in_python.py::test_rpe_dp_tp - Runtime...
======================= 1 failed, 22 warnings in 13.67s ========================
[INFO] DISTRIBUTED(788890,ffff25c6f120,python3.9):2026-05-19-10:54:11.916.441 [mindspore/ccsrc/cluster/rpc/core/communicator/tcp_server.cc:313] Start] Event base dispatch success!
[INFO] DISTRIBUTED(788890,ffff2545f120,python3.9):2026-05-19-10:54:11.916.438 [mindspore/ccsrc/cluster/rpc/core/communicator/tcp_client.cc:357] Start] Event base dispatch success!

Conversation compacted

285	# ---------------------------------------------------------------------------
286	# 4-card tests
287	# ---------------------------------------------------------------------------
288	
289	def test_rpe_dp_tp():
290	    """
291	    Feature: RotaryPositionEmbedding (MindSpore) 2-D dp×tp mesh, B by dp, N by tp.
292	    Description:
293	        - 4-card 2-D mesh (dp=2, tp=2); x/cos/sin all (Shard(0), Shard(1)).
294	        - cos/sin full shape (B, N, S, D) — same placements as x.
295	        - float16, mode=1 (rotate_interleaved).
296	        - Forward and all three gradients verified; all grad placements = (Shard(0), Shard(1)).
297	    Expectation: Distributed fwd and all gradients match standalone within tolerance.
298	    """
299	    mode = 1
300	    x_np = np.random.randn(B, N, S, D_DIM).astype(np.float32)
301	    cos_np = np.random.randn(B, N, S, D_DIM).astype(np.float32)
302	    sin_np = np.random.randn(B, N, S, D_DIM).astype(np.float32)
303	    ref_fwd, ref_grads = _run_standalone(x_np, cos_np, sin_np, mode=mode)
304	
305	    mesh = init_device_mesh(device_type="npu", mesh_shape=(2, 2), mesh_dim_names=("dp", "tp"))
306	    placements = (Shard(0), Shard(1))
307	    dx = distribute_tensor(_np_to_f16(x_np), mesh, placements)
308	    dcos = distribute_tensor(_np_to_f16(cos_np), mesh, placements)
309	    dsin = distribute_tensor(_np_to_f16(sin_np), mesh, placements)
310	    d_inputs = (dx, dcos, dsin)
311	
312	    fn = _make_rpe(mode)
313	    _assert_fwd(fn(*d_inputs), ref_fwd, "dp_tp mode=1")
314	
315	    raw_grads = ms.grad(fn, (0, 1, 2))(*d_inputs)
316	    _assert_bwd(raw_grads, d_inputs, ref_grads, "dp_tp mode=1")
317	
318	
319	def test_rpe_dp_sp():
320	    """
321	    Feature: RotaryPositionEmbedding (MindSpore) 2-D dp×sp mesh, B by dp, S by sp.
322	    Description:
323	        - 4-card 2-D mesh (dp=2, sp=2); x (Shard(0), Shard(2)).
324	        - cos/sin broadcast (1, 1, S, D): dp Replicate, sp Shard(2).

The file /root/workspace/hyper-parallel/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py has been updated successfully. (file state is current in your context — no need to Read it back)

Wasted call — file unchanged since your last Read. Refer to that earlier tool_result instead.

The file /root/workspace/hyper-parallel/.claude/skills/ascend-test/ascend-test-work-rope.log has been updated successfully. (file state is current in your context — no need to Read it back)

diff --git a/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py b/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py
index 18b1dd3..b6f881a 100644
--- a/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py
+++ b/tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py
@@ -292,11 +292,11 @@ def test_rpe_dp_tp():
     Description:
         - 4-card 2-D mesh (dp=2, tp=2); x/cos/sin all (Shard(0), Shard(1)).
         - cos/sin full shape (B, N, S, D) — same placements as x.
-        - float16, mode=1 (rotate_interleaved).
+        - float16, mode=0 (rotate_half).
         - Forward and all three gradients verified; all grad placements = (Shard(0), Shard(1)).
     Expectation: Distributed fwd and all gradients match standalone within tolerance.
     """
-    mode = 1
+    mode = 0
     x_np = np.random.randn(B, N, S, D_DIM).astype(np.float32)
     cos_np = np.random.randn(B, N, S, D_DIM).astype(np.float32)
     sin_np = np.random.randn(B, N, S, D_DIM).astype(np.float32)
@@ -310,10 +310,10 @@ def test_rpe_dp_tp():
     d_inputs = (dx, dcos, dsin)
 
     fn = _make_rpe(mode)
-    _assert_fwd(fn(*d_inputs), ref_fwd, "dp_tp mode=1")
+    _assert_fwd(fn(*d_inputs), ref_fwd, "dp_tp mode=0")
 
     raw_grads = ms.grad(fn, (0, 1, 2))(*d_inputs)
-    _assert_bwd(raw_grads, d_inputs, ref_grads, "dp_tp mode=1")
+    _assert_bwd(raw_grads, d_inputs, ref_grads, "dp_tp mode=0")
 
 
 def test_rpe_dp_sp():

[master_dist_op 01f122f] test: fix test_rpe_dp_tp mode=1 tiling failure, switch to mode=0
 1 file changed, 4 insertions(+), 4 deletions(-)

master_dist_op

Authorized users only. All activities may be monitored and reported.
To 8.92.7.125:/home/hdd/bare-repo-hyper-parallel/
   574a74b..01f122f  master_dist_op -> master_dist_op

Authorized users only. All activities may be monitored and reported.
M	tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py
Already on 'master_dist_op'
HEAD is now at 01f122f test: fix test_rpe_dp_tp mode=1 tiling failure, switch to mode=0

Authorized users only. All activities may be monitored and reported.
01f122f test: fix test_rpe_dp_tp mode=1 tiling failure, switch to mode=0

<persisted-output>
Output too large (535.8KB). Full output saved to: /root/.claude/projects/-root-workspace-hyper-parallel/d0ef256d-0777-4cfc-b653-4343c4a7ae32/tool-results/bjexqx7i7.txt

Preview (first 2KB):
Authorized users only. All activities may be monitored and reported.
============================= test session starts ==============================
platform linux -- Python 3.9.25, pytest-8.4.2, pluggy-1.6.0 -- /home/hdd/miniconda3/envs/ci3.9/bin/python3.9
cachedir: .pytest_cache
rootdir: /home/hdd/hyper-parallel
collecting ... collected 4 items

test_rotary_position_embedding_shard_in_python.py::test_parallel_rpe_ms_group1 /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
  setattr(self, word, getattr(machar, word).flat[0])
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
  return self._float_to_str(self.smallest_subnormal)
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for <class 'numpy.float32'> type is zero.
  setattr(self, word, getattr(machar, word).flat[0])
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for <class 'numpy.float32'> type is zero.
  return self._float_to_str(self.smallest_subnormal)
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:549: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
  setattr(self, word, getattr(machar, word).flat[0])
/home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/numpy/core/getlimits.py:89: UserWarning: The value of the smallest subnormal for <class 'numpy.float64'> type is zero.
  return self._float_to_str(self.smallest_subnormal)
...
</persisted-output>

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perlayer.py:50
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perlayer.py:50: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("minmax_update_perlayer")

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
======================= 1 passed, 22 warnings in 16.79s ========================
../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer.py:54
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer.py:54: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("fake_quant_per_layer")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer_grad.py:81
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer_grad.py:81: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("fake_quant_per_layer_grad")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perchannel.py:50
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perchannel.py:50: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("minmax_update_perchannel")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perlayer.py:50
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perlayer.py:50: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("minmax_update_perlayer")

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
======================= 1 passed, 22 warnings in 16.70s ========================
[INFO] DISTRIBUTED(891459,ffff127cf120,python3.9):2026-05-19-11:03:51.402.169 [mindspore/ccsrc/cluster/rpc/core/communicator/tcp_server.cc:313] Start] Event base dispatch success!
[INFO] DISTRIBUTED(891459,ffff11fbf120,python3.9):2026-05-19-11:03:51.402.169 [mindspore/ccsrc/cluster/rpc/core/communicator/tcp_client.cc:357] Start] Event base dispatch success!
[INFO] DISTRIBUTED(891480,ffff33fff120,python3.9):2026-05-19-11:03:51.403.678 [mindspore/ccsrc/cluster/rpc/core/communicator/tcp_client.cc:357] Start] Event base dispatch success!
[INFO] DISTRIBUTED(891480,ffff48a3f120,python3.9):2026-05-19-11:03:51.403.678 [mindspore/ccsrc/cluster/rpc/core/communicator/tcp_server.cc:313] Start] Event base dispatch success!
../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer.py:54
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer.py:54: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("fake_quant_per_layer")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer_grad.py:81
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/fake_quant_perlayer_grad.py:81: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("fake_quant_per_layer_grad")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perchannel.py:50
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perchannel.py:50: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("minmax_update_perchannel")

../../../../../../miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perlayer.py:50
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/mindspore/ops/_op_impl/_custom_op/minmax_update_perlayer.py:50: DeprecationWarning: te_fusion.fusion_manager.fusion_manager.register is deprecated,please replace it with tbe.common.register.register_op_compute
    @fusion_manager.register("minmax_update_perlayer")

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
======================= 1 passed, 22 warnings in 16.83s ========================
[INFO] DISTRIBUTED(891448,ffff1d44f120,python3.9):2026-05-19-11:03:51.428.052 [mindspore/ccsrc/cluster/rpc/core/communicator/tcp_server.cc:313] Start] Event base dispatch success!
[INFO] DISTRIBUTED(891448,ffff1cc3f120,python3.9):2026-05-19-11:03:51.428.053 [mindspore/ccsrc/cluster/rpc/core/communicator/tcp_client.cc:357] Start] Event base dispatch success!
PASSED

=============================== warnings summary ===============================
../../../../common/mark_utils.py:38
  /home/hdd/hyper-parallel/tests/common/mark_utils.py:38: PytestUnknownMarkWarning: Unknown pytest.mark.platform_ascend910b - is this a typo?  You can register custom marks to avoid this warning - for details, see https://docs.pytest.org/en/stable/how-to/mark.html
    func = getattr(pytest.mark, plat_mark)(func)

../../../../common/mark_utils.py:39
  /home/hdd/hyper-parallel/tests/common/mark_utils.py:39: PytestUnknownMarkWarning: Unknown pytest.mark.level1 - is this a typo?  You can register custom marks to avoid this warning - for details, see https://docs.pytest.org/en/stable/how-to/mark.html
    func = getattr(pytest.mark, level_mark)(func)

../../../../common/mark_utils.py:40
  /home/hdd/hyper-parallel/tests/common/mark_utils.py:40: PytestUnknownMarkWarning: Unknown pytest.mark.allcards - is this a typo?  You can register custom marks to avoid this warning - for details, see https://docs.pytest.org/en/stable/how-to/mark.html
    func = getattr(pytest.mark, card_mark)(func)

../../../../common/mark_utils.py:41
  /home/hdd/hyper-parallel/tests/common/mark_utils.py:41: PytestUnknownMarkWarning: Unknown pytest.mark.essential - is this a typo?  You can register custom marks to avoid this warning - for details, see https://docs.pytest.org/en/stable/how-to/mark.html
    func = getattr(pytest.mark, essential_mark)(func)

tests/mindspore/st/shard/ops/test_rotary_position_embedding_shard_in_python.py::test_parallel_rpe_ms_group1
tests/mindspore/st/shard/ops/test_rotary_position_embedding_shard_in_python.py::test_parallel_rpe_ms_group1
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/torch_npu/utils/collect_env.py:58: UserWarning: Warning: The /usr/local/Ascend/cann-9.0.0 owner does not match the current owner.
    warnings.warn(f"Warning: The {path} owner does not match the current owner.")

tests/mindspore/st/shard/ops/test_rotary_position_embedding_shard_in_python.py::test_parallel_rpe_ms_group1
tests/mindspore/st/shard/ops/test_rotary_position_embedding_shard_in_python.py::test_parallel_rpe_ms_group1
  /home/hdd/miniconda3/envs/ci3.9/lib/python3.9/site-packages/torch_npu/utils/collect_env.py:58: UserWarning: Warning: The /usr/local/Ascend/cann-9.0.0/aarch64-linux/ascend_ops_install.info owner does not match the current owner.
    warnings.warn(f"Warning: The {path} owner does not match the current owner.")

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
================== 4 passed, 8 warnings in 148.78s (0:02:28) ===================

1	**What type of PR is this?**
2	
3	/kind feature
4	
5	----
6	
7	**What does this PR do / why do we need it**:
8	
9	为 HyperParallel 接入 MindSpore MC2（Memory-Communication Coupling）通算融合算子的分布式调度逻辑，包含：
10	- `mindspore.ops.all_gather_matmul`（`AllGatherMatmul`）
11	- `mindspore.ops.matmul_reduce_scatter`（`MatmulReduceScatter`）
12	
13	CANN MC2 算子在内核内部自行完成 AllGather / ReduceScatter 通信，不需要 HyperParallel 插入额外通信原语。HyperParallel 的作用是：
14	1. **DTensor 输入解包**：通过 `preprocess()` 的 `to_local()` 提取本地 tensor 后再传给 CANN。
15	2. **输出 DTensor Layout 推断**：通过 `infer_layout()` 基于输入分片方式推断输出 tensor 的分布状态，使下游算子能正确组合分布式计算图。
16	
17	### AllGatherMatmul
18	
19	| 参数 | 说明 |
20	|------|------|
21	| x1 (m, k) | m 维被 CANN 内部 AllGather 消耗，输出 m 变为全局（Replicate）|
22	| x1 k 维 | 支持 Replicate 或 Shard；k 被切分时输出带 Partial(sum)，调用方需 AllReduce |
23	| x2 (k, n) 或 (n, k) | trans_x2=False 时 k 在 dim 0，trans_x2=True 时 k 在 dim 1；k 维须与 x1 k 维 placement 匹配 |
24	| gather_output | True 时额外返回 AllGather 后的 x1 全局张量；False 时 CANN 返回空 tensor |
25	
26	输出 layout 推断：
27	- output dim 0 (m)：始终 Replicate（-1），因 AllGather 消耗了 m 维分片
28	- output dim 1 (n)：继承 x2 的 n 维 placement
29	- k 被切分时，output 带 `Partial(sum)`（与 `LinearDistributedOp` contract_dim 分片语义一致）
30	- gather_output=False 时 gather_out layout 强制全 Replicate，避免空 tensor 的维度越界
31	
32	### MatmulReduceScatter
33	
34	| 参数 | 说明 |
35	|------|------|
36	| x1 (m, k) | k 维须 Shard（TP），m 维可为 Replicate（纯 TP）或 Shard（DP）|
37	| x2 (k, n) 或 (n, k) | k 维须与 x1 k 维 layout 一致；n 维可任意 |
38	
39	输出 layout 推断：
40	- output dim 0 (m)：ReduceScatter 将 k 的 TP 分片转化为 m 分片；若 x1 m 维有 DP 分片则联合分片（tuple tensor_map）
41	- output dim 1 (n)：继承 x2 的 n 维 placement
42	- **无 Partial 状态**：CANN ReduceScatter 已在内部完成 k 方向的 sum reduce + m 方向 scatter
43	
44	### 测试覆盖
45	
46	| 测试类型 | 文件 | 内容 |
47	|----------|------|------|
48	| **UT** | `test_parallel_all_gather_matmul.py` (15 cases) | normalize_args / YAML 注册 / infer_layout 正例（1D+2D mesh、trans_x2、k 分片带 Partial）/ 错误处理（Partial 输入、k-placement 不匹配、m 维多 mesh 联合分片）|
49	| **UT** | `test_parallel_matmul_reduce_scatter.py` (12 cases) | normalize_args / YAML 注册 / infer_layout 正例 / 错误处理 |
50	| **ST** | `all_gather_matmul_shard_in_python.py` (7 cases) | x1 Shard(0)+x2 Replicate / x1 Shard(0)+x2 Shard(1) / trans_x2=True / gather_output=False / 2D (dp,tp) mesh / 4 卡 k 切分 (mp,tp) / 8 卡 mnk 全切 (mp,np,tp) |
51	| **ST** | `matmul_reduce_scatter_shard_in_python.py` (5 cases) | tp_basic / trans_x2=True / large_m / 2D (dp,tp) mesh / 8 卡 mnk 全切 (mp,np,tp) |
52	
53	----
54	
55	**Which issue(s) this PR fixes**:
56	
57	Fixes [#152](https://gitcode.com/mindspore/hyper-parallel/issues/152)
58	
59	----
60	
61	**Test Plan and Test result：What scenarios were tested, and what were the verification results（Function, performance, reliability, etc.）**：
62	
63	### UT（单元测试）
64	- `tests/ut/core/shard/ops/test_parallel_all_gather_matmul.py` — 15 cases，覆盖参数归一化、YAML 注册、infer_layout（replicate/shard/trans_x2/2D-mesh/**k 分片带 Partial**）、错误处理（Partial 输入、k-placement 不匹配、m 维多 mesh 联合分片）
65	- `tests/ut/core/shard/ops/test_parallel_matmul_reduce_scatter.py` — 12 cases，覆盖参数归一化、YAML 注册、infer_layout（TP/DP+TP/n-shard/trans_x2/联合 tuple m 分片）、错误处理（Partial 输入、k Replicate、k layout 不一致）
66	
67	### ST（系统测试 - distributed, MindSpore）
68	- `tests/mindspore/st/shard/ops/test_all_gather_matmul_shard_in_python.py` — 3 组：group1（4×2 卡）、group2（2×4 卡，含 k 分片场景）、group3（1×8 卡，mnk 三轴全切）
69	- `tests/mindspore/st/shard/ops/test_matmul_reduce_scatter_shard_in_python.py` — 3 组：group1（3×2 卡）、group2（1×4 卡）、group3（1×8 卡，mnk 三轴全切）
70	
71	### 验证结果
72	- 所有 27 条 UT 通过（mock 平台，无 GPU/NPU 依赖）
73	- 所有 AGM group1/group2 ST 在 Ascend NPU 上通过（float16 精度，atol/rtol 满足）
74	- k 分片场景使用 float32 numpy 参考值，atol=0.1/rtol=0.02（与 MRS 一致，因 CANN 跨 rank partial sum 累加顺序与单次全 k matmul 路径不同）
75	- pylint 通过
76	
77	----
78	
79	**Self-checklist**:（**请自检，在[ ]内打上x，我们将检视你的完成情况，否则会导致pr无法合入**）
80	
81	- [x] **设计**：PR对应的方案是否已经经过Maintainer评审，方案检视意见是否均已答复并完成方案修改
82	- [x] **测试**：PR中的代码是否已有UT/ST测试用例进行充分的覆盖，新增测试用例是否随本PR一并上库或已经上库
83	- [x] **验证**：PR描述信息中是否已包含对该PR对应的Feature、Refactor、Bugfix的预期目标达成情况的详细验证结果描述
84	- [ ] **接口**：是否涉及对外接口变更，相应变更已得到接口评审组织的通过，API对应的注释信息已经刷新正确
85	- [ ] **文档**：是否涉及官网文档修改，如果涉及请及时提交资料到Doc仓
86	
87	<!-- **Special notes for your reviewers**: -->
88	<!-- - [ ] 是否导致无法前向兼容 -->
89	<!-- - [ ] 是否涉及依赖的三方库变更 -->
90	

1	# ISSUE: 接入 MindSpore MC2 通算融合算子分布式调度（AllGatherMatmul / MatmulReduceScatter）
2	
3	## 背景
4	
5	在分布式 Tensor Parallelism（TP）训练场景中，CANN MC2 算子（AllGatherMatmul / MatmulReduceScatter）在内核内部自行完成 AllGather / ReduceScatter 通信，性能优于分离式通信+计算方案。然而 HyperParallel 当前没有这两个算子的分布式调度逻辑，用 DTensor 对象调用时会直接失败（CANN 内核无法处理 DTensor），即使绕过也无法推断输出 DTensor 的分布状态，下游算子无法正确组合分布式计算图。
6	
7	## 目标
8	
9	为以下两个 MindSpore MC2 算子提供分布式调度支持：
10	
11	### AllGatherMatmul
12	
13	- CANN 内核内部完成 AllGather（消耗 x1 m 维分片），HyperParallel 负责：
14	  - 通过 `preprocess()` 的 `to_local()` 提取本地 tensor 传给 CANN
15	  - 通过 `infer_layout()` 推断输出 layout：
16	    - m 维始终 Replicate（AllGather 消耗了 m 维分片）
17	    - n 维继承 x2 的 n 维 placement
18	    - k 被切分时输出带 `Partial(sum)`，调用方需 AllReduce（与 `LinearDistributedOp` contract_dim 分片语义一致）
19	    - `gather_output=False` 时 CANN 返回空 tensor，gather_out layout 强制全 Replicate，避免 Shard dim 越界
20	
21	支持的分片模式：
22	- x1 Shard(0) on tp，x2 Replicate
23	- x1 Shard(0) on tp，x2 Shard(1) on n
24	- trans_x2=True
25	- 2D mesh (dp, tp)
26	- k 切分：4 卡 (mp=2, tp=2)，x1 Shard(0)+Shard(1)，x2 Replicate+Shard(0)，输出带 Partial(sum) on tp
27	- k 切分：8 卡 (mp=2, np=2, tp=2)，x1 Shard(0)+Replicate+Shard(1)，x2 Replicate+Shard(1)+Shard(0)，输出带 Partial(sum) on tp
28	
29	### MatmulReduceScatter
30	
31	- CANN 内核内部完成 ReduceScatter（k 方向 sum reduce + m 方向 scatter），HyperParallel 负责：
32	  - 通过 `preprocess()` 的 `to_local()` 提取本地 tensor 传给 CANN
33	  - 通过 `infer_layout()` 推断输出 layout：
34	    - output dim 0 (m)：ReduceScatter 将 k 的 TP 分片转化为 m 分片；若 x1 m 维有 DP 分片则联合分片（tuple tensor_map）
35	    - output dim 1 (n)：继承 x2 的 n 维 placement
36	    - **无 Partial 状态**：CANN ReduceScatter 已在内部完成 k 方向 sum + m 方向 scatter
37	
38	支持的分片模式：
39	- x1 Shard(1) on tp（k 维），x2 Shard(0) on tp（k 维）
40	- trans_x2=True
41	- 2D mesh (dp, tp)，x1 Shard(1) on tp
42	- 8 卡 (mp=2, np=2, tp=2)，m/n/k 三轴全切
43	
44	---
45	
46	两个算子均继承 DistributedOp 基类，实现 `preprocess` 和 `infer_layout`，通过 YAML 注册。
47	不需要 `get_expand_impl` 覆盖（基类默认返回 None）。
48	
49	**Partial 说明**：AllGatherMatmul 在 k 维切分时输出带 `Partial(sum)`，`full_tensor()` 自动触发 AllReduce；MatmulReduceScatter 不需要 Partial（CANN 内核已内部完成 ReduceScatter，结果完整）。
50	
51	约束：
52	
53	| 算子 | 约束 |
54	|------|------|
55	| AllGatherMatmul | x1 k 维与 x2 k 维 placement 必须一致（均 Replicate 或均 Shard 在同一 mesh axis）；k 被切分时输出带 Partial(sum)；x1 m 维不支持多 mesh 联合分片（tuple tensor_map）|
56	| MatmulReduceScatter | x1 k 维必须 Shard（TP）；x2 k 维 placement 必须与 x1 k 维一致 |
57	| 两者共用 | gather_index=0、trans_input=False、bias=None（当前 MindSpore 约束）；不允许 Partial 输入 |
58	
59	UT：`test_parallel_all_gather_matmul.py`（15 cases）、`test_parallel_matmul_reduce_scatter.py`（12 cases）
60	ST：4 个 MindSpore 分布式测试文件（AGM 3 组，MRS 3 组，含 8 卡 mnk 全切场景）
61	
62	## 相关 PR
63	
64	#656
65	

1c7d6e9 feat: add distributed op RotaryPositionEmbedding with UT and MindSpore ST
51c236b !656 feat: add distributed ops AllGatherMatmul and MatmulReduceScatter (MC2)
749d780 !657 bugfix: add CMake pre-build support for mindspore custom_ops
93cdeda feat: add CMake pre-build support for mindspore custom_ops
aa5c8fb feat: add distributed ops all_gather_matmul and matmul_reduce_scatter
cecf192 !588 merge master_dfunction_dsa into master
47233b8 feat: add distributed op npu_dense_lightning_indexer_softmax_lse, npu_dense_lightning_indexer_grad_kl_loss_dist_op, npu_sparse_lightning_indexer_grad_kl_loss_dist_op, npu_mhc_post & npu_mhc_pre_sinkhorn
d84aa1a !648 feat: add dist op InplaceFillScalar & InplaceFillTensor
a914165 feat: add dist op InplaceFillScalar & InplaceFillTensor
89f023b !645 fix(pynative): make DeviceMesh / fully_shard robust under TP+EP+FSDP
8a9cadb fix(pynative): make DeviceMesh / fully_shard robust under TP+EP+FSDP
8d52cab !643 fix(reshape): fix ValueError on Replicate placement with small tensors
ea36632 fix(reshape): fix ValueError on Replicate placement with small tensors
d86e5f2 !603 swap_tensor
6f32e38 swap_tensor
e6b0d1e !599 feat(trainer+models): training stack + qwen3.5 dense/moe/vl-moe aligned
4445ec5 !624 feat(examples/torch/llama3): add 8-card DP+TP+CP+SP+FSDP2 example and llama3 ST coverage
380840b feat(trainer+models): training stack + qwen3.5 dense/moe/vl-moe aligned
df95173 !630 unwrap_cell
a68b934 !629 merge fix/issue-118-fsdp-shape-zero into master
70eaf1e unwrap_cell
37be588 fix: delay requires_grad restore for MindSpore fsdp param
5180fd4 !625 test(torch): 新增 Llama3 风格 TP+FSDP / TP+CP+FSDP 单卡精度对比测试
cb85d7f feat(examples/torch/llama3): add 8-card DP+TP+CP+SP+FSDP2 example and llama3 ST coverage
176fd15 !621 fix(fully_shard): reduce pure tp compat grads
3e614c4 test(torch): add Llama3 accuracy suite for parallel-vs-single-card
2f5ba97 !614 feat(mindspore): Llama3 tensor+sequence parallel example and MindSpore Cell support
561579e !620 fix: isolate CellBackwardHook output wrapping
0822568 !613 feat: add dist op Squeeze
34373d5 fix(fully_shard): reduce pure tp compat grads
f63bb9b !541 feat(pipeline): add dual-pipe comm/compute overlap coordinator
66a81fc test: add CellBackwardHook wrap_output coverage
422a72c feat(pipeline): add dual-pipe comm/compute overlap coordinator
a995770 !618 fix: add symmetric_memory docs and fix multiocore test for different devices.
d50401b fix: isolate CellBackwardHook output wrapping
92c9cdb fix: add symmetric_memory docs and fix multiocore test for different devices.
d7aa233 !616 merge fix/ms-hsdp-queued-div-policy into master
25dc07a fix: preserve MindSpore async reduce div policy
0b6d200 !615 feat(examples): Llama3 Torch FSDP+TP、TP+CP demo and README
504152c feat(examples): Llama3 Torch demos for TP, FSDP+TP, and TP+CP
26c6adc feat(mindspore): Llama3 TP+SP example and Cell parity in TP distribute_module
e6671d5 Add dist op Squeeze
bfe13cf !611 merge pr/20260430_152133 into master
fd4dac1 fix: remove MindSpore runtime param_info patch
59004a0 !610 merge master_dist_op into master
0e75552 !594 merge device_mesh_concate_master into master
140fa30 feat(dtensor): add layout-backed device mesh operations
c300f97 !609 fix: hide MindSpore activation wrapper cell prefix
3b0d596 add dist op Maximum
2ea2bf9 fix: hide MindSpore activation wrapper cell prefix
f02a94f !606 normalize device type
4386aff !607 feat(dcp): record dtype and full_shape when layout is absent or falsy
470b38b !583 fix: make fully_shard prefetch idempotent with replicate_params
85c9279 feat(dcp): record dtype and full_shape when layout is absent or falsy
1afcfa3 normalize device type
345c1c3 fix: handle replicate params in fully_shard transitions
6a68999 !601 fix: update 3rdparty shmem to v1.3.0
5e41778 !592 feat: Native compiling Multicore_moe_ffn
52b8192 !595 refactor(tests): consolidate mindspore and torch ut under tests/ut
7cd47bb !604 merge ops into master
84d40e4 multicore本地编译，依赖ops库代码拉取打patch
6738257 update shmem to v1.3.0
0930877 !602 merge add_llama into master
d029e48 !591 add_mesh_layout
b6b6b5c feat(examples): add Llama3 tensor-parallel demo; fix RowwiseParallel embedding output
8231e18 feat: add DTensorBase.to() for MindSpore backend and UT tests
e58f53a refactor(tests): consolidate mindspore and torch ut under tests/ut
30a6540 !582 merge fix/ms-comm-fusion-prefetch-overlap into master
ca97a6d add_mesh_layout
b91d00d fix some bugs
f8599d8 ops代码合入

.claude/rules/testing.md
.claude/skills/autogit/references/examples.md
.claude/skills/autogit/scripts/commands.py
.claude/skills/code-review/review-checklist.md
.claude/skills/platform-dev/workflows/05-testing.md
.gitignore
.jenkins/check/config/filter_cpplint.txt
.jenkins/check/config/filter_pylint.txt
.jenkins/rules/codespell/codespell.allow
3rdparty/patch/shmem.patch
CLAUDE.md
CODEOWNERS
README.md
docs/pipeline_parallel_overlap_b_f.md
examples/README.md
examples/mindspore/llama3/README.md
examples/mindspore/llama3/__init__.py
examples/mindspore/llama3/model.py
examples/mindspore/llama3/parallelize.py
examples/mindspore/llama3/tensor_parallel_example.py
examples/qwen3_5_0_8b_base/train.yaml
examples/qwen3_5_35b_a3b_base/train.yaml
examples/qwen3_vl_30b_a3b_instruct/train.yaml
examples/torch/llama3/README.md
examples/torch/llama3/__init__.py
examples/torch/llama3/dp_tp_cp_sp_fsdp_example.py
examples/torch/llama3/fsdp_tp_example.py
examples/torch/llama3/model.py
examples/torch/llama3/parallelize.py
examples/torch/llama3/tensor_parallel_example.py
examples/torch/llama3/tp_cp_example.py
examples/torch/pp_overlap/pp_overlap_moe_example.py
hyper_parallel/__init__.py
hyper_parallel/core/activation_checkpoint/__init__.py
hyper_parallel/core/activation_checkpoint/activation_checkpoint.py
hyper_parallel/core/activation_checkpoint/swap.py
hyper_parallel/core/distributed_checkpoint/layout.py
hyper_parallel/core/dtensor/_mesh_layout.py
hyper_parallel/core/dtensor/device_mesh.py
hyper_parallel/core/dtensor/dtensor.py
hyper_parallel/core/dtensor/layout.py
hyper_parallel/core/expert_parallel/expert_parallel.py
hyper_parallel/core/fully_shard/hsdp_scheduler.py
hyper_parallel/core/fully_shard/hsdp_state.py
hyper_parallel/core/multicore/__init__.py
hyper_parallel/core/multicore/doc/README.md
hyper_parallel/core/multicore/modules/common/compute_graph.py
hyper_parallel/core/multicore/modules/moe_ffn/backward/gen_runtime_data.py
hyper_parallel/core/multicore/modules/moe_ffn/common/task_builders.py
hyper_parallel/core/multicore/modules/moe_ffn/forward/gen_runtime_data.py
hyper_parallel/core/multicore/ops/multicore_moe_ffn/CMakeLists.txt
hyper_parallel/core/multicore/ops/multicore_moe_ffn/op_graph/CMakeLists.txt
hyper_parallel/core/multicore/ops/multicore_moe_ffn/op_graph/fusion_pass/.gitkeep
hyper_parallel/core/multicore/ops/multicore_moe_ffn/op_graph/multicore_moe_ffn_proto.h
hyper_parallel/core/multicore/ops/multicore_moe_ffn/op_host/CMakeLists.txt
hyper_parallel/core/multicore/ops/multicore_moe_ffn/op_host/config/ascend910_93/multicore_moe_ffn_binary.json
hyper_parallel/core/multicore/ops/multicore_moe_ffn/op_host/config/ascend910_93/multicore_moe_ffn_simplified_key.ini
hyper_parallel/core/multicore/ops/multicore_moe_ffn/op_host/config/ascend910b/multicore_moe_ffn_binary.json
hyper_parallel/core/multicore/ops/multicore_moe_ffn/op_host/config/ascend910b/multicore_moe_ffn_simplified_key.ini
hyper_parallel/core/multicore/ops/multicore_moe_ffn/op_host/multicore_moe_ffn_def.cpp
hyper_parallel/core/multicore/ops/multicore_moe_ffn/op_host/multicore_moe_ffn_infershape.cpp
hyper_parallel/core/multicore/ops/multicore_moe_ffn/op_host/multicore_moe_ffn_tiling.cpp
hyper_parallel/core/multicore/ops/multicore_moe_ffn/op_host/multicore_moe_ffn_tiling.h
hyper_parallel/core/multicore/ops/multicore_moe_ffn/op_host/op_api/aclnn_multicore_moe_ffn.cpp
hyper_parallel/core/multicore/ops/multicore_moe_ffn/op_host/op_api/aclnn_multicore_moe_ffn.h
hyper_parallel/core/multicore/ops/multicore_moe_ffn/op_host/op_api/multicore_moe_ffn.cpp
hyper_parallel/core/multicore/ops/multicore_moe_ffn/op_host/op_api/multicore_moe_ffn.h
hyper_parallel/core/multicore/ops/multicore_moe_ffn/op_kernel/multicore_moe_ffn.cpp
hyper_parallel/core/multicore/ops/multicore_moe_ffn/op_kernel/multicore_moe_ffn_tiling_key.h
hyper_parallel/core/multicore/ops/multicore_moe_ffn/op_kernel/put_mem_signal/put_mem_signal_kernel.cpp
hyper_parallel/core/multicore/ops/multicore_moe_ffn/op_kernel/runtime_config.hpp
hyper_parallel/core/multicore/ops/multicore_moe_ffn/op_kernel/worker_kernel.cpp
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/CMakeLists.txt
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/op_graph/CMakeLists.txt
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/op_graph/fusion_pass/.gitkeep
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/op_graph/multicore_moe_ffn_grad_proto.h
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/op_host/CMakeLists.txt
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/op_host/config/ascend910_93/multicore_moe_ffn_grad_binary.json
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/op_host/config/ascend910_93/multicore_moe_ffn_grad_simplified_key.ini
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/op_host/config/ascend910b/multicore_moe_ffn_grad_binary.json
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/op_host/config/ascend910b/multicore_moe_ffn_grad_simplified_key.ini
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/op_host/multicore_moe_ffn_grad_def.cpp
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/op_host/multicore_moe_ffn_grad_infershape.cpp
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/op_host/multicore_moe_ffn_grad_tiling.cpp
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/op_host/multicore_moe_ffn_grad_tiling.h
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/op_host/op_api/aclnn_multicore_moe_ffn_grad.cpp
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/op_host/op_api/aclnn_multicore_moe_ffn_grad.h
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/op_host/op_api/multicore_moe_ffn_grad.cpp
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/op_host/op_api/multicore_moe_ffn_grad.h
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/op_kernel/multicore_moe_ffn_grad.cpp
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/op_kernel/multicore_moe_ffn_grad_tiling_key.h
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/op_kernel/put_mem_signal/put_mem_signal_kernel.cpp
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/op_kernel/runtime_config.hpp
hyper_parallel/core/multicore/ops/multicore_moe_ffn_grad/op_kernel/worker_kernel.cpp
hyper_parallel/core/multicore/ops/ops-nn.patch
hyper_parallel/core/multicore/ops/ops-transformer.patch
hyper_parallel/core/multicore/platform/mindspore/CMakeLists.txt
hyper_parallel/core/multicore/platform/mindspore/__init__.py
hyper_parallel/core/multicore/platform/mindspore/c_api/moe_ffn_bwd/moe_ffn_bwd_graph.cc
hyper_parallel/core/multicore/platform/mindspore/c_api/moe_ffn_bwd/moe_ffn_bwd_pynative.cc
hyper_parallel/core/multicore/platform/mindspore/c_api/moe_ffn_fwd/moe_ffn_fwd_pynative.cc
hyper_parallel/core/multicore/platform/torch/__init__.py
hyper_parallel/core/multicore/platform/torch/csrc/moe_ffn_bwd.cpp
hyper_parallel/core/multicore/platform/torch/csrc/moe_ffn_fwd.cpp
hyper_parallel/core/multicore/platform/torch/csrc/registration.cpp
hyper_parallel/core/multicore/platform/torch/setup.py
hyper_parallel/core/multicore/prebuild/multicore_moe_ffn.tar.gz
hyper_parallel/core/multicore/prebuild/multicore_moe_ffn.tar.gz.sha256
hyper_parallel/core/pipeline_parallel/__init__.py
hyper_parallel/core/pipeline_parallel/comm_compute_overlap.py
hyper_parallel/core/pipeline_parallel/hook_coordinator.py
hyper_parallel/core/pipeline_parallel/scheduler.py
hyper_parallel/core/shard/_op_dispatch.py
hyper_parallel/core/shard/ops/parallel_all_gather_matmul.py
hyper_parallel/core/shard/ops/parallel_cell_backward_hook.py
hyper_parallel/core/shard/ops/parallel_lightning_indexer.py
hyper_parallel/core/shard/ops/parallel_matmul_reduce_scatter.py
hyper_parallel/core/shard/ops/parallel_mhc_post.py
hyper_parallel/core/shard/ops/parallel_mhc_pre_sinkhorn.py
hyper_parallel/core/shard/ops/parallel_ms_flash_attention_score.py
hyper_parallel/core/shard/ops/parallel_npu_dense_lightning_indexer_grad_kl_loss.py
hyper_parallel/core/shard/ops/parallel_npu_dense_lightning_indexer_softmax_lse.py
hyper_parallel/core/shard/ops/parallel_npu_flash_attention_score.py
hyper_parallel/core/shard/ops/parallel_npu_sparse_flash_attention.py
hyper_parallel/core/shard/ops/parallel_npu_sparse_lightning_indexer_grad_kl_loss.py
hyper_parallel/core/shard/ops/parallel_one_hot_ext.py
hyper_parallel/core/shard/ops/parallel_ops.py
hyper_parallel/core/shard/ops/parallel_reshape.py
hyper_parallel/core/shard/ops/parallel_rotary_position_embedding.py
hyper_parallel/core/shard/ops/parallel_scaled_dot_product_attention.py
hyper_parallel/core/shard/ops/yaml/all_gather_matmul_ops.yaml
hyper_parallel/core/shard/ops/yaml/element_wise_ops.yaml
hyper_parallel/core/shard/ops/yaml/element_wise_ops_with_shape.yaml
hyper_parallel/core/shard/ops/yaml/lightning_indexer_ops.yaml
hyper_parallel/core/shard/ops/yaml/matmul_reduce_scatter_ops.yaml
hyper_parallel/core/shard/ops/yaml/npu_dense_lightning_indexer_grad_kl_loss_ops.yaml
hyper_parallel/core/shard/ops/yaml/npu_dense_lightning_indexer_softmax_lse_ops.yaml
hyper_parallel/core/shard/ops/yaml/npu_mhc_post_ops.yaml
hyper_parallel/core/shard/ops/yaml/npu_mhc_pre_sinkhorn_ops.yaml
hyper_parallel/core/shard/ops/yaml/npu_sparse_flash_attention_ops.yaml
hyper_parallel/core/shard/ops/yaml/npu_sparse_lightning_indexer_grad_kl_loss_ops.yaml
hyper_parallel/core/shard/ops/yaml/rotary_position_embedding_ops.yaml
hyper_parallel/core/shard/ops/yaml/squeeze.yaml
hyper_parallel/core/symmetric_memory/doc/symmetric_memory.md
hyper_parallel/core/symmetric_memory/platform/torch/torch_bindings.cpp
hyper_parallel/core/tensor_parallel/api.py
hyper_parallel/core/tensor_parallel/style.py
hyper_parallel/custom_ops/__init__.py
hyper_parallel/custom_ops/experimental/__init__.py
hyper_parallel/integration/llamafactory/utils.py
hyper_parallel/models/__init__.py
hyper_parallel/models/modules/__init__.py
hyper_parallel/models/modules/attention.py
hyper_parallel/models/modules/feed_forward.py
hyper_parallel/models/modules/linear_attention.py
hyper_parallel/models/modules/moe.py
hyper_parallel/models/modules/rmsnorm.py
hyper_parallel/models/modules/rope.py
hyper_parallel/models/qwen3_5/__init__.py
hyper_parallel/models/qwen3_5/checkpoint.py
hyper_parallel/models/qwen3_5/model.py
hyper_parallel/models/qwen3_5/parallelize.py
hyper_parallel/models/qwen3_5/state_dict.py
hyper_parallel/models/qwen3_5_moe/__init__.py
hyper_parallel/models/qwen3_5_moe/checkpoint.py
hyper_parallel/models/qwen3_5_moe/model.py
hyper_parallel/models/qwen3_5_moe/parallelize.py
hyper_parallel/models/qwen3_5_moe/state_dict.py
hyper_parallel/models/qwen3_vl_moe/__init__.py
hyper_parallel/models/qwen3_vl_moe/checkpoint.py
hyper_parallel/models/qwen3_vl_moe/model.py
hyper_parallel/models/qwen3_vl_moe/parallelize.py
hyper_parallel/models/qwen3_vl_moe/state_dict.py
hyper_parallel/models/spec/__init__.py
hyper_parallel/models/spec/model_spec.py
hyper_parallel/models/spec/registry.py
hyper_parallel/models/spec/state_dict_adapter.py
hyper_parallel/platform/mindspore/activation_checkpoint/__init__.py
hyper_parallel/platform/mindspore/activation_checkpoint/activation_swap.py
hyper_parallel/platform/mindspore/custom_ops/CMakeLists.txt
hyper_parallel/platform/mindspore/custom_ops/__init__.py
hyper_parallel/platform/mindspore/custom_ops/custom_op_impl.py
hyper_parallel/platform/mindspore/custom_ops/custom_ops.py
hyper_parallel/platform/mindspore/custom_ops/dense_lightning_indexer_grad_kl_loss.cc
hyper_parallel/platform/mindspore/custom_ops/dense_lightning_indexer_softmax_lse.cc
hyper_parallel/platform/mindspore/custom_ops/mhc_post.cc
hyper_parallel/platform/mindspore/custom_ops/mhc_post_backward.cc
hyper_parallel/platform/mindspore/custom_ops/mhc_pre_sinkhorn.cc
hyper_parallel/platform/mindspore/custom_ops/mhc_pre_sinkhorn_backward.cc
hyper_parallel/platform/mindspore/custom_ops/module.cc
hyper_parallel/platform/mindspore/custom_ops/module.h
hyper_parallel/platform/mindspore/custom_ops/sparse_lightning_indexer_grad_kl_loss.cc
hyper_parallel/platform/mindspore/fully_shard/README.md
hyper_parallel/platform/mindspore/fully_shard/README_CN.md
hyper_parallel/platform/mindspore/fully_shard/_version_utils.py
hyper_parallel/platform/mindspore/fully_shard/param.py
hyper_parallel/platform/mindspore/fully_shard/param_group.py
hyper_parallel/platform/mindspore/fully_shard/state.py
hyper_parallel/platform/mindspore/init_weights.py
hyper_parallel/platform/mindspore/multicore/__init__.py
hyper_parallel/platform/mindspore/parameter_param_info_patch.py
hyper_parallel/platform/mindspore/platform.py
hyper_parallel/platform/mindspore/utils.py
hyper_parallel/platform/platform.py
hyper_parallel/platform/torch/activation_checkpoint/__init__.py
hyper_parallel/platform/torch/activation_checkpoint/activation_swap.py
hyper_parallel/platform/torch/common/moe.py
hyper_parallel/platform/torch/custom_ops/__init__.py
hyper_parallel/platform/torch/dtensor.py
hyper_parallel/platform/torch/fully_shard/param.py
hyper_parallel/platform/torch/fully_shard/param_group.py
hyper_parallel/platform/torch/fully_shard/state.py
hyper_parallel/platform/torch/multicore/__init__.py
hyper_parallel/platform/torch/platform.py
hyper_parallel/trainer/__init__.py
hyper_parallel/trainer/base.py
hyper_parallel/trainer/callbacks/__init__.py
hyper_parallel/trainer/callbacks/base.py
hyper_parallel/trainer/config.py
hyper_parallel/trainer/llm_trainer.py
hyper_parallel/trainer/parallel_dims.py
hyper_parallel/trainer/utils/__init__.py
hyper_parallel/trainer/utils/discovery.py
hyper_parallel/trainer/utils/logging.py
hyper_parallel/trainer/utils/loss.py
hyper_parallel/trainer/vl_trainer.py
prebuild/symmetric_memory.tar.gz
prebuild/symmetric_memory.tar.gz.sha256
scripts/build_custom_ops.sh
scripts/build_multicore_local.sh
scripts/build_symmetric_memory.sh
scripts/train_lm.py
scripts/train_vl.py
setup.py
tests/mindspore/st/activation_checkpoint/swap_activation.py
tests/mindspore/st/fully_shard/_test_device_mesh_concatenate.py
tests/mindspore/st/fully_shard/_test_fully_shard_replicate_prefetch_regression.py
tests/mindspore/st/fully_shard/_tp_fully_shard_e2e.py
tests/mindspore/st/fully_shard/test_device_mesh_concatenate.py
tests/mindspore/st/fully_shard/test_fully_shard_replicate_prefetch_regression.py
tests/mindspore/st/fully_shard/test_parameter_param_info_cycle.py
tests/mindspore/st/fully_shard/test_tp_fully_shard_e2e.py
tests/mindspore/st/init_weights/test_init_weights.py
tests/mindspore/st/multicore/moe_ffn.py
tests/mindspore/st/multicore/test_moe_ffn.py
tests/mindspore/st/process_group/device_mesh.py
tests/mindspore/st/shard/ops/all_gather_matmul_shard_in_python.py
tests/mindspore/st/shard/ops/lightning_indexer_shard_in_python.py
tests/mindspore/st/shard/ops/matmul_reduce_scatter_shard_in_python.py
tests/mindspore/st/shard/ops/npu_dense_lightning_indexer_grad_kl_loss_custom_op.py
tests/mindspore/st/shard/ops/npu_dense_lightning_indexer_softmax_lse_custom_op.py
tests/mindspore/st/shard/ops/npu_mhc_post_custom_op.py
tests/mindspore/st/shard/ops/npu_mhc_pre_sinkhorn_custom_op.py
tests/mindspore/st/shard/ops/npu_sparse_flash_attention_shard_in_python.py
tests/mindspore/st/shard/ops/npu_sparse_lightning_indexer_grad_kl_loss_custom_op.py
tests/mindspore/st/shard/ops/rotary_position_embedding_shard_in_python.py
tests/mindspore/st/shard/ops/test_all_gather_matmul_shard_in_python.py
tests/mindspore/st/shard/ops/test_lightning_indexer_shard_in_python.py
tests/mindspore/st/shard/ops/test_matmul_reduce_scatter_shard_in_python.py
tests/mindspore/st/shard/ops/test_npu_dense_lightning_indexer_grad_kl_loss_custom_op.py
tests/mindspore/st/shard/ops/test_npu_dense_lightning_indexer_softmax_lse_custom_op.py
tests/mindspore/st/shard/ops/test_npu_mhc_post_custom_op.py
tests/mindspore/st/shard/ops/test_npu_mhc_pre_sinkhorn_custom_op.py
tests/mindspore/st/shard/ops/test_npu_sparse_flash_attention_shard_in_python.py
tests/mindspore/st/shard/ops/test_npu_sparse_lightning_indexer_grad_kl_loss_custom_op.py
tests/mindspore/st/shard/ops/test_rotary_position_embedding_shard_in_python.py
tests/mindspore/st/symmetric_memory/test_symmetric_memory.py
tests/mindspore/ut/test_dtensor_set_data.py
tests/mindspore/ut/test_mindspore_patches.py
tests/torch/accuracy/README.md
tests/torch/accuracy/__init__.py
tests/torch/accuracy/_test_llama3_accuracy.py
tests/torch/accuracy/model.py
tests/torch/accuracy/parallelize.py
tests/torch/accuracy/test_llama3_accuracy.py
tests/torch/activation_checkpoint/activation_checkpoint_swap.py
tests/torch/fully_shard/_test_device_mesh_concatenate.py
tests/torch/fully_shard/_test_tp_fully_shard_e2e.py
tests/torch/fully_shard/test_device_mesh_concatenate.py
tests/torch/fully_shard/test_tp_fully_shard_e2e.py
tests/torch/llama3_examples/__init__.py
tests/torch/llama3_examples/_test_llama3_examples.py
tests/torch/llama3_examples/test_llama3_examples.py
tests/ut/conftest.py
tests/ut/core/context_parallel/__init__.py
tests/ut/core/context_parallel/test_async_context_parallel_mindspore.py
tests/ut/core/distributed_checkpoint/test_layout.py
tests/ut/core/distributed_checkpoint/test_loader.py
tests/ut/core/distributed_checkpoint/test_saver.py
tests/ut/core/dtensor/test_device_mesh_mindspore.py
tests/ut/core/dtensor/test_device_mesh_pytest.py
tests/ut/core/dtensor/test_dtensor.py
tests/ut/core/dtensor/test_dtensor_to.py
tests/ut/core/shard/ops/test_parallel_activation_with_axis.py
tests/ut/core/shard/ops/test_parallel_all_gather_matmul.py
tests/ut/core/shard/ops/test_parallel_arithmetic.py
tests/ut/core/shard/ops/test_parallel_atleast_1d.py
tests/ut/core/shard/ops/test_parallel_cell_backward_hook.py
tests/ut/core/shard/ops/test_parallel_chunk_view.py
tests/ut/core/shard/ops/test_parallel_elementwise_ops.py
tests/ut/core/shard/ops/test_parallel_gather_d.py
tests/ut/core/shard/ops/test_parallel_lightning_indexer.py
tests/ut/core/shard/ops/test_parallel_linear.py
tests/ut/core/shard/ops/test_parallel_matmul_reduce_scatter.py
tests/ut/core/shard/ops/test_parallel_mhc_post.py
tests/ut/core/shard/ops/test_parallel_mhc_pre_sinkhorn.py
tests/ut/core/shard/ops/test_parallel_npu_dense_lightning_indexer_grad_kl_loss.py
tests/ut/core/shard/ops/test_parallel_npu_dense_lightning_indexer_softmax_lse.py
tests/ut/core/shard/ops/test_parallel_npu_sparse_flash_attention.py
tests/ut/core/shard/ops/test_parallel_npu_sparse_lightning_indexer_grad_kl_loss.py
tests/ut/core/shard/ops/test_parallel_repeat.py
tests/ut/core/shard/ops/test_parallel_reshape.py
tests/ut/core/shard/ops/test_parallel_rotary_position_embedding.py
tests/ut/core/shard/ops/test_parallel_scatter.py
tests/ut/core/shard/ops/test_parallel_slice.py
tests/ut/core/shard/ops/test_parallel_sort.py
tests/ut/core/shard/ops/test_parallel_split.py
tests/ut/core/shard/ops/test_parallel_squeeze.py
tests/ut/core/shard/ops/test_parallel_stack.py
tests/ut/core/shard/ops/test_parallel_zeros_like.py
tests/ut/core/shard/test_op_dispatch.py
tests/ut/core/tensor_parallel/test_rowwise_parallel.py
tests/ut/core/test_device_mesh.py
tests/ut/platform/mindspore/__init__.py
tests/ut/platform/mindspore/_ensure_mindspore_platform.py
tests/ut/platform/mindspore/custom_ops/test_custom_op_impl.py
tests/ut/platform/mindspore/dtensor/__init__.py
tests/ut/platform/mindspore/dtensor/test_dtensor_set_data.py
tests/ut/platform/mindspore/fully_shard/__init__.py
tests/ut/platform/mindspore/fully_shard/test_fully_shard_list_api.py
tests/ut/platform/mindspore/fully_shard/test_fully_shard_register_hook.py
tests/ut/platform/mindspore/fully_shard/test_mixed_precision.py
tests/ut/platform/mindspore/fully_shard/test_pack_utils.py
tests/ut/platform/mindspore/fully_shard/test_param.py
tests/ut/platform/mindspore/fully_shard/test_scheduler.py
tests/ut/platform/mindspore/fully_shard/test_state.py
tests/ut/platform/mindspore/pipeline_parallel/__init__.py
tests/ut/platform/mindspore/pipeline_parallel/test_microbatch.py
tests/ut/platform/mindspore/test_activation_wrapper.py
tests/ut/platform/mindspore/test_autograd_compat.py
tests/ut/platform/torch/fully_shard/test_fully_shard.py
tests/ut/platform/torch/fully_shard/test_fully_shard_register_hook.py
tests/ut/platform/torch/fully_shard/test_fully_shard_state.py

Error: Uncommitted changes detected, please commit first:
  /autogit commit -m "your message"
Or manually: git add -A && git commit -m "message"

ascend-repo	hdd@8.92.7.125:/home/hdd/bare-repo-hyper-parallel/ (fetch)
ascend-repo	hdd@8.92.7.125:/home/hdd/bare-repo-hyper-parallel/ (push)
origin	https://gitcode.com/david-he91/hyper-parallel.git (fetch)
origin	https://gitcode.com/david-he91/hyper-parallel.git (push)
upstream	https://gitcode.com/mindspore/hyper-parallel.git (fetch)
upstream	https://gitcode.com/mindspore/hyper-parallel.git (push)

?? .claude/commands/ascend-test.md
?? .claude/skills/ascend-test/
?? .sisyphus/
?? 1-torch-mindspore-virtual-stonebraker.md
?? 1.patch
?? bcba11b5a0b945fc9a1a3ccf0a50728e46746fa5.patch
?? custom_shard_issue.md
?? d_function_issue.md
?? docs/1-npu-dense-lightning-indexer-softmax-l-cheerful-yeti.md
?? docs/1-torch-mindspore-virtual-stonebraker.md
?? docs/3-commit-1-3-dense-lightning-indexer-gr-snoopy-toast.md
?? docs/ISSUE.md
?? docs/PR_description.md
?? docs/ci310_install_troubleshooting.md
?? docs/distributed_op_migration_audit_report.md
?? docs/distributed_op_migration_guide.md
?? docs/dsa_distributed_op_analysis.md
?? docs/ep_deleted_code.md
?? docs/etp.md
?? docs/hyper-parallel-ep-commit-f0fa78f686caa8f-prancy-donut.md
?? docs/issues/
?? docs/lightning_indexer_implementation_plan.md
?? docs/memoized-seeking-honey.md
?? docs/mhc/
?? docs/moe_ep_refactor_design.md
?? docs/np.md
?? docs/npu-sparse-flash-attention-torch-mindsp-wobbly-sunset.md
?? docs/npu_dense_lightning_indexer_grad_kl_loss_dist_op.md
?? docs/npu_dense_lightning_indexer_softmax_lse_dist_op.md
?? docs/npu_dense_lightning_indexer_softmax_lse_issue.md
?? docs/npu_sparse_flash_attention_dist_op.md
?? docs/npu_sparse_lightning_indexer_grad_kl_loss_bsnd_cp_fix.md
?? docs/npu_sparse_lightning_indexer_grad_kl_loss_dist_op.md
?? docs/rfc-moe-parallel.md
?? docs/rfc-operator-parallel.md
?? docs/rotary_position_embedding_dist_op_plan.md
?? docs/sfa_operator_params_comparison.md
?? docs/shimmering-riding-catmull_1.md
?? docs/shimmering-riding-catmull_2.md
?? docs/shimmering-riding-catmull_3.md
?? docs/velvety-shimmying-cake.md
?? email.sh
?? expert_parallel_pr.md
?? hyper_connection.py
?? install.sh
?? npu_mhc_pre_sinkhorn_custom_op.py
?? pylintrc
?? run_ep_test.sh
?? test_npu.sh
?? test_sfa_torch.py

1	# Copyright 2026 Huawei Technologies Co., Ltd
2	#
3	# Licensed under the Apache License, Version 2.0 (the "License");
4	# you may not use this file except in compliance with the License.
5	# You may obtain a copy of the License at
6	#
7	# http://www.apache.org/licenses/LICENSE-2.0
8	#
9	# Unless required by applicable law or agreed to in writing, software
10	# distributed under the License is distributed on an "AS IS" BASIS,
11	# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12	# See the License for the specific language governing permissions and
13	# limitations under the License.
14	# ============================================================================
15	"""Distributed implementation for RotaryPositionEmbedding operator."""
16	import copy
17	from typing import Optional, Tuple
18	
19	from .parallel_ops import DistributedOp
20	
21	
22	def _normalize_rpe_args(x, cos, sin, mode=0):
23	    """Normalize positional and keyword arguments into a canonical positional tuple.
24	
25	    Args:
26	        x: Input tensor.
27	        cos: Cosine position encoding tensor.
28	        sin: Sine position encoding tensor.
29	        mode: Rotation mode. 0=rotate_half, 1=rotate_interleaved, 2=quarter,
30	            3=interleave-half. Defaults to 0.
31	
32	    Returns:
33	        tuple: (positional_args_tuple, empty_kwargs_dict)
34	    """
35	    return (x, cos, sin, mode), {}
36	
37	
38	class RotaryPositionEmbeddingDistributedOp(DistributedOp):
39	    """Distributed operator for RotaryPositionEmbedding.
40	
41	    Computes rotary position embedding element-wise:
42	        y = x * cos + x_rotate * sin
43	
44	    where x_rotate is obtained by rotating within the last (D) dimension.
45	    Output shape equals x shape exactly.
46	
47	    Sharding constraints:
48	      - D (last dim) must be replicated for x, cos, and sin: the kernel rotates
49	        within D and the operation is indivisible along that axis.
50	      - B, N, S dims are fully independent across positions and can be freely
51	        sharded.
52	      - cos/sin may have any subset of non-D dims replicated (broadcast case),
53	        but if cos/sin is sharded on a dimension, it must match x's sharding
54	        on that dimension.
55	
56	    MODE does not affect layout inference: all modes produce output shape == x
57	    shape and leave B/N/S independence unchanged. MODE is passed through as a
58	    kernel parameter.
59	
60	    Output:
61	      Single tensor with the same shape and layout as x.
62	    """
63	
64	    @staticmethod
65	    def _validate_input_layouts(x_layout, cos_layout, sin_layout) -> None:
66	        """Validate sharding constraints for all input tensors.
67	
68	        Rules (applied to both 4-D BNSD/BSND/SBND and 3-D TND layouts):
69	          - x's last dim (D) must be replicated.
70	          - cos and sin's last dim (D) must be replicated.
71	          - For any non-D dimension d: if cos/sin is sharded there, the mesh
72	            axis must equal x's mesh axis on the same dimension.
73	
74	        Args:
75	            x_layout: Layout of the x tensor.
76	            cos_layout: Layout of the cos tensor.
77	            sin_layout: Layout of the sin tensor.
78	
79	        Raises:
80	            ValueError: If D is sharded for any input, or if cos/sin sharding
81	                is inconsistent with x on any non-D dimension.
82	        """
83	        op = "rotary_position_embedding"
84	        x_tm = x_layout.tensor_map
85	
86	        if x_tm[-1] != -1:
87	            raise ValueError(
88	                f"For {op}, D (last dim) of x must be replicated, "
89	                f"but got tensor_map={x_tm}"
90	            )
91	
92	        for name, layout in [('cos', cos_layout), ('sin', sin_layout)]:
93	            tm = layout.tensor_map
94	            if tm[-1] != -1:
95	                raise ValueError(
96	                    f"For {op}, D (last dim) of {name} must be replicated, "
97	                    f"but got tensor_map={tm}"
98	                )
99	            for d in range(len(tm) - 1):
100	                x_d = x_tm[d] if d < len(x_tm) - 1 else -1
101	                if tm[d] != -1 and tm[d] != x_d:
102	                    raise ValueError(
103	                        f"For {op}, {name} sharding on dim {d} must match x or be replicated, "
104	                        f"but got x={x_d}, {name}={tm[d]}"
105	                    )
106	
107	    def preprocess(self, args: tuple, kwargs: dict) -> Optional[tuple]:
108	        """Extract local tensors and build the layout cache.
109	
110	        All inputs (x, cos, sin) are expected to be DTensors.
111	
112	        Args:
113	            args: Positional arguments, may include DTensors.
114	            kwargs: Keyword arguments.
115	
116	        Returns:
117	            tuple: (local_args, local_kwargs, cache_values) where
118	                local_args = (x_local, cos_local, sin_local, mode),
119	                local_kwargs = {},
120	                cache_values = [x_layout, cos_layout, sin_layout].
121	        """
122	        norm_args, _ = _normalize_rpe_args(*args, **kwargs)
123	        x = norm_args[0]
124	        cos = norm_args[1]
125	        sin = norm_args[2]
126	        mode = norm_args[3]
127	
128	        local_args = (x.to_local(), cos.to_local(), sin.to_local(), mode)
129	        cache_values = [x.layout, cos.layout, sin.layout]
130	        return local_args, {}, cache_values
131	
132	    def infer_layout(self, cache_values: list) -> Tuple[tuple, None]:
133	        """Infer output layout for the single output tensor.
134	
135	        Rules:
136	            1. Partial inputs are rejected.
137	            2. D (last dim) must be replicated for x, cos, and sin.
138	            3. cos/sin sharding on non-D dims must match x or be replicated.
139	            4. Output layout = deep copy of x_layout (output shape == x shape).
140	
141	        Args:
142	            cache_values: [x_layout, cos_layout, sin_layout]
143	
144	        Returns:
145	            tuple: ((output_layout,), None)
146	
147	        Raises:
148	            ValueError: If any input has Partial status, D is sharded,
149	                or cos/sin sharding is inconsistent with x.
150	        """
151	        x_layout = cache_values[0]
152	        cos_layout = cache_values[1]
153	        sin_layout = cache_values[2]
154	
155	        self._check_partial_inputs([x_layout, cos_layout, sin_layout])
156	        self._validate_input_layouts(x_layout, cos_layout, sin_layout)
157	        return (copy.deepcopy(x_layout),), None
158	

1	# Copyright 2026 Huawei Technologies Co., Ltd
2	#
3	# Licensed under the Apache License, Version 2.0 (the "License");
4	# you may not use this file except in compliance with the License.
5	# You may obtain a copy of the License at
6	#
7	# http://www.apache.org/licenses/LICENSE-2.0
8	#
9	# Unless required by applicable law or agreed to in writing, software
10	# distributed under the License is distributed on an "AS IS" BASIS,
11	# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
12	# See the License for the specific language governing permissions and
13	# limitations under the License.
14	# ============================================================================
15	"""Unit tests for RotaryPositionEmbeddingDistributedOp."""
16	import os
17	import unittest
18	from unittest.mock import MagicMock, patch
19	
20	import numpy as np
21	
22	os.environ["HYPER_PARALLEL_PLATFORM"] = "torch"
23	
24	from hyper_parallel.core.dtensor.dtensor import _build_layout, _LAYOUT_CACHE
25	from hyper_parallel.core.dtensor.placement_types import Shard, Replicate
26	from hyper_parallel.core.shard.ops.parallel_rotary_position_embedding import (
27	    RotaryPositionEmbeddingDistributedOp,
28	    _normalize_rpe_args,
29	)
30	from hyper_parallel.core.shard.ops.parallel_ops_register import get_distributed_op
31	from hyper_parallel.core.dtensor.device_mesh import init_device_mesh, _DEVICE_MESH_MAP
32	from hyper_parallel.platform.platform import EXISTING_COMM_GROUPS
33	
34	
35	class TestNormalizeRpeArgs(unittest.TestCase):
36	    """Unit tests for _normalize_rpe_args."""
37	
38	    def test_three_positional_args_mode_defaults_to_zero_1(self):
39	        """
40	        Feature: _normalize_rpe_args fills in default mode=0.
41	        Description: Call with only x, cos, sin positional args.
42	        Expectation: args == (x, cos, sin, 0), kwargs == {}.
43	        """
44	        x, cos, sin = object(), object(), object()
45	        args, kwargs = _normalize_rpe_args(x, cos, sin)
46	        self.assertIs(args[0], x, msg=f"args[0] should be x, got {args[0]}")
47	        self.assertIs(args[1], cos, msg=f"args[1] should be cos, got {args[1]}")
48	        self.assertIs(args[2], sin, msg=f"args[2] should be sin, got {args[2]}")
49	        self.assertEqual(args[3], 0, msg=f"default mode should be 0, got {args[3]}")
50	        self.assertEqual(kwargs, {}, msg=f"kwargs should be empty, got {kwargs}")

22

204:def test_rpe_replicated():
227:def test_rpe_dp_b():
256:def test_rpe_tp_n():
289:def test_rpe_dp_tp():
319:def test_rpe_dp_sp():
343:def test_rpe_tp_sp():
367:def test_rpe_dp_tp_cos_full():
396:def test_rpe_dp_tp_sp():

(Bash completed with no output)

69:            "  Linux/macOS: export GITCODE_TOKEN=<your-token>\n"
70:            "  Windows CMD: set GITCODE_TOKEN=<your-token>\n"
71:            "  Windows PowerShell: $env:GITCODE_TOKEN=\"<your-token>\"\n\n"
1527:  export GITCODE_TOKEN=<your-token>

(Bash completed with no output)

__pycache__
api.py
autogit.py
code_style_guard.py
commands.py
commit_msg_check.py
diff_analysis.py
git_utils.py
lint_check.py
models.py
pr_content.py

32:def get_token() -> Optional[str]:
65:def api_request(method: str, endpoint: str, token: str,
98:def get_pr_info(owner: str, repo: str, pr_number: int,
114:def get_pr_commits(owner: str, repo: str, pr_number: int,
130:def get_pr_files(owner: str, repo: str, pr_number: int,
146:def get_pr_stats(owner: str, repo: str, pr_number: int,
178:def add_reviewers(owner: str, repo: str, pr_number: int, token: str,
200:def get_pr_status_display(owner: str, repo: str, pr_number: int,
267:def create_pr(owner: str, repo: str, token: str, title: str, body: str,
302:def update_pr_description(owner: str, repo: str, pr_number: int, token: str,

65:def api_request(method: str, endpoint: str, token: str,
111:    return api_request("GET", f"/repos/{owner}/{repo}/pulls/{pr_number}", token)
127:    return api_request("GET", f"/repos/{owner}/{repo}/pulls/{pr_number}/commits", token)
143:    return api_request("GET", f"/repos/{owner}/{repo}/pulls/{pr_number}/files", token)
193:    return api_request(
263:    output += f"\nhttps://gitcode.com/{owner}/{repo}/pull/{pr_number}\n"
293:    status, result = api_request("POST", f"/repos/{owner}/{repo}/pulls", token, data)
298:            result["html_url"] = f"https://gitcode.com/{owner}/{repo}/pull/{pr_number}"
318:    return api_request("PATCH", f"/repos/{owner}/{repo}/pulls/{pr_number}", token, data)

1	#!/usr/bin/env python3
2	# Copyright 2026 Huawei Technologies Co., Ltd
3	#
4	# Licensed under the Apache License, Version 2.0 (the "License");
5	# you may not use this file except in compliance with the License.
6	# You may obtain a copy of the License at
7	#
8	# http://www.apache.org/licenses/LICENSE-2.0
9	#
10	# Unless required by applicable law or agreed to in writing, software
11	# distributed under the License is distributed on an "AS IS" BASIS,
12	# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
13	# See the License for the specific language governing permissions and
14	# limitations under the License.
15	# ============================================================================
16	"""Token management and GitCode HTTP API operations for AutoGit."""
17	
18	import json
19	import os
20	import subprocess
21	import urllib.error
22	import urllib.request
23	from typing import Any, Dict, List, Optional, Tuple
24	
25	from models import CREDENTIAL_HOST, GITCODE_API_BASE
26	
27	
28	# ============================================================================
29	# Token Management
30	# ============================================================================
31	
32	def get_token() -> Optional[str]:
33	    """Get the GitCode API token from environment or git credentials.
34	
35	    Checks ``GITCODE_TOKEN`` environment variable first, then falls back
36	    to ``git credential fill``.
37	
38	    Returns:
39	        Token string, or None if not found.
40	    """
41	    token = os.environ.get("GITCODE_TOKEN")
42	    if token:
43	        return token
44	
45	    try:
46	        proc = subprocess.run(
47	            ["git", "credential", "fill"],
48	            input=f"protocol=https\nhost={CREDENTIAL_HOST}\n",
49	            capture_output=True, text=True, timeout=10,
50	            check=False
51	        )
52	        if proc.returncode == 0:
53	            for line in proc.stdout.strip().split("\n"):
54	                if line.startswith("password="):
55	                    return line.split("=", 1)[1]
56	    except Exception:  # pylint: disable=broad-except
57	        pass
58	    return None
59	
60	
61	# ============================================================================
62	# GitCode API
63	# ============================================================================
64	
65	def api_request(method: str, endpoint: str, token: str,
66	                data: Optional[Dict] = None) -> Tuple[int, Any]:
67	    """Send an HTTP request to the GitCode API.
68	
69	    Args:
70	        method: HTTP method (GET, POST, PATCH, etc.).
71	        endpoint: API endpoint path (appended to GITCODE_API_BASE).
72	        token: Bearer token for authorization.
73	        data: Optional JSON-serializable request body.
74	
75	    Returns:
76	        Tuple of (HTTP status code, parsed JSON response).
77	    """
78	    url = f"{GITCODE_API_BASE}{endpoint}"
79	    headers = {
80	        "Content-Type": "application/json",
81	        "Authorization": f"Bearer {token}",
82	    }
83	    body = json.dumps(data).encode("utf-8") if data else None
84	    req = urllib.request.Request(url, data=body, headers=headers, method=method)
85	
86	    try:
87	        with urllib.request.urlopen(req, timeout=30) as resp:
88	            return resp.status, json.loads(resp.read().decode("utf-8"))
89	    except urllib.error.HTTPError as e:
90	        try:
91	            return e.code, json.loads(e.read().decode("utf-8"))
92	        except Exception:  # pylint: disable=broad-except
93	            return e.code, {"error": str(e)}
94	    except Exception as e:  # pylint: disable=broad-except
95	        return 0, {"error": str(e)}
96	
97	
98	def get_pr_info(owner: str, repo: str, pr_number: int,
99	                token: str) -> Tuple[int, Any]:
100	    """Get PR information from GitCode.

27:GITCODE_API_BASE = "https://api.gitcode.com/api/v5"
28:CREDENTIAL_HOST = "api.gitcode.com"

G3WanUqctM

1c7d6e9 feat: add distributed op RotaryPositionEmbedding with UT and MindSpore ST
51c236b !656 feat: add distributed ops AllGatherMatmul and MatmulReduceScatter (MC2)
749d780 !657 bugfix: add CMake pre-build support for mindspore custom_ops
93cdeda feat: add CMake pre-build support for mindspore custom_ops
aa5c8fb feat: add distributed ops all_gather_matmul and matmul_reduce_scatter
cecf192 !588 merge master_dfunction_dsa into master
47233b8 feat: add distributed op npu_dense_lightning_indexer_softmax_lse, npu_dense_lightning_indexer_grad_kl_loss_dist_op, npu_sparse_lightning_indexer_grad_kl_loss_dist_op, npu_mhc_post & npu_mhc_pre_sinkhorn
d84aa1a !648 feat: add dist op InplaceFillScalar & InplaceFillTensor

1c7d6e9 feat: add distributed op RotaryPositionEmbedding with UT and MindSpore ST
51c236b !656 feat: add distributed ops AllGatherMatmul and MatmulReduceScatter (MC2)
749d780 !657 bugfix: add CMake pre-build support for mindspore custom_ops
93cdeda feat: add CMake pre-build support for mindspore custom_ops
aa5c8fb feat: add distributed ops all_gather_matmul and matmul_reduce_scatter

1c7d6e9 feat: add distributed op RotaryPositionEmbedding with UT and MindSpore ST
51c236b !656 feat: add distributed ops AllGatherMatmul and MatmulReduceScatter (MC2)
749d780 !657 bugfix: add CMake pre-build support for mindspore custom_ops
93cdeda feat: add CMake pre-build support for mindspore custom_ops
aa5c8fb feat: add distributed ops all_gather_matmul and matmul_reduce_scatter
cecf192 !588 merge master_dfunction_dsa into master
47233b8 feat: add distributed op npu_dense_lightning_indexer_softmax_lse, npu_dense_lightning_indexer_grad_kl_loss_dist_op, npu_sparse_lightning_indexer_grad_kl_loss_dist_op, npu_mhc_post & npu_mhc_pre_sinkhorn
d84aa1a !648 feat: add dist op InplaceFillScalar & InplaceFillTensor
a914165 feat: add dist op InplaceFillScalar & InplaceFillTensor
89f023b !645 fix(pynative): make DeviceMesh / fully_shard robust under TP+EP+FSDP

master_dist_op
1c7d6e9 feat: add distributed op RotaryPositionEmbedding with UT and MindSpore ST
51c236b !656 feat: add distributed ops AllGatherMatmul and MatmulReduceScatter (MC2)
749d780 !657 bugfix: add CMake pre-build support for mindspore custom_ops
93cdeda feat: add CMake pre-build support for mindspore custom_ops
aa5c8fb feat: add distributed ops all_gather_matmul and matmul_reduce_scatter

1c7d6e9 HEAD@{0}: rebase (finish): returning to refs/heads/master_dist_op
1c7d6e9 HEAD@{1}: rebase (squash): feat: add distributed op RotaryPositionEmbedding with UT and MindSpore ST
583a333 HEAD@{2}: rebase (squash): # This is a combination of 3 commits.
90fdc93 HEAD@{3}: rebase (squash): # This is a combination of 2 commits.
244a8d0 HEAD@{4}: rebase (start): checkout 51c236bd893d04e7d4325c38d1682b062ab5b78d
01f122f HEAD@{5}: commit: test: fix test_rpe_dp_tp mode=1 tiling failure, switch to mode=0
574a74b HEAD@{6}: commit: fix: RPE ST — split fwd/fwd+bwd standalone, restrict mode=2/3 to fwd-only tests
a7add13 HEAD@{7}: commit: refactor: improve RPE ST tests — multi-mode, _get_grad_placements bwd, tighten tol to 1e-3
244a8d0 HEAD@{8}: commit: feat: add distributed op RotaryPositionEmbedding with UT and MindSpore ST
51c236b HEAD@{9}: rebase (finish): returning to refs/heads/master_dist_op
51c236b HEAD@{10}: rebase (start): checkout upstream/master
a914165 HEAD@{11}: checkout: moving from master_custom_op_cmake to master_dist_op
93cdeda HEAD@{12}: commit (amend): feat: add CMake pre-build support for mindspore custom_ops
c7aadcd HEAD@{13}: commit (amend): feat: add CMake pre-build support for mindspore custom_ops
0cc6224 HEAD@{14}: checkout: moving from master_dist_op_mc2 to master_custom_op_cmake

commit 1c7d6e93df8fabcc59c40015ce43fde5bae43854
Author: david-he91 <hedongdong@huawei.com>
Date:   Tue May 19 10:29:49 2026 +0800

    feat: add distributed op RotaryPositionEmbedding with UT and MindSpore ST

 .../ops/parallel_rotary_position_embedding.py      | 157 +++++++
 .../ops/yaml/rotary_position_embedding_ops.yaml    |   4 +
 .../rotary_position_embedding_shard_in_python.py   | 418 +++++++++++++++++
 ...st_rotary_position_embedding_shard_in_python.py |  91 ++++
 .../ops/test_parallel_rotary_position_embedding.py | 507 +++++++++++++++++++++
 5 files changed, 1177 insertions(+)

remote: Start Git Hooks Checking 					[PASSED]        
remote: 
remote: To create a merge request for master_dist_op, visit:        
remote:   https://gitcode.com/david-he91/hyper-parallel/merge_requests/new?source_branch=master_dist_op        
remote: 
To https://gitcode.com/david-he91/hyper-parallel.git
   a914165..1c7d6e9  master_dist_op -> master_dist_op

Exit code 1
Traceback (most recent call last):
  File "<string>", line 4, in <module>
ImportError: cannot import name 'get_upstream_repo_info' from 'git_utils' (/root/workspace/hyper-parallel/.claude/skills/autogit/scripts/git_utils.py)

['get_upstream_default_branch']

https://gitcode.com/mindspore/hyper-parallel.git

Token obtained: 24 chars

Status: 200
Issue #: 155
URL: https://gitcode.com/mindspore/hyper-parallel/issues/155

Status: 200
PR #: 1
URL: https://gitcode.com/mindspore/hyper-parallel/pull/1

267	def create_pr(owner: str, repo: str, token: str, title: str, body: str,
268	              head: str, base: str, fork_owner: str,
269	              fork_repo: str) -> Tuple[int, Any]:
270	    """Create a PR on GitCode.
271	
272	    Args:
273	        owner: Upstream repository owner.
274	        repo: Upstream repository name.
275	        token: API token.
276	        title: PR title.
277	        body: PR description body.
278	        head: Head branch reference (e.g. 'fork_owner:branch').
279	        base: Target branch name.
280	        fork_owner: Fork repository owner.
281	        fork_repo: Fork repository name.
282	
283	    Returns:
284	        Tuple of (status_code, response_data).
285	    """
286	    data = {
287	        "title": title,
288	        "body": body,
289	        "head": head,
290	        "base": base,
291	        "fork_path": f"{fork_owner}/{fork_repo}"
292	    }
293	    status, result = api_request("POST", f"/repos/{owner}/{repo}/pulls", token, data)
294	
295	    if status in [200, 201] and isinstance(result, dict):
296	        pr_number = result.get("number") or result.get("iid")
297	        if pr_number:
298	            result["html_url"] = f"https://gitcode.com/{owner}/{repo}/pull/{pr_number}"
299	    return status, result
300	
301	
302	def update_pr_description(owner: str, repo: str, pr_number: int, token: str,
303	                          title: str, body: str) -> Tuple[int, Any]:
304	    """Update a PR title and description.
305	
306	    Args:
307	        owner: Repository owner.
308	        repo: Repository name.
309	        pr_number: Pull request number.
310	        token: API token.
311	        title: New PR title.
312	        body: New PR description body.
313	
314	    Returns:
315	        Tuple of (status_code, response_data).
316	    """

Status: 400
Error: {
  "error_code": 404,
  "error_code_name": "UN_KNOW",
  "error_message": "Can not find the branch: master_dist_op in project: hyper-parallel",
  "trace_id": "eea6d42503b32a2e6d5137b3ee1d0aec"
}

Status: 200
PR #: 661
URL: https://gitcode.com/mindspore/hyper-parallel/pull/661

Status: 200
ISSUE #155 updated successfully

RotaryPositionEmbedding 分布式算子开发和 ST 验证已全部完成，所有 4 组测试（2/4/8 卡）在 Ascend 910B 通过，ISSUE #155 和 PR #661 已创建。无待处理事项。 (disable recaps in /config)

RotaryPositionEmbedding 分布式算子已开发完成并通过全部 ST 验证（4 组共 8 个用例），ISSUE #155 和 PR #661 已创建。任务完成，无待处理事项。 (disable recaps in /config)