nihui
|
f7652ec72d
|
pnnx fuse chinese clip mha and write need_weights=False (#4745)
|
3 years ago |
nihui
|
1d6bfdca38
|
fix pnnx pass on fp16 weight, common fp16 conversion routines (#4743)
|
3 years ago |
nihui
|
f893d2440d
|
innerproduct allow 1 height gemm (#4730)
|
3 years ago |
nihui
|
d4046b4ae9
|
pnnx fuse transformer clip attention and diffusers attentionblock (#4727)
* pnnx fuse transformer clip attention
* skip fuse mha for 1.8
* select one method other than forward
* pnnx fuse diffusers attentionblock
|
3 years ago |
nihui
|
c038b8227b
|
pnnx convert sdpa (#4722)
* pnnx convert sdpa
* pnnx fuse diffuser attention2
|
3 years ago |
Yoh
|
2d9ec410f4
|
fix pnnx build bug (#4721)
|
3 years ago |
nihui
|
1fa38fe5ac
|
pnnx convert torch std (#4715)
* pnnx convert torch std
* fix multiple fuse pass on torch 2.0
* fuse vit pytorch mha pattern
|
3 years ago |
nihui
|
05ad0c52c6
|
pnnx fuse gelu (#4702)
|
3 years ago |
nihui
|
a37a83d850
|
clip gelu mish tanh 4d (#4695)
|
3 years ago |
nihui
|
cd5a6098a2
|
sigmoid and swish 4d (#4692)
|
3 years ago |
Hillwood Yang
|
d878f0e16d
|
Fix build by gcc 13 (#4684)
|
3 years ago |
nihui
|
c28c8c04a1
|
multiheadattention attn mask (#4668)
|
3 years ago |
Zhuo Zhang
|
737a114242
|
fix typo in PNNX readme (#4670)
|
3 years ago |
Zhuo Zhang
|
c07ff1ce4f
|
fix typo in PNNX groupnorm error message (#4669)
|
3 years ago |
nihui
|
b640574b88
|
rough vulkan gemm and multiheadattention (#4618)
|
3 years ago |
Yukie
|
3a64083756
|
Fix spelling errors, ncn -> ncnn. (#4665)
|
3 years ago |
Chen Xin
|
9fe843fa39
|
fix wrong param id of Squeeze for modelwriter (#4658)
|
3 years ago |
nihui
|
f7af84f001
|
pnnx reset maxpool return_indices if only output found, drop convtransposed output_size arg for static output padding (#4654)
|
3 years ago |
He Yang
|
9a5b6ce20e
|
fix fprintf type mismatch. (#4647)
|
3 years ago |
nihui
|
91cc508a23
|
pnnx recognize foldable scalar and fuse it into expression (#4639)
|
3 years ago |
nihui
|
5b5e9ea537
|
fix some pnnx build warnings (#4634)
|
3 years ago |
nihui
|
f8e32aba9c
|
fix pnnx gru rnn with optional output, fix #4608 (#4631)
|
3 years ago |
nihui
|
8066c76bc5
|
pnnx complex data type and torch.stft family (#4627)
|
3 years ago |
nihui
|
8049623d31
|
pnnx convert torch.mm (#4589)
|
3 years ago |
nihui
|
693535afc1
|
pnnx torch 2.0 (#4579)
* fix build with torch-2.0
* torch 2.0 new patterns
* add torch 2.0 ci
|
3 years ago |
nihui
|
ae4f630467
|
pnnx fuse multiheadattention (#4544)
* torch baddbmm
* always convert to fp32 for shape inference
* silence info on nonetype and devicetype
|
3 years ago |
nihui
|
c68266efd0
|
pnnx eliminate reshape shape expression for only one dynamic dimsize (#4548)
|
3 years ago |
nihui
|
afc9310c62
|
update new operators for modelwriter (#4540)
|
3 years ago |
nihui
|
4fc977d7fd
|
fold scalar constant, write scalar attribute correctly (#4533)
|
3 years ago |
nihui
|
1832da8292
|
concat 4d (#4528)
|
3 years ago |
nihui
|
32e2de015e
|
slice 4d (#4525)
|
3 years ago |
nihui
|
fc6ce4a641
|
copyto operator (#4522)
|
3 years ago |
nihui
|
242e775d21
|
pnnx convert torch log10, pow 2 as square (#4518)
|
3 years ago |
nihui
|
246e71c526
|
implement atan2 (#4516)
|
3 years ago |
Fangjun Kuang
|
92e75105c9
|
Support torch.cumsum (#4505)
|
3 years ago |
nihui
|
ab4cfbf5b0
|
enrich ncnn binary broadcast rules (#4513)
|
3 years ago |
Yoh
|
529c83b8a5
|
Fuse pixel unshuffle (#4504)
|
3 years ago |
nihui
|
ce6b80a16b
|
pnnx flatten input tuple list (#4498)
|
3 years ago |
向阳
|
6af0f3777f
|
Remove redundant include (#4497)
|
3 years ago |
li mengyang
|
ff4d05a713
|
fix typo (#4482)
|
3 years ago |
Yoh
|
0fffefd3c9
|
fix crop bug and eliminate Tensor.clone (#4416)
* fix crop bug and eliminate Tensor.clone
* fix crop slice bug on msvc
* fix slice bug on msvc
|
3 years ago |
inisis
|
62fc16d157
|
pnnx readme remove duplicate space (#4428)
|
3 years ago |
nihui
|
aed05aa851
|
pnnx fuse more function to module (#4351)
* pnnx fuse more function to module
* rename some pass name
* fuse adjacent reshape, fuse pad conv2d
* fuse pad conv1d
|
3 years ago |
nihui
|
ec1b07c9fe
|
pnnx fp16 option for ncnn and onnx weight type (#4350)
|
3 years ago |
nihui
|
6967baaccc
|
pnnx convert torch bitwise left_shift right_shift (#4349)
|
3 years ago |
nihui
|
eceac35a7f
|
implement MultiheadAttention kdim vdim (#4347)
|
3 years ago |
nihui
|
498ca7341b
|
squeeze and expanddims 4d (#4346)
|
3 years ago |
Lry89757
|
6a47f8d15c
|
gridsample op support (#4288)
Co-authored-by: LRY89757 <LRY89757@users.noreply.github.com>
Co-authored-by: nihuini <nihuini@tencent.com>
Co-authored-by: nihui <shuizhuyuanluo@126.com>
|
3 years ago |
nihui
|
a2af6369d9
|
match inplace slice copy pattern, rewrite copy uses (#4338)
|
3 years ago |
nihui
|
a7e3c62a1b
|
save foldable constants in file for reducing memory usage (#4337)
|
3 years ago |