Frank Seide
4c8c1b820f
Merged PR 14437: minor fixes to make VS happy after latest changes
...
minor fixes to make VS happy after latest changes
2020-07-29 18:00:59 +00:00
Young Jin Kim
ae7cae2760
Merged PR 13049: Remove repeated memory allocation for fbgemm temp scratch space
...
Repeated memory allocation/deallocation in a multi-threaded environment caused heap contention problem which all the threads are waiting each other to synchronously allocate/deallocate memory. To resolve this, several unnecessary memory allocations are removed.
1. Remove repeated memory allocation and deallocation for the fbgemm's scratch space.
2. Remove unnecessary memory allocation in fbgemm submodule (https://github.com/marian-nmt/FBGEMM/pull/3/files )
3. Make "USE_ONNX" false as a default option in vsproj windows build
4. Improve the variable naming in int8 fbgemm code for a better readability
2020-05-26 20:22:38 +00:00
Martin Junczys-Dowmunt
0dc318e993
Merged PR 13052: Fix FAISS on Windows
...
* adds the correct definitions for FINTEGER for Linux and Windows
* moves things around a bit in the CMakeLists.txt files to keep things more local
* fixes a few more warnings in 3rd party code
2020-05-24 02:22:13 +00:00
Frank Seide
f9029638ff
Merged PR 13039: FAISS builds with Windows
...
FAISS builds with Windows
2020-05-22 02:49:04 +00:00
Frank Seide
77a420740c
Merged PR 12958: ONNX support
...
This branch adds functionality to export ONNX models (with limitations).
2020-05-21 05:51:18 +00:00
Frank Seide
1d2a137e14
Merged PR 12957: fixes to VS Solution
...
fixes to VS Solution
2020-05-16 20:46:23 +00:00
Young Jin Kim
d06018d98c
Merged PR 12505: Fix windows compile errors
...
Fix windows compile errors
- Add new files to the vs project files
- Fix a warning
- Add namespace specifiers for ambiguous types
2020-04-17 08:38:37 +00:00
Young Jin Kim
d2b4f3803e
Merged PR 11831: Change the weight matrix quantization to use 7-bit min/max quantization to avoid overflow
...
1. Change the weight matrix quantization to use 7-bit min/max quantization
-> This resolves all the overflow issue, because weight and activations are quantized by min/max range.
2. Clip fp16 quantization to avoid overflow
3. Fix windows build errors (cmake options, vcproj file)
4. int8 pack model (encoder -> fp16)
2020-03-25 02:52:17 +00:00
Young Jin Kim
9c9a240354
Merged PR 10266: FBGEMM based Int8 model
...
FBGEMM based Int8 model - working with the master
1. Added int8 implementation into packed_gemm.h/cpp with FBGEMM
2. Update FBGEMM library to make it work on windows
3. Split 'ispacked' into packed8 and packed16
4. Change all names for PackFp32 to PackFp16 which is more accurate
2019-12-03 19:14:18 +00:00
Young Jin Kim
5fb31b28d2
Merged PR 10415: Fix windows build errors
...
1. Added template definition for 'uint64_t', Windows has different definition of 'long' type.
2. Fixed warnings on windows.
2019-11-13 07:37:46 +00:00
Young Jin Kim
5050aad8db
Fix vs project warnings
2019-11-01 11:03:56 -07:00
Young Jin Kim
5ed441f843
Merged PR 9284: Packed model support in production
...
This PR enables loading and serving of offline packed models.
* Enable support for FBGEMM offline-packed models
* Added different parameters sets for different parameter element types (currently inference only)
* New types (packed*) are used for the packed memory format.
* Added CPU instruction query into the quicksand API.
Related work items: #102572
2019-11-01 17:24:49 +00:00
Hieu Hoang
bd0118cc66
const NormalizeTempPrefix(). Retain all streamBuf so we can delete them afterwards
2019-10-25 07:30:49 -07:00
Hieu Hoang
9d31fb58dd
merge with previous pull request
2019-10-24 14:57:45 -07:00
Hieu Hoang
c5c1a4c449
replace all OutputFileStream with OutputFileStreamNew
2019-10-22 17:50:32 -07:00
Hieu Hoang
c33074ce54
move code .h -> .cpp
2019-10-21 12:17:35 -07:00
Hieu Hoang
fef5c301ca
figure out where temp file class is used
2019-10-21 10:48:14 -07:00
Frank Seide
2e4f04e308
downgraded to AVX build, since AVX2 builds do not run on older machines even if no AVX2 math is used
2019-09-13 17:26:35 -07:00
Frank Seide
c5eeca2490
removed custom stdint.h, as it is not necessary
2019-09-12 10:09:59 -07:00
Frank Seide
eea9473786
Visual Studio Project builds again
2019-09-12 10:08:12 -07:00
Frank Seide
e40491f44a
sorted out AVX2 build problems on Windows
2019-09-11 14:38:22 -07:00
Young Jin Kim
c0a0f4add9
Fix windows build error
2019-09-04 15:01:05 -07:00
Young Jin Kim
4c41fec0ae
QuickSAND env fix, Remove binary dependecy of FBGEMM and compile FBGEMM together, Update FBGEMM to the latest.
2019-06-21 09:59:52 -07:00
Young Jin Kim
56092425ed
Add missing FBGEMM lib files
2019-06-19 10:25:42 -07:00
Young Jin Kim
d2d8ec041d
Enable FBGEMM based packed GEMM on windows
2019-06-18 16:59:11 -07:00
Frank Seide
75034f1a20
Now builds with VS 2017
2019-06-03 10:23:53 -07:00
Frank Seide
865f863485
VS Project now builds with CUDA
2019-05-31 15:33:51 -07:00
Frank Seide
356f804293
minor bug fixes, to get last few changes to pass tests
2019-05-15 11:43:20 -07:00
Frank Seide
eacfd2de11
weird mode change back
2019-04-29 19:01:29 -07:00
Frank Seide
cd329be400
weird mode change
2019-04-29 17:52:23 -07:00
Frank Seide
b570638533
weirdo change of access permissions
2019-04-26 14:59:40 -07:00
Frank Seide
4351229899
(updated the VS project)
2019-04-12 09:41:11 -07:00
Frank Seide
e7dbe93691
merged with fseide/criterionfunction
2019-02-22 15:23:02 -08:00
Frank Seide
23ece0040a
ModelBase::build() now returns an Expr, not a RationalLoss
2019-02-22 13:21:45 -08:00
Frank Seide
b7d245945f
added factored_vocab stubs
2019-02-05 11:45:55 -08:00
Frank Seide
de69efea79
fix after merge
2019-02-04 17:32:53 -08:00
Frank Seide
1b9dad403d
renamed VocabBase to IVocab, since it is a pure interface
2019-02-04 17:09:34 -08:00
Frank Seide
bcbe247a57
auto-merged pieces of last merge
2019-02-04 12:45:25 -08:00
Frank Seide
f9e63da391
removed auto-conversion between Word and WordIndex, and dealt with fallout
2019-01-31 22:24:26 -08:00
Frank Seide
acee183116
towards fixing the bug that bert.h includes transformer.h
2019-01-31 14:55:45 -08:00
Frank Seide
7b649c00e0
In VS Project, changed transformer.h from CPP file to a true header, and added the stub, to fix a compilation warning
2019-01-31 14:45:51 -08:00
Frank Seide
93763c30e6
now always normalizes factors
2019-01-28 10:02:53 -08:00
Marcin Junczys-Dowmunt
50d64de62c
Merge pull request #415 from marian-nmt/chmod-x
...
C++ source should not have execute bits
2019-01-25 08:52:30 -08:00
Kenneth Heafield
f100a8649a
C++ source is should not have execute bits
2019-01-23 23:20:46 +00:00
Frank Seide
2e331b254d
merged dot_csr() change
2019-01-15 16:45:49 -08:00
Frank Seide
31f4f721b4
scaffolding for new operation dot_csr(); renamed prod.cu to prod.cpp, since it no longer has CUDA code
2019-01-15 13:25:36 -08:00
Frank Seide
c5839f07e0
towards factored embeddings; towards csr_dot() operation
2019-01-10 21:14:55 -08:00
Frank Seide
664e29ff6e
removed operator-> for Factories
2019-01-08 17:59:05 -08:00
Frank Seide
9023bf8b90
fixed build break due to MPI change in examples
2019-01-08 17:15:39 -08:00
Frank Seide
f6de599980
generalized step() to narrow() and sliceView(), new class Slice;
...
bug fix: SliceViewNodeOp should use correct size for memory piece;
new operation stopGradient()
2018-12-26 16:35:56 -08:00