Separate vision projector device from LLM GPU split

inference.py

@@ -158,6 +158,7 @@ class LlamaWorker:
             if params['split_mode']=='none':selected=selected[:1]
         if not selected or any(g['processes'] or g['free_mib']<2048 for g in selected):raise InferenceError('Benötigte GPU ist durch einen anderen Dienst belegt. Deck stoppt keine fremden Prozesse.')
         mem={line.split(':')[0]:int(line.split()[1])*1024 for line in Path('/proc/meminfo').read_text().splitlines() if line.startswith('MemAvailable:')}
+        model_devices=list(selected)
         projector=None;vision_args=[];vision_bytes=0;vision_device=params.get('vision_device','cpu')
         if params.get('vision_projector'):
             projector=self.catalog.entry(params['vision_projector'])
@@ -166,7 +167,11 @@ class LlamaWorker:
             vision_args=['--mmproj',str((self.catalog.root/projector['id']/'model.gguf').resolve())]
             if vision_device=='cpu':vision_args+=['--no-mmproj-offload']
             else:
-                if vision_device not in [g['uuid'] for g in selected]:raise InferenceError('Projektor-GPU ist nicht ausgewählt.')
+                if vision_device not in [g['uuid'] for g in selected]:
+                    target=next((g for g in devices if g['uuid']==vision_device),None)
+                    if target is None:raise InferenceError('Projektor-GPU ist nicht verfügbar.')
+                    if target['processes'] or target['free_mib']<2048:raise InferenceError('Projektor-GPU ist durch einen anderen Dienst belegt.')
+                    selected.append(target)
                 vision_args+=['--mmproj-device','CUDA'+str([g['uuid'] for g in selected].index(vision_device)),'--mmproj-offload']
         staging=entry['size']+4*GIB+vision_bytes
         if mem.get('MemAvailable',0)<staging+4*GIB:raise InferenceError('Zu wenig freier RAM für Modellladen und Host-Reserve.')
@@ -177,7 +182,8 @@ class LlamaWorker:
         env.update(CUDA_VISIBLE_DEVICES=','.join(g['uuid'] for g in selected),HF_HUB_OFFLINE='1',OMP_NUM_THREADS=str(params['threads']))
         reserve_mode=params.get('gpu_reserve_mode','auto')
         margins=[0 if reserve_mode=='none' else params.get('gpu_reserve_mib',{}).get(g['uuid'],512) if reserve_mode=='manual' else max(512,round(g['total_mib']*.025)) for g in selected]
-        common=['--model',str(model),'--ctx-size',str(params['context']),'--parallel',str(params['slots']),'--batch-size',str(params['batch']),'--ubatch-size',str(params['ubatch']),'--cache-type-k',params.get('cache_type_k','q4_0'),'--cache-type-v',params.get('cache_type_v','q4_0'),'--flash-attn','on','--split-mode',params['split_mode'],'--fit-target',','.join(map(str,margins))]
+        common=['--model',str(model),'--ctx-size',str(params['context']),'--parallel',str(params['slots']),'--batch-size',str(params['batch']),'--ubatch-size',str(params['ubatch']),'--cache-type-k',params.get('cache_type_k','q4_0'),'--cache-type-v',params.get('cache_type_v','q4_0'),'--flash-attn','on','--split-mode',params['split_mode'],'--fit-target',','.join(map(str,margins[:len(model_devices)]))]
+        common+=['--device',','.join('CUDA'+str(i) for i in range(len(model_devices)))]
         if params['tensor_split']:common+=['--tensor-split',','.join(map(str,params['tensor_split']))]
         tool=str(directory/'build/bin/llama-fit-params')
         mtp_fit=['--deck-mtp'] if mtp else []
@@ -188,7 +194,8 @@ class LlamaWorker:
             for line in output.splitlines():
                 parts=line.split()
                 if len(parts)==4 and all(x.isdigit() for x in parts[1:]):rows[parts[0]]=sum(map(int,parts[1:]))
-            if 'Host' not in rows or any('CUDA'+str(i) not in rows for i in range(len(selected))):raise InferenceError('Speicherprognose unvollständig; Start abgebrochen.')
+            if 'Host' not in rows or any('CUDA'+str(i) not in rows for i in range(len(model_devices))):raise InferenceError('Speicherprognose unvollständig; Start abgebrochen.')
+            for i in range(len(model_devices),len(selected)):rows.setdefault('CUDA'+str(i),0)
             if projector:
                 target='Host' if vision_device=='cpu' else 'CUDA'+str([g['uuid'] for g in selected].index(vision_device))
                 rows[target]+=round(vision_bytes/1024**2)
@@ -226,7 +233,7 @@ class LlamaWorker:
             layers=int(fit['-ngl'])
             if '-ts' in fit:
                 parts=fit['-ts'].split(',')
-                if len(parts)!=len(selected) or any(not x.replace('.','',1).isdigit() for x in parts):raise InferenceError('Ungültige automatische GPU-Verteilung.')
+                if len(parts)!=len(model_devices) or any(not x.replace('.','',1).isdigit() for x in parts):raise InferenceError('Ungültige automatische GPU-Verteilung.')
                 extra=['--tensor-split',fit['-ts']]
             fits,memory=estimate(layers,extra)
             if not fits:raise InferenceError('Speicherprognose überschreitet die GPU-/RAM-Reserve.')