v0.6.2.4

Rocblas and library for hip sdk 6.2.4 (for Windows)

Some prebuild with rocm6.2 package also works on 6.2.4 .

ROCBLAS packages based on GPU Arches:

By GPU Arches (Alphabetical Order)

• gfx1010:xnack-; gfx1011:xnack-; gfx1012:xnack- :
  • rocm.gfx1010-xnack-gfx1011-xnack-gfx1012-xnack-.for.hip.sdk.6.2.4.7z
• gfx1031:
  • rocm.gfx1031.for.hip.sdk.6.2.4.littlewu.s.logic.7z build with little wu's tensile logic
• gfx1032:
  • rocm.gfx1032.for.hip.sdk.6.2.4.navi21.logic.7z
• gfx1034; gfx1035; gfx1036:
  • rocm.gfx1034-gfx1035-gfx1036.for.hip.sdk.6.2.4.7z
• gfx1103
  • rocm gfx1103 AMD 780M phoenix V5.0 for hip sdk 6.2.4.7z
• gfx1150
  • rocm.gfx1150.for.hip.skd.6.2.4.7z

Below list the file in Packages .
(library from ROCm official Linux version ,able to work on Windows. work with HIP SDK 6.2.4 default rocblas.dll, not fully test ,you may build the libs by the guide available in wiki )

• rocBLAS-HIP6.2.4.7z:
  For some unsupport cards(gfx803;gfx900;gfx1010;gfx1030;gfx1031;gfx1032;gfx1100;gfx1101;gfx1102) ,the rocmlibs build with rocm6.2.0, should also work with hip 6.2.4 . you may test koboldcpp-rocm build at https://github.com/YellowRoseCx/koboldcpp-rocm/releases/tag/deps-v6.2.0

v0.6.1.2

Rocblas and library for hip sdk 6.1.2 (for Windows)

ROCBLAS packages based on GPU models:

By GPU Model (Alphabetical Order)

• gfx803:（reported not working)

• rocm.gfx803.optic.vega10.logic.hip.sdk.6.1.2.7z

• gfx900-xnack-（reported not working)

  • rocm-gfx900-xnack-rocm6.2.7z

• gfx902: （Possible not working, not confirmed yet)

  • rocm.gfx902.for.hip.sdk.6.1.2.7z

• gfx90c:（reported not working)

  • rocm.gfx90c.for.hip.sdk.6.1.2.7z

• rocm.gfx90c-xnack-.for.hip.sdk.6.1.2.7z

• gfx906:xnack- and gfx1100

  • gfx906xnack-.gfx1100.for.hip.sdk.6.1.2.7z

• gfx1010:xnack-:

  • rocm.gfx1010-xnack-.for.hip.sdk.6.1.2.7z

• gfx1011:

  • rocm.gfx1011.for.hip.sdk.6.1.2.7z

• gfx1012:xnack-:

  • rocm.gfx1012-xnack-.for.hip.sdk.6.1.2.7z

• gfx1031:

  • rocm gfx1031 for hip sdk 6.1.2 optimized with little wu's logic.7z build with little wu's tensile logic
  • rocm gfx1031 for hip sdk 6.1.2.7z

• gfx1032:

  • rocm gfx1032 for hip sdk 6.1.2.7z
  • rocm.gfx1032.for.hip.sdk.6.1.2.optimized.Fremont.Dango.Version.7z shared by Fremont.Dango on SD.Next Discords.

• gfx1034:

  • rocm gfx1034 for hip sdk 6.1.2.7z

• gfx1035:

  • rocm.gfx1035.for.hip.sdk.6.1.2.7z

• gfx1036:

  • rocm gfx1036 for hip sdk 6.1.2.7z

• gfx1103

  • rocm gfx1103 AMD 780M phoenix V4.0 for hip sdk 6.1.2.7z

• gfx1150

  • rocm-gfx1150-890M-rocm6.2.7z （ build with with rocm6.2, can be working for hip6.1.2 ,experimental only.)

Below list the file in Packages .
(library from ROCm official Linux version ,able to work on Windows. work with HIP SDK 6.1.2 default rocblas.dll, not fully test ,you may build the libs by the guide available in wiki )

• rocBLAS-HIP6.1.2.7z:
  • rocBLAS-HIP6.1.2.7z (gfx900,gfx90a-xnack+ , gfx90a-xnack-, gfx906-xnack-,gfx908-xnack-,gfx940,gfx941,gfx942,gfx1010(fallback logic),gfx1012（fallback logic), gfx1030,gfx1100,gfx1101,gfx1102)
  • gfx906xnack-.gfx1100.for.hip.sdk.6.1.2.7z ( gfx906-xnack-,gfx1100)

For some unsupport cards(gfx803;gfx900;gfx1010;gfx1030;gfx1031;gfx1032;gfx1100;gfx1101;gfx1102) ,the rocmlibs build with rocm6.2.0, should also work with hip 6.1.2 . you may test koboldcpp-rocm build at https://github.com/YellowRoseCx/koboldcpp-rocm/releases/tag/deps-v6.2.0

v0.5.7

rocblas 3.1.0 for ROCm 5.7.0 ( for Windows)

Place rocblas.dll under rocblas\library folder (eg, C:\Program Files\AMD\ROCm\5.7\bin) replace the origial one , replace the library with rocblas\library.
ROCBLAS packages based on GPU models:

By GPU Model (Alphabetical Order)

• gfx803:
  • rocblas for gfx803 (override with vega10).7z(fallback logic )
  • rocm.gfx803.optic.test.version.7z (new, report ,this libs fix the memory error in ollama)
• gfx900:xnack-
  • rocm-gfx900-xnack--HIP.SDK.5.7.7z
• gfx902:
  • rocblas for gfx902.7z
• gfx90c:
  • rocblas for gfx90c workable.7z
  • rocblas for gfx90c:xnack- V2.7z
  • rocblas for gfx90c xnack- V2 with building guide.7z
• gfx906:
  • rocblas for gfx906.7z
• gfx940, gfx941, gfx942:
  • rocblas for gfx940_ gfx941_ gfx942.7z
• gfx1010:
  • rocblas for gfx1010:xnack- with building guide.7z
• gfx1012:
  • rocblas for gfx1012:xnack- with building guide.7z
• gfx1031:
  • rocm gfx1031 for hip sdk 5.7 optimized with little wu logic and I8II support.7z
• gfx1032:
  • rocm gfx1032 for hip sdk 5.7.7z
• gfx1035:
  • rocblas for gfx1035 AMD680M rembrandt V2.7z
• gfx1103
  • rocm gfx1103 AMD780M phoenix V3.7z (gfx1103)

Below list the file in Packages .( mostly used fallback logic , no optimized logic ,slow ,please use the sepearte didicted rocblas package first )

• gfx1031, gfx1032, gfx1033, gfx1034, gfx1035, gfx1036:
  • rocBLAS-HIP5.7.1-win.rar (add more gpu for gfx803、gfx900、gfx1010、gfx1031、gfx1032)
  • rocBLAS-HIP5.7.0-V2.7z (add more gpu for gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1034;gfx1035;gfx1036; gfx1100;gfx1102; gfx1103)
• gfx1100, gfx1101, gfx1102, gfx1103:
  • rocBLAS-HIP5.7.0-win.7z (add more gpu forgfx1010;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1103)

Additional Notes:

• Multiple GPUs in a Package: Some packages contain files for multiple GPU models. List them separately
  within each package's entry.
• File Types: Include the file extension (.7z, .zip, .rar) when listing the package names.

v0.1.0-alpha

rocblas4.1.0 for ROCm 6.1.0( available for Linux gfx1103,gfx1100)
Install as normal deb or replace the file available in zip file .