feat(serve): set APEX I-Compact as default, harden benchmark workflow

Serving:
- make serve now launches Claude-distilled APEX 35B-A3B (16GB) with 2
  parallel slots and 256K context as the daily driver
- add serve-custom for ad-hoc model testing
- add flush-gpu to reclaim unified memory after stuck runs

Benchmarks:
- default Vulkan-only backends (ROCm trails at long context)
- add --backends filter to run-baseline.sh
- fix backend filter substring bug (grep -qFx for exact line match)
- fix model filter regex metacharacter bug (grep -qiF for literal)
- respect --tg in long-context tests instead of hardcoded n=32

ROCm bump to 7.2.1 (kernel 6.18.4+ patch); keep 7.2 as optional.

Catalog:
- add mudler APEX I-Compact (Claude-distilled 35B, 17GB)
- add 0xSero REAP-40 (pruned 122B-A10B, 46GB)
- update download instructions: hf download (huggingface-cli is gone)

Makefile

@@ -39,12 +39,23 @@ benchmark-compare: ## Compare two benchmark runs (usage: make benchmark-compare
 	@bash bin/benchmark compare $(BEFORE) $(AFTER)
 
 # --- Serve ---
-serve: ## Launch llama-server with optimized settings (ARGS="-m MODEL.gguf")
+serve: ## Launch APEX I-Compact daily driver (2 slots, 256K ctx)
+	@bash bin/serve -m Qwen3.5-35B-A3B-Claude-Distilled-APEX-I-Compact.gguf --parallel 2 --ctx 262144 $(ARGS)
+
+serve-custom: ## Launch llama-server with custom model (ARGS="-m MODEL.gguf")
 	@bash bin/serve $(ARGS)
 
 serve-ngram: ## Launch with n-gram speculative decoding (ARGS="-m MODEL.gguf")
 	@bash bin/serve --ngram $(ARGS)
 
+flush-gpu: ## Kill llama-server/bench processes and drop kernel caches to free unified VRAM
+	-@pkill -x llama-server 2>/dev/null || true
+	-@pkill -x llama-bench 2>/dev/null || true
+	-@pkill -x llama-cli 2>/dev/null || true
+	-@podman ps --filter name=llama --format '{{.Names}}' | xargs -r podman stop
+	@sync && sudo sysctl vm.drop_caches=3
+	@echo "VRAM usage:" && cat /sys/class/drm/card*/device/mem_info_vram_used 2>/dev/null | awk '{printf "  %.2f MiB\n", $$1/1048576}'
+
 # --- Hardware Info ---
 hw-bandwidth: ## Measure GPU memory bandwidth and compute (clpeak)
 	@clpeak 2>&1

configs/models.conf

@@ -2,7 +2,7 @@
 # Format: NAME|HF_REPO|FILE|SIZE_GB|CATEGORY|DESCRIPTION
 #
 # Categories: smoke, standard, moe, dense
-# Download with: huggingface-cli download REPO FILE --local-dir /data/models/llms/REPO
+# Download with: hf download REPO FILE --local-dir /data/models/llms/REPO
 
 # ── Smoke tests (quick, small) ───────────────────────────
 qwen2.5-0.5b-q8|lmstudio-community/Qwen2.5-0.5B-Instruct-GGUF|Qwen2.5-0.5B-Instruct-Q8_0.gguf|0.4|smoke|Tiny Qwen2.5, Q8
@@ -26,11 +26,15 @@ qwen3.5-27b-opus-distill|Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distille
 # ── MoE models (fast generation, best for 64GB) ─────────
 qwen3.5-35b-a3b-q4|unsloth/Qwen3.5-35B-A3B-GGUF|Qwen3.5-35B-A3B-UD-Q4_K_XL.gguf|21|moe|MoE 35B, 3B active, Unsloth dynamic XL
 qwen3.5-35b-a3b-q8|unsloth/Qwen3.5-35B-A3B-GGUF|Qwen3.5-35B-A3B-Q8_0.gguf|37|moe|MoE 35B Q8, near-full precision
+qwen3.5-35b-a3b-apex-compact|mudler/Qwen3.5-35B-A3B-Claude-Distilled-APEX-GGUF|Qwen3.5-35B-A3B-Claude-Distilled-APEX-I-Compact.gguf|17|moe|MoE 35B Claude-distilled APEX, I-Compact quant
 nemotron-cascade2-q8|bartowski/nvidia_Nemotron-Cascade-2-30B-A3B-GGUF|nvidia_Nemotron-Cascade-2-30B-A3B-Q8_0.gguf|31|moe|Nemotron Cascade 2, Mamba-2 hybrid (replaces Nano)
 
 # ── Coding models ─────────────────────────────────────────
 qwen3-coder-30b-a3b-q6|unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF|Qwen3-Coder-30B-A3B-Instruct-UD-Q6_K_XL.gguf|26|moe|Agentic coding MoE, pure Transformer
 qwen3-coder-next-q3|unsloth/Qwen3-Coder-Next-GGUF|Qwen3-Coder-Next-UD-Q3_K_XL.gguf|34|moe|80B MoE coder, >70% SWE-bench, hybrid DeltaNet
 
+# ── Pruned MoE (REAP expert pruning) ─────────────────────
+qwen3.5-122b-a10b-reap40-q4|0xSero/Qwen3.5-122B-A10B-REAP-40-GGUF|Qwen3.5-122B-A10B-REAP-40-Q4_K_M.gguf|46|moe|122B MoE pruned to 40 experts, 10B active, Q4_K_M
+
 # ── Draft models (speculative decoding) ───────────────────
 qwen3.5-0.8b-q8-draft|unsloth/Qwen3.5-0.8B-GGUF|Qwen3.5-0.8B-Q8_0.gguf|0.8|draft|Draft for Qwen3.5 speculative decoding

docs/benchmarking.md

@@ -41,7 +41,8 @@ The `-fa 1 -mmp 0 -ngl 99` flags are **mandatory** on Strix Halo to avoid crashe
 | `llama-vulkan-radv` | Mesa RADV | Vulkan | Most stable, recommended default |
 | `llama-vulkan-amdvlk` | AMDVLK | Vulkan | Fastest when it works, 2GB buffer limit |
 | `llama-rocm-6.4.4` | ROCm 6.4.4 | HIP | Proven stable |
-| `llama-rocm-7.2` | ROCm 7.2 | HIP | Latest, compiler fixes applied |
+| `llama-rocm-7.2.1` | ROCm 7.2.1 | HIP | Current stable (kernel 6.18.4+ patch) |
+| `llama-rocm-7.2` | ROCm 7.2 | HIP | Deprecated — use 7.2.1 |
 | `llama-rocm7-nightlies` | ROCm 7 nightly | HIP | Experimental/development builds |
 
 Containers are from [kyuz0/amd-strix-halo-toolboxes](https://github.com/kyuz0/amd-strix-halo-toolboxes). Set up with `make benchmark-setup`.

scripts/benchmark/run-baseline.sh

@@ -23,12 +23,14 @@ PP_TOKENS=512
 TG_TOKENS=128
 BATCH_SIZE=""    # Batch size override (-b flag, empty = llama-bench default 2048)
 KV_TYPES_RAW=""  # Comma-separated KV cache types to sweep (e.g. f16,q8_0,q4_0 or q4_0:q8_0)
+BACKENDS_FILTER="llama-vulkan-radv"  # Default to Vulkan; use --backends to override
 
 while [[ $# -gt 0 ]]; do
     case "$1" in
         --skip-longctx)  SKIP_LONGCTX=true; shift ;;
         --max-size|-s)   MAX_SIZE_GB="$2"; shift 2 ;;
         --category|-c)   CATEGORY_FILTER="$2"; shift 2 ;;
+        --backends)      BACKENDS_FILTER="$2"; shift 2 ;;
         --reps|-r)       REPS_STANDARD="$2"; shift 2 ;;
         --context|-d)    CTX_DEPTH="$2"; shift 2 ;;
         --pp)            PP_TOKENS="$2"; shift 2 ;;
@@ -42,6 +44,7 @@ while [[ $# -gt 0 ]]; do
             echo "  --skip-longctx       Skip long-context tests"
             echo "  --max-size GB        Only bench models up to this file size in GB"
             echo "  --category LIST      Comma-separated: smoke,dense,moe (from models.conf)"
+            echo "  --backends LIST      Comma-separated backends (default: llama-vulkan-radv)"
             echo "  --reps N             Standard test repetitions (default: 5)"
             echo "  --context N          Long-context depth in tokens (default: 32768)"
             echo "  --pp N               Prompt processing tokens (default: 512)"
@@ -94,15 +97,18 @@ declare -A BENCH_PATHS=(
     [llama-vulkan-amdvlk]="/usr/sbin/llama-bench"
     [llama-rocm-6.4.4]="/usr/local/bin/llama-bench"
     [llama-rocm-7.2]="/usr/local/bin/llama-bench"
+    [llama-rocm-7.2.1]="/usr/local/bin/llama-bench"
     [llama-rocm7-nightlies]="/usr/local/bin/llama-bench"
 )
 
 available_backends=()
 for tb in "${!BENCH_PATHS[@]}"; do
     if echo "$existing" | grep -q "^${tb}$"; then
+        if [[ -z "$BACKENDS_FILTER" ]] || echo "$BACKENDS_FILTER" | tr ',' '\n' | grep -qFx "$tb"; then
             available_backends+=("$tb")
             log_success "Backend: $tb"
         fi
+    fi
 done
 
 if (( ${#available_backends[@]} == 0 )); then
@@ -269,7 +275,7 @@ for MODEL_PATH in "${MODEL_PATHS[@]}"; do
 
                 CMD_LC=(toolbox run -c "$BACKEND" -- "${ENV_ARGS[@]}" "$BENCH_BIN"
                     -ngl 99 -mmp 0 -m "$TOOLBOX_MODEL_PATH" -fa 1
-                    -p "$CTX_PROMPT" -n 32 -d "$CTX_DEPTH" -ub "$UB_SIZE"
+                    -p "$CTX_PROMPT" -n "$TG_TOKENS" -d "$CTX_DEPTH" -ub "$UB_SIZE"
                     -r "$REPS_LONGCTX" "${BATCH_ARGS[@]}" "${KV_ARGS[@]}")
 
                 printf "  cmd: %s\n" "${CMD_LC[*]}"

scripts/benchmark/run-suite.sh

@@ -9,7 +9,7 @@ source "$SCRIPT_DIR/../../lib/format.sh"
 
 MODEL_DIR="$(data_dir models)"
 TAG="run"
-BACKENDS_FILTER=""
+BACKENDS_FILTER="llama-vulkan-radv"
 MODELS_FILTER=""
 SKIP_LONGCTX=false
 MAX_SIZE_GB=0
@@ -99,13 +99,14 @@ declare -A BENCH_PATHS=(
     [llama-vulkan-amdvlk]="/usr/sbin/llama-bench"
     [llama-rocm-6.4.4]="/usr/local/bin/llama-bench"
     [llama-rocm-7.2]="/usr/local/bin/llama-bench"
+    [llama-rocm-7.2.1]="/usr/local/bin/llama-bench"
     [llama-rocm7-nightlies]="/usr/local/bin/llama-bench"
 )
 
 available_backends=()
 for tb in "${!BENCH_PATHS[@]}"; do
     if echo "$existing" | grep -q "^${tb}$"; then
-        if [[ -z "$BACKENDS_FILTER" ]] || echo "$BACKENDS_FILTER" | tr ',' '\n' | grep -q "$tb"; then
+        if [[ -z "$BACKENDS_FILTER" ]] || echo "$BACKENDS_FILTER" | tr ',' '\n' | grep -qFx "$tb"; then
             available_backends+=("$tb")
         fi
     fi
@@ -130,7 +131,7 @@ for p in "${ALL_MODEL_PATHS[@]}"; do
 
     # Name filter
     if [[ -n "$MODELS_FILTER" ]]; then
-        if ! echo "$MODELS_FILTER" | tr ',' '\n' | grep -qi "$local_name"; then
+        if ! echo "$MODELS_FILTER" | tr ',' '\n' | grep -qiF "$local_name"; then
             continue
         fi
     fi
@@ -252,7 +253,7 @@ for MODEL_PATH in "${MODEL_PATHS[@]}"; do
                 UB_SIZE=2048; [[ "$BACKEND" == *vulkan* ]] && UB_SIZE=512
                 CMD_LC=(toolbox run -c "$BACKEND" -- "${ENV_ARGS[@]}" "$BENCH_BIN"
                     -ngl 99 -mmp 0 -m "$TOOLBOX_MODEL_PATH" -fa 1
-                    -p "$CTX_PROMPT" -n 32 -d "$CTX_DEPTH" -ub "$UB_SIZE" -r "$REPS_LONGCTX" "${BATCH_ARGS[@]}" "${KV_ARGS[@]}")
+                    -p "$CTX_PROMPT" -n "$TG_TOKENS" -d "$CTX_DEPTH" -ub "$UB_SIZE" -r "$REPS_LONGCTX" "${BATCH_ARGS[@]}" "${KV_ARGS[@]}")
                 if "${CMD_LC[@]}" > "$OUT_LC" 2>&1; then
                     log_success "Done"; tail -3 "$OUT_LC"
                 else

scripts/benchmark/setup.sh

@@ -15,8 +15,8 @@ log_header "Benchmark Setup"
 # ── 1. Check toolbox containers ──────────────────────────
 log_info "Checking toolbox containers..."
 
-REQUIRED_TOOLBOXES=("llama-vulkan-radv" "llama-rocm-7.2")
+REQUIRED_TOOLBOXES=("llama-vulkan-radv" "llama-rocm-7.2.1")
-OPTIONAL_TOOLBOXES=("llama-rocm-6.4.4" "llama-vulkan-amdvlk")
+OPTIONAL_TOOLBOXES=("llama-rocm-7.2" "llama-rocm-6.4.4" "llama-vulkan-amdvlk")
 
 existing=$(detect_toolbox_names 2>/dev/null || true)
 missing=()

scripts/serve/launch.sh

@@ -79,6 +79,7 @@ declare -A SERVER_PATHS=(
     [llama-vulkan-amdvlk]="/usr/sbin/llama-server"
     [llama-rocm-6.4.4]="/usr/local/bin/llama-server"
     [llama-rocm-7.2]="/usr/local/bin/llama-server"
+    [llama-rocm-7.2.1]="/usr/local/bin/llama-server"
     [llama-rocm7-nightlies]="/usr/local/bin/llama-server"
 )