increase default context length to 4096 (#10364)

* increase default context length to 4096

We lower the default numParallel from 4 to 2 and use these "savings" to
double the default context length from 2048 to 4096.

We're memory neutral in cases when we previously would've used
numParallel == 4, but we add the following mitigation to handle some
cases where we would have previously fallen back to 1x2048 due to low
VRAM: we decide between 2048 and 4096 using a runtime check, choosing
2048 if we're on a one GPU system with total VRAM of <= 4 GB. We
purposefully don't check the available VRAM because we don't want the
context window size to change unexpectedly based on the available VRAM.

We plan on making the default even larger, but this is a relatively
low-risk change we can make to quickly double it.

* fix tests

add an explicit context length so they don't get truncated. The code
that converts -1 from being a signal for doing a runtime check isn't
running as part of these tests.

* tweak small gpu message

* clarify context length default

also make it actually show up in `ollama serve --help`

server/routes_generate_test.go

@@ -299,6 +299,9 @@ func TestGenerateChat(t *testing.T) {
 				{Role: "user", Content: "Hello!"},
 			},
 			Stream: &stream,
+			Options: map[string]any{
+				"num_ctx": 1024,
+			},
 		})
 
 		if w.Code != http.StatusOK {
@@ -321,6 +324,9 @@ func TestGenerateChat(t *testing.T) {
 				{Role: "user", Content: "Hello!"},
 			},
 			Stream: &stream,
+			Options: map[string]any{
+				"num_ctx": 1024,
+			},
 		})
 
 		if w.Code != http.StatusOK {
@@ -344,6 +350,9 @@ func TestGenerateChat(t *testing.T) {
 				{Role: "user", Content: "Help me write tests."},
 			},
 			Stream: &stream,
+			Options: map[string]any{
+				"num_ctx": 1024,
+			},
 		})
 
 		if w.Code != http.StatusOK {

server/sched.go

@@ -58,7 +58,7 @@ var defaultModelsPerGPU = 3
 // Default automatic value for parallel setting
 // Model will still need to fit in VRAM.  If this setting won't fit
 // we'll back off down to 1 to try to get it to fit
-var defaultParallel = 4
+var defaultParallel = 2
 
 var ErrMaxQueue = errors.New("server busy, please try again.  maximum pending requests exceeded")
 
@@ -81,10 +81,6 @@ 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 *api.Duration) (chan *runnerRef, chan error) {
-	if opts.NumCtx < 4 {
-		opts.NumCtx = 4
-	}
-
 	req := &LlmRequest{
 		ctx:             c,
 		model:           model,
@@ -114,6 +110,11 @@ func (s *Scheduler) Run(ctx context.Context) {
 	}()
 }
 
+const (
+	defaultContextLength  = 4096
+	smallGpuContextLength = 2048
+)
+
 func (s *Scheduler) processPending(ctx context.Context) {
 	for {
 		select {
@@ -166,6 +167,17 @@ func (s *Scheduler) processPending(ctx context.Context) {
 						gpus = s.getGpuFn()
 					}
 
+					if pending.origNumCtx == -1 {
+						if len(gpus) == 1 && gpus[0].Library != "cpu" && gpus[0].TotalMemory <= 4096*1024*1024 {
+							slog.Info("GPU is small, limiting default context window", "num_ctx", smallGpuContextLength)
+							pending.opts.NumCtx = smallGpuContextLength
+							pending.origNumCtx = smallGpuContextLength
+						} else {
+							pending.opts.NumCtx = defaultContextLength
+							pending.origNumCtx = defaultContextLength
+						}
+					}
+
 					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

server/sched_test.go

@@ -148,6 +148,7 @@ func newScenarioRequest(t *testing.T, ctx context.Context, modelName string, est
 		successCh:       make(chan *runnerRef, 1),
 		errCh:           make(chan error, 1),
 	}
+	b.req.opts.NumCtx = 4096
 	b.srv = &mockLlm{estimatedVRAM: estimatedVRAM, estimatedVRAMByGPU: map[string]uint64{"": estimatedVRAM}}
 	return b
 }