1. 16 11月, 2021 1 次提交
    • Z
      [cherry-pick-2.2.1]fix fused_transformer_encoder_layer bug (#37229) · 36dd295e
      zhangkaihuo 提交于
      修复了fused_transformer_encoder_layer fine-tune过程发现的一些问题:
      
          fused_attention_op添加attn_mask=None的支持:PR
          pre_layer_norm处理问题:PR
          参数处理,计算错误的问题:PR
          add_bias计算错误问题:PR
          添加pure fp16的支持:PR
      36dd295e
  2. 28 10月, 2021 1 次提交
  3. 27 10月, 2021 2 次提交
    • Z
      [cherry-pick]Fused transformer encoder layer and fused feedforward layer #36776 · e1b5b1da
      zhangkaihuo 提交于
      本PR是fused_transformer的layer层代码,包含FusedFeedForward的layer层代码和FusedTransformerEncoderLayer的代码。
      e1b5b1da
    • L
      Add fused attention op backward and python layer. (#36498) (#36752) · 64643d50
      Li Min 提交于
      功能:本PR的目标是提高attention模块的计算性能。
      为了减少框架层对op的调度开销,本PR通过在C++层手动实现attention模块,对外提供attention 大op;
      为了减少防存开销,本PR采取了两种优化方法:
      (1)在q,k,v计算时通过共享输入X,将该处的gemm,transpose和bias add从三次调用减少为一次;
      (2)使用kernel融合优化技术,在不同cuda kernel之间通过寄存器传输数据;
      64643d50
  4. 06 9月, 2021 1 次提交
    • F
      replase pass with error exception (#35367) · 5675042d
      Feng Xing 提交于
      This PR adds error exception in fused transformer python interface.
      The function body are not implemented (will be implemented later).
      Following zhiqiu's comment in previous PR-35206 (merged already), it is better to raise an exception instead of using "pass".
      5675042d
  5. 31 8月, 2021 1 次提交
    • F
      transformer opt python files (#35206) · e2991555
      Feng Xing 提交于
      This PR adds fused transformer python related files. It defines interface of fused transformer.
      
      Fused transformer implements an optimized version of transformer layer (in python/paddle/nn/layer/transformer.py). In this PR, four layers (functions) are defined:
      (1) FusedMultiHeadAttention: multi-head attention layer
      (2) FusedFeedForward: feed forward layer
      (3) FusedTransformerEncoderLayer: transformer encoder layer
      (4) FusedTransformer: transformer layer
      e2991555