ollama/ollama

Three canonical flows: (1) serve startup: main -> cobra serve -> initializeKeypair -> net.Listen -> server.Serve -> blob prune -> GenerateRoutes -> InitScheduler.Run; (2) `ollama run X prompt`: CLI heartbeat/pull -> api.Client.Chat -> ChatHandler -> GetModel -> getRunner -> load -> llama-server subprocess -> streamed parse -> NDJSON; (3) pull: PullHandler -> PullModel -> blob downloads -> manifest write. Safetensors models diverge at sched.load into the mlxrunner subprocess.

• Startup order: main() -> NewCLI() -> serveCmd -> RunServer -> initializeKeypair (ed25519 at ~/.ollama) -> net.Listen(envconfig.Host) -> server.Serve(ln)
  main.gocmd/cmd.go#L2101-L2117cmd/cmd.go#L2119-L2163
• server.Serve sequence: slog logger init -> log cloud-disabled status -> llm.PruneSplitModelDirs on blobs dir -> fixBlobs -> PruneLayers + manifest.PruneDirectory (unless OLLAMA_NOPRUNE) -> GenerateRoutes -> InitScheduler -> sched.Run goroutines -> http.Serve
  server/routes.go#L1969-L2031server/sched.go#L91-L106server/sched.go#L219-L220
• CLI run flow: RunHandler parses --think/--keepalive/--format flags -> api.ClientFromEnvironment -> showOrPullModel (pulls if absent) -> inferThinkingOption -> isEmbeddingModel branch -> generateChat/generateInteractive -> client.Chat -> POST /api/chat
  cmd/cmd.go#L814-L953api/client.go#L74-L79
• Server chat flow: ChatHandler binds JSON -> parseAndValidateModelRef -> cloud refs proxied via proxyCloudJSONRequest -> GetModel -> keep_alive=0 unloads via sched.expireRunner -> remote models proxied to RemoteHost -> otherwise modelOptions + scheduleRunner
  server/routes.go#L2465-L2553server/routes.go#L2555-L2621server/routes.go#L205-L253
• Scheduling flow: getRunner builds LlmRequest -> loaded-map fast path via schedulerModelKey (modelPaths joined with NUL, or digest:/name: fallback) -> else enqueue pendingReqCh (ErrMaxQueue when full) -> scheduler picks GPUs -> load() spawns runner -> useLoadedRunner
  server/sched.go#L108-L132server/sched.go#L174-L216server/sched.go#L520-L632
• GGUF runner launch: sched.load -> llm.LoadModel (gguf.ReadModel metadata) -> PredictServerVRAM -> selectLlamaServerPlacement -> newServerFn = llm.NewLlamaServer -> NewLlamaServerRunner -> startLlamaServer -> exec.Command(llama-server, params) -> WaitUntilRunning polls /health
  server/sched.go#L528-L600llm/server.go#L93-L111llm/llama_server.go#L348llm/llama_server.go#L424llm/llama_server.go#L851
• MLX runner launch: sched.load -> m.IsMLX() -> mlxrunner.NewClient(modelName, numCtx) -> exec.Command(self, "runner", --model, --port) -> hidden cobra `runner` cmd -> mlxrunner.Execute -> HTTP server inside subprocess
  server/sched.go#L591-L594mlxrunner/client.go#L349cmd/cmd.go#L2569-L2576
• Pull trace: PullHandler binds api.PullRequest -> parseNormalizePullModelRef (normalizes :cloud) -> PullModel in goroutine -> per-layer blobDownload (chunked parts with backoff) -> manifest write -> progress channel -> streamResponse emits api.ProgressResponse NDJSON
  server/routes.go#L1135-L1193server/images.go#L1002server/download.go#L187-L333
• Create trace (Modelfile): CreateHandler -> readCreateModelfile -> parser.ParseFile -> reject adapter/typical_p -> modelfile.CreateRequest -> createBlob uploads per file (or sharedBlobStore local shortcut) -> client.Create -> /api/create -> server assembles layers + manifest
  cmd/cmd.go#L220-L241cmd/cmd.go#L317-L400parser/parser.go#L56-L120server/routes.go#L1917
• Unload/keep-alive: empty chat request with keep_alive=0 triggers sched.expireRunner and returns done_reason=unload without inference; runners expire via expiredCh after sessionDuration (OLLAMA_KEEP_ALIVE default 5m)
  server/routes.go#L2541-L2553server/sched.go#L61-L82envconfig/config.go#L126-L144
• OOM resilience: on llm.IsOutOfMemory during Load, scheduler reduces automatically-derived num_ctx/num_batch once (oomRetryAttempted) and retries; evict-all-and-retry is guarded against infinite loops
  server/sched.go#L39-L42server/sched.go#L632-L659