llm: New memory management

This changes the memory allocation strategy from upfront estimation to
tracking actual allocations done by the engine and reacting to that. The
goal is avoid issues caused by both under-estimation (crashing) and
over-estimation (low performance due to under-utilized GPUs).

It is currently opt-in and can be enabled for models running on the
Ollama engine by setting OLLAMA_NEW_ESTIMATES=1. Behavior in other
cases is unchanged and will continue to use the existing estimates.

server/routes.go

@@ -1477,14 +1477,14 @@ func (s *Server) PsHandler(c *gin.Context) {
 		mr := api.ProcessModelResponse{
 			Model:     model.ShortName,
 			Name:      model.ShortName,
-			Size:      int64(v.estimatedTotal),
-			SizeVRAM:  int64(v.estimatedVRAM),
+			Size:      int64(v.totalSize),
+			SizeVRAM:  int64(v.vramSize),
 			Digest:    model.Digest,
 			Details:   modelDetails,
 			ExpiresAt: v.expiresAt,
 		}
 		if v.Options != nil {
-			mr.ContextLength = v.Options.NumCtx / v.numParallel
+			mr.ContextLength = v.Options.NumCtx
 		}
 		// The scheduler waits to set expiresAt, so if a model is loading it's
 		// possible that it will be set to the unix epoch. For those cases, just