已合并
[bugfix]cp冗余资料回退 #36645
TrHan创建于 5月25日
[bugfix]cp冗余资料回退 #36645
已合并
共 1 个文件变更+0-23
| @@ -32,8 +32,6 @@ PyTorch Ascend Adapter named torch_npu is committed to maintaining compatibility | |||
| 32 | 32 | ||
| 33 | - **[torch.distributed.optim](#torchdistributedoptim)** | 33 | - **[torch.distributed.optim](#torchdistributedoptim)** |
| 34 | 34 | ||
| 35 | -- **[torch.distributed.tensor.experimental](#torchdistributedtensorexperimental)** | ||
| 36 | - | ||
| 37 | - **[torch.distributed.tensor.parallel](#torchdistributedtensorparallel)** | 35 | - **[torch.distributed.tensor.parallel](#torchdistributedtensorparallel)** |
| 38 | 36 | ||
| 39 | - **[torch.distributed.checkpoint](#torchdistributedcheckpoint)** | 37 | - **[torch.distributed.checkpoint](#torchdistributedcheckpoint)** |
| @@ -11894,27 +11892,6 @@ Just like torch.cuda above, you only need to replace ``torch.cuda.amp.xxx`` with | |||
| 11894 | </tbody> | 11892 | </tbody> |
| 11895 | </table> | 11893 | </table> |
| 11896 | 11894 | ||
| 11897 | -## torch.distributed.tensor.experimental | ||
| 11898 | - | ||
| 11899 | -<a name="table-context-parallel-experimental"></a> | ||
| 11900 | -<table><thead align="left"><tr id="row-context-parallel-experimental-head"><th class="cellrowborder" valign="top" width="44.84%" id="mcps-context-parallel-experimental-api"><p id="p-context-parallel-experimental-api"><a name="p-context-parallel-experimental-api"></a><a name="p-context-parallel-experimental-api"></a>PyTorch API</p> | ||
| 11901 | -</th> | ||
| 11902 | -<th class="cellrowborder" valign="top" width="27.58%" id="mcps-context-parallel-experimental-compat"><p id="p-context-parallel-experimental-compat"><a name="p-context-parallel-experimental-compat"></a><a name="p-context-parallel-experimental-compat"></a>Compatibility</p> | ||
| 11903 | -</th> | ||
| 11904 | -<th class="cellrowborder" valign="top" width="27.58%" id="mcps-context-parallel-experimental-limit"><p id="p-context-parallel-experimental-limit"><a name="p-context-parallel-experimental-limit"></a><a name="p-context-parallel-experimental-limit"></a>Limitations</p> | ||
| 11905 | -</th> | ||
| 11906 | -</tr> | ||
| 11907 | -</thead> | ||
| 11908 | -<tbody><tr id="row-context-parallel-experimental-context-parallel"><td class="cellrowborder" valign="top" width="44.84%" headers="mcps-context-parallel-experimental-api "><p id="p-context-parallel-experimental-context-parallel"><a name="p-context-parallel-experimental-context-parallel"></a><a name="p-context-parallel-experimental-context-parallel"></a>torch.distributed.tensor.experimental.context_parallel</p> | ||
| 11909 | -</td> | ||
| 11910 | -<td class="cellrowborder" valign="top" width="27.58%" headers="mcps-context-parallel-experimental-compat "><p id="p-context-parallel-experimental-context-parallel-compat"><a name="p-context-parallel-experimental-context-parallel-compat"></a><a name="p-context-parallel-experimental-context-parallel-compat"></a>Y</p> | ||
| 11911 | -</td> | ||
| 11912 | -<td class="cellrowborder" valign="top" width="27.58%" headers="mcps-context-parallel-experimental-limit "><p id="p-context-parallel-experimental-context-parallel-limit"><a name="p-context-parallel-experimental-context-parallel-limit"></a><a name="p-context-parallel-experimental-context-parallel-limit"></a>Only supports the NPU fused SDPA path. The q/k/v tensors must use BNSD layout. pse, padding_mask, prefix, actual_seq_qlen, actual_seq_kvlen, sink, and arbitrary non-causal attention masks are not supported. Load balancing requires causal attention.</p> | ||
| 11913 | -</td> | ||
| 11914 | -</tr> | ||
| 11915 | -</tbody> | ||
| 11916 | -</table> | ||
| 11917 | - | ||
| 11918 | ## torch.distributed.tensor.parallel | 11895 | ## torch.distributed.tensor.parallel |
| 11919 | 11896 | ||
| 11920 | <a name="table42392735619"></a> | 11897 | <a name="table42392735619"></a> |