hip : substituted bpermute ops with swizzle ops (gfx906, maybe all AMD)

SCRIPT_compile_MI50.sh

@@ -0,0 +1,202 @@
+#!/bin/bash
+#
+# SCRIPT MI50 Compilation Script for llama.cpp
+# Optimized build for AMD MI50 (gfx906) with ROCm/HIP support
+# 
+# This script compiles llama.cpp with maximum optimizations for the MI50 GPU
+# including server support, flash attention, and all performance features
+#
+
+set -e  # Exit on any error
+
+# Colors for output
+RED='\033[0;31m'
+GREEN='\033[0;32m'
+YELLOW='\033[1;33m'
+BLUE='\033[0;34m'
+NC='\033[0m' # No Color
+
+echo -e "${BLUE}======================================${NC}"
+echo -e "${BLUE} SCRIPT MI50 llama.cpp Builder  ${NC}"
+echo -e "${BLUE}======================================${NC}"
+
+# Check if we're in the right directory
+if [[ ! -f "CMakeLists.txt" ]]; then
+    echo -e "${RED}Error: Not in llama.cpp root directory${NC}"
+    echo "Please run this script from the llama.cpp root directory"
+    exit 1
+fi
+
+# Verify ROCm installation
+echo -e "${YELLOW}Checking ROCm installation...${NC}"
+if ! command -v rocm_agent_enumerator &> /dev/null; then
+    echo -e "${RED}Error: ROCm not found. Please install ROCm first.${NC}"
+    exit 1
+fi
+
+# Check for gfx906 support
+GPUS=$(rocm_agent_enumerator)
+if [[ ! "$GPUS" =~ "gfx906" ]]; then
+    echo -e "${RED}Warning: gfx906 (MI50) not detected in system${NC}"
+    echo "Available GPUs: $GPUS"
+    read -p "Continue anyway? (y/N): " -n 1 -r
+    echo
+    if [[ ! $REPLY =~ ^[Yy]$ ]]; then
+        exit 1
+    fi
+fi
+
+echo -e "${GREEN}✓ ROCm installation verified${NC}"
+echo -e "${GREEN}✓ Available GPUs: $GPUS${NC}"
+
+# Set ROCm environment variables for optimal gfx906 compilation
+echo -e "${YELLOW}Setting ROCm environment variables for gfx906...${NC}"
+export ROCM_PATH=${ROCM_PATH:-/opt/rocm}
+export HCC_AMDGPU_TARGET=gfx906
+export HSA_OVERRIDE_GFX_VERSION=9.0.6
+export AMDGPU_TARGETS=gfx906
+export GPU_TARGETS=gfx906
+
+# Clean previous build
+echo -e "${YELLOW}Cleaning previous build...${NC}"
+rm -rf build
+mkdir -p build
+
+# Configure with maximum optimizations
+echo -e "${YELLOW}Configuring CMake with MI50 optimizations...${NC}"
+cd build
+
+cmake .. \
+    -DCMAKE_BUILD_TYPE=Release \
+    -DCMAKE_C_COMPILER=gcc \
+    -DCMAKE_CXX_COMPILER=g++ \
+    -DCMAKE_HIP_COMPILER_FORCED=1 \
+    -DCMAKE_HIP_ARCHITECTURES=gfx906 \
+    -DCMAKE_C_FLAGS="-O3 -march=native -mtune=native -DNDEBUG -ffast-math -fno-finite-math-only -ffp-contract=fast" \
+    -DCMAKE_CXX_FLAGS="-O3 -march=native -mtune=native -DNDEBUG -DGGML_HIP_GFX906_OPTIMIZED -ffast-math -fno-finite-math-only -ffp-contract=fast" \
+    -DCMAKE_HIP_FLAGS=" --offload-arch=gfx906 -DGGML_HIP_GFX906_OPTIMIZED -Wno-ignored-attributes -Wno-cuda-compat -Wno-unused-result -mllvm -amdgpu-simplify-libcall -mllvm -amdgpu-internalize-symbols -mllvm -amdgpu-enable-lower-module-lds=false -mllvm -amdgpu-early-inline-all=true -mllvm -amdgpu-function-calls=false -ffast-math -ffp-contract=fast" \
+    -DGGML_HIP=ON \
+    -DGGML_HIP_MMQ_MFMA=ON \
+    -DGGML_HIP_GRAPHS=ON \
+    -DGGML_HIP_NO_VMM=ON \
+    -DGGML_HIP_EXPORT_METRICS=ON \
+    -DGGML_HIP_GFX906_OPTIMIZED=ON \
+    -DGGML_NATIVE=ON \
+    -DGGML_CUDA_FA=ON \
+    -DGGML_CUDA_FA_ALL_QUANTS=ON \
+    -DGGML_CUDA_FORCE_MMQ=OFF \
+    -DGGML_CUDA_FORCE_CUBLAS=OFF \
+    -DGGML_CUDA_NO_PEER_COPY=ON \
+    -DLLAMA_BUILD_SERVER=ON \
+    -DLLAMA_BUILD_EXAMPLES=ON \
+    -DLLAMA_BUILD_TOOLS=ON \
+    -DLLAMA_BUILD_TESTS=OFF \
+    -DLLAMA_CURL=ON \
+    -DLLAMA_STATIC=OFF
+
+if [[ $? -ne 0 ]]; then
+    echo -e "${RED}✗ CMake configuration failed${NC}"
+    exit 1
+fi
+
+echo -e "${GREEN}✓ CMake configuration successful${NC}"
+
+# Compile with all CPU cores and dump detailed logs
+NPROC=$(nproc)
+LOG_FILE="compilation_log.txt"
+echo -e "${YELLOW}Compiling with $NPROC cores...${NC}"
+echo -e "${YELLOW}This may take several minutes...${NC}"
+echo -e "${YELLOW}Detailed compilation log will be saved to: $LOG_FILE${NC}"
+
+# Clear previous log
+> $LOG_FILE
+
+# Run make with detailed output and save to log file
+make -j$NPROC 2>&1 | tee $LOG_FILE
+
+if [[ ${PIPESTATUS[0]} -ne 0 ]]; then
+    echo -e "${RED}✗ Compilation failed${NC}"
+    echo -e "${RED}Check $LOG_FILE for detailed error information${NC}"
+    exit 1
+fi
+
+echo -e "${GREEN}✓ Compilation successful!${NC}"
+
+# Verify the build
+echo -e "${YELLOW}Verifying build...${NC}"
+
+# Check if main executables were built
+EXECUTABLES=(
+    "bin/llama-cli"
+    "bin/llama-server"
+    "bin/llama-bench"
+    "bin/libggml-hip.so"
+)
+
+ALL_GOOD=true
+for exec in "${EXECUTABLES[@]}"; do
+    if [[ -f "$exec" ]]; then
+        echo -e "${GREEN}✓ $exec built successfully${NC}"
+
+        # Check HIP linking for executables (not libraries)
+        if [[ "$exec" =~ ^bin/llama- && ! "$exec" =~ \.so$ ]]; then
+            if ldd "$exec" | grep -q "libggml-hip.so"; then
+                echo -e "${GREEN}  ✓ HIP backend linked${NC}"
+            else
+                echo -e "${RED}  ✗ HIP backend not linked${NC}"
+                ALL_GOOD=false
+            fi
+        fi
+    else
+        echo -e "${RED}✗ $exec not found${NC}"
+        ALL_GOOD=false
+    fi
+done
+
+if [[ "$ALL_GOOD" = false ]]; then
+    echo -e "${RED}✗ Build verification failed${NC}"
+    exit 1
+fi
+
+# Display ROCm libraries linked
+echo -e "${YELLOW}ROCm libraries linked:${NC}"
+ldd bin/llama-cli | grep -E "(hip|roc)" | head -5
+
+# Quick functionality test
+echo -e "${YELLOW}Testing HIP backend availability...${NC}"
+if ./bin/llama-cli --help 2>/dev/null | grep -q "backend"; then
+    echo -e "${GREEN}✓ llama-cli responding correctly${NC}"
+else
+    echo -e "${RED}✗ llama-cli test failed${NC}"
+fi
+
+# Success message
+echo
+echo -e "${GREEN}======================================${NC}"
+echo -e "${GREEN}    ✓ BUILD COMPLETED SUCCESSFULLY    ${NC}"
+echo -e "${GREEN}======================================${NC}"
+echo
+echo -e "${BLUE}Built executables:${NC}"
+echo "  • CLI:    ./build/bin/llama-cli"
+echo "  • Server: ./build/bin/llama-server" 
+echo "  • Bench:  ./build/bin/llama-bench"
+echo
+echo -e "${BLUE}Optimizations enabled:${NC}"
+echo "  • Target GPU: AMD MI50 (gfx906)"
+echo "  • HIP/ROCm backend with MFMA support"
+echo "  • Flash Attention kernels"
+echo "  • All quantization formats"
+echo "  • Performance metrics export"
+echo "  • Native CPU optimizations"
+echo "  • Optimization 5: GFX906 compiler flags (-ffast-math, early-inline, function-calls=false)"
+echo
+echo -e "${BLUE}Ready to run:${NC}"
+echo "  ./SCRIPT_launch_server_MI50.sh <model.gguf>"
+echo
+echo -e "${YELLOW}Note: Make sure to set proper ROCm environment variables before running!${NC}"
+echo
+echo -e "${BLUE}For debugging with maximum HIP logging:${NC}"
+echo "  export AMD_LOG_LEVEL=8"
+echo "  export AMD_LOG_MASK=0xFFFFFFFF" 
+echo "  export AMD_SERIALIZE_KERNEL=3"
+echo "  ./SCRIPT_launch_server_MI50.sh <model.gguf> 2>&1 | tee hip_debug.log"

SCRIPT_launch_server_MI50.sh

@@ -0,0 +1,61 @@
+#!/bin/bash
+#
+# Launch llama.cpp server with AMD MI50 ROCm support
+# Built for gfx906 architecture
+#
+
+# Set ROCm environment variables for MI50 ONLY (optimal configuration)
+export HSA_OVERRIDE_GFX_VERSION=9.0.6
+export HIP_VISIBLE_DEVICES=0           # ONLY MI50 (Device 0)
+export CUDA_VISIBLE_DEVICES=0          # Additional CUDA compatibility
+export ROCR_VISIBLE_DEVICES=0          # ROCr runtime device selection
+export GGML_BACKEND_HIP=1
+export HCC_AMDGPU_TARGET=gfx906
+
+# Path to your model file - update this to your actual model path
+ MODEL_PATH="/home/iacopo/Downloads/Qwen3-30B-A3B-Thinking-2507-Q4_0.gguf"
+
+PARAMS=(
+    -m "$MODEL_PATH"
+    -ngl 99                    # Offload all layers to GPU
+    -c 32000                    # Context size
+    -np 1                      # Parallel requests
+    -t $(nproc)                # Use all CPU threads
+    --port 8090                # Server port
+    --host 0.0.0.0            # Listen on all interfaces
+    #--mlock                    # Lock model in memory
+    #--no-mmap                  # Don't use memory mapping
+    -b 512                       # Batch size
+    #--cont-batching            # Enable continuous batching
+    --flash-attn on              # Enable flash attention
+    --cache-type-k q8_0        # q8_0 quantized K cache (50% memory savings)
+    --cache-type-v q8_0        # q8_0 quantized V cache (50% memory savings)
+    --main-gpu 0               # Force MI50 as main GPU
+    --device "ROCm0"           # Explicit ROCm device
+    # --no-warmup                # Skip warmup for consistent profiling
+)
+
+# Check if model file exists
+if [ ! -f "$MODEL_PATH" ]; then
+    echo "Error: Model file not found at: $MODEL_PATH"
+    echo "Usage: $0 [model_path] [additional_args...]"
+    echo ""
+    echo "Example: $0 ./models/llama-2-7b-chat.q4_0.gguf --ctx-size 8192"
+    exit 1
+fi
+
+# Display GPU info
+echo "=== ROCm GPU Information ==="
+rocm-smi --showproductname --showtemp --showmeminfo --showuse --showpower
+echo ""
+
+# Launch llama.cpp server
+echo "=== Launching llama.cpp server with MI50 optimization ==="
+echo "Model: $MODEL_PATH"
+echo "GPU: MI50 (gfx906)"
+echo "Server will be available at: http://localhost:8080"
+echo "Parameters: ${PARAMS[*]} ${@:2}"
+echo ""
+
+cd "$(dirname "$0")"
+./build/bin/llama-server "${PARAMS[@]}" "${@:2}"