fix: SentenceTransformer meta tensor avec accelerate + torch 2.10

low_cpu_mem_usage=False évite les meta tensors lors du chargement
de l'embedding (sentence-transformers 5.x + accelerate 1.12).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

src/medical/rag_search.py

@@ -35,21 +35,27 @@ def _get_embed_model():
     """Charge le modèle d'embedding (singleton).
 
     Tente CUDA d'abord, fallback CPU si OOM (Ollama peut occuper la VRAM).
+    low_cpu_mem_usage=False évite les meta tensors (accelerate + sentence-transformers 5.x).
     """
     global _embed_model
     if _embed_model is None:
         from sentence_transformers import SentenceTransformer
         import torch
         _device = "cuda" if torch.cuda.is_available() else "cpu"
+        _model_kwargs = {"low_cpu_mem_usage": False}
         try:
             logger.info("Chargement du modèle d'embedding (%s)...", _device)
-            _embed_model = SentenceTransformer(EMBEDDING_MODEL, device=_device)
+            _embed_model = SentenceTransformer(
+                EMBEDDING_MODEL, device=_device, model_kwargs=_model_kwargs,
+            )
         except (torch.OutOfMemoryError, torch.cuda.CudaError, torch.AcceleratorError, RuntimeError) as exc:
             exc_msg = str(exc).lower()
             if _device == "cuda" and ("memory" in exc_msg or "meta tensor" in exc_msg):
                 logger.warning("CUDA erreur pour l'embedding — fallback CPU : %s", exc)
                 torch.cuda.empty_cache()
-                _embed_model = SentenceTransformer(EMBEDDING_MODEL, device="cpu")
+                _embed_model = SentenceTransformer(
+                    EMBEDDING_MODEL, device="cpu", model_kwargs=_model_kwargs,
+                )
             else:
                 raise
         _embed_model.max_seq_length = 512