mlxrunner: Report actual memory usage from runner

The MLX runner previously reported a static VRAM estimate that was
computed at load time and consisted only of the weights. This is
strictly less than the actual memory usage, as it does not include
the KV cache or compute graph.

server/routes.go

@@ -1952,6 +1952,9 @@ func (s *Server) PsHandler(c *gin.Context) {
 		}
 		if v.llama != nil {
 			mr.ContextLength = v.llama.ContextLength()
+			total, vram := v.llama.MemorySize()
+			mr.Size = int64(total)
+			mr.SizeVRAM = int64(vram)
 		}
 		// 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

server/sched.go

@@ -536,6 +536,7 @@ iGPUScan:
 		}
 	}
 
+	totalSize, vramSize := llama.MemorySize()
 	runner := &runnerRef{
 		model:           req.model,
 		modelPath:       req.model.ModelPath,
@@ -545,8 +546,8 @@ iGPUScan:
 		sessionDuration: sessionDuration,
 		gpus:            gpuIDs,
 		discreteGPUs:    discreteGPUs,
-		vramSize:        llama.VRAMSize(),
-		totalSize:       llama.TotalSize(),
+		totalSize:       totalSize,
+		vramSize:        vramSize,
 		loading:         true,
 		pid:             llama.Pid(),
 	}
@@ -619,6 +620,7 @@ func (s *Scheduler) loadMLX(req *LlmRequest) bool {
 		sessionDuration = req.sessionDuration.Duration
 	}
 
+	totalSize, vramSize := server.MemorySize()
 	runner := &runnerRef{
 		model:           req.model,
 		modelPath:       req.model.ModelPath,
@@ -628,8 +630,8 @@ func (s *Scheduler) loadMLX(req *LlmRequest) bool {
 		loading:         false,
 		isImagegen:      isImagegen,
 		sessionDuration: sessionDuration,
-		totalSize:       server.TotalSize(),
-		vramSize:        server.VRAMSize(),
+		totalSize:       totalSize,
+		vramSize:        vramSize,
 	}
 
 	s.loadedMu.Lock()

server/sched_test.go

@@ -861,8 +861,7 @@ func (s *mockLlm) Close() error {
 	s.closeCalled = true
 	return s.closeResp
 }
-func (s *mockLlm) VRAMSize() uint64                                   { return s.vramSize }
-func (s *mockLlm) TotalSize() uint64                                  { return s.totalSize }
+func (s *mockLlm) MemorySize() (uint64, uint64)                       { return s.totalSize, s.vramSize }
 func (s *mockLlm) VRAMByGPU(id ml.DeviceID) uint64                    { return s.vramByGPU[id] }
 func (s *mockLlm) Pid() int                                           { return -1 }
 func (s *mockLlm) GetPort() int                                       { return -1 }