harness: llama-server lifecycle in pre-clean and health guard

autoresearch.sh

@@ -38,7 +38,8 @@ rsync -a "$REPO/scripts/frank-large.env" "$REPO/scripts/serve-frank-large.sh" \
 # 2. Free the GPU and the port: stop the deployed unit (mask is unnecessary—
 #    the unit only starts via systemd, and we never enable more of it).
 ssh "$HOST" 'sudo -n systemctl stop frank-large.service;
-             podman rm -f frank-large-mimo 2>/dev/null; true' >/dev/null
+             podman rm -f frank-large-mimo 2>/dev/null;
+             pkill -x llama-server 2>/dev/null; true' >/dev/null
 
 # 3. Boot the candidate detached; the launcher's own output IS the serve
 #    stage output (gate failures print there before any container exists).
@@ -55,7 +56,7 @@ if ! ssh "$HOST" "for i in \$(seq 1 $HEALTH_TRIES); do
               http://127.0.0.1:7770/health 2>/dev/null || echo 000)
         [ \"\$code\" = 200 ] && echo UP && exit 0
         st=\$(podman inspect -f '{{.State.Status}}' frank-large-mimo 2>/dev/null || echo gone)
-        if [ \"\$st\" = exited ] || { [ \"\$st\" = gone ] && ! pgrep -f serve-frank-large >/dev/null; }; then echo DOWN; exit 1; fi
+        if [ "\$st" = exited ] || { [ "\$st" = gone ] && ! pgrep -f serve-frank-large >/dev/null && ! pgrep -x llama-server >/dev/null; }; then echo DOWN; exit 1; fi
         sleep 5
     done; exit 1" | grep -q UP; then
     echo "ERROR: candidate never became healthy; serve.log tail:" >&2

scripts/frank-large.env

@@ -88,11 +88,17 @@ FRANK_LARGE_UTIL=0.45
 FRANK_LARGE_SPEC=
 FRANK_LARGE_SPEC_DIR=/home/dank/models/Qwen3.5-9B
 FRANK_LARGE_SPEC_TOKENS=2
-# AUTORESEARCH: torchao int8 weight-only (quark int8 + all fp8 kernels
-# rejected on gfx1151; torchao is a separate Triton dequant stack)
-FRANK_LARGE_QUANT=torchao
-FRANK_LARGE_QUANT_CONFIG='{"torchao_config":"int8_weight_only"}'
 
+# --- AUTORESEARCH arm: llama.cpp runtime -------------------------------------
+# master build (qwen35 GDN kernels, HIP gfx1151) on Q6_K GGUF: 7.79 GiB
+# weights = 0.51x the BF16 bytes, so the LPDDR roof doubles. Cross-model
+# DFlash draft (qwen3.5-9b, Q8_0). --reasoning-format deepseek emits
+# reasoning_content, NOT vLLM's reasoning.
+FRANK_LARGE_RUNTIME=llama
+FRANK_LARGE_LLAMA_BIN=/home/dank/GitPackages/llama-master/build-hip/bin/llama-server
+FRANK_LARGE_GGUF=/home/dank/models/MiMo-Q6K/MiMo-V2.6-Distill-Qwen-9B-Q6_K.gguf
+FRANK_LARGE_DRAFT_GGUF=/home/dank/models/DFlash-Qwen35/qwen3.5-9b-dflash-Q8_0.gguf
+FRANK_LARGE_LLAMA_CTX=1048576
 # --- client contract ----------------------------------------------------------
 # vLLM emits the thinking text in `reasoning` (NOT llama.cpp/gufo's
 # `reasoning_content`) — the harness models.yml compat block for this