Add missing hipBLASLt library folder and fix some building instructions (#22)

* Updating llama.cpp is not needed anymore. Please check: https://github.com/ggml-org/llama.cpp/blob/master/ggml/src/ggml-cuda/vendors/hip.h#L145

* Fix gfx targets, RDNA and commercial names

* Add missing hipblaslt library folder

* Update HIP version check is not needed anymore. Please check https://github.com/ggml-org/llama.cpp/blob/master/ggml/src/ggml-cuda/vendors/hip.h#L145

* Add missing hipblaslt library folder

* Another fix for RDNA3/4 commercial names

* GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12 flag is not needed any more due to llama.cpp rocWMMA updates: ggml-org/llama.cpp#16221

* Delete GGML_HIP_FORCE_ROCWMMA_FATTN_GFX12 flag

* Add gfx1101 and gfx1102 targets

---------

Co-authored-by: Daniel Holanda <[email protected]>

Author: slojosic-amd

.github/workflows/build-llamacpp-rocm.yml

@@ -266,19 +266,6 @@ jobs:
         # Show current commit info
         Write-Host "Current llama.cpp commit:"
         git log --oneline -1
-        
-    - name: Update HIP version check
-      run: |
-        cd llama.cpp
-        $hipFile = "ggml\src\ggml-cuda\vendors\hip.h"
-        if (Test-Path $hipFile) {
-          $content = Get-Content $hipFile -Raw
-          $updatedContent = $content -replace 'HIP_VERSION >= 70000000', 'HIP_VERSION >= 50600000'
-          Set-Content $hipFile $updatedContent -NoNewline
-          Write-Host "Updated HIP version check in $hipFile"
-        } else {
-          Write-Host "Warning: $hipFile not found"
-        }
 
     # - name: Apply rocWMMA patch (non-interactive)
     #   shell: bash
@@ -296,7 +283,7 @@ jobs:
         echo Input target: %current_target%
         
         if "%current_target%"=="gfx110X" (
-          set "mapped_target=gfx1100"
+          set "mapped_target=gfx1100;gfx1101;gfx1102"
         ) else if "%current_target%"=="gfx1151" (
           set "mapped_target=gfx1151"
         ) else if "%current_target%"=="gfx120X" (
@@ -334,7 +321,6 @@ jobs:
           -DGGML_OPENMP=OFF ^
           -DGGML_CUDA_FORCE_CUBLAS=OFF ^
           -DGGML_HIP_ROCWMMA_FATTN=OFF ^
-          -DGGML_HIP_FORCE_ROCWMMA_FATTN_GFX12=OFF ^
           -DLLAMA_CURL=OFF ^
           -DGGML_NATIVE=OFF ^
           -DGGML_STATIC=OFF ^
@@ -393,6 +379,21 @@ jobs:
           } else {
             Write-Host "Warning: rocblas\library folder not found at: $rocblasLibPath"
           }
+
+          # Copy the hipblaslt\library folder and all its contents
+          $hipblasltLibPath = Join-Path $rocmBinPath "hipblaslt\library"
+          if (Test-Path $hipblasltLibPath) {
+            Write-Host "Copying hipblaslt\library folder and all contents..."
+            $destHipblasltPath = Join-Path $buildBinPath "hipblaslt\library"
+            Copy-Item -Path $hipblasltLibPath -Destination $destHipblasltPath -Recurse -Force
+            Write-Host "Copied: hipblaslt\library folder with all contents"
+
+            # List the contents of the copied hipblaslt\library folder
+            Write-Host "Contents of hipblaslt\library:"
+            Get-ChildItem $destHipblasltPath -Recurse | Select-Object FullName, Length | Format-Table -AutoSize
+          } else {
+            Write-Host "Warning: rocblas\library folder not found at: $rocblasLibPath"
+          }
           
           Write-Host "ROCm core files successfully copied to build directory"
         } else {
@@ -654,17 +655,6 @@ jobs:
         # Show current commit info
         echo "Current llama.cpp commit:"
         git log --oneline -1
-        
-    - name: Update HIP version check
-      run: |
-        cd llama.cpp
-        hip_file="ggml/src/ggml-cuda/vendors/hip.h"
-        if [ -f "$hip_file" ]; then
-          sed -i 's/HIP_VERSION >= 70000000/HIP_VERSION >= 50600000/g' "$hip_file"
-          echo "Updated HIP version check in $hip_file"
-        else
-          echo "Warning: $hip_file not found"
-        fi
 
     #- name: Apply rocWMMA patch (non-interactive)
     #  run: |
@@ -679,11 +669,11 @@ jobs:
         echo "Input target: $current_target"
         
         if [ "$current_target" = "gfx110X" ]; then
-          mapped_target="gfx1100"
+          mapped_target="gfx1100;gfx1101;gfx1102"
         elif [ "$current_target" = "gfx1151" ]; then
           mapped_target="gfx1151"
         elif [ "$current_target" = "gfx120X" ]; then
-          mapped_target="gfx1200,gfx1201"
+          mapped_target="gfx1200;gfx1201"
         else
           mapped_target="$current_target"
         fi
@@ -707,7 +697,6 @@ jobs:
           -DGGML_OPENMP=OFF \
           -DGGML_CUDA_FORCE_CUBLAS=OFF \
           -DGGML_HIP_ROCWMMA_FATTN=OFF \
-          -DGGML_HIP_FORCE_ROCWMMA_FATTN_GFX12=OFF \
           -DLLAMA_CURL=OFF \
           -DGGML_NATIVE=OFF \
           -DGGML_STATIC=OFF \
@@ -736,6 +725,22 @@ jobs:
         else
           echo "Warning: rocblas/library folder not found at: $rocblas_lib_path"
         fi
+
+        # Copy the hipblaslt/library folder and all its contents
+        hipblaslt_lib_path="/opt/rocm/lib/hipblaslt/library"
+        if [ -d "$hipblaslt_lib_path" ]; then
+          echo "Copying hipblaslt/library folder and all contents..."
+          dest_hipblaslt_path="$build_bin_path/hipblaslt/library"
+          mkdir -p "$(dirname "$dest_hipblaslt_path")"
+          cp -r "$hipblaslt_lib_path" "$(dirname "$dest_hipblaslt_path")/"
+          echo "Copied: hipblaslt/library folder with all contents"
+
+          # List the contents of the copied hipblaslt/library folder
+          echo "Contents of hipblaslt/library:"
+          find "$dest_hipblaslt_path" -type f -exec ls -la {} \; | head -20
+        else
+          echo "Warning: hipblaslt/library folder not found at: $hipblaslt_lib_path"
+        fi
  
         # Copy required ROCm libraries to build directory
         # If artifacts from ROCm or Llama.cpp change, you may need to update this list

README.md

@@ -32,8 +32,8 @@ We provide nightly builds of **llama.cpp** with **AMD ROCm™ 7** acceleration b
 
 This build specifically targets the following GPU architectures:
 - **gfx1151** (STX Halo GPUs) - Ryzen AI MAX+ Pro 395
-- **gfx120X** (RDNA4 GPUs) - includes AMD Radeon AI PRO R9700, RX 9070 XT/GRE/9070, RX 9060 XT
-- **gfx110X** (RDNA3 GPUs) - includes AMD Radeon PRO W7900/W7800/W7700/V710, RX 7900 XTX/XT/GRE, RX 7800 XT, RX 7700 XT
+- **gfx120X** (RDNA4 GPUs) - includes AMD Radeon RX 9070 XT/GRE/9070, RX 9060 XT/9060
+- **gfx110X** (RDNA3 GPUs) - includes AMD Radeon PRO W7900/W7800/W7700/W7600, RX 7900 XTX/XT/GRE, RX 7800 XT, RX 7700 XT/7700, RX 7600 XT/7600
 
 **All builds include ROCm™ 7 built-in** - no separate ROCm™ installation required!
 

docs/manual_instructions.md

@@ -42,11 +42,7 @@ I used chocolatey, but you can also install those manually.
 * Setp 6: Add `C:\opt\rocm\lib\llvm\bin` to path
 * Step 7: clone llamacpp
 
-### Part 3: Updating llama.cpp
-
-Open `C:\<YOUR_LLAMACPP_PATH>\ggml\src\ggml-cuda\vendors\hip.h` and replace `HIP_VERSION >= 70000000` with `HIP_VERSION >= 50600000`
-
-### Part 4: Building Llama.cpp + ROCm
+### Part 3: Building Llama.cpp + ROCm
 
 Open `x64 Native Tools Command Prompt` and run the following commands:
 
@@ -115,17 +111,7 @@ sudo apt install -y cmake ninja-build git wget
   git clone https://github.com/ggerganov/llama.cpp.git
   ```
 
-### Part 3: Updating llama.cpp
-
-> **Note**: This step is identical to the [Windows Part 3](#part-3-updating-llamacpp) above, with the same file modification.
-
-Navigate to your llama.cpp directory and update the HIP version check:
-```bash
-cd llama.cpp
-sed -i 's/HIP_VERSION >= 70000000/HIP_VERSION >= 50600000/g' ggml/src/ggml-cuda/vendors/hip.h
-```
-
-### Part 4: Building Llama.cpp + ROCm
+### Part 3: Building Llama.cpp + ROCm
 
 Run the following commands to build llama.cpp with ROCm support:
 
@@ -150,7 +136,6 @@ cmake .. -G Ninja \
   -DGGML_OPENMP=OFF \
   -DGGML_CUDA_FORCE_CUBLAS=OFF \
   -DGGML_HIP_ROCWMMA_FATTN=ON \
-  -DGGML_HIP_FORCE_ROCWMMA_FATTN_GFX12=OFF \
   -DLLAMA_CURL=OFF \
   -DGGML_NATIVE=OFF \
   -DGGML_STATIC=OFF \
@@ -162,7 +147,7 @@ cmake --build . -j $(nproc)
 
 > **Note**: Adjust the `-DAMDGPU_TARGETS="gfx1151"` parameter for your specific GPU. See the [GPU Target Reference](#gpu-target-reference) section for details.
 
-### Part 5: Copy required ROCm libraries
+### Part 4: Copy required ROCm libraries
 
 After successful compilation, copy the required ROCm libraries to the build directory:
 
@@ -182,6 +167,10 @@ cp -v /opt/rocm/lib/rocm_sysdeps/lib/*.so* .
 # Copy the rocblas library folder
 mkdir -p rocblas
 cp -r /opt/rocm/lib/rocblas/library rocblas/
+
+# Copy the hipblaslt library folder
+mkdir -p hipblaslt
+cp -r /opt/rocm/lib/hipblaslt/library hipblaslt/
 ```
 
 If you see no errors during the build process, llama.cpp has been successfully compiled and all files are available in your `build/bin` folder.
@@ -192,9 +181,9 @@ If you see no errors during the build process, llama.cpp has been successfully c
 
 When building llama.cpp with ROCm, the `-DAMDGPU_TARGETS` parameter must be set based on your specific GPU architecture. Our automated workflow uses generic targets that get mapped to specific architectures:
 
-- **`gfx120X`** maps to `gfx1200,gfx1201` (RDNA 3 series like RX 7900 XT/XTX)
-- **`gfx110X`** maps to `gfx1100` (RDNA 2 series like RX 6000 series)  
-- **`gfx1151`** remains as `gfx1151` (specific for RX 7600/7700 XT)
+- **`gfx120X`** maps to `gfx1200, gfx1201` (RDNA4 series like: RX 9070 XT/GRE/9070, RX 9060 XT/9060)
+- **`gfx110X`** maps to `gfx1100, gfx1101, gfx1102` (RDNA3 series like: PRO W7900/W7800/W7700/W7600, RX 7900 XTX/XT/GRE, RX 7800 XT, RX 7700 XT/7700, RX 7600 XT/7600)  
+- **`gfx1151`** remains as `gfx1151` (Strix Halo)
 
 For a complete list of GPU targets and their mappings, see the [automated workflow](../.github/workflows/build-llamacpp-rocm.yml).
 
@@ -203,12 +192,12 @@ For a complete list of GPU targets and their mappings, see the [automated workfl
 Replace the `-DAMDGPU_TARGETS="gfx1151"` parameter in your cmake command with the appropriate target for your GPU:
 
 ```bash
-# For RDNA 3 series (RX 7900 XT/XTX)
--DAMDGPU_TARGETS="gfx1200,gfx1201"
+# For RDNA4 series (RX 9070 XT/GRE/9070, RX 9060 XT/9060)
+-DAMDGPU_TARGETS="gfx1200, gfx1201"
 
-# For RDNA 2 series (RX 6000 series) 
+# For RDNA3 series (PRO W7900/W7800/W7700/W7600, RX 7900 XTX/XT/GRE, RX 7800 XT, RX 7700 XT/7700, RX 7600 XT/7600) 
 -DAMDGPU_TARGETS="gfx1100"
 
-# For RX 7600/7700 XT
+# For Strix Halo
 -DAMDGPU_TARGETS="gfx1151"
 ```