Add low memory mode, model unloading between stages, and a CPU thread cap

app.py

@@ -82,6 +82,9 @@ GLOBAL_DEFAULTS = {
     "target_seconds": 60,
     "clip_seconds": 8,
     "scene_timing": "narration",
+    "low_memory": True,
+    "free_models_between_stages": True,
+    "cpu_threads": 0,
     "error_cooldown_minutes": 5,
     "work_hours_enabled": False,
     "work_days": "mon,tue,wed,thu,fri,sat,sun",
@@ -956,6 +959,21 @@ def ollama_pull(s, model):
     return True
 
 
+def unload_ollama(s, model):
+    model = (model or "").strip()
+    if not model or not s.get("free_models_between_stages", True):
+        return
+    try:
+        requests.post(
+            f"{s['ollama_url'].rstrip('/')}/api/generate",
+            json={"model": model, "prompt": "", "keep_alive": 0, "stream": False},
+            timeout=20,
+        )
+        logger.debug("Asked Ollama to release %s", model)
+    except Exception as e:
+        logger.debug("Could not release %s: %s", model, e)
+
+
 def ensure_local_deps(s):
     if not ensure_module(s, "torch", "torch") or not ensure_module(s, "diffusers", "diffusers accelerate safetensors transformers".split()[0]):
         raise RuntimeError("Could not install the image packages. Run ./install-local.sh")
@@ -1470,10 +1488,21 @@ def load_pipeline(s):
             device, dtype = "cuda", torch.float16
         else:
             device, dtype = "cpu", torch.float32
+        threads = int(s.get("cpu_threads", 0) or 0)
+        if threads > 0:
+            torch.set_num_threads(threads)
+            logger.info("Limiting image generation to %d CPU threads", threads)
         logger.info("Loading image model %s on %s (first run downloads several GB)", model, device)
         pipe = AutoPipelineForText2Image.from_pretrained(model, torch_dtype=dtype, variant="fp16" if dtype == torch.float16 else None)
         pipe = pipe.to(device)
         pipe.set_progress_bar_config(disable=True)
+        if s.get("low_memory", True):
+            for enable in ("enable_attention_slicing", "enable_vae_slicing", "enable_vae_tiling"):
+                try:
+                    getattr(pipe, enable)()
+                except Exception:
+                    pass
+            logger.info("Low memory mode is on for image generation")
         pipeline["id"] = model
         pipeline["obj"] = pipe
         logger.info("Image model ready")
@@ -1800,6 +1829,8 @@ def run_job(s):
             logger.info("Loaded %d existing channel titles to avoid repeats", len(s["_channel_titles"]))
         script, tags, compliance = produce_script(s, src)
         remember_story(s["profile_id"], script)
+        if engine == "local":
+            unload_ollama(s, s["ollama_model"])
         compliance["engine"] = engine
         update_history(job_id, title=script["title"], notes="; ".join(compliance.get("reviewer_notes", []))[:400])
         write_json(job_dir / "script.json", {"profile": s["name"], "source": src, "script": script, "tags": tags})
@@ -1909,6 +1940,8 @@ def run_job(s):
                     build_scene(i, script["scenes"][i - 1], clips[i - 1])
                 pending = failed
 
+        if kids:
+            unload_ollama(s, s["vision_model"])
         anatomy_note = f"Possible anatomy glitches kept in scenes {', '.join(str(i) for i in anatomy_flags)}" if anatomy_flags else ""
         if kids:
             compliance["vision_checks"] = vision_log

static/index.html

@@ -319,6 +319,10 @@ const FIELDS=[
 ["piper_model","Piper voice file","wide","Blank uses the first voice found in the voices folder under Models","g"],
 ["gemini_api_key","Gemini API key (Veo)","password","Only needed for the Veo engine. Billing must be enabled","g"],
 ["veo_model","Veo model","text","e.g. veo-3.1-fast-generate-preview","g"],
+["","Performance","header","Shared by all profiles. Lower settings here if the Mac becomes slow or unstable",""],
+["low_memory","Low memory mode","checkbox","Generates images in smaller pieces. A little slower, much less memory. Leave on for Macs with 16 GB or less","g"],
+["free_models_between_stages","Free models between stages","checkbox","Unloads the writing model before images are drawn and the vision model after checks","g"],
+["cpu_threads","CPU threads","select:0=Automatic|2|4|6|8|10","Caps how many cores image generation uses so the rest of the Mac stays responsive","g"],
 ["","Writing and timing","header","Shared by all profiles",""],
 ["ollama_url","Ollama URL","text","","g"],
 ["ollama_model","Ollama model","model:text","Writes scripts and tags. Models not installed yet are downloaded automatically","g"],
@@ -332,7 +336,7 @@ const FIELDS=[
 ];
 const DAYS=[["mon","Mon"],["tue","Tue"],["wed","Wed"],["thu","Thu"],["fri","Fri"],["sat","Sat"],["sun","Sun"]];
 const TIMES=[];for(let h=0;h<24;h++)for(const m of [0,30]){const v=String(h).padStart(2,"0")+":"+String(m).padStart(2,"0");const hr=h%12||12;TIMES.push([v,hr+":"+String(m).padStart(2,"0")+(h<12?" AM":" PM")])}
-const NUMERIC=new Set(["sd_steps","sd_guidance","sd_width","sd_height","trending_days","trending_count","videos_per_day","minutes_between_videos","target_seconds","clip_seconds","ollama_timeout","error_cooldown_minutes","tts_rate"]);
+const NUMERIC=new Set(["sd_steps","sd_guidance","sd_width","sd_height","trending_days","trending_count","videos_per_day","minutes_between_videos","target_seconds","clip_seconds","ollama_timeout","error_cooldown_minutes","tts_rate","cpu_threads"]);
 const $=s=>document.querySelector(s);
 const esc=s=>String(s??"").replace(/[&<>"]/g,c=>({"&":"&amp;","<":"&lt;",">":"&gt;",'"':"&quot;"}[c]));
 async function api(path,opts={}){const r=await fetch(path,{headers:{"Content-Type":"application/json"},...opts});const j=await r.json().catch(()=>({}));if(!r.ok)throw new Error(j.detail||r.statusText);return j}