Update memory profile script and troubleshooting guide for multimodal models

docs/OLLAMA_ROCM_TROUBLESHOOTING.md

@@ -34,6 +34,7 @@ Typical log pattern:
 ### Why this happens
 
 Ollama can detect ROCm correctly but still OOM during model graph/KV allocation, often due to high context size or aggressive defaults.
+If logs show a multimodal architecture such as `qwen3vl` and stack frames under `multimodal.go`, a small-parameter model can still fail due to multimodal graph reservation.
 
 ### Fix
 
@@ -59,6 +60,8 @@ Preset selection quick guide:
 | 20B to 32B | `balanced` | Best first choice for stability on large models. |
 | 30B+ with load failures/OOM | `safe` | Uses lower context, `KEEP_ALIVE=0`, and higher GPU overhead reserve. |
 
+For troubleshooting, test a text-only model first (for example `qwen3:8b`) before VL/multimodal models.
+
 If a model fails to load, step down from `max` → `balanced` → `safe` before manual tuning.
 
 Manual method:

scripts/set_ollama_memory_profile_in_ct.sh

@@ -69,12 +69,12 @@ done
 
 case "$PRESET" in
   safe)
-    PRESET_CONTEXT_LENGTH="4096"
+    PRESET_CONTEXT_LENGTH="2048"
     PRESET_NUM_PARALLEL="1"
     PRESET_MAX_LOADED_MODELS="1"
     PRESET_FLASH_ATTENTION="false"
     PRESET_KEEP_ALIVE="0"
-    PRESET_GPU_OVERHEAD_BYTES="2147483648"
+    PRESET_GPU_OVERHEAD_BYTES="4294967296"
     ;;
   balanced)
     PRESET_CONTEXT_LENGTH="8192"
@@ -202,6 +202,7 @@ set -euo pipefail
 install -d /etc/systemd/system/ollama.service.d
 cat >/etc/systemd/system/ollama.service.d/20-memory-tuning.conf <<EOF
 [Service]
+LimitMEMLOCK=infinity
 Environment=OLLAMA_CONTEXT_LENGTH=${CONTEXT_LENGTH}
 Environment=OLLAMA_MAX_LOADED_MODELS=${MAX_LOADED_MODELS}
 Environment=OLLAMA_NUM_PARALLEL=${NUM_PARALLEL}