Add acceleration/distillation support (Lightning/Turbo/LCM/Hyper-SD)

Per-model `acceleration` config block fuses a distillation LoRA into the
pipeline at load and supplies low step-count / guidance defaults at
generation time, for a 5-10x speedup. Covers video (Wan), image diffusers
(SD/SDXL), and sd.cpp (step/cfg defaults + <lora:> prompt injection).

- New codai/models/acceleration.py: preset catalog (ACCEL_PRESETS),
  resolve_acceleration(), apply_accel_to_pipeline() (load->fuse->unload so
  it stays orthogonal to per-request character/env LoRAs), accel_call_defaults().
- video.py: fuse accel LoRA after load; _generate_video / _generate_sdcpp_video
  use preset steps/guidance (request always wins).
- images.py: _apply_image_acceleration on both diffusers load paths;
  _generate_image and _generate_with_sdcpp honour preset steps/guidance.
- main.py: surface `acceleration` as a first-class runtime kwarg.
- admin: persist `acceleration`; new GET /admin/api/accel-presets; models.html
  Acceleration/Distillation card (preset dropdown + manual override).

Also fix a latent null-trap: float(cfg.get('balanced_gpu_percent', 80))
crashed when the config stored an explicit null (written by the admin UI for
blank fields) since .get(key, default) returns the stored None. Use `or 80`.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

codai/admin/routes.py

@@ -1939,7 +1939,7 @@ async def api_model_configure(request: Request, username: str = Depends(require_
                 "lora_train_base_model",
                 "max_vram", "sdcpp_flash_attn", "sdcpp_diffusion_flash_attn", "vae_tiling",
                 "component_quantization", "output_crf", "force_vram_update",
-                "balanced_gpu_percent"):
+                "balanced_gpu_percent", "acceleration"):
         if key in data:
             entry[key] = data[key]
 
@@ -1970,6 +1970,18 @@ async def api_model_configure(request: Request, username: str = Depends(require_
     return {"success": True, "applied_live": applied}
 
 
+@router.get("/admin/api/accel-presets")
+async def api_accel_presets(username: str = Depends(require_admin)):
+    """Return the acceleration/distillation preset catalog (Lightning / Turbo /
+    LCM / Hyper-SD) so the model-config UI dropdown stays in sync with the Python
+    source of truth in codai/models/acceleration.py."""
+    try:
+        from codai.models.acceleration import ACCEL_PRESETS
+        return {"presets": ACCEL_PRESETS}
+    except Exception as e:
+        raise HTTPException(status_code=500, detail=str(e))
+
+
 # --- System endpoints ---
 
 @router.post("/admin/api/system/reload")

codai/admin/templates/models.html

@@ -689,6 +689,49 @@ window.__DEFAULT_WHISPER_SERVER_PATH__ = {{ default_whisper_server_path|tojson }
         <span class="form-hint">When the model won't fit entirely in VRAM, fill this % of free VRAM then spill the rest to CPU RAM. Also the cap used by auto-balanced fallback.</span>
       </div>
 
+      <!-- Acceleration / distillation (Lightning / Turbo / LCM / Hyper-SD) — image & video -->
+      <div id="cfg-accel-section" style="display:none">
+        <div class="card-title" style="margin-top:1.25rem">Acceleration / Distillation
+          <span class="muted" style="font-weight:normal">(Lightning / Turbo / LCM / Hyper-SD — image &amp; video)</span></div>
+        <label style="display:flex;align-items:center;gap:.5rem;cursor:pointer;font-size:13px;margin:.4rem 0">
+          <input type="checkbox" id="cfg-accel-enabled" onchange="onAccelToggle()"> Enable acceleration
+          <span class="muted">fuse a distill LoRA and run in few steps at low guidance</span></label>
+        <div id="cfg-accel-fields" style="display:none">
+          <div class="form-row" style="max-width:420px">
+            <label class="form-label">Preset</label>
+            <select id="cfg-accel-preset" class="form-input" onchange="onAccelPreset()">
+              <option value="custom">Custom (set fields manually)</option>
+            </select>
+          </div>
+          <div class="form-row" style="max-width:560px">
+            <label class="form-label">Distill LoRA <span class="muted">(path or HF repo, optionally repo:weight_name.safetensors; blank for turbo full-models)</span></label>
+            <input type="text" id="cfg-accel-lora" class="form-input" placeholder="e.g. ByteDance/SDXL-Lightning:sdxl_lightning_4step_lora.safetensors">
+          </div>
+          <div style="display:flex;gap:1rem;flex-wrap:wrap">
+            <div class="form-row" style="max-width:130px">
+              <label class="form-label">LoRA weight</label>
+              <input type="number" id="cfg-accel-weight" class="form-input" min="0" max="2" step="0.05" placeholder="1.0">
+            </div>
+            <div class="form-row" style="max-width:120px">
+              <label class="form-label">Steps</label>
+              <input type="number" id="cfg-accel-steps" class="form-input" min="1" max="50" step="1" placeholder="4">
+            </div>
+            <div class="form-row" style="max-width:130px">
+              <label class="form-label">Guidance</label>
+              <input type="number" id="cfg-accel-guidance" class="form-input" min="0" max="15" step="0.5" placeholder="1.0">
+            </div>
+            <div class="form-row" style="max-width:150px">
+              <label class="form-label">Flow shift <span class="muted">(Wan)</span></label>
+              <input type="number" id="cfg-accel-flowshift" class="form-input" min="0" max="20" step="0.5" placeholder="(none)">
+            </div>
+            <div class="form-row" style="max-width:200px">
+              <label class="form-label">Scheduler <span class="muted">(optional)</span></label>
+              <input type="text" id="cfg-accel-scheduler" class="form-input" placeholder="e.g. LCMScheduler">
+            </div>
+          </div>
+        </div>
+      </div>
+
       <!-- components -->
       <div class="card-title" style="margin-top:1.25rem">Components</div>
       <div class="form-row">
@@ -2225,6 +2268,10 @@ async function refreshLocal(){
 loadGlobalSettings();
 refreshLocal();
 
+// Toggle the acceleration section as image/video model types are checked/unchecked.
+document.querySelectorAll('.cfg-type-cb').forEach(cb =>
+  cb.addEventListener('change', () => _refreshAccelVisibility()));
+
 // ── Deep-link from Studio: /admin/models?tab=search&q=...&pipeline=...&gguf=...
 // ── or: /admin/models?local_cap=CAPABILITY — highlight local models with that capability
 (function applyDeepLink(){
@@ -2630,9 +2677,97 @@ function openCfgModal(idx, cfgIdx){
   _setCompField('lora',       s.lora_path       || '');
   document.getElementById('cfg-lora-dir').value = s.lora_model_dir || '';
   document.getElementById('cfg-lora-train-base').value = s.lora_train_base_model || '';
+  _populateAccel(s.acceleration);
   openModal('cfg-modal');
 }
 
+// ---- Acceleration / distillation (Lightning / Turbo / LCM / Hyper-SD) ----
+let _accelPresets = null;
+async function _loadAccelPresets(){
+  if (_accelPresets) return _accelPresets;
+  try {
+    const r = await fetch(ROOT_PATH + '/admin/api/accel-presets');
+    const d = await r.json();
+    _accelPresets = d.presets || {};
+  } catch(e){ _accelPresets = {}; }
+  return _accelPresets;
+}
+function _accelAppliesTo(){
+  // Which families the current model-type selection allows (image / video).
+  const types = new Set([...document.querySelectorAll('.cfg-type-cb:checked')].map(cb=>cb.value));
+  const out = new Set();
+  if (types.has('image_models')) out.add('image');
+  if (types.has('video_models')) out.add('video');
+  return out;
+}
+function _refreshAccelVisibility(){
+  const section = document.getElementById('cfg-accel-section');
+  if (!section) return;
+  const applies = _accelAppliesTo();
+  section.style.display = applies.size ? '' : 'none';
+  // Rebuild the preset dropdown filtered to the applicable families.
+  const sel = document.getElementById('cfg-accel-preset');
+  if (!sel) return;
+  const cur = sel.value;
+  sel.innerHTML = '<option value="custom">Custom (set fields manually)</option>';
+  const presets = _accelPresets || {};
+  Object.keys(presets).forEach(k=>{
+    const p = presets[k];
+    const at = p.applies_to || [];
+    if (!applies.size || at.some(a=>applies.has(a))){
+      const o = document.createElement('option');
+      o.value = k; o.textContent = p.label || k;
+      sel.appendChild(o);
+    }
+  });
+  // Preserve the previously selected preset if still present.
+  if ([...sel.options].some(o=>o.value===cur)) sel.value = cur;
+}
+function onAccelToggle(){
+  const on = document.getElementById('cfg-accel-enabled').checked;
+  document.getElementById('cfg-accel-fields').style.display = on ? '' : 'none';
+}
+function onAccelPreset(){
+  const key = document.getElementById('cfg-accel-preset').value;
+  const p = (_accelPresets || {})[key];
+  if (!p) return;  // "custom" — leave fields as-is
+  document.getElementById('cfg-accel-lora').value = p.lora || '';
+  document.getElementById('cfg-accel-weight').value = p.lora_weight != null ? p.lora_weight : '';
+  document.getElementById('cfg-accel-steps').value = p.steps != null ? p.steps : '';
+  document.getElementById('cfg-accel-guidance').value = p.guidance_scale != null ? p.guidance_scale : '';
+  document.getElementById('cfg-accel-flowshift').value = p.flow_shift != null ? p.flow_shift : '';
+  document.getElementById('cfg-accel-scheduler').value = p.scheduler || '';
+}
+async function _populateAccel(a){
+  await _loadAccelPresets();
+  _refreshAccelVisibility();
+  a = a || {};
+  document.getElementById('cfg-accel-enabled').checked = !!a.enabled;
+  const sel = document.getElementById('cfg-accel-preset');
+  sel.value = [...sel.options].some(o=>o.value===(a.preset||'')) ? a.preset : 'custom';
+  document.getElementById('cfg-accel-lora').value = a.lora || '';
+  document.getElementById('cfg-accel-weight').value = a.lora_weight != null ? a.lora_weight : '';
+  document.getElementById('cfg-accel-steps').value = a.steps != null ? a.steps : '';
+  document.getElementById('cfg-accel-guidance').value = a.guidance_scale != null ? a.guidance_scale : '';
+  document.getElementById('cfg-accel-flowshift').value = a.flow_shift != null ? a.flow_shift : '';
+  document.getElementById('cfg-accel-scheduler').value = a.scheduler || '';
+  onAccelToggle();
+}
+function _collectAccel(){
+  if (!document.getElementById('cfg-accel-enabled').checked) return null;
+  const num = id => { const v = document.getElementById(id).value.trim(); return v === '' ? null : parseFloat(v); };
+  return {
+    enabled: true,
+    preset: document.getElementById('cfg-accel-preset').value || 'custom',
+    lora: document.getElementById('cfg-accel-lora').value.trim() || null,
+    lora_weight: num('cfg-accel-weight'),
+    steps: num('cfg-accel-steps'),
+    guidance_scale: num('cfg-accel-guidance'),
+    flow_shift: num('cfg-accel-flowshift'),
+    scheduler: document.getElementById('cfg-accel-scheduler').value.trim() || '',
+  };
+}
+
 function _updatePreloadAllVisibility() {
   const loadMode = document.getElementById('cfg-load-mode').value;
   const maxInst = parseInt(document.getElementById('cfg-max-instances').value) || 1;
@@ -2730,6 +2865,7 @@ async function saveModelConfig(){
                           ? null : parseInt(document.getElementById('cfg-output-crf').value)),
     balanced_gpu_percent: (document.getElementById('cfg-balanced-gpu-pct').value.trim() === ''
                           ? null : parseFloat(document.getElementById('cfg-balanced-gpu-pct').value)),
+    acceleration:      _collectAccel(),
   };
   try{
     const r = await fetch(ROOT_PATH + '/admin/api/model-configure',{

codai/api/images.py

@@ -331,6 +331,22 @@ def _disable_safety_checker(pipe):
     return pipe
 
 
+def _apply_image_acceleration(pipeline, model_config):
+    """Fuse a configured acceleration/distillation LoRA (Lightning / Turbo / LCM /
+    Hyper-SD) into a freshly loaded diffusers image pipeline. No-op when no
+    acceleration is configured. Failures are caught inside apply_accel_to_pipeline."""
+    try:
+        from codai.models.acceleration import resolve_acceleration, apply_accel_to_pipeline
+        accel = resolve_acceleration(model_config)
+        if accel:
+            print(f"  [image][accel] applying {accel.get('preset')} "
+                  f"(steps={accel.get('steps')}, guidance={accel.get('guidance_scale')})")
+            apply_accel_to_pipeline(pipeline, accel)
+    except Exception as e:
+        print(f"  [image][accel] skipped: {e}")
+    return pipeline
+
+
 def _load_diffusers_pipeline(model_name: str, global_args, model_config: dict = None):
     """
     Try to load a model using the diffusers library.
@@ -443,7 +459,7 @@ def _load_diffusers_pipeline(model_name: str, global_args, model_config: dict =
                 if _img_quant_config is None:
                     raise  # only quantized pipelines may reject .to()
             print(f"--no-ram: Diffusers model loaded on {cuda_device}")
-            return pipeline
+            return _apply_image_acceleration(pipeline, _mc)
         except Exception as e:
             raise RuntimeError(
                 f"--no-ram: Failed to load diffusers model entirely on GPU ({cuda_device}). "
@@ -591,7 +607,7 @@ def _load_diffusers_pipeline(model_name: str, global_args, model_config: dict =
                     raise
                 pipeline = None
 
-    return pipeline
+    return _apply_image_acceleration(pipeline, _mc)
 
 
 async def _apply_vae_override(pipeline, vae_model_id: str):
@@ -725,9 +741,17 @@ async def _generate_with_diffusers(pipeline, request, global_args, http_request=
         generator = torch.Generator(device=pipeline.device).manual_seed(seed)
 
     quality = request.quality or "standard"
-    num_steps = request.steps if request.steps else (30 if quality == "standard" else 50)
+    # Acceleration/distillation defaults (Lightning / Turbo / LCM / Hyper-SD): when
+    # the loaded pipeline has a fused distill LoRA, default to its low step-count /
+    # guidance. The request always wins if it specified steps/guidance.
+    _accel = getattr(pipeline, '_coderai_accel', None)
+    _accel_steps = _accel.get('steps') if _accel else None
+    _accel_cfg = _accel.get('guidance_scale') if _accel else None
+    num_steps = request.steps if request.steps else (
+        _accel_steps if _accel_steps else (30 if quality == "standard" else 50))
     cfg_scale = request.guidance_scale if request.guidance_scale else (
-        getattr(global_args, 'image_cfg_scale', 7.5) if quality == "standard" else 9.0
+        _accel_cfg if _accel_cfg is not None else
+        (getattr(global_args, 'image_cfg_scale', 7.5) if quality == "standard" else 9.0)
     )
 
     _progress_reset(num_steps)
@@ -950,7 +974,8 @@ async def _generate_with_diffusers(pipeline, request, global_args, http_request=
     }
 
 
-async def _generate_with_sdcpp(sd_model, request, global_args, http_request=None):
+async def _generate_with_sdcpp(sd_model, request, global_args, http_request=None,
+                               model_config=None):
     """Generate images using stable-diffusion-cpp-python."""
     import time
 
@@ -965,8 +990,27 @@ async def _generate_with_sdcpp(sd_model, request, global_args, http_request=None
             except ValueError:
                 pass
 
+    # Acceleration/distillation defaults (Lightning / Turbo / LCM): sd.cpp can't
+    # fuse a diffusers LoRA, but honour the preset's low step-count / guidance and
+    # inject the distill LoRA via the "<lora:name:weight>" prompt syntax when a
+    # lora_model_dir is configured.
+    from codai.models.acceleration import resolve_acceleration
+    _accel = resolve_acceleration(model_config)
+    _accel_steps = _accel.get('steps') if _accel else None
+    _accel_cfg = _accel.get('guidance_scale') if _accel else None
     # Use default steps for fast generation
-    steps = request.steps if request.steps else 4
+    steps = request.steps if request.steps else (_accel_steps or 4)
+    cfg_scale = request.guidance_scale or _accel_cfg or get_cfg_scale()
+
+    prompt = request.prompt
+    if _accel and _accel.get('lora') and (model_config or {}).get('lora_model_dir'):
+        from codai.models.acceleration import _split_lora_ref
+        _repo, _wn = _split_lora_ref(_accel['lora'])
+        _lname = (_wn or _repo).rsplit('/', 1)[-1]
+        for _suf in ('.safetensors', '.ckpt', '.pt', '.bin'):
+            if _lname.endswith(_suf):
+                _lname = _lname[: -len(_suf)]
+        prompt = f"{prompt} <lora:{_lname}:{_accel.get('lora_weight') or 1.0}>"
 
     _progress_reset(steps)
 
@@ -979,11 +1023,11 @@ async def _generate_with_sdcpp(sd_model, request, global_args, http_request=None
     try:
         result = await asyncio.to_thread(
             sd_model.generate_image,
-            prompt=request.prompt,
+            prompt=prompt,
             negative_prompt='',
             width=width,
             height=height,
-            cfg_scale=get_cfg_scale(),
+            cfg_scale=cfg_scale,
             sample_steps=steps,
             seed=seed if seed is not None else 42,
             batch_count=request.n if request.n else 1,
@@ -992,11 +1036,11 @@ async def _generate_with_sdcpp(sd_model, request, global_args, http_request=None
     except TypeError:
         result = await asyncio.to_thread(
             sd_model.generate_image,
-            prompt=request.prompt,
+            prompt=prompt,
             negative_prompt='',
             width=width,
             height=height,
-            cfg_scale=get_cfg_scale(),
+            cfg_scale=cfg_scale,
             sample_steps=steps,
             seed=seed if seed is not None else 42,
             batch_count=request.n if request.n else 1,
@@ -1266,7 +1310,10 @@ async def create_image_generation(request: ImageGenerationRequest, http_request:
         if pipeline is not None:
             if is_sdcpp:
                 print(f"Using cached sd.cpp model for generation")
-                return await _generate_with_sdcpp(pipeline, request, global_args, http_request)
+                _sdcpp_cfg = (multi_model_manager.config.get(model_key)
+                              or multi_model_manager.config.get(model_name) or {})
+                return await _generate_with_sdcpp(pipeline, request, global_args,
+                                                  http_request, model_config=_sdcpp_cfg)
             else:
                 # Assume it's a diffusers pipeline
                 print(f"Using cached diffusers pipeline for generation")
@@ -1335,7 +1382,8 @@ async def create_image_generation(request: ImageGenerationRequest, http_request:
                         pass
                     print(f"Loaded sd.cpp model: {model_name}")
                     
-                    return await _generate_with_sdcpp(sd_model, request, global_args, http_request)
+                    return await _generate_with_sdcpp(sd_model, request, global_args,
+                                                      http_request, model_config=cfg)
             else:
                 sdcpp_error = f"Model '{model_name}' is not a local file, cannot use sd.cpp"
                 print(sdcpp_error)

codai/api/video.py

@@ -603,7 +603,27 @@ def _generate_sdcpp_video(sd_model, request, model_cfg=None):
     mode = request.mode or 't2v'
     fps = request.fps or 8
     num_frames = request.num_frames or 25
-    steps = request.num_inference_steps or 20
+
+    # Acceleration/distillation defaults (sd.cpp can't fuse a diffusers LoRA, but
+    # we can still honour the preset's low step-count / guidance, and inject the
+    # distill LoRA via sd.cpp's "<lora:name:weight>" prompt syntax when a
+    # lora_model_dir is configured).
+    from codai.models.acceleration import resolve_acceleration
+    _accel = resolve_acceleration(model_cfg)
+    _accel_steps = _accel.get('steps') if _accel else None
+    _accel_cfg = _accel.get('guidance_scale') if _accel else None
+    steps = request.num_inference_steps or _accel_steps or 20
+    cfg_scale = request.guidance_scale or _accel_cfg or 7.0
+
+    prompt = request.prompt or ''
+    if _accel and _accel.get('lora') and (model_cfg or {}).get('lora_model_dir'):
+        from codai.models.acceleration import _split_lora_ref
+        _repo, _wn = _split_lora_ref(_accel['lora'])
+        _lname = (_wn or _repo).rsplit('/', 1)[-1]
+        for _suf in ('.safetensors', '.ckpt', '.pt', '.bin'):
+            if _lname.endswith(_suf):
+                _lname = _lname[: -len(_suf)]
+        prompt = f"{prompt} <lora:{_lname}:{_accel.get('lora_weight') or 1.0}>"
 
     _vid_progress_reset(steps)
 
@@ -611,13 +631,13 @@ def _generate_sdcpp_video(sd_model, request, model_cfg=None):
         _vid_progress_step(step)
 
     kw = {
-        'prompt':          request.prompt or '',
+        'prompt':          prompt,
         'negative_prompt': request.negative_prompt or '',
         'width':           request.width or 512,
         'height':          request.height or 512,
         'video_frames':    num_frames,
         'sample_steps':    steps,
-        'cfg_scale':       request.guidance_scale or 7.0,
+        'cfg_scale':       cfg_scale,
         'seed':            request.seed if request.seed is not None else -1,
         'progress_callback': _progress_cb,
     }
@@ -933,7 +953,7 @@ def _load_video_pipeline(model_name: str, device: str, mode: str, offload: str =
         # RAM (and disk if needed). This is the preferred strategy when the
         # model won't fit entirely in VRAM but should maximise GPU utilisation.
         # `gpu_percent` (0–100) controls what fraction of FREE VRAM to occupy.
-        _gpu_pct = float((model_cfg or {}).get('balanced_gpu_percent', 80))
+        _gpu_pct = float((model_cfg or {}).get('balanced_gpu_percent') or 80)
         try:
             if torch.cuda.is_available():
                 _free_v, _ = torch.cuda.mem_get_info()
@@ -1346,6 +1366,16 @@ def _generate_video(pipe, request: VideoGenerationRequest):
                              else 'v2v' if request.video else 't2v')
     fps = request.fps or 8
     kw = _build_call_kwargs(request)
+    # Acceleration/distillation defaults (Lightning / Lightx2v): when the model has
+    # a fused distill LoRA, default to its low step-count / guidance instead of the
+    # standard 25 steps / 7.5 CFG. The request always wins if it set these — note
+    # _build_call_kwargs only populates them when the request specified them, so
+    # setdefault below correctly leaves an explicit request value untouched.
+    _accel = getattr(pipe, '_coderai_accel', None)
+    if _accel:
+        from codai.models.acceleration import accel_call_defaults
+        for _k, _v in accel_call_defaults(_accel).items():
+            kw.setdefault(_k, _v)
     kw.setdefault('num_inference_steps', 25)
     kw.setdefault('guidance_scale', 7.5)
     kw.setdefault('num_frames', 16)
@@ -1940,7 +1970,7 @@ async def video_generations(request: VideoGenerationRequest,
                     _need_gb = multi_model_manager._get_model_used_vram_gb(
                         model_key, model_name)
                     if _need_gb > 0 and _free_gb < _need_gb:
-                        _gpu_pct = float(_model_cfg.get('balanced_gpu_percent', 80))
+                        _gpu_pct = float(_model_cfg.get('balanced_gpu_percent') or 80)
                         print(f"  VRAM insufficient for full-GPU load "
                               f"({_need_gb:.1f} GB needed, {_free_gb:.1f} GB free) "
                               f"— auto-selecting balanced strategy "
@@ -1961,6 +1991,24 @@ async def video_generations(request: VideoGenerationRequest,
                     "video model. Restart coderai to recover. "
                     f"Original error: {str(e).splitlines()[0]}"))
             raise HTTPException(status_code=500, detail=f"Failed to load video model: {e}")
+        # Fuse any configured acceleration/distillation LoRA (Lightning / Lightx2v /
+        # LCM) into the freshly loaded pipeline. Done once at load; cached pipes keep
+        # it. No-op for sd.cpp pipes and when no acceleration is configured.
+        try:
+            from codai.models.acceleration import resolve_acceleration, apply_accel_to_pipeline
+            _accel = resolve_acceleration(_model_cfg)
+            _accel_is_sdcpp = False
+            try:
+                from stable_diffusion_cpp import StableDiffusion as _SDc
+                _accel_is_sdcpp = isinstance(pipe, _SDc)
+            except ImportError:
+                pass
+            if _accel and not _accel_is_sdcpp:
+                print(f"  [video][accel] applying {_accel.get('preset')} "
+                      f"(steps={_accel.get('steps')}, guidance={_accel.get('guidance_scale')})")
+                apply_accel_to_pipeline(pipe, _accel)
+        except Exception as _e:
+            print(f"  [video][accel] skipped: {_e}")
         multi_model_manager.models[model_key] = pipe
         multi_model_manager.current_model_key = model_key
         # Record the real VRAM used. record_vram_delta only persists when no

codai/main.py

@@ -160,6 +160,7 @@ def build_runtime_kwargs(model_cfg, model_type):
         kwargs['seed'] = model_cfg.get('seed')
         kwargs['vae_tiling'] = model_cfg.get('vae_tiling', False)
         kwargs['clip_on_cpu'] = model_cfg.get('clip_on_cpu', False)
+        kwargs['acceleration'] = model_cfg.get('acceleration')
     elif model_type == "audio":
         kwargs['ctx'] = model_cfg.get('context_ms')
         kwargs['offload'] = model_cfg.get('offload') or model_cfg.get('offload_strategy')
@@ -172,6 +173,7 @@ def build_runtime_kwargs(model_cfg, model_type):
         kwargs['vae_tiling'] = model_cfg.get('vae_tiling', True)
         kwargs['balanced_gpu_percent'] = model_cfg.get('balanced_gpu_percent', 80)
         kwargs['output_crf'] = model_cfg.get('output_crf')
+        kwargs['acceleration'] = model_cfg.get('acceleration')
     return kwargs