0.11.0

Release 0.11.0

主要依赖库版本描述

• python 3.9 或者更高
• [torch-mlu] 2.7.1；
• [torch-npu] 2.7.1；
• pytorch与torch-mlu或torch-npu对应；
• [triton-x] 3.2.0.post9 或者更高；

主要特性与改进

• 增加GraphRunner用于提供基于torch.nn.Module的NpuGraph与MluGraph模块封装，使用样例：

      import torch
      from xpu_graph import GraphRunner, Target
      model = torch.nn.Sequential(*[torch.nn.Linear(1024, 1024)] * 3).npu()
      input_tensor = torch.empty(1024, 1024).uniform_(10, 100).npu()
  
      golden = model(input_tensor)
      device_graph = GraphRunner[Target.npu](
          model,
          lambda input_tensor: input_tensor,
          lambda input_buffer, input_tensor: (input_buffer.copy_(input_tensor), True),
      )
      device_graph.capture(torch.empty_like(input_tensor).uniform_(1, 2))
  
      assert torch.allclose(device_graph(input_tensor), golden)

Bug修复与其他改动

• 修复LayerNorm相关pass对不带weight和bias的场景的分支处理。
• 修复CustomBatchMatmul调用pass对bias非连续场景的支持。

What's Changed

Exciting New Features 🎉

• feat(xpu_graph): add device graph. by @LeoLau94 in #478

Other Changes

• feat(tensorify_scalars): tensorify python scalars after dispatch by @yfchen-byted in #475
• fix(mlu_compiler/artifact): fix returned artifact of mlu_compile by @yfchen-byted in #474
• fix(combo_bmm): add dtype restrict, fix shape bug and adjust matching requirements by @jyjyjyjyjyjyjyj in #468
• fix: remove --pid=host flag to improve container isolation by @YYYYimo in #487
• chore(pre-commit): replace black and isort with ruff by @YYYYimo in #489
• fix(device_graph_runner): ensure stream sync after graph replay by @YYYYimo in #491
• fix(mlu): fix layernorm and batchmatmul by @jyjyjyjyjyjyjyj in #492

New Contributors

• @YYYYimo made their first contribution in #483

Full Changelog: 0.10.1...0.11.0