Add per-model GPU and context profiles

backend/app/routes/chat.py

@@ -42,7 +42,7 @@ async def chat(req: ChatRequest) -> StreamingResponse:
         raise HTTPException(404, "session introuvable")
 
     model = req.model or session.get("model") or settings.default_model
-    cfg = agent_config.get_config()
+    cfg = agent_config.get_config(model)
 
     # Premier message : titre la session avec un extrait.
     if not db.list_messages(req.session_id):

backend/app/routes/config.py

@@ -1,11 +1,10 @@
-"""Routes de configuration de l'agent (lecture / mise à jour / auto-réglage)."""
+"""Routes de lecture et mise à jour de la configuration de l'agent."""
 from __future__ import annotations
 
 from fastapi import APIRouter
 from pydantic import BaseModel
 
-from .. import agent_config, autotune
-from ..config import settings
+from .. import agent_config
 
 router = APIRouter(prefix="/api/config", tags=["config"])
 
@@ -17,48 +16,21 @@ class ConfigPatch(BaseModel):
     top_k: int | None = None
     max_tokens: int | None = None
     num_ctx: int | None = None
+    num_gpu: int | None = None
+    num_batch: int | None = None
     tools: dict[str, bool] | None = None
     confirm_shell: bool | None = None
 
 
-class AutoTuneRequest(BaseModel):
-    model: str
-    apply: bool = True
-
-
 @router.get("")
-async def get_config() -> dict:
+async def get_config(model: str | None = None) -> dict:
     return {
-        "config": agent_config.get_config(),
+        "config": agent_config.get_config(model),
         "available_tools": agent_config.AVAILABLE_TOOLS,
     }
 
 
 @router.put("")
-async def put_config(patch: ConfigPatch) -> dict:
-    cfg = agent_config.save_config(patch.model_dump(exclude_none=True))
+async def put_config(patch: ConfigPatch, model: str | None = None) -> dict:
+    cfg = agent_config.save_config(patch.model_dump(exclude_none=True), model)
     return {"config": cfg}
-
-
-@router.post("/auto")
-async def auto_tune(req: AutoTuneRequest) -> dict:
-    """Détecte le GPU + le modèle, calcule les réglages optimaux, et (par
-    défaut) les applique. Renvoie le détail de la détection pour l'UI."""
-    reco = await autotune.recommend(req.model)
-    placement = await autotune.placement(req.model)
-
-    applied = None
-    if req.apply:
-        applied = agent_config.save_config(
-            {
-                "num_ctx": reco["recommended"]["num_ctx"],
-                "max_tokens": reco["recommended"]["max_tokens"],
-            }
-        )
-
-    return {
-        "detection": reco,
-        "placement": placement,
-        "config": applied or agent_config.get_config(),
-        "vram_override": settings.gpu_vram_mb,
-    }