Claude/vigilant keller 5iuff1

bench.sh

@@ -0,0 +1,147 @@
+#!/bin/bash
+set -x
+
+CCACHE_FLAGS="-DCMAKE_C_COMPILER_LAUNCHER=ccache -DCMAKE_CXX_COMPILER_LAUNCHER=ccache"
+
+# ── Detectar GPU: AMD (ROCm/HIP) o NVIDIA (CUDA) ────────────────
+detect_gpu_flags() {
+    # Preferir variable de entorno si ya está fijada
+    if [ -n "$GPU_FLAGS" ]; then
+        return
+    fi
+
+    # Detectar AMD: rocm-smi o /dev/kfd presente
+    if command -v rocm-smi &>/dev/null || [ -e /dev/kfd ]; then
+        # Intentar obtener la arquitectura automáticamente
+        local arch=""
+        if command -v rocminfo &>/dev/null; then
+            arch=$(rocminfo 2>/dev/null | awk '/gfx[0-9]/{print $NF; exit}')
+        fi
+        if [ -n "$arch" ]; then
+            GPU_FLAGS="-DGGML_HIP=ON -DCMAKE_HIP_ARCHITECTURES=${arch}"
+        else
+            # Fallback: RDNA3 es la más común hoy en dia; ajustar si hace falta
+            GPU_FLAGS="-DGGML_HIP=ON -DCMAKE_HIP_ARCHITECTURES=gfx1100"
+        fi
+        GPU_VENDOR="AMD"
+    # Detectar NVIDIA: nvidia-smi o /dev/nvidia0 presente
+    elif command -v nvidia-smi &>/dev/null || [ -e /dev/nvidia0 ]; then
+        GPU_FLAGS="-DGGML_CUDA=ON"
+        GPU_VENDOR="NVIDIA"
+    else
+        echo "ADVERTENCIA: No se detectó GPU. Compilando solo CPU."
+        GPU_FLAGS=""
+        GPU_VENDOR="CPU"
+    fi
+
+    echo "=== GPU detectada: ${GPU_VENDOR} | Flags: ${GPU_FLAGS} ==="
+}
+
+build_if_changed() {
+    local branch="$1"
+    local build_dir="$2"
+    local hash_file=".last_build_hash_${build_dir}"
+
+    git checkout "$branch"
+    local current_hash
+    current_hash=$(git rev-parse HEAD)
+
+    if [ -f "$hash_file" ] && [ "$(cat "$hash_file")" = "$current_hash" ]; then
+        echo "=== $branch: sin cambios ($(git log -1 --format='%h %s')), saltando compilación ==="
+    else
+        cmake -B "$build_dir" $GPU_FLAGS $CCACHE_FLAGS
+        cmake --build "$build_dir" --config Release --parallel 6
+        echo "$current_hash" > "$hash_file"
+    fi
+}
+
+detect_gpu_flags
+
+# ── Compilar ambas ramas (solo si hubo cambios) ─────────────────
+build_if_changed feature/triattention-paged build-tri
+build_if_changed feature/supermerge         build-super
+
+# ── Correctitud (una vez por rama) ─────────────────────────────
+./build-tri/bin/test-quantize-fns  > triattention_results.txt     2>&1 || true
+./build-tri/bin/test-backend-ops   > triattention_backend_ops.txt  2>&1 || true
+./build-super/bin/test-quantize-fns > supermerge_results.txt       2>&1 || true
+./build-super/bin/test-backend-ops  > supermerge_backend_ops.txt   2>&1 || true
+
+# ── Benchmarks por modelo ───────────────────────────────────────
+OLLAMA_MANIFESTS_DIR="$HOME/.ollama/models/manifests"
+
+if [ ! -d "$OLLAMA_MANIFESTS_DIR" ]; then
+    echo "No se encontró el directorio de manifiestos de Ollama."
+    exit 1
+fi
+
+find "$OLLAMA_MANIFESTS_DIR" -type f | while read -r manifest; do
+    model_path="${manifest#$OLLAMA_MANIFESTS_DIR/}"
+    model_name=$(echo "$model_path" | tr '/' '_')
+
+    # Buscar el layer de tipo "model" (no license/template/params)
+    digest=$(python3 -c "
+import json, sys
+try:
+    data = json.load(open('$manifest'))
+    for layer in data.get('layers', []):
+        if layer.get('mediaType','').endswith('.model'):
+            print(layer['digest']); break
+except: pass
+" 2>/dev/null)
+
+    if [ -z "$digest" ]; then
+        echo "Sin blob de modelo en $model_name, omitiendo."
+        continue
+    fi
+
+    blob_file="$HOME/.ollama/models/blobs/${digest/:/-}"
+    if [ ! -f "$blob_file" ]; then
+        echo "Blob no encontrado: $blob_file"
+        continue
+    fi
+
+    echo "=== $model_name ==="
+
+    ./build-tri/bin/llama-bench \
+      -m "$blob_file" --cache-type-v turbo3 --cache-type-k q8_0 -r 3 \
+      > "triattention_bench_${model_name}.txt" 2>&1
+
+    ./build-super/bin/llama-bench \
+      -m "$blob_file" --cache-type-v turbo3 --cache-type-k q8_0 -r 3 \
+      > "supermerge_bench_${model_name}.txt" 2>&1
+
+    TRIATTENTION_STATS="${model_name}.triattention"
+    if [ -f "$TRIATTENTION_STATS" ]; then
+        ./build-super/bin/llama-bench \
+          -m "$blob_file" --cache-type-v turbo3 --cache-type-k q8_0 \
+          --triattention-stats "$TRIATTENTION_STATS" \
+          --triattention-budget 2048 -r 3 \
+          > "supermerge_bench_triattention_${model_name}.txt" 2>&1
+    fi
+done
+
+echo "=== Local Qwen2.5-Coder-1.5B ==="
+LOCAL_BLOB="qwen2.5-coder-1.5b-bf16.gguf"
+LOCAL_STATS="qwen2.5-coder-1.5b.triattention"
+if [ -f "$LOCAL_BLOB" ]; then
+    ./build-tri/bin/llama-bench \
+      -m "$LOCAL_BLOB" --cache-type-v turbo3 --cache-type-k q8_0 -r 3 \
+      > "triattention_bench_local_qwen.txt" 2>&1
+
+    ./build-super/bin/llama-bench \
+      -m "$LOCAL_BLOB" --cache-type-v turbo3 --cache-type-k q8_0 -r 3 \
+      > "supermerge_bench_local_qwen.txt" 2>&1
+
+    if [ -f "$LOCAL_STATS" ]; then
+        ./build-super/bin/llama-bench \
+          -m "$LOCAL_BLOB" --cache-type-v turbo3 --cache-type-k q8_0 \
+          --triattention-stats "$LOCAL_STATS" \
+          --triattention-budget 1024 -r 3 \
+          > "supermerge_bench_triattention_local_qwen.txt" 2>&1
+    fi
+fi
+
+echo "Listo."
+ls -lh *bench*.txt *results*.txt *backend*.txt 2>/dev/null
+zip results.zip *bench*.txt *results*.txt *backend*.txt

ggml/src/ggml-turbo-quant.c

@@ -15,6 +15,10 @@
 #include <assert.h>
 #include <stdlib.h>
 
+#ifndef M_PI
+#define M_PI 3.14159265358979323846
+#endif
+
 /* Global: WHT group size for CPU quantize path (set by CPU SET_ROWS handler) */
 int turbo3_cpu_wht_group_size = 0;
 
@@ -79,9 +83,6 @@ static void turbo_init_rotation(void) {
         turbo_rotation[i] = (float)turbo_prng_normal();
     }
 
-    /* QR decomposition via modified Gram-Schmidt */
-    /* Q stored column-major in turbo_rotation */
-
     for (int j = 0; j < d; j++) {
         /* Normalize column j */
         float norm = 0.0f;

src/llama-kv-cache.cpp

@@ -266,8 +266,10 @@ llama_kv_cache::llama_kv_cache(
         layers.push_back({ il, k, v, k_stream, v_stream, });
 
         // TurboQuant: create rotation matrix tensors (once, shared across layers)
+        // NOTE: check BOTH type_k and type_v for asymmetric configs (e.g., -ctk q8_0 -ctv turbo3)
         if (turbo_rotation == nullptr &&
-            (type_k == GGML_TYPE_TURBO3_0 || type_k == GGML_TYPE_TURBO4_0 || type_k == GGML_TYPE_TURBO2_0)) {
+            ((type_k == GGML_TYPE_TURBO3_0 || type_k == GGML_TYPE_TURBO4_0 || type_k == GGML_TYPE_TURBO2_0) ||
+             (type_v == GGML_TYPE_TURBO3_0 || type_v == GGML_TYPE_TURBO4_0 || type_v == GGML_TYPE_TURBO2_0))) {
             turbo_rotation = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 128, 128);
             ggml_format_name(turbo_rotation, "turbo_rotation");  // R^T
             turbo_rotation_inv = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 128, 128);