From 17b7186cd759337fa98b626e82de150f3789b040 Mon Sep 17 00:00:00 2001 From: Daniel Hiltgen Date: Mon, 6 May 2024 17:47:52 -0700 Subject: [PATCH 01/31] Enable concurrency by default This adjusts our default settings to enable multiple models and parallel requests to a single model. Users can still override these by the same env var settings as before. Parallel has a direct impact on num_ctx, which in turn can have a significant impact on small VRAM GPUs so this change also refines the algorithm so that when parallel is not explicitly set by the user, we try to find a reasonable default that fits the model on their GPU(s). As before, multiple models will only load concurrently if they fully fit in VRAM. --- envconfig/config.go | 16 ++++---- llm/server.go | 13 ++---- server/sched.go | 98 +++++++++++++++++++++++++++++++++----------- server/sched_test.go | 80 +++++++++++++++++++++++------------- 4 files changed, 135 insertions(+), 72 deletions(-) diff --git a/envconfig/config.go b/envconfig/config.go index e86f72e6a..cb456448c 100644 --- a/envconfig/config.go +++ b/envconfig/config.go @@ -85,13 +85,13 @@ func AsMap() map[string]EnvVar { "OLLAMA_HOST": {"OLLAMA_HOST", Host, "IP Address for the ollama server (default 127.0.0.1:11434)"}, "OLLAMA_KEEP_ALIVE": {"OLLAMA_KEEP_ALIVE", KeepAlive, "The duration that models stay loaded in memory (default \"5m\")"}, "OLLAMA_LLM_LIBRARY": {"OLLAMA_LLM_LIBRARY", LLMLibrary, "Set LLM library to bypass autodetection"}, - "OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners, "Maximum number of loaded models (default 1)"}, + "OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners, "Maximum number of loaded models per GPU (default 4)"}, "OLLAMA_MAX_QUEUE": {"OLLAMA_MAX_QUEUE", MaxQueuedRequests, "Maximum number of queued requests"}, "OLLAMA_MAX_VRAM": {"OLLAMA_MAX_VRAM", MaxVRAM, "Maximum VRAM"}, "OLLAMA_MODELS": {"OLLAMA_MODELS", ModelsDir, "The path to the models directory"}, "OLLAMA_NOHISTORY": {"OLLAMA_NOHISTORY", NoHistory, "Do not preserve readline history"}, "OLLAMA_NOPRUNE": {"OLLAMA_NOPRUNE", NoPrune, "Do not prune model blobs on startup"}, - "OLLAMA_NUM_PARALLEL": {"OLLAMA_NUM_PARALLEL", NumParallel, "Maximum number of parallel requests (default 1)"}, + "OLLAMA_NUM_PARALLEL": {"OLLAMA_NUM_PARALLEL", NumParallel, "Maximum number of parallel requests"}, "OLLAMA_ORIGINS": {"OLLAMA_ORIGINS", AllowOrigins, "A comma separated list of allowed origins"}, "OLLAMA_RUNNERS_DIR": {"OLLAMA_RUNNERS_DIR", RunnersDir, "Location for runners"}, "OLLAMA_SCHED_SPREAD": {"OLLAMA_SCHED_SPREAD", SchedSpread, "Always schedule model across all GPUs"}, @@ -129,8 +129,8 @@ func clean(key string) string { func init() { // default values - NumParallel = 1 - MaxRunners = 1 + NumParallel = 0 + MaxRunners = 4 MaxQueuedRequests = 512 LoadConfig() @@ -205,8 +205,8 @@ func LoadConfig() { if onp := clean("OLLAMA_NUM_PARALLEL"); onp != "" { val, err := strconv.Atoi(onp) - if err != nil || val <= 0 { - slog.Error("invalid setting must be greater than zero", "OLLAMA_NUM_PARALLEL", onp, "error", err) + if err != nil { + slog.Error("invalid setting, ignoring", "OLLAMA_NUM_PARALLEL", onp, "error", err) } else { NumParallel = val } @@ -251,7 +251,7 @@ func LoadConfig() { if maxRunners != "" { m, err := strconv.Atoi(maxRunners) if err != nil { - slog.Error("invalid setting", "OLLAMA_MAX_LOADED_MODELS", maxRunners, "error", err) + slog.Error("invalid setting, ignoring", "OLLAMA_MAX_LOADED_MODELS", maxRunners, "error", err) } else { MaxRunners = m } @@ -260,7 +260,7 @@ func LoadConfig() { if onp := os.Getenv("OLLAMA_MAX_QUEUE"); onp != "" { p, err := strconv.Atoi(onp) if err != nil || p <= 0 { - slog.Error("invalid setting", "OLLAMA_MAX_QUEUE", onp, "error", err) + slog.Error("invalid setting, ignoring", "OLLAMA_MAX_QUEUE", onp, "error", err) } else { MaxQueuedRequests = p } diff --git a/llm/server.go b/llm/server.go index da83416ee..3cb5ac1f0 100644 --- a/llm/server.go +++ b/llm/server.go @@ -77,7 +77,7 @@ func LoadModel(model string) (*GGML, error) { // NewLlamaServer will run a server for the given GPUs // The gpu list must be a single family. -func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, projectors []string, opts api.Options) (LlamaServer, error) { +func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, projectors []string, opts api.Options, numParallel int) (LlamaServer, error) { var err error var cpuRunner string var estimate MemoryEstimate @@ -213,8 +213,10 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr // Windows CUDA should not use mmap for best performance // Linux with a model larger than free space, mmap leads to thrashing + // For CPU loads we want the memory to be allocated, not FS cache if (runtime.GOOS == "windows" && gpus[0].Library == "cuda" && opts.UseMMap == api.TriStateUndefined) || (runtime.GOOS == "linux" && systemFreeMemory < estimate.TotalSize && opts.UseMMap == api.TriStateUndefined) || + (gpus[0].Library == "cpu" && opts.UseMMap == api.TriStateUndefined) || opts.UseMMap == api.TriStateFalse { params = append(params, "--no-mmap") } @@ -227,15 +229,6 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr params = append(params, "--numa") } - numParallel := envconfig.NumParallel - - // TODO (jmorganca): multimodal models don't support parallel yet - // see https://github.com/ollama/ollama/issues/4165 - if len(projectors) > 0 { - numParallel = 1 - slog.Warn("multimodal models don't support parallel requests yet") - } - params = append(params, "--parallel", fmt.Sprintf("%d", numParallel)) if estimate.TensorSplit != "" { diff --git a/server/sched.go b/server/sched.go index 424395544..31ef560f5 100644 --- a/server/sched.go +++ b/server/sched.go @@ -23,6 +23,7 @@ type LlmRequest struct { ctx context.Context //nolint:containedctx model *Model opts api.Options + origNumCTX int // Track the initial ctx request sessionDuration time.Duration successCh chan *runnerRef errCh chan error @@ -38,8 +39,8 @@ type Scheduler struct { loaded map[string]*runnerRef loadedMu sync.Mutex - loadFn func(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) - newServerFn func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options) (llm.LlamaServer, error) + loadFn func(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel int) + newServerFn func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error) getGpuFn func() gpu.GpuInfoList getCpuFn func() gpu.GpuInfoList reschedDelay time.Duration @@ -65,13 +66,10 @@ func InitScheduler(ctx context.Context) *Scheduler { // context must be canceled to decrement ref count and release the runner func (s *Scheduler) GetRunner(c context.Context, model *Model, opts api.Options, sessionDuration time.Duration) (chan *runnerRef, chan error) { - // allocate a large enough kv cache for all parallel requests if opts.NumCtx < 4 { opts.NumCtx = 4 } - opts.NumCtx *= envconfig.NumParallel - req := &LlmRequest{ ctx: c, model: model, @@ -102,6 +100,7 @@ func (s *Scheduler) Run(ctx context.Context) { } func (s *Scheduler) processPending(ctx context.Context) { + maxRunnerFactor := 1 // number of GPUs or 1 for { select { case <-ctx.Done(): @@ -110,11 +109,25 @@ func (s *Scheduler) processPending(ctx context.Context) { case pending := <-s.pendingReqCh: // Block other requests until we get this pending request running pending.schedAttempts++ + if pending.origNumCTX == 0 { + pending.origNumCTX = pending.opts.NumCtx + } if pending.ctx.Err() != nil { slog.Debug("pending request cancelled or timed out, skipping scheduling") continue } + numParallel := envconfig.NumParallel + // TODO (jmorganca): multimodal models don't support parallel yet + // see https://github.com/ollama/ollama/issues/4165 + if len(pending.model.ProjectorPaths) > 0 && numParallel != 1 { + numParallel = 1 + slog.Warn("multimodal models don't support parallel requests yet") + } + // Keep NumCtx and numParallel in sync + if numParallel > 1 { + pending.opts.NumCtx = pending.origNumCTX * numParallel + } for { var runnerToExpire *runnerRef @@ -130,7 +143,7 @@ func (s *Scheduler) processPending(ctx context.Context) { pending.useLoadedRunner(runner, s.finishedReqCh) break } - } else if envconfig.MaxRunners > 0 && loadedCount >= envconfig.MaxRunners { + } else if envconfig.MaxRunners > 0 && loadedCount >= (maxRunnerFactor*envconfig.MaxRunners) { slog.Debug("max runners achieved, unloading one to make room", "runner_count", loadedCount) runnerToExpire = s.findRunnerToUnload() } else { @@ -142,6 +155,7 @@ func (s *Scheduler) processPending(ctx context.Context) { } else { gpus = s.getGpuFn() } + maxRunnerFactor = max(len(gpus), 1) // Load model for fitting ggml, err := llm.LoadModel(pending.model.ModelPath) @@ -152,26 +166,32 @@ func (s *Scheduler) processPending(ctx context.Context) { // Evaluate if the model will fit in the available system memory, or if we should unload a model first if len(gpus) == 1 && gpus[0].Library == "cpu" { + // simplifying assumption of defaultParallel when in CPU mode + if numParallel <= 0 { + numParallel = defaultParallel + pending.opts.NumCtx = pending.origNumCTX * numParallel + } + if loadedCount == 0 { slog.Debug("cpu mode with first model, loading") - s.loadFn(pending, ggml, gpus) + s.loadFn(pending, ggml, gpus, numParallel) break } runnerToExpire = s.maybeFindCPURunnerToUnload(pending, ggml, gpus) if runnerToExpire == nil { slog.Debug("cpu mode with available system memory or first model, loading") - s.loadFn(pending, ggml, gpus) + s.loadFn(pending, ggml, gpus, numParallel) break } // else we need to expire a runner } else if loadedCount == 0 { // No models loaded. Load the model but prefer the best fit. slog.Debug("loading first model", "model", pending.model.ModelPath) - g := pickBestFitGPUs(pending, ggml, gpus) + g := pickBestFitGPUs(pending, ggml, gpus, &numParallel) if g != nil { gpus = g } - s.loadFn(pending, ggml, gpus) + s.loadFn(pending, ggml, gpus, numParallel) break } @@ -186,10 +206,10 @@ func (s *Scheduler) processPending(ctx context.Context) { // Update free memory from currently loaded models s.updateFreeSpace(availGpus) - fitGpus := pickBestFitGPUs(pending, ggml, availGpus) + fitGpus := pickBestFitGPUs(pending, ggml, availGpus, &numParallel) if fitGpus != nil { slog.Debug("new model fits with existing models, loading") - s.loadFn(pending, ggml, fitGpus) + s.loadFn(pending, ggml, fitGpus, numParallel) break } @@ -350,8 +370,11 @@ func (pending *LlmRequest) useLoadedRunner(runner *runnerRef, finished chan *Llm }() } -func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) { - llama, err := s.newServerFn(gpus, req.model.ModelPath, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts) +func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel int) { + if numParallel < 1 { + numParallel = 1 + } + llama, err := s.newServerFn(gpus, req.model.ModelPath, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts, numParallel) if err != nil { // some older models are not compatible with newer versions of llama.cpp // show a generalized compatibility error until there is a better way to @@ -375,6 +398,7 @@ func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) loading: true, refCount: 1, } + runner.numParallel = numParallel runner.refMu.Lock() s.loadedMu.Lock() @@ -483,8 +507,9 @@ type runnerRef struct { expireTimer *time.Timer expiresAt time.Time - model *Model - modelPath string + model *Model + modelPath string + numParallel int *api.Options } @@ -525,6 +550,9 @@ func (runner *runnerRef) needsReload(ctx context.Context, req *LlmRequest) bool optsNew.NumGPU = -1 } + // Normalize the NumCtx for parallelism + optsExisting.NumCtx = optsExisting.NumCtx / runner.numParallel + ctx, cancel := context.WithTimeout(ctx, timeout) defer cancel() if !reflect.DeepEqual(runner.model.AdapterPaths, req.model.AdapterPaths) || // have the adapters changed? @@ -611,22 +639,38 @@ func (a ByDuration) Less(i, j int) bool { // pickBestFitGPUs will try to find the optimal placement of the model in the available GPUs where the model fully fits // If the model can not be fit fully within the available GPU(s) nil is returned -func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) gpu.GpuInfoList { +// If numParallel is <= 0, this will attempt try to optimize parallism based on available VRAM, and adjust +// opts.NumCtx accordingly +func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel *int) gpu.GpuInfoList { var estimatedVRAM uint64 + + var numParallelToTry []int + if *numParallel <= 0 { + // If no specific parallel setting was provided, try larger then smaller, always end with 1 + numParallelToTry = append(numParallelToTry, 4, 1) + } else { + numParallelToTry = []int{*numParallel} + } + for _, gl := range gpus.ByLibrary() { var ok bool sgl := append(make(gpu.GpuInfoList, 0, len(gl)), gl...) // TODO - potentially sort by performance capability, existing models loaded, etc. + // TODO - Eliminate any GPUs that already have envconfig.MaxRunners loaded on them // Note: at present, this will favor more VRAM over faster GPU speed in mixed setups sort.Sort(sort.Reverse(gpu.ByFreeMemory(sgl))) // First attempt to fit the model into a single GPU - if !envconfig.SchedSpread { - for _, g := range sgl { - if ok, estimatedVRAM = llm.PredictServerFit([]gpu.GpuInfo{g}, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok { - slog.Debug("new model will fit in available VRAM in single GPU, loading", "model", req.model.ModelPath, "gpu", g.ID, "available", g.FreeMemory, "required", format.HumanBytes2(estimatedVRAM)) - return []gpu.GpuInfo{g} + for _, p := range numParallelToTry { + req.opts.NumCtx = req.origNumCTX * p + if !envconfig.SchedSpread { + for _, g := range sgl { + if ok, estimatedVRAM = llm.PredictServerFit([]gpu.GpuInfo{g}, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok { + slog.Info("new model will fit in available VRAM in single GPU, loading", "model", req.model.ModelPath, "gpu", g.ID, "parallel", p, "available", g.FreeMemory, "required", format.HumanBytes2(estimatedVRAM)) + *numParallel = p + return []gpu.GpuInfo{g} + } } } } @@ -636,9 +680,13 @@ func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) gpu. // - try subsets of GPUs instead of just falling back to 1 or all in a family // Now try all the GPUs - if ok, estimatedVRAM = llm.PredictServerFit(sgl, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok { - slog.Debug("new model will fit in available VRAM, loading", "model", req.model.ModelPath, "library", sgl[0].Library, "required", format.HumanBytes2(estimatedVRAM)) - return sgl + for _, p := range numParallelToTry { + req.opts.NumCtx = req.origNumCTX * p + if ok, estimatedVRAM = llm.PredictServerFit(sgl, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok { + slog.Info("new model will fit in available VRAM, loading", "model", req.model.ModelPath, "library", sgl[0].Library, "parallel", p, "required", format.HumanBytes2(estimatedVRAM)) + *numParallel = p + return sgl + } } } return nil diff --git a/server/sched_test.go b/server/sched_test.go index 953288347..5e5913a7c 100644 --- a/server/sched_test.go +++ b/server/sched_test.go @@ -47,11 +47,11 @@ func TestLoad(t *testing.T) { sessionDuration: 2, } // Fail to load model first - s.newServerFn = func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options) (llm.LlamaServer, error) { + s.newServerFn = func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error) { return nil, fmt.Errorf("something failed to load model blah") } gpus := gpu.GpuInfoList{} - s.load(req, ggml, gpus) + s.load(req, ggml, gpus, 0) require.Empty(t, req.successCh) require.Len(t, req.errCh, 1) s.loadedMu.Lock() @@ -61,10 +61,10 @@ func TestLoad(t *testing.T) { require.Contains(t, err.Error(), "this model may be incompatible") server := &mockLlm{estimatedVRAM: 10, estimatedVRAMByGPU: map[string]uint64{}} - s.newServerFn = func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options) (llm.LlamaServer, error) { + s.newServerFn = func(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error) { return server, nil } - s.load(req, ggml, gpus) + s.load(req, ggml, gpus, 0) select { case err := <-req.errCh: require.NoError(t, err) @@ -78,12 +78,12 @@ func TestLoad(t *testing.T) { req.model.ModelPath = "dummy_model_path" server.waitResp = fmt.Errorf("wait failure") - s.load(req, ggml, gpus) + s.load(req, ggml, gpus, 0) select { case err := <-req.errCh: require.Contains(t, err.Error(), "wait failure") case resp := <-req.successCh: - t.Errorf("unexpected success %v", resp) + t.Fatalf("unexpected success %v", resp) } s.loadedMu.Lock() runner := s.loaded["dummy_model_path"] @@ -102,7 +102,7 @@ type bundle struct { ggml *llm.GGML } -func (scenario *bundle) newServer(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options) (llm.LlamaServer, error) { +func (scenario *bundle) newServer(gpus gpu.GpuInfoList, model string, ggml *llm.GGML, adapters []string, projectors []string, opts api.Options, numParallel int) (llm.LlamaServer, error) { return scenario.srv, nil } @@ -200,7 +200,7 @@ func TestRequests(t *testing.T) { require.Empty(t, s.pendingReqCh) require.Empty(t, scenario1a.req.errCh) case <-ctx.Done(): - t.Errorf("timeout") + t.Fatal("timeout") } // Same runner as first request due to not needing a reload @@ -213,7 +213,7 @@ func TestRequests(t *testing.T) { require.Empty(t, s.pendingReqCh) require.Empty(t, scenario1b.req.errCh) case <-ctx.Done(): - t.Errorf("timeout") + t.Fatal("timeout") } // Trigger a reload @@ -231,7 +231,7 @@ func TestRequests(t *testing.T) { require.Empty(t, s.pendingReqCh) require.Empty(t, scenario2a.req.errCh) case <-ctx.Done(): - t.Errorf("timeout") + t.Fatal("timeout") } envconfig.MaxRunners = 1 @@ -247,7 +247,7 @@ func TestRequests(t *testing.T) { require.Empty(t, s.pendingReqCh) require.Empty(t, scenario3a.req.errCh) case <-ctx.Done(): - t.Errorf("timeout") + t.Fatal("timeout") } s.loadedMu.Lock() require.Len(t, s.loaded, 1) @@ -263,7 +263,7 @@ func TestRequests(t *testing.T) { require.Empty(t, s.pendingReqCh) require.Empty(t, scenario3b.req.errCh) case <-ctx.Done(): - t.Errorf("timeout") + t.Fatal("timeout") } s.loadedMu.Lock() require.Len(t, s.loaded, 2) @@ -279,7 +279,7 @@ func TestRequests(t *testing.T) { require.Empty(t, s.pendingReqCh) require.Empty(t, scenario3c.req.errCh) case <-ctx.Done(): - t.Errorf("timeout") + t.Fatal("timeout") } s.loadedMu.Lock() require.Len(t, s.loaded, 3) @@ -306,7 +306,7 @@ func TestRequests(t *testing.T) { require.Empty(t, s.pendingReqCh) require.Empty(t, scenario3d.req.errCh) case <-ctx.Done(): - t.Errorf("timeout") + t.Fatal("timeout") } s.loadedMu.Lock() require.Len(t, s.loaded, 2) @@ -349,7 +349,7 @@ func TestGetRunner(t *testing.T) { require.Empty(t, s.pendingReqCh) require.Empty(t, errCh1a) case <-ctx.Done(): - t.Errorf("timeout") + t.Fatal("timeout") } scenario1a.ctxDone() s.loadedMu.Lock() @@ -400,7 +400,7 @@ func TestPrematureExpired(t *testing.T) { slog.Info("sending premature expired event now") s.expiredCh <- resp // Shouldn't happen in real life, but make sure its safe case <-ctx.Done(): - t.Errorf("timeout") + t.Fatal("timeout") } time.Sleep(scenario1a.req.sessionDuration) scenario1a.ctxDone() @@ -427,7 +427,7 @@ func TestUseLoadedRunner(t *testing.T) { } finished := make(chan *LlmRequest) llm1 := &mockLlm{estimatedVRAMByGPU: map[string]uint64{}} - r1 := &runnerRef{llama: llm1, sessionDuration: 1} + r1 := &runnerRef{llama: llm1, sessionDuration: 1, numParallel: 1} req.useLoadedRunner(r1, finished) require.Equal(t, uint(1), r1.refCount) require.Equal(t, time.Duration(2), r1.sessionDuration) @@ -435,7 +435,7 @@ func TestUseLoadedRunner(t *testing.T) { case success := <-req.successCh: require.Equal(t, r1, success) case <-ctx.Done(): - t.Errorf("timeout") + t.Fatal("timeout") } done() fin := <-finished @@ -461,8 +461,8 @@ func TestUpdateFreeSpace(t *testing.T) { gpus[1].FreeMemory = 1900 llm1 := &mockLlm{estimatedVRAMByGPU: map[string]uint64{"1": 50, "2": 50}} llm2 := &mockLlm{estimatedVRAMByGPU: map[string]uint64{"1": 125, "2": 75}} - r1 := &runnerRef{llama: llm1, gpus: gpus} - r2 := &runnerRef{llama: llm2, gpus: gpus} + r1 := &runnerRef{llama: llm1, gpus: gpus, numParallel: 1} + r2 := &runnerRef{llama: llm2, gpus: gpus, numParallel: 1} s := InitScheduler(ctx) s.loadedMu.Lock() @@ -513,8 +513,8 @@ func TestFindRunnerToUnload(t *testing.T) { ctx, done := context.WithTimeout(context.Background(), 100*time.Millisecond) defer done() - r1 := &runnerRef{refCount: 1, sessionDuration: 1} - r2 := &runnerRef{sessionDuration: 2} + r1 := &runnerRef{refCount: 1, sessionDuration: 1, numParallel: 1} + r2 := &runnerRef{sessionDuration: 2, numParallel: 1} s := InitScheduler(ctx) s.loadedMu.Lock() @@ -536,9 +536,13 @@ func TestNeedsReload(t *testing.T) { llm := &mockLlm{estimatedVRAMByGPU: map[string]uint64{}} do := api.DefaultOptions() runner := &runnerRef{ - model: &Model{AdapterPaths: []string{"adapter1"}, ProjectorPaths: []string{"projector1"}}, - Options: &do, - llama: llm, + model: &Model{ + AdapterPaths: []string{"adapter1"}, + ProjectorPaths: []string{"projector1"}, + }, + Options: &do, + llama: llm, + numParallel: 1, } req := &LlmRequest{ model: &Model{ @@ -581,8 +585,8 @@ func TestUnloadAllRunners(t *testing.T) { s := InitScheduler(ctx) s.unloadAllRunners() - r1 := &runnerRef{llama: llm1} - r2 := &runnerRef{llama: llm2} + r1 := &runnerRef{llama: llm1, numParallel: 1} + r2 := &runnerRef{llama: llm2, numParallel: 1} s.loadedMu.Lock() s.loaded["a"] = r1 @@ -596,14 +600,32 @@ func TestUnloadAllRunners(t *testing.T) { func TestUnload(t *testing.T) { llm1 := &mockLlm{estimatedVRAMByGPU: map[string]uint64{}} - r1 := &runnerRef{llama: llm1} - r2 := &runnerRef{model: &Model{AdapterPaths: []string{"A"}}} + r1 := &runnerRef{llama: llm1, numParallel: 1} + r2 := &runnerRef{model: &Model{AdapterPaths: []string{"A"}}, numParallel: 1} r1.unload() require.True(t, llm1.closeCalled) r2.unload() require.Nil(t, r2.model) } +func TestAlreadyCanceled(t *testing.T) { + ctx, done := context.WithTimeout(context.Background(), 500*time.Millisecond) + defer done() + dctx, done2 := context.WithCancel(ctx) + done2() + scenario1a := newScenario(t, dctx, "ollama-model-1", 10) + scenario1a.req.sessionDuration = 0 + s := InitScheduler(ctx) + slog.Info("scenario1a") + s.pendingReqCh <- scenario1a.req + require.Len(t, s.pendingReqCh, 1) + s.Run(ctx) + time.Sleep(5 * time.Millisecond) + require.Empty(t, s.pendingReqCh) + require.Empty(t, scenario1a.req.errCh) + require.Empty(t, scenario1a.req.successCh) +} + type mockLlm struct { pingResp error waitResp error From 9929751cc8b415e7b83d5151742dad734e8b5efc Mon Sep 17 00:00:00 2001 From: Daniel Hiltgen Date: Wed, 19 Jun 2024 13:35:38 -0700 Subject: [PATCH 02/31] Disable concurrency for AMD + Windows Until ROCm v6.2 ships, we wont be able to get accurate free memory reporting on windows, which makes automatic concurrency too risky. Users can still opt-in but will need to pay attention to model sizes otherwise they may thrash/page VRAM or cause OOM crashes. All other platforms and GPUs have accurate VRAM reporting wired up now, so we can turn on concurrency by default. --- envconfig/config.go | 8 ++++---- gpu/amd_windows.go | 5 +++-- gpu/types.go | 5 +++++ server/sched.go | 36 ++++++++++++++++++++++++++++++++---- 4 files changed, 44 insertions(+), 10 deletions(-) diff --git a/envconfig/config.go b/envconfig/config.go index cb456448c..0f0f7f058 100644 --- a/envconfig/config.go +++ b/envconfig/config.go @@ -85,13 +85,13 @@ func AsMap() map[string]EnvVar { "OLLAMA_HOST": {"OLLAMA_HOST", Host, "IP Address for the ollama server (default 127.0.0.1:11434)"}, "OLLAMA_KEEP_ALIVE": {"OLLAMA_KEEP_ALIVE", KeepAlive, "The duration that models stay loaded in memory (default \"5m\")"}, "OLLAMA_LLM_LIBRARY": {"OLLAMA_LLM_LIBRARY", LLMLibrary, "Set LLM library to bypass autodetection"}, - "OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners, "Maximum number of loaded models per GPU (default 4)"}, + "OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners, "Maximum number of loaded models per GPU (default auto)"}, "OLLAMA_MAX_QUEUE": {"OLLAMA_MAX_QUEUE", MaxQueuedRequests, "Maximum number of queued requests"}, "OLLAMA_MAX_VRAM": {"OLLAMA_MAX_VRAM", MaxVRAM, "Maximum VRAM"}, "OLLAMA_MODELS": {"OLLAMA_MODELS", ModelsDir, "The path to the models directory"}, "OLLAMA_NOHISTORY": {"OLLAMA_NOHISTORY", NoHistory, "Do not preserve readline history"}, "OLLAMA_NOPRUNE": {"OLLAMA_NOPRUNE", NoPrune, "Do not prune model blobs on startup"}, - "OLLAMA_NUM_PARALLEL": {"OLLAMA_NUM_PARALLEL", NumParallel, "Maximum number of parallel requests"}, + "OLLAMA_NUM_PARALLEL": {"OLLAMA_NUM_PARALLEL", NumParallel, "Maximum number of parallel requests (default auto)"}, "OLLAMA_ORIGINS": {"OLLAMA_ORIGINS", AllowOrigins, "A comma separated list of allowed origins"}, "OLLAMA_RUNNERS_DIR": {"OLLAMA_RUNNERS_DIR", RunnersDir, "Location for runners"}, "OLLAMA_SCHED_SPREAD": {"OLLAMA_SCHED_SPREAD", SchedSpread, "Always schedule model across all GPUs"}, @@ -129,8 +129,8 @@ func clean(key string) string { func init() { // default values - NumParallel = 0 - MaxRunners = 4 + NumParallel = 0 // Autoselect + MaxRunners = 0 // Autoselect MaxQueuedRequests = 512 LoadConfig() diff --git a/gpu/amd_windows.go b/gpu/amd_windows.go index 21585277a..8b6fabebb 100644 --- a/gpu/amd_windows.go +++ b/gpu/amd_windows.go @@ -115,8 +115,6 @@ func AMDGetGPUInfo() []RocmGPUInfo { continue } - // TODO revisit this once ROCm v6 is available on windows. - // v5.7 only reports VRAM used by this process, so it's completely wrong and unusable slog.Debug("amdgpu memory", "gpu", i, "total", format.HumanBytes2(totalMemory)) slog.Debug("amdgpu memory", "gpu", i, "available", format.HumanBytes2(freeMemory)) gpuInfo := RocmGPUInfo{ @@ -126,6 +124,9 @@ func AMDGetGPUInfo() []RocmGPUInfo { TotalMemory: totalMemory, FreeMemory: freeMemory, }, + // Free memory reporting on Windows is not reliable until we bump to ROCm v6.2 + UnreliableFreeMemory: true, + ID: strconv.Itoa(i), // TODO this is probably wrong if we specify visible devices DependencyPath: libDir, MinimumMemory: rocmMinimumMemory, diff --git a/gpu/types.go b/gpu/types.go index 9920db5ff..2eaa9bae9 100644 --- a/gpu/types.go +++ b/gpu/types.go @@ -29,6 +29,11 @@ type GpuInfo struct { // Extra environment variables specific to the GPU as list of [key,value] EnvWorkarounds [][2]string `json:"envs,omitempty"` + // Set to true if we can NOT reliably discover FreeMemory. A value of true indicates + // the FreeMemory is best effort, and may over or under report actual memory usage + // False indicates FreeMemory can generally be trusted on this GPU + UnreliableFreeMemory bool + // GPU information ID string `json:"gpu_id"` // string to use for selection of this specific GPU Name string `json:"name"` // user friendly name if available diff --git a/server/sched.go b/server/sched.go index 31ef560f5..de8c9d281 100644 --- a/server/sched.go +++ b/server/sched.go @@ -46,6 +46,16 @@ type Scheduler struct { reschedDelay time.Duration } +// Default automatic value for number of models we allow per GPU +// Model will still need to fit in VRAM, but loading many small models +// on a large GPU can cause stalling +var defaultModelsPerGPU = 3 + +// Default automatic value for parallel setting +// Model will still need to fit in VRAM. If this setting wont fit +// we'll back off down to 1 to try to get it to fit +var defaultParallel = 4 + var ErrMaxQueue = fmt.Errorf("server busy, please try again. maximum pending requests exceeded") func InitScheduler(ctx context.Context) *Scheduler { @@ -100,7 +110,6 @@ func (s *Scheduler) Run(ctx context.Context) { } func (s *Scheduler) processPending(ctx context.Context) { - maxRunnerFactor := 1 // number of GPUs or 1 for { select { case <-ctx.Done(): @@ -143,7 +152,7 @@ func (s *Scheduler) processPending(ctx context.Context) { pending.useLoadedRunner(runner, s.finishedReqCh) break } - } else if envconfig.MaxRunners > 0 && loadedCount >= (maxRunnerFactor*envconfig.MaxRunners) { + } else if envconfig.MaxRunners > 0 && loadedCount >= envconfig.MaxRunners { slog.Debug("max runners achieved, unloading one to make room", "runner_count", loadedCount) runnerToExpire = s.findRunnerToUnload() } else { @@ -155,7 +164,26 @@ func (s *Scheduler) processPending(ctx context.Context) { } else { gpus = s.getGpuFn() } - maxRunnerFactor = max(len(gpus), 1) + + if envconfig.MaxRunners <= 0 { + // No user specified MaxRunners, so figure out what automatic setting to use + // If all GPUs have reliable free memory reporting, defaultModelsPerGPU * the number of GPUs + // if any GPU has unreliable free memory reporting, 1x the number of GPUs + allReliable := true + for _, gpu := range gpus { + if gpu.UnreliableFreeMemory { + allReliable = false + break + } + } + if allReliable { + envconfig.MaxRunners = defaultModelsPerGPU * len(gpus) + slog.Debug("updating default concurrency", "OLLAMA_MAX_LOADED_MODELS", envconfig.MaxRunners, "gpu_count", len(gpus)) + } else { + slog.Info("one or more GPUs detected that are unable to accurately report free memory - disabling default concurrency") + envconfig.MaxRunners = len(gpus) + } + } // Load model for fitting ggml, err := llm.LoadModel(pending.model.ModelPath) @@ -647,7 +675,7 @@ func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numP var numParallelToTry []int if *numParallel <= 0 { // If no specific parallel setting was provided, try larger then smaller, always end with 1 - numParallelToTry = append(numParallelToTry, 4, 1) + numParallelToTry = append(numParallelToTry, defaultParallel, 1) } else { numParallelToTry = []int{*numParallel} } From 642cee13426c994f90d5a95376025fe9a223891a Mon Sep 17 00:00:00 2001 From: Daniel Hiltgen Date: Fri, 21 Jun 2024 15:59:41 -0700 Subject: [PATCH 03/31] Sort the ps output Provide consistent ordering for the ps command - longest duration listed first --- server/routes.go | 5 +++++ 1 file changed, 5 insertions(+) diff --git a/server/routes.go b/server/routes.go index 3d112e9f1..a7f72edc2 100644 --- a/server/routes.go +++ b/server/routes.go @@ -1224,6 +1224,11 @@ func (s *Server) ProcessHandler(c *gin.Context) { models = append(models, mr) } + slices.SortStableFunc(models, func(i, j api.ProcessModelResponse) int { + // longest duration remaining listed first + return cmp.Compare(j.ExpiresAt.Unix(), i.ExpiresAt.Unix()) + }) + c.JSON(http.StatusOK, api.ProcessResponse{Models: models}) } From 4d311eb731bb59512bcd17f1f33d60f3d9022837 Mon Sep 17 00:00:00 2001 From: Jeffrey Morgan Date: Wed, 26 Jun 2024 21:38:12 -0700 Subject: [PATCH 04/31] llm: architecture patch (#5316) --- llm/patches/07-gemma.diff | 305 ++++++++++++++++++++++++++++++++++++++ 1 file changed, 305 insertions(+) create mode 100644 llm/patches/07-gemma.diff diff --git a/llm/patches/07-gemma.diff b/llm/patches/07-gemma.diff new file mode 100644 index 000000000..86eac3d17 --- /dev/null +++ b/llm/patches/07-gemma.diff @@ -0,0 +1,305 @@ +From 5cadb45f39d001ffbad95b690d6cf0abcb4a6d96 Mon Sep 17 00:00:00 2001 +From: Ollama maintainers +Date: Wed, 26 Jun 2024 16:18:09 -0700 +Subject: [PATCH] Architecture support + +--- + llama.cpp | 194 +++++++++++++++++++++++++++++++++++++++++++++++++++++- + 1 file changed, 193 insertions(+), 1 deletion(-) + +diff --git a/llama.cpp b/llama.cpp +index 61948751..3b4196f5 100644 +--- a/llama.cpp ++++ b/llama.cpp +@@ -217,6 +217,7 @@ enum llm_arch { + LLM_ARCH_INTERNLM2, + LLM_ARCH_MINICPM, + LLM_ARCH_GEMMA, ++ LLM_ARCH_GEMMA2, + LLM_ARCH_STARCODER2, + LLM_ARCH_MAMBA, + LLM_ARCH_XVERSE, +@@ -255,6 +256,7 @@ static const std::map LLM_ARCH_NAMES = { + { LLM_ARCH_INTERNLM2, "internlm2" }, + { LLM_ARCH_MINICPM, "minicpm" }, + { LLM_ARCH_GEMMA, "gemma" }, ++ { LLM_ARCH_GEMMA2, "gemma2" }, + { LLM_ARCH_STARCODER2, "starcoder2" }, + { LLM_ARCH_MAMBA, "mamba" }, + { LLM_ARCH_XVERSE, "xverse" }, +@@ -464,10 +466,12 @@ enum llm_tensor { + LLM_TENSOR_ATTN_NORM, + LLM_TENSOR_ATTN_NORM_2, + LLM_TENSOR_ATTN_OUT_NORM, ++ LLM_TENSOR_ATTN_POST_NORM, + LLM_TENSOR_ATTN_ROT_EMBD, + LLM_TENSOR_FFN_GATE_INP, + LLM_TENSOR_FFN_GATE_INP_SHEXP, + LLM_TENSOR_FFN_NORM, ++ LLM_TENSOR_FFN_POST_NORM, + LLM_TENSOR_FFN_GATE, + LLM_TENSOR_FFN_DOWN, + LLM_TENSOR_FFN_UP, +@@ -960,6 +964,24 @@ static const std::map> LLM_TENSOR_NA + { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, + }, + }, ++ { ++ LLM_ARCH_GEMMA2, ++ { ++ { LLM_TENSOR_TOKEN_EMBD, "token_embd" }, ++ { LLM_TENSOR_OUTPUT_NORM, "output_norm" }, ++ { LLM_TENSOR_ATTN_NORM, "blk.%d.attn_norm" }, ++ { LLM_TENSOR_ATTN_Q, "blk.%d.attn_q" }, ++ { LLM_TENSOR_ATTN_K, "blk.%d.attn_k" }, ++ { LLM_TENSOR_ATTN_V, "blk.%d.attn_v" }, ++ { LLM_TENSOR_ATTN_OUT, "blk.%d.attn_output" }, ++ { LLM_TENSOR_ATTN_POST_NORM, "blk.%d.post_attention_norm" }, ++ { LLM_TENSOR_FFN_NORM, "blk.%d.ffn_norm" }, ++ { LLM_TENSOR_FFN_GATE, "blk.%d.ffn_gate" }, ++ { LLM_TENSOR_FFN_DOWN, "blk.%d.ffn_down" }, ++ { LLM_TENSOR_FFN_UP, "blk.%d.ffn_up" }, ++ { LLM_TENSOR_FFN_POST_NORM, "blk.%d.post_ffw_norm" }, ++ }, ++ }, + { + LLM_ARCH_STARCODER2, + { +@@ -1941,6 +1963,8 @@ enum e_model { + MODEL_8x22B, + MODEL_16x12B, + MODEL_10B_128x3_66B, ++ MODEL_9B, ++ MODEL_27B, + }; + + static const size_t kiB = 1024; +@@ -2114,6 +2138,7 @@ struct llama_layer { + struct ggml_tensor * attn_out_norm_b; + struct ggml_tensor * attn_q_a_norm; + struct ggml_tensor * attn_kv_a_norm; ++ struct ggml_tensor * attn_post_norm; + + // attention + struct ggml_tensor * wq; +@@ -2136,6 +2161,7 @@ struct llama_layer { + // normalization + struct ggml_tensor * ffn_norm; + struct ggml_tensor * ffn_norm_b; ++ struct ggml_tensor * ffn_post_norm; + struct ggml_tensor * layer_out_norm; + struct ggml_tensor * layer_out_norm_b; + struct ggml_tensor * ffn_norm_exps; +@@ -4529,6 +4555,16 @@ static void llm_load_hparams( + } + } break; + case LLM_ARCH_GEMMA: ++ { ++ ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); ++ ++ switch (hparams.n_layer) { ++ case 18: model.type = e_model::MODEL_9B; break; ++ case 28: model.type = e_model::MODEL_27B; break; ++ default: model.type = e_model::MODEL_UNKNOWN; ++ } ++ } break; ++ case LLM_ARCH_GEMMA2: + { + ml.get_key(LLM_KV_ATTENTION_LAYERNORM_RMS_EPS, hparams.f_norm_rms_eps); + +@@ -6305,6 +6341,40 @@ static bool llm_load_tensors( + layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}); + } + } break; ++ case LLM_ARCH_GEMMA2: ++ { ++ model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); ++ ++ // output ++ model.output_norm = ml.create_tensor(ctx_output, tn(LLM_TENSOR_OUTPUT_NORM, "weight"), {n_embd}); ++ model.output = ml.create_tensor(ctx_output, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}, llama_model_loader::TENSOR_DUPLICATED); // same as tok_embd, duplicated to allow offloading ++ ++ const int64_t n_ff = hparams.n_ff; ++ const int64_t n_embd_head_k = hparams.n_embd_head_k; ++ const int64_t n_embd_k_gqa = hparams.n_embd_k_gqa(); ++ const int64_t n_embd_v_gqa = hparams.n_embd_v_gqa(); ++ ++ for (uint32_t i = 0; i < n_layer; ++i) { ++ ggml_context * ctx_layer = ctx_for_layer(i); ++ ggml_context * ctx_split = ctx_for_layer_split(i); ++ ++ auto & layer = model.layers[i]; ++ ++ layer.attn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_ATTN_NORM, "weight", i), {n_embd}); ++ ++ layer.wq = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_Q, "weight", i), {n_embd, n_embd_head_k * hparams.n_head}); ++ layer.wk = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_K, "weight", i), {n_embd, n_embd_k_gqa}); ++ layer.wv = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_V, "weight", i), {n_embd, n_embd_v_gqa}); ++ layer.wo = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_OUT, "weight", i), {n_embd_head_k * hparams.n_head, n_embd}); ++ layer.attn_post_norm = ml.create_tensor(ctx_split, tn(LLM_TENSOR_ATTN_POST_NORM, "weight", i), {n_embd}); ++ ++ layer.ffn_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_NORM, "weight", i), {n_embd}); ++ layer.ffn_gate = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_GATE, "weight", i), {n_embd, n_ff}); ++ layer.ffn_up = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_UP, "weight", i), {n_embd, n_ff}); ++ layer.ffn_down = ml.create_tensor(ctx_split, tn(LLM_TENSOR_FFN_DOWN, "weight", i), { n_ff, n_embd}); ++ layer.ffn_post_norm = ml.create_tensor(ctx_layer, tn(LLM_TENSOR_FFN_POST_NORM, "weight", i), {n_embd}); ++ } ++ } break; + case LLM_ARCH_STARCODER2: + { + model.tok_embd = ml.create_tensor(ctx_input, tn(LLM_TENSOR_TOKEN_EMBD, "weight"), {n_embd, n_vocab}); +@@ -10614,6 +10684,123 @@ struct llm_build_context { + return gf; + } + ++ struct ggml_cgraph * build_gemma2() { ++ struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); ++ ++ const int64_t n_embd_head_k = hparams.n_embd_head_k; ++ ++ struct ggml_tensor * cur; ++ struct ggml_tensor * inpL; ++ ++ inpL = llm_build_inp_embd(ctx0, lctx, hparams, batch, model.tok_embd, cb); ++ ++ inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd)); ++ cb(inpL, "inp_scaled", -1); ++ ++ // inp_pos - contains the positions ++ struct ggml_tensor * inp_pos = build_inp_pos(); ++ ++ // KQ_mask (mask for 1 head, it will be broadcasted to all heads) ++ struct ggml_tensor * KQ_mask = build_inp_KQ_mask(); ++ ++ for (int il = 0; il < n_layer; ++il) { ++ // norm ++ cur = llm_build_norm(ctx0, inpL, hparams, ++ model.layers[il].attn_norm, NULL, ++ LLM_NORM_RMS, cb, il); ++ cb(cur, "attn_norm", il); ++ ++ // self-attention ++ { ++ // compute Q and K and RoPE them ++ struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur); ++ cb(Qcur, "Qcur", il); ++ ++ struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur); ++ cb(Kcur, "Kcur", il); ++ ++ struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur); ++ cb(Vcur, "Vcur", il); ++ ++ Qcur = ggml_rope_ext( ++ ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head_k, n_head, n_tokens), inp_pos, nullptr, ++ n_embd_head_k, rope_type, n_ctx_orig, freq_base, freq_scale, ++ ext_factor, attn_factor, beta_fast, beta_slow); ++ cb(Qcur, "Qcur", il); ++ ++ Qcur = ggml_scale(ctx0, Qcur, 1.0f / sqrtf(float(n_embd_head_k))); ++ cb(Qcur, "Qcur_scaled", il); ++ ++ Kcur = ggml_rope_ext( ++ ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head_k, n_head_kv, n_tokens), inp_pos, nullptr, ++ n_embd_head_k, rope_type, n_ctx_orig, freq_base, freq_scale, ++ ext_factor, attn_factor, beta_fast, beta_slow); ++ cb(Kcur, "Kcur", il); ++ ++ cur = llm_build_kv(ctx0, model, hparams, cparams, kv_self, gf, ++ model.layers[il].wo, NULL, ++ Kcur, Vcur, Qcur, KQ_mask, n_tokens, kv_head, n_kv, 1.0f, cb, il); ++ } ++ ++ if (il == n_layer - 1) { ++ // skip computing output for unused tokens ++ struct ggml_tensor * inp_out_ids = build_inp_out_ids(); ++ cur = ggml_get_rows(ctx0, cur, inp_out_ids); ++ inpL = ggml_get_rows(ctx0, inpL, inp_out_ids); ++ } ++ ++ cur = llm_build_norm(ctx0, cur, hparams, ++ model.layers[il].attn_post_norm, NULL, ++ LLM_NORM_RMS, cb, il); ++ cb(cur, "attn_post_norm", il); ++ ++ struct ggml_tensor * sa_out = ggml_add(ctx0, cur, inpL); ++ cb(sa_out, "sa_out", il); ++ ++ cur = llm_build_norm(ctx0, sa_out, hparams, ++ model.layers[il].ffn_norm, NULL, ++ LLM_NORM_RMS, cb, il); ++ cb(cur, "ffn_norm", il); ++ ++ // feed-forward network ++ { ++ cur = llm_build_ffn(ctx0, cur, ++ model.layers[il].ffn_up, NULL, ++ model.layers[il].ffn_gate, NULL, ++ model.layers[il].ffn_down, NULL, ++ NULL, ++ LLM_FFN_GELU, LLM_FFN_PAR, cb, il); ++ cb(cur, "ffn_out", il); ++ } ++ ++ cur = llm_build_norm(ctx0, cur, hparams, ++ model.layers[il].ffn_post_norm, NULL, ++ LLM_NORM_RMS, cb, -1); ++ cb(cur, "ffn_post_norm", -1); ++ ++ cur = ggml_add(ctx0, cur, sa_out); ++ cb(cur, "l_out", il); ++ ++ // input for next layer ++ inpL = cur; ++ } ++ ++ cur = inpL; ++ ++ cur = llm_build_norm(ctx0, cur, hparams, ++ model.output_norm, NULL, ++ LLM_NORM_RMS, cb, -1); ++ cb(cur, "result_norm", -1); ++ ++ // lm_head ++ cur = ggml_mul_mat(ctx0, model.output, cur); ++ cb(cur, "result_output", -1); ++ ++ ggml_build_forward_expand(gf, cur); ++ ++ return gf; ++ } ++ + struct ggml_cgraph * build_starcoder2() { + struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false); + +@@ -11847,6 +12034,10 @@ static struct ggml_cgraph * llama_build_graph( + { + result = llm.build_gemma(); + } break; ++ case LLM_ARCH_GEMMA2: ++ { ++ result = llm.build_gemma2(); ++ } break; + case LLM_ARCH_STARCODER2: + { + result = llm.build_starcoder2(); +@@ -16671,6 +16862,7 @@ enum llama_rope_type llama_rope_type(const struct llama_model * model) { + case LLM_ARCH_PHI2: + case LLM_ARCH_PHI3: + case LLM_ARCH_GEMMA: ++ case LLM_ARCH_GEMMA2: + case LLM_ARCH_STARCODER2: + case LLM_ARCH_GPTNEOX: + return LLAMA_ROPE_TYPE_NEOX; +@@ -18551,7 +18743,7 @@ static int32_t llama_chat_apply_template_internal( + if (add_ass) { + ss << "assistant\n"; + } +- } else if (tmpl == "gemma" || tmpl.find("") != std::string::npos) { ++ } else if (tmpl == "gemma" || tmpl == "gemma2" || tmpl.find("") != std::string::npos) { + // google/gemma-7b-it + std::string system_prompt = ""; + for (auto message : chat) { +-- +2.45.2 + From 123a722a6f541e300bc8e34297ac378ebe23f527 Mon Sep 17 00:00:00 2001 From: Michael Yang Date: Wed, 26 Jun 2024 21:38:21 -0700 Subject: [PATCH 05/31] zip: prevent extracting files into parent dirs (#5314) --- cmd/cmd.go | 6 +-- server/model.go | 57 ++++++++++++++++++--------- server/model_test.go | 92 ++++++++++++++++++++++++++++++++++++++++++++ 3 files changed, 133 insertions(+), 22 deletions(-) create mode 100644 server/model_test.go diff --git a/cmd/cmd.go b/cmd/cmd.go index 89b551f40..909e8e4b2 100644 --- a/cmd/cmd.go +++ b/cmd/cmd.go @@ -162,9 +162,6 @@ func tempZipFiles(path string) (string, error) { } defer tempfile.Close() - zipfile := zip.NewWriter(tempfile) - defer zipfile.Close() - detectContentType := func(path string) (string, error) { f, err := os.Open(path) if err != nil { @@ -233,6 +230,9 @@ func tempZipFiles(path string) (string, error) { files = append(files, tks...) } + zipfile := zip.NewWriter(tempfile) + defer zipfile.Close() + for _, file := range files { f, err := os.Open(file) if err != nil { diff --git a/server/model.go b/server/model.go index 055ffd63a..d56e641ba 100644 --- a/server/model.go +++ b/server/model.go @@ -11,6 +11,7 @@ import ( "net/http" "os" "path/filepath" + "strings" "github.com/ollama/ollama/api" "github.com/ollama/ollama/convert" @@ -77,62 +78,80 @@ func parseFromModel(ctx context.Context, name model.Name, fn func(api.ProgressRe return layers, nil } -func parseFromZipFile(_ context.Context, file *os.File, digest string, fn func(api.ProgressResponse)) (layers []*layerGGML, err error) { +func extractFromZipFile(p string, file *os.File, fn func(api.ProgressResponse)) error { stat, err := file.Stat() if err != nil { - return nil, err + return err } r, err := zip.NewReader(file, stat.Size()) if err != nil { - return nil, err + return err } - tempdir, err := os.MkdirTemp(filepath.Dir(file.Name()), "") - if err != nil { - return nil, err - } - defer os.RemoveAll(tempdir) - fn(api.ProgressResponse{Status: "unpacking model metadata"}) for _, f := range r.File { + n := filepath.Join(p, f.Name) + if !strings.HasPrefix(n, p) { + slog.Warn("skipped extracting file outside of context", "name", f.Name) + continue + } + + if err := os.MkdirAll(filepath.Dir(n), 0o750); err != nil { + return err + } + // TODO(mxyng): this should not write out all files to disk - outfile, err := os.Create(filepath.Join(tempdir, f.Name)) + outfile, err := os.Create(n) if err != nil { - return nil, err + return err } defer outfile.Close() infile, err := f.Open() if err != nil { - return nil, err + return err } defer infile.Close() if _, err = io.Copy(outfile, infile); err != nil { - return nil, err + return err } if err := outfile.Close(); err != nil { - return nil, err + return err } if err := infile.Close(); err != nil { - return nil, err + return err } } - mf, err := convert.GetModelFormat(tempdir) + return nil +} + +func parseFromZipFile(_ context.Context, file *os.File, digest string, fn func(api.ProgressResponse)) (layers []*layerGGML, err error) { + tempDir, err := os.MkdirTemp(filepath.Dir(file.Name()), "") + if err != nil { + return nil, err + } + defer os.RemoveAll(tempDir) + + if err := extractFromZipFile(tempDir, file, fn); err != nil { + return nil, err + } + + mf, err := convert.GetModelFormat(tempDir) if err != nil { return nil, err } - params, err := mf.GetParams(tempdir) + params, err := mf.GetParams(tempDir) if err != nil { return nil, err } - mArch, err := mf.GetModelArch("", tempdir, params) + mArch, err := mf.GetModelArch("", tempDir, params) if err != nil { return nil, err } @@ -150,7 +169,7 @@ func parseFromZipFile(_ context.Context, file *os.File, digest string, fn func(a // TODO(mxyng): this should write directly into a layer // e.g. NewLayer(arch.Reader(), "application/vnd.ollama.image.model") - temp, err := os.CreateTemp(tempdir, "fp16") + temp, err := os.CreateTemp(tempDir, "fp16") if err != nil { return nil, err } diff --git a/server/model_test.go b/server/model_test.go new file mode 100644 index 000000000..c3023eb2b --- /dev/null +++ b/server/model_test.go @@ -0,0 +1,92 @@ +package server + +import ( + "archive/zip" + "bytes" + "io" + "os" + "path/filepath" + "slices" + "testing" + + "github.com/ollama/ollama/api" +) + +func createZipFile(t *testing.T, name string) *os.File { + t.Helper() + + f, err := os.CreateTemp(t.TempDir(), "") + if err != nil { + t.Fatal(err) + } + + zf := zip.NewWriter(f) + defer zf.Close() + + zh, err := zf.CreateHeader(&zip.FileHeader{Name: name}) + if err != nil { + t.Fatal(err) + } + + if _, err := io.Copy(zh, bytes.NewReader([]byte(""))); err != nil { + t.Fatal(err) + } + + return f +} + +func TestExtractFromZipFile(t *testing.T) { + cases := []struct { + name string + expect []string + }{ + { + name: "good", + expect: []string{"good"}, + }, + { + name: filepath.Join("..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "bad"), + }, + } + + for _, tt := range cases { + t.Run(tt.name, func(t *testing.T) { + f := createZipFile(t, tt.name) + defer f.Close() + + tempDir := t.TempDir() + if err := extractFromZipFile(tempDir, f, func(api.ProgressResponse) {}); err != nil { + t.Fatal(err) + } + + var matches []string + if err := filepath.Walk(tempDir, func(p string, fi os.FileInfo, err error) error { + if err != nil { + return err + } + + if !fi.IsDir() { + matches = append(matches, p) + } + + return nil + }); err != nil { + t.Fatal(err) + } + + var actual []string + for _, match := range matches { + rel, err := filepath.Rel(tempDir, match) + if err != nil { + t.Error(err) + } + + actual = append(actual, rel) + } + + if !slices.Equal(actual, tt.expect) { + t.Fatalf("expected %d files, got %d", len(tt.expect), len(matches)) + } + }) + } +} From 2cc7d050124929ae4745633fddf053585a22f0a2 Mon Sep 17 00:00:00 2001 From: Michael Date: Thu, 27 Jun 2024 12:45:16 -0400 Subject: [PATCH 06/31] update readme for gemma 2 (#5333) * update readme for gemma 2 --- README.md | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index 978625731..72ed8fa5e 100644 --- a/README.md +++ b/README.md @@ -53,8 +53,8 @@ Here are some example models that can be downloaded: | Llama 3 | 70B | 40GB | `ollama run llama3:70b` | | Phi 3 Mini | 3.8B | 2.3GB | `ollama run phi3` | | Phi 3 Medium | 14B | 7.9GB | `ollama run phi3:medium` | -| Gemma | 2B | 1.4GB | `ollama run gemma:2b` | -| Gemma | 7B | 4.8GB | `ollama run gemma:7b` | +| Gemma 2 | 9B | 5.5GB | `ollama run gemma2` | +| Gemma 2 | 27B | 16GB | `ollama run gemma2:27b` | | Mistral | 7B | 4.1GB | `ollama run mistral` | | Moondream 2 | 1.4B | 829MB | `ollama run moondream` | | Neural Chat | 7B | 4.1GB | `ollama run neural-chat` | From 4e986a823ca47eb16f563d15a6fe4cc393a00715 Mon Sep 17 00:00:00 2001 From: Josh Yan Date: Thu, 27 Jun 2024 10:59:15 -0700 Subject: [PATCH 07/31] unquote, trimp space --- parser/parser.go | 9 ++++++++- parser/parser_test.go | 25 +++++++++++++++++++++++++ 2 files changed, 33 insertions(+), 1 deletion(-) diff --git a/parser/parser.go b/parser/parser.go index 686a1e695..fa60ebc0f 100644 --- a/parser/parser.go +++ b/parser/parser.go @@ -125,6 +125,7 @@ func ParseFile(r io.Reader) (*File, error) { // pass case stateValue: s, ok := unquote(b.String()) + if !ok || isSpace(r) { if _, err := b.WriteRune(r); err != nil { return nil, err @@ -158,7 +159,13 @@ func ParseFile(r io.Reader) (*File, error) { case stateComment, stateNil: // pass; nothing to flush case stateValue: - s, ok := unquote(b.String()) + var s string + var ok bool + if cmd.Name == "model" { + s, ok = unquote(strings.TrimSpace(b.String())) + } else { + s, ok = unquote(b.String()) + } if !ok { return nil, io.ErrUnexpectedEOF } diff --git a/parser/parser_test.go b/parser/parser_test.go index 7123e53bf..35556515d 100644 --- a/parser/parser_test.go +++ b/parser/parser_test.go @@ -48,6 +48,26 @@ func TestParseFileFrom(t *testing.T) { expected []Command err error }{ + { + "FROM \"FOO BAR \"", + []Command{{Name: "model", Args: "FOO BAR "}}, + nil, + }, + { + "FROM \"FOO BAR\"\nPARAMETER param1 value1", + []Command{{Name: "model", Args: "FOO BAR"}, {Name: "param1", Args: "value1"}}, + nil, + }, + { + "FROM FOOO BAR ", + []Command{{Name: "model", Args: "FOOO BAR"}}, + nil, + }, + { + "FROM /what/is/the path ", + []Command{{Name: "model", Args: "/what/is/the path"}}, + nil, + }, { "FROM foo", []Command{{Name: "model", Args: "foo"}}, @@ -86,6 +106,11 @@ func TestParseFileFrom(t *testing.T) { []Command{{Name: "param1", Args: "value1"}, {Name: "model", Args: "foo"}}, nil, }, + { + "PARAMETER what the \nFROM lemons make lemonade ", + []Command{{Name: "what", Args: "the "}, {Name: "model", Args: "lemons make lemonade"}}, + nil, + }, } for _, c := range cases { From 9bd00041fa1c82881299f34a5950f9edc2a7e66c Mon Sep 17 00:00:00 2001 From: Josh Yan Date: Thu, 27 Jun 2024 11:18:38 -0700 Subject: [PATCH 08/31] trim all params --- parser/parser.go | 11 ++--------- parser/parser_test.go | 4 ++-- 2 files changed, 4 insertions(+), 11 deletions(-) diff --git a/parser/parser.go b/parser/parser.go index fa60ebc0f..7f566da4e 100644 --- a/parser/parser.go +++ b/parser/parser.go @@ -124,8 +124,7 @@ func ParseFile(r io.Reader) (*File, error) { case stateComment, stateNil: // pass case stateValue: - s, ok := unquote(b.String()) - + s, ok := unquote(strings.TrimSpace(b.String())) if !ok || isSpace(r) { if _, err := b.WriteRune(r); err != nil { return nil, err @@ -159,13 +158,7 @@ func ParseFile(r io.Reader) (*File, error) { case stateComment, stateNil: // pass; nothing to flush case stateValue: - var s string - var ok bool - if cmd.Name == "model" { - s, ok = unquote(strings.TrimSpace(b.String())) - } else { - s, ok = unquote(b.String()) - } + s, ok := unquote(strings.TrimSpace(b.String())) if !ok { return nil, io.ErrUnexpectedEOF } diff --git a/parser/parser_test.go b/parser/parser_test.go index 35556515d..3dc592239 100644 --- a/parser/parser_test.go +++ b/parser/parser_test.go @@ -108,7 +108,7 @@ func TestParseFileFrom(t *testing.T) { }, { "PARAMETER what the \nFROM lemons make lemonade ", - []Command{{Name: "what", Args: "the "}, {Name: "model", Args: "lemons make lemonade"}}, + []Command{{Name: "what", Args: "the"}, {Name: "model", Args: "lemons make lemonade"}}, nil, }, } @@ -424,7 +424,7 @@ func TestParseFileParameters(t *testing.T) { "mirostat_eta 1.0": {"mirostat_eta", "1.0"}, "penalize_newline true": {"penalize_newline", "true"}, "stop ### User:": {"stop", "### User:"}, - "stop ### User: ": {"stop", "### User: "}, + "stop ### User: ": {"stop", "### User:"}, "stop \"### User:\"": {"stop", "### User:"}, "stop \"### User: \"": {"stop", "### User: "}, "stop \"\"\"### User:\"\"\"": {"stop", "### User:"}, From de2163dafd19b5ba2bed3d459354179662cc524d Mon Sep 17 00:00:00 2001 From: Michael Yang Date: Thu, 27 Jun 2024 10:52:25 -0700 Subject: [PATCH 09/31] gemma2 graph --- llm/ggml.go | 15 ++++++++++++--- 1 file changed, 12 insertions(+), 3 deletions(-) diff --git a/llm/ggml.go b/llm/ggml.go index d0d0b6ddc..cfead450d 100644 --- a/llm/ggml.go +++ b/llm/ggml.go @@ -366,9 +366,18 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui 4*batch*(1+2*embedding+context*(1+heads))+embedding*(6*context*headsKV/heads+embedding*9/16), ) } - case "gemma": - fullOffload = 4 * batch * (embedding + vocab) - partialOffload = 4*batch*(2*embedding+vocab+1) + embedding*vocab*105/128 + case "gemma", "gemma2": + fullOffload = max( + 4*batch*(embedding+vocab), + 4*batch*(2+context+context*heads+2*embedding+2*embeddingHeadsK*heads), + ) + + partialOffload = max( + 4*embedding*batch+embedding*vocab*105/128+4*vocab*batch, + 4*batch*(2*embedding+1+2*embeddingHeadsK*heads+context+context*heads)+ + 4*embeddingHeadsK*context*8+ + embedding*embeddingHeadsK*heads*9/16, + ) case "command-r": fullOffload = max( 4*batch*(embedding+vocab), From 6d4219083c56ec4b031f0fda67e9ef2c09ad9888 Mon Sep 17 00:00:00 2001 From: royjhan <65097070+royjhan@users.noreply.github.com> Date: Fri, 28 Jun 2024 09:58:14 -0700 Subject: [PATCH 10/31] Update docs (#5312) --- docs/openai.md | 1 - 1 file changed, 1 deletion(-) diff --git a/docs/openai.md b/docs/openai.md index 59e7d6405..81b967eb7 100644 --- a/docs/openai.md +++ b/docs/openai.md @@ -104,7 +104,6 @@ curl http://localhost:11434/v1/chat/completions \ #### Notes -- `finish_reason` will always be `stop` - `usage.prompt_tokens` will be 0 for completions where prompt evaluation is cached ## Models From b910fa90101038d09ca9cbbea16701831fafaffb Mon Sep 17 00:00:00 2001 From: royjhan <65097070+royjhan@users.noreply.github.com> Date: Fri, 28 Jun 2024 11:30:16 -0700 Subject: [PATCH 11/31] Ollama Show: Check for Projector Type (#5307) * Check exists projtype * Maintain Ordering --- cmd/cmd.go | 12 +++++++++--- 1 file changed, 9 insertions(+), 3 deletions(-) diff --git a/cmd/cmd.go b/cmd/cmd.go index 909e8e4b2..debb39218 100644 --- a/cmd/cmd.go +++ b/cmd/cmd.go @@ -672,11 +672,17 @@ func ShowHandler(cmd *cobra.Command, args []string) error { projectorData := [][]string{ {"arch", "clip"}, {"parameters", format.HumanNumber(uint64(resp.ProjectorInfo["general.parameter_count"].(float64)))}, - {"projector type", resp.ProjectorInfo["clip.projector_type"].(string)}, - {"embedding length", fmt.Sprintf("%v", resp.ProjectorInfo["clip.vision.embedding_length"].(float64))}, - {"projection dimensionality", fmt.Sprintf("%v", resp.ProjectorInfo["clip.vision.projection_dim"].(float64))}, } + if projectorType, ok := resp.ProjectorInfo["clip.projector_type"]; ok { + projectorData = append(projectorData, []string{"projector type", projectorType.(string)}) + } + + projectorData = append(projectorData, + []string{"embedding length", fmt.Sprintf("%v", resp.ProjectorInfo["clip.vision.embedding_length"].(float64))}, + []string{"projection dimensionality", fmt.Sprintf("%v", resp.ProjectorInfo["clip.vision.projection_dim"].(float64))}, + ) + mainTableData = append(mainTableData, []string{"Projector"}, []string{renderSubTable(projectorData, false)}, From 5f034f5b63cab3a5eb61104118727b088cceea21 Mon Sep 17 00:00:00 2001 From: royjhan <65097070+royjhan@users.noreply.github.com> Date: Fri, 28 Jun 2024 13:15:52 -0700 Subject: [PATCH 12/31] Include Show Info in Interactive (#5342) --- cmd/cmd.go | 24 +++++++++++------------- cmd/interactive.go | 10 +--------- 2 files changed, 12 insertions(+), 22 deletions(-) diff --git a/cmd/cmd.go b/cmd/cmd.go index debb39218..c898c7db6 100644 --- a/cmd/cmd.go +++ b/cmd/cmd.go @@ -624,13 +624,13 @@ func ShowHandler(cmd *cobra.Command, args []string) error { return errors.New("only one of '--license', '--modelfile', '--parameters', '--system', or '--template' can be specified") } - if flagsSet == 1 { - req := api.ShowRequest{Name: args[0]} - resp, err := client.Show(cmd.Context(), &req) - if err != nil { - return err - } + req := api.ShowRequest{Name: args[0]} + resp, err := client.Show(cmd.Context(), &req) + if err != nil { + return err + } + if flagsSet == 1 { switch showType { case "license": fmt.Println(resp.License) @@ -647,12 +647,12 @@ func ShowHandler(cmd *cobra.Command, args []string) error { return nil } - req := api.ShowRequest{Name: args[0]} - resp, err := client.Show(cmd.Context(), &req) - if err != nil { - return err - } + showInfo(resp) + return nil +} + +func showInfo(resp *api.ShowResponse) { arch := resp.ModelInfo["general.architecture"].(string) modelData := [][]string{ @@ -711,8 +711,6 @@ func ShowHandler(cmd *cobra.Command, args []string) error { } table.Render() - - return nil } func renderSubTable(data [][]string, file bool) string { diff --git a/cmd/interactive.go b/cmd/interactive.go index 0a2f429b6..9214f2db5 100644 --- a/cmd/interactive.go +++ b/cmd/interactive.go @@ -404,15 +404,7 @@ func generateInteractive(cmd *cobra.Command, opts runOptions) error { switch args[1] { case "info": - fmt.Println("Model details:") - if len(resp.Details.Families) > 0 { - fmt.Printf("Family %s\n", strings.Join(resp.Details.Families, ", ")) - } else if resp.Details.Family != "" { - fmt.Printf("Family %s\n", resp.Details.Family) - } - fmt.Printf("Parameter Size %s\n", resp.Details.ParameterSize) - fmt.Printf("Quantization Level %s\n", resp.Details.QuantizationLevel) - fmt.Println("") + showInfo(resp) case "license": if resp.License == "" { fmt.Println("No license was specified for this model.") From aae56abb7cc96b8495a1c761a08b92cfd136d9d2 Mon Sep 17 00:00:00 2001 From: Daniel Hiltgen Date: Fri, 28 Jun 2024 13:15:57 -0700 Subject: [PATCH 13/31] Document concurrent behavior and settings --- docs/faq.md | 14 ++++++++++++++ 1 file changed, 14 insertions(+) diff --git a/docs/faq.md b/docs/faq.md index b50a3138c..841f1d13d 100644 --- a/docs/faq.md +++ b/docs/faq.md @@ -257,3 +257,17 @@ If you wish to override the `OLLAMA_KEEP_ALIVE` setting, use the `keep_alive` AP ## How do I manage the maximum number of requests the Ollama server can queue? If too many requests are sent to the server, it will respond with a 503 error indicating the server is overloaded. You can adjust how many requests may be queue by setting `OLLAMA_MAX_QUEUE`. + +## How does Ollama handle concurrent requests? + +Ollama supports two levels of concurrent processing. If your system has sufficient available memory (system memory when using CPU inference, or VRAM for GPU inference) then multiple models can be loaded at the same time. For a given model, if there is sufficient available memory when the model is loaded, it is configured to allow parallel request processing. + +If there is insufficient available memory to load a new model request while one or more models are already loaded, all new requests will be queued until the new model can be loaded. As prior models become idle, one or more will be unloaded to make room for the new model. Queued requests will be processed in order. When using GPU inference new models must be able to completely fit in VRAM to allow concurrent model loads. + +Parallel request processing for a given model results in increasing the context size by the number of parallel requests. For example, a 2K context with 4 parallel requests will result in an 8K context and additional memory allocation. + +The following server settings may be used to adjust how Ollama handles concurrent requests: + +- `OLLAMA_MAX_LOADED_MODELS` - The maximum number of models that can be loaded concurrently provided they fit in available memory. The default is 3 * the number of GPUs or 3 for CPU inference. +- `OLLAMA_NUM_PARALLEL` - The maximum number of parallel requests each model will process at the same time. The default will auto-select either 4 or 1 based on available memory. +- `OLLAMA_MAX_QUEUE` - The maximum number of requests Ollama will queue when busy before rejecting additional requests. The default is 512 From 717f7229eb4f9220d4070aae617923950643d327 Mon Sep 17 00:00:00 2001 From: Jeffrey Morgan Date: Fri, 28 Jun 2024 19:39:31 -0700 Subject: [PATCH 14/31] Do not shift context for sliding window models (#5368) * Do not shift context for sliding window models * truncate prompt > 2/3 tokens * only target gemma2 --- llm/ext_server/server.cpp | 46 +++++++++++++++++++++++++++++++-------- 1 file changed, 37 insertions(+), 9 deletions(-) diff --git a/llm/ext_server/server.cpp b/llm/ext_server/server.cpp index 492126a4f..3bc012521 100644 --- a/llm/ext_server/server.cpp +++ b/llm/ext_server/server.cpp @@ -1650,26 +1650,41 @@ struct llama_server_context } slot.params.n_keep = std::min(slot.n_ctx - 4, slot.params.n_keep); + char buf[256]; + llama_model_meta_val_str(model, "general.architecture", buf, 256); + bool gemma2 = strcmp(buf, "gemma2") == 0; + + int32_t truncate_at = slot.n_ctx; + + // truncate at 2/3 of the context length for gemma2 models + // as they do not support context shifts (from the sliding window implementation). + // this way, prompts that almost fit the context length can still generate a full + // response without a sudden stop from hitting the context limit + if (gemma2) { + truncate_at = 2 * slot.n_ctx / 3; + } + // if input prompt is too big, truncate it, if group attention self-extend is disabled - if (slot.ga_n == 1 && slot.n_prompt_tokens >= slot.n_ctx) + if (slot.ga_n == 1 && slot.n_prompt_tokens >= truncate_at) { const int n_left = slot.n_ctx - slot.params.n_keep; - const int n_block_size = n_left / 2; - const int erased_blocks = (slot.n_prompt_tokens - slot.params.n_keep - n_block_size) / n_block_size; + const int n_shift = n_left / 2; + const int n_erase = slot.n_prompt_tokens - slot.params.n_keep - n_shift; std::vector new_tokens( prompt_tokens.begin(), prompt_tokens.begin() + slot.params.n_keep); new_tokens.insert( new_tokens.end(), - prompt_tokens.begin() + slot.params.n_keep + erased_blocks * n_block_size, + prompt_tokens.begin() + slot.params.n_keep + n_erase, prompt_tokens.end()); - LOG_VERBOSE("input truncated", { - {"n_ctx", slot.n_ctx}, - {"n_keep", slot.params.n_keep}, - {"n_left", n_left}, - {"new_tokens", tokens_to_str(ctx, new_tokens.cbegin(), new_tokens.cend())}, + LOG_INFO("input truncated", { + {"n_ctx", slot.n_ctx}, + {"n_keep", slot.params.n_keep}, + {"n_left", n_left}, + {"n_shift", n_shift}, + {"n_erase", n_erase}, }); slot.truncated = true; prompt_tokens = new_tokens; @@ -1678,6 +1693,19 @@ struct llama_server_context GGML_ASSERT(slot.n_prompt_tokens < slot.n_ctx); } + // Models with sliding window attention do not work with context shifts, so + // limit their prediction to the context length + if (gemma2) { + int32_t limit = slot.n_ctx - slot.n_prompt_tokens; + slot.n_predict = limit; + slot.params.n_predict = limit; + LOG_INFO("model does not support sliding window, limiting generation", { + {"n_ctx", slot.n_ctx}, + {"n_prompt_tokens", slot.n_prompt_tokens}, + {"n_predict", slot.n_predict} + }); + } + if (!slot.params.cache_prompt) { llama_sampling_reset(slot.ctx_sampling); From c1218199cfe82eda35f5e4a8031eee28f01ebf75 Mon Sep 17 00:00:00 2001 From: Jeffrey Morgan Date: Sat, 29 Jun 2024 16:22:49 -0700 Subject: [PATCH 15/31] Update api.md --- docs/api.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/docs/api.md b/docs/api.md index 107b5211f..c577bb1a5 100644 --- a/docs/api.md +++ b/docs/api.md @@ -26,7 +26,7 @@ All durations are returned in nanoseconds. ### Streaming responses -Certain endpoints stream responses as JSON objects and can optional return non-streamed responses. +Certain endpoints stream responses as JSON objects. Streaming can be disabled by providing `{"stream": false}` for these endpoints. ## Generate a completion From 27402cb7a28555a3efcaa5af054b1ce2d18e5442 Mon Sep 17 00:00:00 2001 From: Eduard Date: Mon, 1 Jul 2024 03:48:51 +0200 Subject: [PATCH 16/31] Update gpu.md (#5382) Runs fine on a NVIDIA GeForce GTX 1050 Ti --- docs/gpu.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/docs/gpu.md b/docs/gpu.md index 55c41c9de..80f276c3b 100644 --- a/docs/gpu.md +++ b/docs/gpu.md @@ -18,7 +18,7 @@ Check your compute compatibility to see if your card is supported: | | Quadro | `RTX 8000` `RTX 6000` `RTX 5000` `RTX 4000` | | 7.0 | NVIDIA | `TITAN V` `V100` `Quadro GV100` | | 6.1 | NVIDIA TITAN | `TITAN Xp` `TITAN X` | -| | GeForce GTX | `GTX 1080 Ti` `GTX 1080` `GTX 1070 Ti` `GTX 1070` `GTX 1060` `GTX 1050` | +| | GeForce GTX | `GTX 1080 Ti` `GTX 1080` `GTX 1070 Ti` `GTX 1070` `GTX 1060` `GTX 1050 Ti` `GTX 1050` | | | Quadro | `P6000` `P5200` `P4200` `P3200` `P5000` `P4000` `P3000` `P2200` `P2000` `P1000` `P620` `P600` `P500` `P520` | | | Tesla | `P40` `P4` | | 6.0 | NVIDIA | `Tesla P100` `Quadro GP100` | From 1963c00201958da7165a40f9d2f22b28e11be718 Mon Sep 17 00:00:00 2001 From: RAPID ARCHITECT <126218667+rapidarchitect@users.noreply.github.com> Date: Sun, 30 Jun 2024 21:00:57 -0500 Subject: [PATCH 17/31] Update README.md (#5214) * Update README.md Added Mesop example to web & desktop * Update README.md --------- Co-authored-by: Jeffrey Morgan --- README.md | 1 + 1 file changed, 1 insertion(+) diff --git a/README.md b/README.md index 72ed8fa5e..62f5cd65c 100644 --- a/README.md +++ b/README.md @@ -292,6 +292,7 @@ See the [API documentation](./docs/api.md) for all endpoints. - [Olpaka](https://github.com/Otacon/olpaka) (User-friendly Flutter Web App for Ollama) - [OllamaSpring](https://github.com/CrazyNeil/OllamaSpring) (Ollama Client for macOS) - [LLocal.in](https://github.com/kartikm7/llocal) (Easy to use Electron Desktop Client for Ollama) +- [Ollama with Google Mesop](https://github.com/rapidarchitect/ollama_mesop/) (Mesop Chat Client implementation with Ollama) ### Terminal From 26e4e66faff20a94bb8fee9ec2bc3e17a07fb19e Mon Sep 17 00:00:00 2001 From: Josh Yan Date: Mon, 1 Jul 2024 09:43:49 -0700 Subject: [PATCH 18/31] updated parsefile test --- parser/parser_test.go | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-) diff --git a/parser/parser_test.go b/parser/parser_test.go index 3dc592239..171bd4206 100644 --- a/parser/parser_test.go +++ b/parser/parser_test.go @@ -22,7 +22,13 @@ ADAPTER adapter1 LICENSE MIT PARAMETER param1 value1 PARAMETER param2 value2 -TEMPLATE template1 +TEMPLATE """{{ if .System }}<|start_header_id|>system<|end_header_id|> + +{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|> + +{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|> + +{{ .Response }}<|eot_id|>""" ` reader := strings.NewReader(input) @@ -36,7 +42,7 @@ TEMPLATE template1 {Name: "license", Args: "MIT"}, {Name: "param1", Args: "value1"}, {Name: "param2", Args: "value2"}, - {Name: "template", Args: "template1"}, + {Name: "template", Args: "{{ if .System }}<|start_header_id|>system<|end_header_id|>\n\n{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>\n\n{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>\n\n{{ .Response }}<|eot_id|>"}, } assert.Equal(t, expectedCommands, modelfile.Commands) From cff3f44f4a4097de864d70d9a95f31c62e8ecdfa Mon Sep 17 00:00:00 2001 From: Daniel Hiltgen Date: Mon, 1 Jul 2024 09:43:59 -0700 Subject: [PATCH 19/31] Fix case for NumCtx --- server/sched.go | 14 +++++++------- 1 file changed, 7 insertions(+), 7 deletions(-) diff --git a/server/sched.go b/server/sched.go index 87da1db47..71b535ae2 100644 --- a/server/sched.go +++ b/server/sched.go @@ -23,7 +23,7 @@ type LlmRequest struct { ctx context.Context //nolint:containedctx model *Model opts api.Options - origNumCTX int // Track the initial ctx request + origNumCtx int // Track the initial ctx request sessionDuration time.Duration successCh chan *runnerRef errCh chan error @@ -118,8 +118,8 @@ func (s *Scheduler) processPending(ctx context.Context) { case pending := <-s.pendingReqCh: // Block other requests until we get this pending request running pending.schedAttempts++ - if pending.origNumCTX == 0 { - pending.origNumCTX = pending.opts.NumCtx + if pending.origNumCtx == 0 { + pending.origNumCtx = pending.opts.NumCtx } if pending.ctx.Err() != nil { @@ -135,7 +135,7 @@ func (s *Scheduler) processPending(ctx context.Context) { } // Keep NumCtx and numParallel in sync if numParallel > 1 { - pending.opts.NumCtx = pending.origNumCTX * numParallel + pending.opts.NumCtx = pending.origNumCtx * numParallel } for { @@ -197,7 +197,7 @@ func (s *Scheduler) processPending(ctx context.Context) { // simplifying assumption of defaultParallel when in CPU mode if numParallel <= 0 { numParallel = defaultParallel - pending.opts.NumCtx = pending.origNumCTX * numParallel + pending.opts.NumCtx = pending.origNumCtx * numParallel } if loadedCount == 0 { @@ -691,7 +691,7 @@ func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numP // First attempt to fit the model into a single GPU for _, p := range numParallelToTry { - req.opts.NumCtx = req.origNumCTX * p + req.opts.NumCtx = req.origNumCtx * p if !envconfig.SchedSpread { for _, g := range sgl { if ok, estimatedVRAM = llm.PredictServerFit([]gpu.GpuInfo{g}, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok { @@ -709,7 +709,7 @@ func pickBestFitGPUs(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numP // Now try all the GPUs for _, p := range numParallelToTry { - req.opts.NumCtx = req.origNumCTX * p + req.opts.NumCtx = req.origNumCtx * p if ok, estimatedVRAM = llm.PredictServerFit(sgl, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts); ok { slog.Info("new model will fit in available VRAM, loading", "model", req.model.ModelPath, "library", sgl[0].Library, "parallel", p, "required", format.HumanBytes2(estimatedVRAM)) *numParallel = p From 173b5504381a77b042f3957226a23c0569406aca Mon Sep 17 00:00:00 2001 From: Daniel Hiltgen Date: Mon, 1 Jul 2024 09:48:05 -0700 Subject: [PATCH 20/31] Remove default auto from help message This may confuse users thinking "auto" is an acceptable string - it must be numeric --- envconfig/config.go | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/envconfig/config.go b/envconfig/config.go index 0f0f7f058..c02c4878e 100644 --- a/envconfig/config.go +++ b/envconfig/config.go @@ -85,13 +85,13 @@ func AsMap() map[string]EnvVar { "OLLAMA_HOST": {"OLLAMA_HOST", Host, "IP Address for the ollama server (default 127.0.0.1:11434)"}, "OLLAMA_KEEP_ALIVE": {"OLLAMA_KEEP_ALIVE", KeepAlive, "The duration that models stay loaded in memory (default \"5m\")"}, "OLLAMA_LLM_LIBRARY": {"OLLAMA_LLM_LIBRARY", LLMLibrary, "Set LLM library to bypass autodetection"}, - "OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners, "Maximum number of loaded models per GPU (default auto)"}, + "OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners, "Maximum number of loaded models per GPU"}, "OLLAMA_MAX_QUEUE": {"OLLAMA_MAX_QUEUE", MaxQueuedRequests, "Maximum number of queued requests"}, "OLLAMA_MAX_VRAM": {"OLLAMA_MAX_VRAM", MaxVRAM, "Maximum VRAM"}, "OLLAMA_MODELS": {"OLLAMA_MODELS", ModelsDir, "The path to the models directory"}, "OLLAMA_NOHISTORY": {"OLLAMA_NOHISTORY", NoHistory, "Do not preserve readline history"}, "OLLAMA_NOPRUNE": {"OLLAMA_NOPRUNE", NoPrune, "Do not prune model blobs on startup"}, - "OLLAMA_NUM_PARALLEL": {"OLLAMA_NUM_PARALLEL", NumParallel, "Maximum number of parallel requests (default auto)"}, + "OLLAMA_NUM_PARALLEL": {"OLLAMA_NUM_PARALLEL", NumParallel, "Maximum number of parallel requests"}, "OLLAMA_ORIGINS": {"OLLAMA_ORIGINS", AllowOrigins, "A comma separated list of allowed origins"}, "OLLAMA_RUNNERS_DIR": {"OLLAMA_RUNNERS_DIR", RunnersDir, "Location for runners"}, "OLLAMA_SCHED_SPREAD": {"OLLAMA_SCHED_SPREAD", SchedSpread, "Always schedule model across all GPUs"}, From 3f0b309ad4c49c0d87839e50fe6a46163902aba0 Mon Sep 17 00:00:00 2001 From: Michael Yang Date: Mon, 10 Jun 2024 08:47:13 -0700 Subject: [PATCH 21/31] remove ManifestV2 --- server/images.go | 17 +++++------------ server/manifest.go | 20 +++++++++++--------- server/manifest_test.go | 2 +- 3 files changed, 17 insertions(+), 22 deletions(-) diff --git a/server/images.go b/server/images.go index e949fb18a..447a63a69 100644 --- a/server/images.go +++ b/server/images.go @@ -135,13 +135,6 @@ type Message struct { Content string `json:"content"` } -type ManifestV2 struct { - SchemaVersion int `json:"schemaVersion"` - MediaType string `json:"mediaType"` - Config *Layer `json:"config"` - Layers []*Layer `json:"layers"` -} - type ConfigV2 struct { ModelFormat string `json:"model_format"` ModelFamily string `json:"model_family"` @@ -160,7 +153,7 @@ type RootFS struct { DiffIDs []string `json:"diff_ids"` } -func GetManifest(mp ModelPath) (*ManifestV2, string, error) { +func GetManifest(mp ModelPath) (*Manifest, string, error) { fp, err := mp.GetManifestPath() if err != nil { return nil, "", err @@ -170,7 +163,7 @@ func GetManifest(mp ModelPath) (*ManifestV2, string, error) { return nil, "", err } - var manifest *ManifestV2 + var manifest *Manifest bts, err := os.ReadFile(fp) if err != nil { @@ -822,7 +815,7 @@ func PushModel(ctx context.Context, name string, regOpts *registryOptions, fn fu func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn func(api.ProgressResponse)) error { mp := ParseModelPath(name) - var manifest *ManifestV2 + var manifest *Manifest var err error var noprune string @@ -929,7 +922,7 @@ func PullModel(ctx context.Context, name string, regOpts *registryOptions, fn fu return nil } -func pullModelManifest(ctx context.Context, mp ModelPath, regOpts *registryOptions) (*ManifestV2, error) { +func pullModelManifest(ctx context.Context, mp ModelPath, regOpts *registryOptions) (*Manifest, error) { requestURL := mp.BaseURL().JoinPath("v2", mp.GetNamespaceRepository(), "manifests", mp.Tag) headers := make(http.Header) @@ -940,7 +933,7 @@ func pullModelManifest(ctx context.Context, mp ModelPath, regOpts *registryOptio } defer resp.Body.Close() - var m *ManifestV2 + var m *Manifest if err := json.NewDecoder(resp.Body).Decode(&m); err != nil { return nil, err } diff --git a/server/manifest.go b/server/manifest.go index 61dd1ab4e..726bb48d8 100644 --- a/server/manifest.go +++ b/server/manifest.go @@ -14,7 +14,10 @@ import ( ) type Manifest struct { - ManifestV2 + SchemaVersion int `json:"schemaVersion"` + MediaType string `json:"mediaType"` + Config *Layer `json:"config"` + Layers []*Layer `json:"layers"` filepath string fi os.FileInfo @@ -66,7 +69,7 @@ func ParseNamedManifest(n model.Name) (*Manifest, error) { p := filepath.Join(manifests, n.Filepath()) - var m ManifestV2 + var m Manifest f, err := os.Open(p) if err != nil { return nil, err @@ -83,12 +86,11 @@ func ParseNamedManifest(n model.Name) (*Manifest, error) { return nil, err } - return &Manifest{ - ManifestV2: m, - filepath: p, - fi: fi, - digest: fmt.Sprintf("%x", sha256sum.Sum(nil)), - }, nil + m.filepath = p + m.fi = fi + m.digest = fmt.Sprintf("%x", sha256sum.Sum(nil)) + + return &m, nil } func WriteManifest(name model.Name, config *Layer, layers []*Layer) error { @@ -108,7 +110,7 @@ func WriteManifest(name model.Name, config *Layer, layers []*Layer) error { } defer f.Close() - m := ManifestV2{ + m := Manifest{ SchemaVersion: 2, MediaType: "application/vnd.docker.distribution.manifest.v2+json", Config: config, diff --git a/server/manifest_test.go b/server/manifest_test.go index ceee31d88..ca6c3d2e9 100644 --- a/server/manifest_test.go +++ b/server/manifest_test.go @@ -25,7 +25,7 @@ func createManifest(t *testing.T, path, name string) { } defer f.Close() - if err := json.NewEncoder(f).Encode(ManifestV2{}); err != nil { + if err := json.NewEncoder(f).Encode(Manifest{}); err != nil { t.Fatal(err) } } From 58e3fff311f9e7abec20cdfe20fa43958e447aeb Mon Sep 17 00:00:00 2001 From: Michael Yang Date: Mon, 10 Jun 2024 14:54:42 -0700 Subject: [PATCH 22/31] rename templates to template --- server/images.go | 26 ++- server/model.go | 4 +- server/prompt.go | 18 +- server/prompt_test.go | 15 +- server/routes.go | 26 ++- {templates => template}/alfred.gotmpl | 0 {templates => template}/alpaca.gotmpl | 0 {templates => template}/chatml.gotmpl | 0 {templates => template}/chatqa.gotmpl | 0 .../codellama-70b-instruct.gotmpl | 0 .../falcon-instruct.gotmpl | 0 {templates => template}/gemma-instruct.gotmpl | 0 .../granite-instruct.gotmpl | 0 {templates => template}/index.json | 0 {templates => template}/llama2-chat.gotmpl | 0 .../llama3-instruct.gotmpl | 0 {templates => template}/magicoder.gotmpl | 0 .../mistral-instruct.gotmpl | 0 {templates => template}/openchat.gotmpl | 0 {templates => template}/phi-3.gotmpl | 0 {templates => template}/solar-instruct.gotmpl | 0 .../starcoder2-instruct.gotmpl | 0 template/template.go | 158 ++++++++++++++++++ template/template_test.go | 89 ++++++++++ .../testdata/templates.jsonl | 0 {templates => template}/vicuna.gotmpl | 0 {templates => template}/zephyr.gotmpl | 0 templates/template.go | 70 -------- templates/template_test.go | 59 ------- 29 files changed, 301 insertions(+), 164 deletions(-) rename {templates => template}/alfred.gotmpl (100%) rename {templates => template}/alpaca.gotmpl (100%) rename {templates => template}/chatml.gotmpl (100%) rename {templates => template}/chatqa.gotmpl (100%) rename {templates => template}/codellama-70b-instruct.gotmpl (100%) rename {templates => template}/falcon-instruct.gotmpl (100%) rename {templates => template}/gemma-instruct.gotmpl (100%) rename {templates => template}/granite-instruct.gotmpl (100%) rename {templates => template}/index.json (100%) rename {templates => template}/llama2-chat.gotmpl (100%) rename {templates => template}/llama3-instruct.gotmpl (100%) rename {templates => template}/magicoder.gotmpl (100%) rename {templates => template}/mistral-instruct.gotmpl (100%) rename {templates => template}/openchat.gotmpl (100%) rename {templates => template}/phi-3.gotmpl (100%) rename {templates => template}/solar-instruct.gotmpl (100%) rename {templates => template}/starcoder2-instruct.gotmpl (100%) create mode 100644 template/template.go create mode 100644 template/template_test.go rename {templates => template}/testdata/templates.jsonl (100%) rename {templates => template}/vicuna.gotmpl (100%) rename {templates => template}/zephyr.gotmpl (100%) delete mode 100644 templates/template.go delete mode 100644 templates/template_test.go diff --git a/server/images.go b/server/images.go index 447a63a69..65ed51c76 100644 --- a/server/images.go +++ b/server/images.go @@ -28,6 +28,7 @@ import ( "github.com/ollama/ollama/format" "github.com/ollama/ollama/llm" "github.com/ollama/ollama/parser" + "github.com/ollama/ollama/template" "github.com/ollama/ollama/types/errtypes" "github.com/ollama/ollama/types/model" "github.com/ollama/ollama/version" @@ -48,12 +49,13 @@ type Model struct { ParentModel string AdapterPaths []string ProjectorPaths []string - Template string System string License []string Digest string Options map[string]interface{} Messages []Message + + Template *template.Template } func (m *Model) IsEmbedding() bool { @@ -82,10 +84,10 @@ func (m *Model) String() string { }) } - if m.Template != "" { + if m.Template != nil { modelfile.Commands = append(modelfile.Commands, parser.Command{ Name: "template", - Args: m.Template, + Args: m.Template.String(), }) } @@ -191,8 +193,7 @@ func GetModel(name string) (*Model, error) { Name: mp.GetFullTagname(), ShortName: mp.GetShortTagname(), Digest: digest, - Template: "{{ .Prompt }}", - License: []string{}, + Template: template.DefaultTemplate, } filename, err := GetBlobsPath(manifest.Config.Digest) @@ -228,13 +229,17 @@ func GetModel(name string) (*Model, error) { model.AdapterPaths = append(model.AdapterPaths, filename) case "application/vnd.ollama.image.projector": model.ProjectorPaths = append(model.ProjectorPaths, filename) - case "application/vnd.ollama.image.template": + case "application/vnd.ollama.image.prompt", + "application/vnd.ollama.image.template": bts, err := os.ReadFile(filename) if err != nil { return nil, err } - model.Template = string(bts) + model.Template, err = template.Parse(string(bts)) + if err != nil { + return nil, err + } case "application/vnd.ollama.image.system": bts, err := os.ReadFile(filename) if err != nil { @@ -242,13 +247,6 @@ func GetModel(name string) (*Model, error) { } model.System = string(bts) - case "application/vnd.ollama.image.prompt": - bts, err := os.ReadFile(filename) - if err != nil { - return nil, err - } - - model.Template = string(bts) case "application/vnd.ollama.image.params": params, err := os.Open(filename) if err != nil { diff --git a/server/model.go b/server/model.go index d56e641ba..6abb5b392 100644 --- a/server/model.go +++ b/server/model.go @@ -16,7 +16,7 @@ import ( "github.com/ollama/ollama/api" "github.com/ollama/ollama/convert" "github.com/ollama/ollama/llm" - "github.com/ollama/ollama/templates" + "github.com/ollama/ollama/template" "github.com/ollama/ollama/types/model" ) @@ -258,7 +258,7 @@ func parseFromFile(ctx context.Context, file *os.File, digest string, fn func(ap func detectChatTemplate(layers []*layerGGML) ([]*layerGGML, error) { for _, layer := range layers { if s := layer.GGML.KV().ChatTemplate(); s != "" { - if t, err := templates.NamedTemplate(s); err != nil { + if t, err := template.Named(s); err != nil { slog.Debug("template detection", "error", err) } else { tmpl, err := NewLayer(t.Reader(), "application/vnd.ollama.image.template") diff --git a/server/prompt.go b/server/prompt.go index 604e69717..bfc319a50 100644 --- a/server/prompt.go +++ b/server/prompt.go @@ -4,10 +4,11 @@ import ( "fmt" "log/slog" "strings" - "text/template" + "text/template/parse" "github.com/ollama/ollama/api" + "github.com/ollama/ollama/template" ) // isResponseNode checks if the node contains .Response @@ -53,13 +54,8 @@ func formatTemplateForResponse(tmpl *template.Template, generate bool) { // Prompt renders a prompt from a template. If generate is set to true, // the response and parts of the template following it are not rendered -func Prompt(tmpl, system, prompt, response string, generate bool) (string, error) { - parsed, err := template.New("").Option("missingkey=zero").Parse(tmpl) - if err != nil { - return "", err - } - - formatTemplateForResponse(parsed, generate) +func Prompt(tmpl *template.Template, system, prompt, response string, generate bool) (string, error) { + formatTemplateForResponse(tmpl, generate) vars := map[string]any{ "System": system, @@ -68,14 +64,14 @@ func Prompt(tmpl, system, prompt, response string, generate bool) (string, error } var sb strings.Builder - if err := parsed.Execute(&sb, vars); err != nil { + if err := tmpl.Execute(&sb, vars); err != nil { return "", err } return sb.String(), nil } -func countTokens(tmpl string, system string, prompt string, response string, encode func(string) ([]int, error)) (int, error) { +func countTokens(tmpl *template.Template, system string, prompt string, response string, encode func(string) ([]int, error)) (int, error) { rendered, err := Prompt(tmpl, system, prompt, response, false) if err != nil { return 0, err @@ -91,7 +87,7 @@ func countTokens(tmpl string, system string, prompt string, response string, enc } // ChatPrompt builds up a prompt from a series of messages, truncating based on context window size -func ChatPrompt(tmpl string, messages []api.Message, window int, encode func(string) ([]int, error)) (string, error) { +func ChatPrompt(tmpl *template.Template, messages []api.Message, window int, encode func(string) ([]int, error)) (string, error) { type prompt struct { System string Prompt string diff --git a/server/prompt_test.go b/server/prompt_test.go index a7e18a70f..7df58d0bd 100644 --- a/server/prompt_test.go +++ b/server/prompt_test.go @@ -5,6 +5,7 @@ import ( "testing" "github.com/ollama/ollama/api" + "github.com/ollama/ollama/template" ) func TestPrompt(t *testing.T) { @@ -61,7 +62,12 @@ func TestPrompt(t *testing.T) { for _, tc := range tests { t.Run(tc.name, func(t *testing.T) { - got, err := Prompt(tc.template, tc.system, tc.prompt, tc.response, tc.generate) + tmpl, err := template.Parse(tc.template) + if err != nil { + t.Fatal(err) + } + + got, err := Prompt(tmpl, tc.system, tc.prompt, tc.response, tc.generate) if err != nil { t.Errorf("error = %v", err) } @@ -192,7 +198,12 @@ func TestChatPrompt(t *testing.T) { for _, tc := range tests { t.Run(tc.name, func(t *testing.T) { - got, err := ChatPrompt(tc.template, tc.messages, tc.window, encode) + tmpl, err := template.Parse(tc.template) + if err != nil { + t.Fatal(err) + } + + got, err := ChatPrompt(tmpl, tc.messages, tc.window, encode) if err != nil { t.Errorf("error = %v", err) } diff --git a/server/routes.go b/server/routes.go index 76ead072f..d8a4a67e7 100644 --- a/server/routes.go +++ b/server/routes.go @@ -31,6 +31,7 @@ import ( "github.com/ollama/ollama/llm" "github.com/ollama/ollama/openai" "github.com/ollama/ollama/parser" + "github.com/ollama/ollama/template" "github.com/ollama/ollama/types/errtypes" "github.com/ollama/ollama/types/model" "github.com/ollama/ollama/version" @@ -161,6 +162,12 @@ func (s *Server) GenerateHandler(c *gin.Context) { return } + tmpl, err := template.Parse(req.Template) + if err != nil { + c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()}) + return + } + checkpointLoaded := time.Now() var prompt string @@ -169,7 +176,11 @@ func (s *Server) GenerateHandler(c *gin.Context) { prompt = req.Prompt case req.Prompt != "": if req.Template == "" { - req.Template = model.Template + model.Template, err = template.Parse(req.Template) + if err != nil { + c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()}) + return + } } if req.System == "" { @@ -187,7 +198,7 @@ func (s *Server) GenerateHandler(c *gin.Context) { sb.WriteString(req.Prompt) - p, err := Prompt(req.Template, req.System, sb.String(), "", true) + p, err := Prompt(tmpl, req.System, sb.String(), "", true) if err != nil { c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()}) return @@ -242,7 +253,7 @@ func (s *Server) GenerateHandler(c *gin.Context) { resp.LoadDuration = checkpointLoaded.Sub(checkpointStart) if !req.Raw { - p, err := Prompt(req.Template, req.System, req.Prompt, generated.String(), false) + p, err := Prompt(tmpl, req.System, req.Prompt, generated.String(), false) if err != nil { c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()}) return @@ -680,7 +691,10 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) { } if req.Template != "" { - m.Template = req.Template + m.Template, err = template.Parse(req.Template) + if err != nil { + return nil, err + } } msgs := make([]api.Message, 0) @@ -701,7 +715,7 @@ func GetModelInfo(req api.ShowRequest) (*api.ShowResponse, error) { resp := &api.ShowResponse{ License: strings.Join(m.License, "\n"), System: m.System, - Template: m.Template, + Template: m.Template.String(), Details: modelDetails, Messages: msgs, ModifiedAt: manifest.fi.ModTime(), @@ -1246,7 +1260,7 @@ func (s *Server) ProcessHandler(c *gin.Context) { } // ChatPrompt builds up a prompt from a series of messages for the currently `loaded` model -func chatPrompt(ctx context.Context, runner *runnerRef, template string, messages []api.Message, numCtx int) (string, error) { +func chatPrompt(ctx context.Context, runner *runnerRef, template *template.Template, messages []api.Message, numCtx int) (string, error) { encode := func(s string) ([]int, error) { return runner.llama.Tokenize(ctx, s) } diff --git a/templates/alfred.gotmpl b/template/alfred.gotmpl similarity index 100% rename from templates/alfred.gotmpl rename to template/alfred.gotmpl diff --git a/templates/alpaca.gotmpl b/template/alpaca.gotmpl similarity index 100% rename from templates/alpaca.gotmpl rename to template/alpaca.gotmpl diff --git a/templates/chatml.gotmpl b/template/chatml.gotmpl similarity index 100% rename from templates/chatml.gotmpl rename to template/chatml.gotmpl diff --git a/templates/chatqa.gotmpl b/template/chatqa.gotmpl similarity index 100% rename from templates/chatqa.gotmpl rename to template/chatqa.gotmpl diff --git a/templates/codellama-70b-instruct.gotmpl b/template/codellama-70b-instruct.gotmpl similarity index 100% rename from templates/codellama-70b-instruct.gotmpl rename to template/codellama-70b-instruct.gotmpl diff --git a/templates/falcon-instruct.gotmpl b/template/falcon-instruct.gotmpl similarity index 100% rename from templates/falcon-instruct.gotmpl rename to template/falcon-instruct.gotmpl diff --git a/templates/gemma-instruct.gotmpl b/template/gemma-instruct.gotmpl similarity index 100% rename from templates/gemma-instruct.gotmpl rename to template/gemma-instruct.gotmpl diff --git a/templates/granite-instruct.gotmpl b/template/granite-instruct.gotmpl similarity index 100% rename from templates/granite-instruct.gotmpl rename to template/granite-instruct.gotmpl diff --git a/templates/index.json b/template/index.json similarity index 100% rename from templates/index.json rename to template/index.json diff --git a/templates/llama2-chat.gotmpl b/template/llama2-chat.gotmpl similarity index 100% rename from templates/llama2-chat.gotmpl rename to template/llama2-chat.gotmpl diff --git a/templates/llama3-instruct.gotmpl b/template/llama3-instruct.gotmpl similarity index 100% rename from templates/llama3-instruct.gotmpl rename to template/llama3-instruct.gotmpl diff --git a/templates/magicoder.gotmpl b/template/magicoder.gotmpl similarity index 100% rename from templates/magicoder.gotmpl rename to template/magicoder.gotmpl diff --git a/templates/mistral-instruct.gotmpl b/template/mistral-instruct.gotmpl similarity index 100% rename from templates/mistral-instruct.gotmpl rename to template/mistral-instruct.gotmpl diff --git a/templates/openchat.gotmpl b/template/openchat.gotmpl similarity index 100% rename from templates/openchat.gotmpl rename to template/openchat.gotmpl diff --git a/templates/phi-3.gotmpl b/template/phi-3.gotmpl similarity index 100% rename from templates/phi-3.gotmpl rename to template/phi-3.gotmpl diff --git a/templates/solar-instruct.gotmpl b/template/solar-instruct.gotmpl similarity index 100% rename from templates/solar-instruct.gotmpl rename to template/solar-instruct.gotmpl diff --git a/templates/starcoder2-instruct.gotmpl b/template/starcoder2-instruct.gotmpl similarity index 100% rename from templates/starcoder2-instruct.gotmpl rename to template/starcoder2-instruct.gotmpl diff --git a/template/template.go b/template/template.go new file mode 100644 index 000000000..d15f7156f --- /dev/null +++ b/template/template.go @@ -0,0 +1,158 @@ +package template + +import ( + "bytes" + "embed" + "encoding/json" + "errors" + "io" + "math" + "slices" + "strings" + "sync" + "text/template" + "text/template/parse" + + "github.com/agnivade/levenshtein" + "golang.org/x/exp/maps" +) + +//go:embed index.json +var indexBytes []byte + +//go:embed *.gotmpl +var templatesFS embed.FS + +var templatesOnce = sync.OnceValues(func() ([]*named, error) { + var templates []*named + if err := json.Unmarshal(indexBytes, &templates); err != nil { + return nil, err + } + + for _, t := range templates { + bts, err := templatesFS.ReadFile(t.Name + ".gotmpl") + if err != nil { + return nil, err + } + + // normalize line endings + t.Bytes = bytes.ReplaceAll(bts, []byte("\r\n"), []byte("\n")) + } + + return templates, nil +}) + +type named struct { + Name string `json:"name"` + Template string `json:"template"` + Bytes []byte +} + +func (t named) Reader() io.Reader { + return bytes.NewReader(t.Bytes) +} + +func Named(s string) (*named, error) { + templates, err := templatesOnce() + if err != nil { + return nil, err + } + + var template *named + score := math.MaxInt + for _, t := range templates { + if s := levenshtein.ComputeDistance(s, t.Template); s < score { + score = s + template = t + } + } + + if score < 100 { + return template, nil + } + + return nil, errors.New("no matching template found") +} + +type Template struct { + *template.Template + raw string +} + +func (t *Template) String() string { + return t.raw +} + +var DefaultTemplate, _ = Parse("{{ .Prompt }}") + +func Parse(s string) (*Template, error) { + t, err := template.New("").Option("missingkey=zero").Parse(s) + if err != nil { + return nil, err + } + + return &Template{Template: t, raw: s}, nil +} + +func (t *Template) Vars() []string { + var vars []string + for _, n := range t.Tree.Root.Nodes { + vars = append(vars, parseNode(n)...) + } + + set := make(map[string]struct{}) + for _, n := range vars { + set[strings.ToLower(n)] = struct{}{} + } + + vars = maps.Keys(set) + slices.Sort(vars) + return vars +} + +func parseNode(n parse.Node) []string { + switch n := n.(type) { + case *parse.ActionNode: + return parseNode(n.Pipe) + case *parse.IfNode: + names := parseNode(n.Pipe) + names = append(names, parseNode(n.List)...) + if n.ElseList != nil { + names = append(names, parseNode(n.ElseList)...) + } + return names + case *parse.RangeNode: + names := parseNode(n.Pipe) + names = append(names, parseNode(n.List)...) + if n.ElseList != nil { + names = append(names, parseNode(n.ElseList)...) + } + return names + case *parse.WithNode: + names := parseNode(n.Pipe) + names = append(names, parseNode(n.List)...) + if n.ElseList != nil { + names = append(names, parseNode(n.ElseList)...) + } + return names + case *parse.PipeNode: + var names []string + for _, c := range n.Cmds { + for _, a := range c.Args { + names = append(names, parseNode(a)...) + } + } + return names + case *parse.ListNode: + var names []string + for _, n := range n.Nodes { + names = append(names, parseNode(n)...) + } + + return names + case *parse.FieldNode: + return n.Ident + } + + return nil +} diff --git a/template/template_test.go b/template/template_test.go new file mode 100644 index 000000000..e5405bdb4 --- /dev/null +++ b/template/template_test.go @@ -0,0 +1,89 @@ +package template + +import ( + "bufio" + "bytes" + "encoding/json" + "io" + "os" + "path/filepath" + "slices" + "testing" + "text/template" + + "github.com/ollama/ollama/llm" +) + +func TestNamed(t *testing.T) { + f, err := os.Open(filepath.Join("testdata", "templates.jsonl")) + if err != nil { + t.Fatal(err) + } + defer f.Close() + + scanner := bufio.NewScanner(f) + for scanner.Scan() { + var ss map[string]string + if err := json.Unmarshal(scanner.Bytes(), &ss); err != nil { + t.Fatal(err) + } + + for k, v := range ss { + t.Run(k, func(t *testing.T) { + kv := llm.KV{"tokenizer.chat_template": v} + s := kv.ChatTemplate() + r, err := Named(s) + if err != nil { + t.Fatal(err) + } + + if r.Name != k { + t.Errorf("expected %q, got %q", k, r.Name) + } + + var b bytes.Buffer + if _, err := io.Copy(&b, r.Reader()); err != nil { + t.Fatal(err) + } + + tmpl, err := template.New(s).Parse(b.String()) + if err != nil { + t.Fatal(err) + } + + if tmpl.Tree.Root.String() == "" { + t.Errorf("empty %s template", k) + } + }) + } + } +} + +func TestParse(t *testing.T) { + cases := []struct { + template string + capabilities []string + }{ + {"{{ .Prompt }}", []string{"prompt"}}, + {"{{ .System }} {{ .Prompt }}", []string{"prompt", "system"}}, + {"{{ .System }} {{ .Prompt }} {{ .Response }}", []string{"prompt", "response", "system"}}, + {"{{ with .Tools }}{{ . }}{{ end }} {{ .System }} {{ .Prompt }}", []string{"prompt", "system", "tools"}}, + {"{{ range .Messages }}{{ .Role }} {{ .Content }}{{ end }}", []string{"content", "messages", "role"}}, + {"{{ range .Messages }}{{ if eq .Role \"system\" }}SYSTEM: {{ .Content }}{{ else if eq .Role \"user\" }}USER: {{ .Content }}{{ else if eq .Role \"assistant\" }}ASSISTANT: {{ .Content }}{{ end }}{{ end }}", []string{"content", "messages", "role"}}, + {"{{ .Prompt }} {{ .Suffix }}", []string{"prompt", "suffix"}}, + } + + for _, tt := range cases { + t.Run("", func(t *testing.T) { + tmpl, err := Parse(tt.template) + if err != nil { + t.Fatal(err) + } + + vars := tmpl.Vars() + if !slices.Equal(tt.capabilities, vars) { + t.Errorf("expected %v, got %v", tt.capabilities, vars) + } + }) + } +} diff --git a/templates/testdata/templates.jsonl b/template/testdata/templates.jsonl similarity index 100% rename from templates/testdata/templates.jsonl rename to template/testdata/templates.jsonl diff --git a/templates/vicuna.gotmpl b/template/vicuna.gotmpl similarity index 100% rename from templates/vicuna.gotmpl rename to template/vicuna.gotmpl diff --git a/templates/zephyr.gotmpl b/template/zephyr.gotmpl similarity index 100% rename from templates/zephyr.gotmpl rename to template/zephyr.gotmpl diff --git a/templates/template.go b/templates/template.go deleted file mode 100644 index 72bd69e9d..000000000 --- a/templates/template.go +++ /dev/null @@ -1,70 +0,0 @@ -package templates - -import ( - "bytes" - "embed" - "encoding/json" - "errors" - "io" - "math" - "sync" - - "github.com/agnivade/levenshtein" -) - -//go:embed index.json -var indexBytes []byte - -//go:embed *.gotmpl -var templatesFS embed.FS - -var templatesOnce = sync.OnceValues(func() ([]*Template, error) { - var templates []*Template - if err := json.Unmarshal(indexBytes, &templates); err != nil { - return nil, err - } - - for _, t := range templates { - bts, err := templatesFS.ReadFile(t.Name + ".gotmpl") - if err != nil { - return nil, err - } - - // normalize line endings - t.Bytes = bytes.ReplaceAll(bts, []byte("\r\n"), []byte("\n")) - } - - return templates, nil -}) - -type Template struct { - Name string `json:"name"` - Template string `json:"template"` - Bytes []byte -} - -func (t Template) Reader() io.Reader { - return bytes.NewReader(t.Bytes) -} - -func NamedTemplate(s string) (*Template, error) { - templates, err := templatesOnce() - if err != nil { - return nil, err - } - - var template *Template - score := math.MaxInt - for _, t := range templates { - if s := levenshtein.ComputeDistance(s, t.Template); s < score { - score = s - template = t - } - } - - if score < 100 { - return template, nil - } - - return nil, errors.New("no matching template found") -} diff --git a/templates/template_test.go b/templates/template_test.go deleted file mode 100644 index 61bc78374..000000000 --- a/templates/template_test.go +++ /dev/null @@ -1,59 +0,0 @@ -package templates - -import ( - "bufio" - "bytes" - "encoding/json" - "io" - "os" - "path/filepath" - "testing" - "text/template" - - "github.com/ollama/ollama/llm" -) - -func TestKVChatTemplate(t *testing.T) { - f, err := os.Open(filepath.Join("testdata", "templates.jsonl")) - if err != nil { - t.Fatal(err) - } - defer f.Close() - - scanner := bufio.NewScanner(f) - for scanner.Scan() { - var ss map[string]string - if err := json.Unmarshal(scanner.Bytes(), &ss); err != nil { - t.Fatal(err) - } - - for k, v := range ss { - t.Run(k, func(t *testing.T) { - kv := llm.KV{"tokenizer.chat_template": v} - s := kv.ChatTemplate() - r, err := NamedTemplate(s) - if err != nil { - t.Fatal(err) - } - - if r.Name != k { - t.Errorf("expected %q, got %q", k, r.Name) - } - - var b bytes.Buffer - if _, err := io.Copy(&b, r.Reader()); err != nil { - t.Fatal(err) - } - - tmpl, err := template.New(s).Parse(b.String()) - if err != nil { - t.Fatal(err) - } - - if tmpl.Tree.Root.String() == "" { - t.Errorf("empty %s template", k) - } - }) - } - } -} From a30915bde166b2f392a0ff72c61c9ac53189a962 Mon Sep 17 00:00:00 2001 From: Michael Yang Date: Tue, 11 Jun 2024 14:03:42 -0700 Subject: [PATCH 23/31] add capabilities --- server/images.go | 20 ++++++++++++++++++-- server/routes.go | 8 ++++---- template/template_test.go | 8 ++++---- 3 files changed, 26 insertions(+), 10 deletions(-) diff --git a/server/images.go b/server/images.go index 65ed51c76..5cd0a7a53 100644 --- a/server/images.go +++ b/server/images.go @@ -34,6 +34,10 @@ import ( "github.com/ollama/ollama/version" ) +type Capability string + +const CapabilityCompletion = Capability("completion") + type registryOptions struct { Insecure bool Username string @@ -58,8 +62,20 @@ type Model struct { Template *template.Template } -func (m *Model) IsEmbedding() bool { - return slices.Contains(m.Config.ModelFamilies, "bert") || slices.Contains(m.Config.ModelFamilies, "nomic-bert") +func (m *Model) Has(caps ...Capability) bool { + for _, cap := range caps { + switch cap { + case CapabilityCompletion: + if slices.Contains(m.Config.ModelFamilies, "bert") || slices.Contains(m.Config.ModelFamilies, "nomic-bert") { + return false + } + default: + slog.Error("unknown capability", "capability", cap) + return false + } + } + + return true } func (m *Model) String() string { diff --git a/server/routes.go b/server/routes.go index d8a4a67e7..8ca6dcc89 100644 --- a/server/routes.go +++ b/server/routes.go @@ -122,8 +122,8 @@ func (s *Server) GenerateHandler(c *gin.Context) { return } - if model.IsEmbedding() { - c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "embedding models do not support generate"}) + if !model.Has(CapabilityCompletion) { + c.JSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("%s does not support generate", req.Model)}) return } @@ -1308,8 +1308,8 @@ func (s *Server) ChatHandler(c *gin.Context) { return } - if model.IsEmbedding() { - c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "embedding models do not support chat"}) + if !model.Has(CapabilityCompletion) { + c.JSON(http.StatusBadRequest, gin.H{"error": fmt.Sprintf("%s does not support chat", req.Model)}) return } diff --git a/template/template_test.go b/template/template_test.go index e5405bdb4..eda4634f4 100644 --- a/template/template_test.go +++ b/template/template_test.go @@ -61,8 +61,8 @@ func TestNamed(t *testing.T) { func TestParse(t *testing.T) { cases := []struct { - template string - capabilities []string + template string + vars []string }{ {"{{ .Prompt }}", []string{"prompt"}}, {"{{ .System }} {{ .Prompt }}", []string{"prompt", "system"}}, @@ -81,8 +81,8 @@ func TestParse(t *testing.T) { } vars := tmpl.Vars() - if !slices.Equal(tt.capabilities, vars) { - t.Errorf("expected %v, got %v", tt.capabilities, vars) + if !slices.Equal(tt.vars, vars) { + t.Errorf("expected %v, got %v", tt.vars, vars) } }) } From da8e2a04479f96ad9c57eaf25ed26b79b239b05c Mon Sep 17 00:00:00 2001 From: Michael Yang Date: Fri, 14 Jun 2024 14:57:49 -0700 Subject: [PATCH 24/31] use kvs to detect embedding models --- server/images.go | 16 +++++++++++++++- 1 file changed, 15 insertions(+), 1 deletion(-) diff --git a/server/images.go b/server/images.go index 5cd0a7a53..a62991f16 100644 --- a/server/images.go +++ b/server/images.go @@ -66,7 +66,21 @@ func (m *Model) Has(caps ...Capability) bool { for _, cap := range caps { switch cap { case CapabilityCompletion: - if slices.Contains(m.Config.ModelFamilies, "bert") || slices.Contains(m.Config.ModelFamilies, "nomic-bert") { + f, err := os.Open(m.ModelPath) + if err != nil { + slog.Error("couldn't open model file", "error", err) + continue + } + defer f.Close() + + // TODO(mxyng): decode the GGML into model to avoid doing this multiple times + ggml, _, err := llm.DecodeGGML(f, 0) + if err != nil { + slog.Error("couldn't decode ggml", "error", err) + continue + } + + if _, ok := ggml.KV()[fmt.Sprintf("%s.pooling_type", ggml.KV().Architecture())]; ok { return false } default: From 7e571f95f0306f90e4f754e34df96ebc36f93626 Mon Sep 17 00:00:00 2001 From: Josh Yan Date: Mon, 1 Jul 2024 11:07:48 -0700 Subject: [PATCH 25/31] trimspace test case --- parser/parser_test.go | 33 +++++++++++++++++++++++++++++++++ 1 file changed, 33 insertions(+) diff --git a/parser/parser_test.go b/parser/parser_test.go index 171bd4206..2b5c4c888 100644 --- a/parser/parser_test.go +++ b/parser/parser_test.go @@ -48,6 +48,39 @@ TEMPLATE """{{ if .System }}<|start_header_id|>system<|end_header_id|> assert.Equal(t, expectedCommands, modelfile.Commands) } +func TestParseFileTrimSpace(t *testing.T) { + input := ` +FROM " model 1" +ADAPTER adapter3 +LICENSE "MIT " +PARAMETER param1 value1 +PARAMETER param2 value2 +TEMPLATE """ {{ if .System }}<|start_header_id|>system<|end_header_id|> + +{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|> + +{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|> + +{{ .Response }}<|eot_id|> """ +` + + reader := strings.NewReader(input) + + modelfile, err := ParseFile(reader) + require.NoError(t, err) + + expectedCommands := []Command{ + {Name: "model", Args: " model 1"}, + {Name: "adapter", Args: "adapter3"}, + {Name: "license", Args: "MIT "}, + {Name: "param1", Args: "value1"}, + {Name: "param2", Args: "value2"}, + {Name: "template", Args: " {{ if .System }}<|start_header_id|>system<|end_header_id|>\n\n{{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>\n\n{{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>\n\n{{ .Response }}<|eot_id|> "}, + } + + assert.Equal(t, expectedCommands, modelfile.Commands) +} + func TestParseFileFrom(t *testing.T) { var cases = []struct { input string From 88bcd79bb9a4b2baa739efe2ccabcbcf3c89bdb5 Mon Sep 17 00:00:00 2001 From: Michael Yang Date: Sun, 30 Jun 2024 11:10:40 -0700 Subject: [PATCH 26/31] err on insecure path --- server/model.go | 8 +++----- server/model_test.go | 24 ++++++++++++++++++++++-- 2 files changed, 25 insertions(+), 7 deletions(-) diff --git a/server/model.go b/server/model.go index d56e641ba..7d5957a18 100644 --- a/server/model.go +++ b/server/model.go @@ -11,7 +11,6 @@ import ( "net/http" "os" "path/filepath" - "strings" "github.com/ollama/ollama/api" "github.com/ollama/ollama/convert" @@ -91,12 +90,11 @@ func extractFromZipFile(p string, file *os.File, fn func(api.ProgressResponse)) fn(api.ProgressResponse{Status: "unpacking model metadata"}) for _, f := range r.File { - n := filepath.Join(p, f.Name) - if !strings.HasPrefix(n, p) { - slog.Warn("skipped extracting file outside of context", "name", f.Name) - continue + if !filepath.IsLocal(f.Name) { + return fmt.Errorf("%w: %s", zip.ErrInsecurePath, f.Name) } + n := filepath.Join(p, f.Name) if err := os.MkdirAll(filepath.Dir(n), 0o750); err != nil { return err } diff --git a/server/model_test.go b/server/model_test.go index c3023eb2b..a383b7e72 100644 --- a/server/model_test.go +++ b/server/model_test.go @@ -3,10 +3,12 @@ package server import ( "archive/zip" "bytes" + "errors" "io" "os" "path/filepath" "slices" + "strings" "testing" "github.com/ollama/ollama/api" @@ -39,13 +41,31 @@ func TestExtractFromZipFile(t *testing.T) { cases := []struct { name string expect []string + err error }{ { name: "good", expect: []string{"good"}, }, { - name: filepath.Join("..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "bad"), + name: strings.Join([]string{"path", "..", "to", "good"}, string(os.PathSeparator)), + expect: []string{filepath.Join("to", "good")}, + }, + { + name: strings.Join([]string{"path", "..", "to", "..", "good"}, string(os.PathSeparator)), + expect: []string{"good"}, + }, + { + name: strings.Join([]string{"path", "to", "..", "..", "good"}, string(os.PathSeparator)), + expect: []string{"good"}, + }, + { + name: strings.Join([]string{"..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "..", "bad"}, string(os.PathSeparator)), + err: zip.ErrInsecurePath, + }, + { + name: strings.Join([]string{"path", "..", "..", "to", "bad"}, string(os.PathSeparator)), + err: zip.ErrInsecurePath, }, } @@ -55,7 +75,7 @@ func TestExtractFromZipFile(t *testing.T) { defer f.Close() tempDir := t.TempDir() - if err := extractFromZipFile(tempDir, f, func(api.ProgressResponse) {}); err != nil { + if err := extractFromZipFile(tempDir, f, func(api.ProgressResponse) {}); !errors.Is(err, tt.err) { t.Fatal(err) } From 33a65e3ba3ad5666d6ba8430efbccfa6d642d1de Mon Sep 17 00:00:00 2001 From: Josh Yan Date: Mon, 1 Jul 2024 16:04:13 -0700 Subject: [PATCH 27/31] error --- llm/server.go | 3 +++ llm/status.go | 1 + 2 files changed, 4 insertions(+) diff --git a/llm/server.go b/llm/server.go index 61346069e..8b63cfbd5 100644 --- a/llm/server.go +++ b/llm/server.go @@ -560,6 +560,9 @@ func (s *llmServer) WaitUntilRunning(ctx context.Context) error { if s.status != nil && s.status.LastErrMsg != "" { msg = s.status.LastErrMsg } + if strings.Contains(msg, "unknown model") { + return fmt.Errorf("this model is not supported by your version of Ollama. You may need to upgrade") + } return fmt.Errorf("llama runner process has terminated: %v %s", err, msg) default: } diff --git a/llm/status.go b/llm/status.go index 8a49bd55a..0f56b7f99 100644 --- a/llm/status.go +++ b/llm/status.go @@ -25,6 +25,7 @@ var errorPrefixes = []string{ "CUDA error", "cudaMalloc failed", "\"ERR\"", + "architecture", } func (w *StatusWriter) Write(b []byte) (int, error) { From 4f67b39d262b1997aa96c47585f1d8e8443d0f90 Mon Sep 17 00:00:00 2001 From: Daniel Hiltgen Date: Tue, 2 Jul 2024 09:22:17 -0700 Subject: [PATCH 28/31] Centos 7 EOL broke mirrors As of July 1st 2024: Could not resolve host: mirrorlist.centos.org This is expected due to EOL dates. --- scripts/rh_linux_deps.sh | 11 +++++++++++ 1 file changed, 11 insertions(+) diff --git a/scripts/rh_linux_deps.sh b/scripts/rh_linux_deps.sh index ed60e4304..81648d68e 100644 --- a/scripts/rh_linux_deps.sh +++ b/scripts/rh_linux_deps.sh @@ -6,10 +6,21 @@ set -ex MACHINE=$(uname -m) if grep -i "centos" /etc/system-release >/dev/null; then + # As of 7/1/2024 mirrorlist.centos.org has been taken offline, so adjust accordingly + sed -i s/mirror.centos.org/vault.centos.org/g /etc/yum.repos.d/*.repo + sed -i s/^#.*baseurl=http/baseurl=http/g /etc/yum.repos.d/*.repo + sed -i s/^mirrorlist=http/#mirrorlist=http/g /etc/yum.repos.d/*.repo + # Centos 7 derivatives have too old of a git version to run our generate script # uninstall and ignore failures yum remove -y git yum -y install epel-release centos-release-scl + + # The release packages reinstate the mirrors, undo that again + sed -i s/mirror.centos.org/vault.centos.org/g /etc/yum.repos.d/*.repo + sed -i s/^#.*baseurl=http/baseurl=http/g /etc/yum.repos.d/*.repo + sed -i s/^mirrorlist=http/#mirrorlist=http/g /etc/yum.repos.d/*.repo + yum -y install dnf if [ "${MACHINE}" = "x86_64" ]; then yum -y install https://repo.ius.io/ius-release-el7.rpm From 020bd60ab2f156661b072515cd2c27d59b956535 Mon Sep 17 00:00:00 2001 From: Daniel Hiltgen Date: Tue, 2 Jul 2024 10:23:05 -0700 Subject: [PATCH 29/31] Switch amd container image base to rocky 8 The centos 7 arm mirrors have disappeared due to the EOL 2 days ago, and the vault sed workaround which works for x86 doesn't work for arm. --- Dockerfile | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/Dockerfile b/Dockerfile index 98a3ddfd2..b2c5c4a2f 100644 --- a/Dockerfile +++ b/Dockerfile @@ -70,12 +70,12 @@ RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx" sh gen_linux.sh FROM --platform=linux/amd64 cpu-builder-amd64 AS cpu_avx2-build-amd64 RUN OLLAMA_SKIP_STATIC_GENERATE=1 OLLAMA_CPU_TARGET="cpu_avx2" sh gen_linux.sh -FROM --platform=linux/arm64 centos:7 AS cpu-builder-arm64 +FROM --platform=linux/arm64 rockylinux:8 AS cpu-builder-arm64 ARG CMAKE_VERSION ARG GOLANG_VERSION COPY ./scripts/rh_linux_deps.sh / RUN CMAKE_VERSION=${CMAKE_VERSION} GOLANG_VERSION=${GOLANG_VERSION} sh /rh_linux_deps.sh -ENV PATH /opt/rh/devtoolset-10/root/usr/bin:$PATH +ENV PATH /opt/rh/gcc-toolset-10/root/usr/bin:$PATH COPY --from=llm-code / /go/src/github.com/ollama/ollama/ ARG OLLAMA_CUSTOM_CPU_DEFS ARG CGO_CFLAGS From 996bb1b85e0c1b3ae64246a50ea412dc2a2e30d8 Mon Sep 17 00:00:00 2001 From: royjhan <65097070+royjhan@users.noreply.github.com> Date: Tue, 2 Jul 2024 11:50:56 -0700 Subject: [PATCH 30/31] OpenAI: /v1/models and /v1/models/{model} compatibility (#5007) * OpenAI v1 models * Refactor Writers * Add Test Co-Authored-By: Attila Kerekes * Credit Co-Author Co-Authored-By: Attila Kerekes <439392+keriati@users.noreply.github.com> * Empty List Testing * Use Namespace for Ownedby * Update Test * Add back envconfig * v1/models docs * Use ModelName Parser * Test Names * Remove Docs * Clean Up * Test name Co-authored-by: Jeffrey Morgan * Add Middleware for Chat and List * Testing Cleanup * Test with Fatal * Add functionality to chat test * OpenAI: /v1/models/{model} compatibility (#5028) * Retrieve Model * OpenAI Delete Model * Retrieve Middleware * Remove Delete from Branch * Update Test * Middleware Test File * Function name * Cleanup * Test Update * Test Update --------- Co-authored-by: Attila Kerekes <439392+keriati@users.noreply.github.com> Co-authored-by: Jeffrey Morgan --- api/types.go | 7 ++ docs/openai.md | 1 + openai/openai.go | 163 ++++++++++++++++++++++++++++++++++++---- openai/openai_test.go | 170 ++++++++++++++++++++++++++++++++++++++++++ server/routes.go | 4 +- server/routes_test.go | 56 ++++++++++++++ 6 files changed, 387 insertions(+), 14 deletions(-) create mode 100644 openai/openai_test.go diff --git a/api/types.go b/api/types.go index 95ed5d37e..428281ba6 100644 --- a/api/types.go +++ b/api/types.go @@ -345,6 +345,13 @@ type ProcessModelResponse struct { SizeVRAM int64 `json:"size_vram"` } +type RetrieveModelResponse struct { + Id string `json:"id"` + Object string `json:"object"` + Created int64 `json:"created"` + OwnedBy string `json:"owned_by"` +} + type TokenResponse struct { Token string `json:"token"` } diff --git a/docs/openai.md b/docs/openai.md index 81b967eb7..9dda05c3a 100644 --- a/docs/openai.md +++ b/docs/openai.md @@ -65,6 +65,7 @@ curl http://localhost:11434/v1/chat/completions \ } ] }' + ``` ## Endpoints diff --git a/openai/openai.go b/openai/openai.go index 706d31aa2..01da44409 100644 --- a/openai/openai.go +++ b/openai/openai.go @@ -12,6 +12,7 @@ import ( "github.com/gin-gonic/gin" "github.com/ollama/ollama/api" + "github.com/ollama/ollama/types/model" ) type Error struct { @@ -85,6 +86,18 @@ type ChatCompletionChunk struct { Choices []ChunkChoice `json:"choices"` } +type Model struct { + Id string `json:"id"` + Object string `json:"object"` + Created int64 `json:"created"` + OwnedBy string `json:"owned_by"` +} + +type ListCompletion struct { + Object string `json:"object"` + Data []Model `json:"data"` +} + func NewError(code int, message string) ErrorResponse { var etype string switch code { @@ -145,7 +158,33 @@ func toChunk(id string, r api.ChatResponse) ChatCompletionChunk { } } -func fromRequest(r ChatCompletionRequest) api.ChatRequest { +func toListCompletion(r api.ListResponse) ListCompletion { + var data []Model + for _, m := range r.Models { + data = append(data, Model{ + Id: m.Name, + Object: "model", + Created: m.ModifiedAt.Unix(), + OwnedBy: model.ParseName(m.Name).Namespace, + }) + } + + return ListCompletion{ + Object: "list", + Data: data, + } +} + +func toModel(r api.ShowResponse, m string) Model { + return Model{ + Id: m, + Object: "model", + Created: r.ModifiedAt.Unix(), + OwnedBy: model.ParseName(m).Namespace, + } +} + +func fromChatRequest(r ChatCompletionRequest) api.ChatRequest { var messages []api.Message for _, msg := range r.Messages { messages = append(messages, api.Message{Role: msg.Role, Content: msg.Content}) @@ -208,13 +247,26 @@ func fromRequest(r ChatCompletionRequest) api.ChatRequest { } } -type writer struct { - stream bool - id string +type BaseWriter struct { gin.ResponseWriter } -func (w *writer) writeError(code int, data []byte) (int, error) { +type ChatWriter struct { + stream bool + id string + BaseWriter +} + +type ListWriter struct { + BaseWriter +} + +type RetrieveWriter struct { + BaseWriter + model string +} + +func (w *BaseWriter) writeError(code int, data []byte) (int, error) { var serr api.StatusError err := json.Unmarshal(data, &serr) if err != nil { @@ -230,7 +282,7 @@ func (w *writer) writeError(code int, data []byte) (int, error) { return len(data), nil } -func (w *writer) writeResponse(data []byte) (int, error) { +func (w *ChatWriter) writeResponse(data []byte) (int, error) { var chatResponse api.ChatResponse err := json.Unmarshal(data, &chatResponse) if err != nil { @@ -270,7 +322,7 @@ func (w *writer) writeResponse(data []byte) (int, error) { return len(data), nil } -func (w *writer) Write(data []byte) (int, error) { +func (w *ChatWriter) Write(data []byte) (int, error) { code := w.ResponseWriter.Status() if code != http.StatusOK { return w.writeError(code, data) @@ -279,7 +331,92 @@ func (w *writer) Write(data []byte) (int, error) { return w.writeResponse(data) } -func Middleware() gin.HandlerFunc { +func (w *ListWriter) writeResponse(data []byte) (int, error) { + var listResponse api.ListResponse + err := json.Unmarshal(data, &listResponse) + if err != nil { + return 0, err + } + + w.ResponseWriter.Header().Set("Content-Type", "application/json") + err = json.NewEncoder(w.ResponseWriter).Encode(toListCompletion(listResponse)) + if err != nil { + return 0, err + } + + return len(data), nil +} + +func (w *ListWriter) Write(data []byte) (int, error) { + code := w.ResponseWriter.Status() + if code != http.StatusOK { + return w.writeError(code, data) + } + + return w.writeResponse(data) +} + +func (w *RetrieveWriter) writeResponse(data []byte) (int, error) { + var showResponse api.ShowResponse + err := json.Unmarshal(data, &showResponse) + if err != nil { + return 0, err + } + + // retrieve completion + w.ResponseWriter.Header().Set("Content-Type", "application/json") + err = json.NewEncoder(w.ResponseWriter).Encode(toModel(showResponse, w.model)) + if err != nil { + return 0, err + } + + return len(data), nil +} + +func (w *RetrieveWriter) Write(data []byte) (int, error) { + code := w.ResponseWriter.Status() + if code != http.StatusOK { + return w.writeError(code, data) + } + + return w.writeResponse(data) +} + +func ListMiddleware() gin.HandlerFunc { + return func(c *gin.Context) { + w := &ListWriter{ + BaseWriter: BaseWriter{ResponseWriter: c.Writer}, + } + + c.Writer = w + + c.Next() + } +} + +func RetrieveMiddleware() gin.HandlerFunc { + return func(c *gin.Context) { + var b bytes.Buffer + if err := json.NewEncoder(&b).Encode(api.ShowRequest{Name: c.Param("model")}); err != nil { + c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error())) + return + } + + c.Request.Body = io.NopCloser(&b) + + // response writer + w := &RetrieveWriter{ + BaseWriter: BaseWriter{ResponseWriter: c.Writer}, + model: c.Param("model"), + } + + c.Writer = w + + c.Next() + } +} + +func ChatMiddleware() gin.HandlerFunc { return func(c *gin.Context) { var req ChatCompletionRequest err := c.ShouldBindJSON(&req) @@ -294,17 +431,17 @@ func Middleware() gin.HandlerFunc { } var b bytes.Buffer - if err := json.NewEncoder(&b).Encode(fromRequest(req)); err != nil { + if err := json.NewEncoder(&b).Encode(fromChatRequest(req)); err != nil { c.AbortWithStatusJSON(http.StatusInternalServerError, NewError(http.StatusInternalServerError, err.Error())) return } c.Request.Body = io.NopCloser(&b) - w := &writer{ - ResponseWriter: c.Writer, - stream: req.Stream, - id: fmt.Sprintf("chatcmpl-%d", rand.Intn(999)), + w := &ChatWriter{ + BaseWriter: BaseWriter{ResponseWriter: c.Writer}, + stream: req.Stream, + id: fmt.Sprintf("chatcmpl-%d", rand.Intn(999)), } c.Writer = w diff --git a/openai/openai_test.go b/openai/openai_test.go new file mode 100644 index 000000000..1f335b965 --- /dev/null +++ b/openai/openai_test.go @@ -0,0 +1,170 @@ +package openai + +import ( + "bytes" + "encoding/json" + "io" + "net/http" + "net/http/httptest" + "testing" + "time" + + "github.com/gin-gonic/gin" + "github.com/ollama/ollama/api" + "github.com/stretchr/testify/assert" +) + +func TestMiddleware(t *testing.T) { + type testCase struct { + Name string + Method string + Path string + TestPath string + Handler func() gin.HandlerFunc + Endpoint func(c *gin.Context) + Setup func(t *testing.T, req *http.Request) + Expected func(t *testing.T, resp *httptest.ResponseRecorder) + } + + testCases := []testCase{ + { + Name: "chat handler", + Method: http.MethodPost, + Path: "/api/chat", + TestPath: "/api/chat", + Handler: ChatMiddleware, + Endpoint: func(c *gin.Context) { + var chatReq api.ChatRequest + if err := c.ShouldBindJSON(&chatReq); err != nil { + c.JSON(http.StatusBadRequest, gin.H{"error": "invalid request"}) + return + } + + userMessage := chatReq.Messages[0].Content + var assistantMessage string + + switch userMessage { + case "Hello": + assistantMessage = "Hello!" + default: + assistantMessage = "I'm not sure how to respond to that." + } + + c.JSON(http.StatusOK, api.ChatResponse{ + Message: api.Message{ + Role: "assistant", + Content: assistantMessage, + }, + }) + }, + Setup: func(t *testing.T, req *http.Request) { + body := ChatCompletionRequest{ + Model: "test-model", + Messages: []Message{{Role: "user", Content: "Hello"}}, + } + + bodyBytes, _ := json.Marshal(body) + + req.Body = io.NopCloser(bytes.NewReader(bodyBytes)) + req.Header.Set("Content-Type", "application/json") + }, + Expected: func(t *testing.T, resp *httptest.ResponseRecorder) { + var chatResp ChatCompletion + if err := json.NewDecoder(resp.Body).Decode(&chatResp); err != nil { + t.Fatal(err) + } + + if chatResp.Object != "chat.completion" { + t.Fatalf("expected chat.completion, got %s", chatResp.Object) + } + + if chatResp.Choices[0].Message.Content != "Hello!" { + t.Fatalf("expected Hello!, got %s", chatResp.Choices[0].Message.Content) + } + }, + }, + { + Name: "list handler", + Method: http.MethodGet, + Path: "/api/tags", + TestPath: "/api/tags", + Handler: ListMiddleware, + Endpoint: func(c *gin.Context) { + c.JSON(http.StatusOK, api.ListResponse{ + Models: []api.ListModelResponse{ + { + Name: "Test Model", + }, + }, + }) + }, + Expected: func(t *testing.T, resp *httptest.ResponseRecorder) { + var listResp ListCompletion + if err := json.NewDecoder(resp.Body).Decode(&listResp); err != nil { + t.Fatal(err) + } + + if listResp.Object != "list" { + t.Fatalf("expected list, got %s", listResp.Object) + } + + if len(listResp.Data) != 1 { + t.Fatalf("expected 1, got %d", len(listResp.Data)) + } + + if listResp.Data[0].Id != "Test Model" { + t.Fatalf("expected Test Model, got %s", listResp.Data[0].Id) + } + }, + }, + { + Name: "retrieve model", + Method: http.MethodGet, + Path: "/api/show/:model", + TestPath: "/api/show/test-model", + Handler: RetrieveMiddleware, + Endpoint: func(c *gin.Context) { + c.JSON(http.StatusOK, api.ShowResponse{ + ModifiedAt: time.Date(2024, 6, 17, 13, 45, 0, 0, time.UTC), + }) + }, + Expected: func(t *testing.T, resp *httptest.ResponseRecorder) { + var retrieveResp Model + if err := json.NewDecoder(resp.Body).Decode(&retrieveResp); err != nil { + t.Fatal(err) + } + + if retrieveResp.Object != "model" { + t.Fatalf("Expected object to be model, got %s", retrieveResp.Object) + } + + if retrieveResp.Id != "test-model" { + t.Fatalf("Expected id to be test-model, got %s", retrieveResp.Id) + } + }, + }, + } + + gin.SetMode(gin.TestMode) + router := gin.New() + + for _, tc := range testCases { + t.Run(tc.Name, func(t *testing.T) { + router = gin.New() + router.Use(tc.Handler()) + router.Handle(tc.Method, tc.Path, tc.Endpoint) + req, _ := http.NewRequest(tc.Method, tc.TestPath, nil) + + if tc.Setup != nil { + tc.Setup(t, req) + } + + resp := httptest.NewRecorder() + router.ServeHTTP(resp, req) + + assert.Equal(t, http.StatusOK, resp.Code) + + tc.Expected(t, resp) + }) + } +} diff --git a/server/routes.go b/server/routes.go index 76ead072f..ad2364507 100644 --- a/server/routes.go +++ b/server/routes.go @@ -1039,7 +1039,9 @@ func (s *Server) GenerateRoutes() http.Handler { r.GET("/api/ps", s.ProcessHandler) // Compatibility endpoints - r.POST("/v1/chat/completions", openai.Middleware(), s.ChatHandler) + r.POST("/v1/chat/completions", openai.ChatMiddleware(), s.ChatHandler) + r.GET("/v1/models", openai.ListMiddleware(), s.ListModelsHandler) + r.GET("/v1/models/:model", openai.RetrieveMiddleware(), s.ShowModelHandler) for _, method := range []string{http.MethodGet, http.MethodHead} { r.Handle(method, "/", func(c *gin.Context) { diff --git a/server/routes_test.go b/server/routes_test.go index 5a5c0fbba..50eaf7e97 100644 --- a/server/routes_test.go +++ b/server/routes_test.go @@ -20,6 +20,7 @@ import ( "github.com/ollama/ollama/api" "github.com/ollama/ollama/envconfig" "github.com/ollama/ollama/llm" + "github.com/ollama/ollama/openai" "github.com/ollama/ollama/parser" "github.com/ollama/ollama/types/model" "github.com/ollama/ollama/version" @@ -105,6 +106,24 @@ func Test_Routes(t *testing.T) { assert.Empty(t, len(modelList.Models)) }, }, + { + Name: "openai empty list", + Method: http.MethodGet, + Path: "/v1/models", + Expected: func(t *testing.T, resp *http.Response) { + contentType := resp.Header.Get("Content-Type") + assert.Equal(t, "application/json", contentType) + body, err := io.ReadAll(resp.Body) + require.NoError(t, err) + + var modelList openai.ListCompletion + err = json.Unmarshal(body, &modelList) + require.NoError(t, err) + + assert.Equal(t, "list", modelList.Object) + assert.Empty(t, modelList.Data) + }, + }, { Name: "Tags Handler (yes tags)", Method: http.MethodGet, @@ -128,6 +147,25 @@ func Test_Routes(t *testing.T) { assert.Equal(t, "test-model:latest", modelList.Models[0].Name) }, }, + { + Name: "openai list models with tags", + Method: http.MethodGet, + Path: "/v1/models", + Expected: func(t *testing.T, resp *http.Response) { + contentType := resp.Header.Get("Content-Type") + assert.Equal(t, "application/json", contentType) + body, err := io.ReadAll(resp.Body) + require.NoError(t, err) + + var modelList openai.ListCompletion + err = json.Unmarshal(body, &modelList) + require.NoError(t, err) + + assert.Len(t, modelList.Data, 1) + assert.Equal(t, "test-model:latest", modelList.Data[0].Id) + assert.Equal(t, "library", modelList.Data[0].OwnedBy) + }, + }, { Name: "Create Model Handler", Method: http.MethodPost, @@ -216,6 +254,24 @@ func Test_Routes(t *testing.T) { assert.InDelta(t, 0, showResp.ModelInfo["general.parameter_count"], 1e-9, "Parameter count should be 0") }, }, + { + Name: "openai retrieve model handler", + Method: http.MethodGet, + Path: "/v1/models/show-model", + Expected: func(t *testing.T, resp *http.Response) { + contentType := resp.Header.Get("Content-Type") + assert.Equal(t, "application/json", contentType) + body, err := io.ReadAll(resp.Body) + require.NoError(t, err) + + var retrieveResp api.RetrieveModelResponse + err = json.Unmarshal(body, &retrieveResp) + require.NoError(t, err) + + assert.Equal(t, "show-model", retrieveResp.Id) + assert.Equal(t, "library", retrieveResp.OwnedBy) + }, + }, } t.Setenv("OLLAMA_MODELS", t.TempDir()) From 69c04eecc4b969149e43d6941f06a7d60dc5d191 Mon Sep 17 00:00:00 2001 From: Daniel Hiltgen Date: Tue, 2 Jul 2024 12:46:14 -0700 Subject: [PATCH 31/31] Add windows radeon concurreny note --- docs/faq.md | 4 +++- 1 file changed, 3 insertions(+), 1 deletion(-) diff --git a/docs/faq.md b/docs/faq.md index 841f1d13d..574112461 100644 --- a/docs/faq.md +++ b/docs/faq.md @@ -266,8 +266,10 @@ If there is insufficient available memory to load a new model request while one Parallel request processing for a given model results in increasing the context size by the number of parallel requests. For example, a 2K context with 4 parallel requests will result in an 8K context and additional memory allocation. -The following server settings may be used to adjust how Ollama handles concurrent requests: +The following server settings may be used to adjust how Ollama handles concurrent requests on most platforms: - `OLLAMA_MAX_LOADED_MODELS` - The maximum number of models that can be loaded concurrently provided they fit in available memory. The default is 3 * the number of GPUs or 3 for CPU inference. - `OLLAMA_NUM_PARALLEL` - The maximum number of parallel requests each model will process at the same time. The default will auto-select either 4 or 1 based on available memory. - `OLLAMA_MAX_QUEUE` - The maximum number of requests Ollama will queue when busy before rejecting additional requests. The default is 512 + +Note: Windows with Radeon GPUs currently default to 1 model maximum due to limitations in ROCm v5.7 for available VRAM reporting. Once ROCm v6 is available, Windows Radeon will follow the defaults above. You may enable concurrent model loads on Radeon on Windows, but ensure you don't load more models than will fit into your GPUs VRAM. \ No newline at end of file