lower default NUM_PARALLEL to 2

this is in part to "pay" for #10452, which doubled the default context length. The combination isn't fully neutral though, because even though the old 4x2k limit and the new 2x4k limit are memory equivalent, the 1x fallback is larger with 4k

cmd/cmd.go

@@ -1407,7 +1407,6 @@ func NewCLI() *cobra.Command {
 				envVars["OLLAMA_LLM_LIBRARY"],
 				envVars["OLLAMA_GPU_OVERHEAD"],
 				envVars["OLLAMA_LOAD_TIMEOUT"],
-				envVars["OLLAMA_CONTEXT_LENGTH"],
 			})
 		default:
 			appendEnvDocs(cmd, envs)

docs/faq.md

@@ -20,7 +20,7 @@ Please refer to the [GPU docs](./gpu.md).
 
 ## How can I specify the context window size?
 
-By default, Ollama uses a context window size of 4096 tokens, unless you have a single GPU with <= 4 GB of VRAM, in which case it will default to 2048 tokens. 
+By default, Ollama uses a context window size of 4096 tokens. 
 
 This can be overridden with the `OLLAMA_CONTEXT_LENGTH` environment variable. For example, to set the default context window to 8K, use: 
 
@@ -31,7 +31,7 @@ OLLAMA_CONTEXT_LENGTH=8192 ollama serve
 To change this when using `ollama run`, use `/set parameter`:
 
 ```shell
-/set parameter num_ctx 8192
+/set parameter num_ctx 4096
 ```
 
 When using the API, specify the `num_ctx` parameter:
@@ -41,7 +41,7 @@ curl http://localhost:11434/api/generate -d '{
   "model": "llama3.2",
   "prompt": "Why is the sky blue?",
   "options": {
-    "num_ctx": 8192
+    "num_ctx": 4096
   }
 }'
 ```

envconfig/config.go

@@ -169,7 +169,7 @@ var (
 	// Enable the new Ollama engine
 	NewEngine = Bool("OLLAMA_NEW_ENGINE")
 	// ContextLength sets the default context length
-	ContextLength = Int64("OLLAMA_CONTEXT_LENGTH", -1)
+	ContextLength = Uint("OLLAMA_CONTEXT_LENGTH", 4096)
 )
 
 func String(s string) func() string {
@@ -227,20 +227,6 @@ func Uint64(key string, defaultValue uint64) func() uint64 {
 	}
 }
 
-func Int64(key string, defaultValue int64) func() int64 {
-	return func() int64 {
-		if s := Var(key); s != "" {
-			if n, err := strconv.ParseInt(s, 10, 64); err != nil {
-				slog.Warn("invalid environment variable, using default", "key", key, "value", s, "default", defaultValue)
-			} else {
-				return n
-			}
-		}
-
-		return defaultValue
-	}
-}
-
 // Set aside VRAM per GPU
 var GpuOverhead = Uint64("OLLAMA_GPU_OVERHEAD", 0)
 
@@ -269,7 +255,7 @@ func AsMap() map[string]EnvVar {
 		"OLLAMA_ORIGINS":           {"OLLAMA_ORIGINS", AllowedOrigins(), "A comma separated list of allowed origins"},
 		"OLLAMA_SCHED_SPREAD":      {"OLLAMA_SCHED_SPREAD", SchedSpread(), "Always schedule model across all GPUs"},
 		"OLLAMA_MULTIUSER_CACHE":   {"OLLAMA_MULTIUSER_CACHE", MultiUserCache(), "Optimize prompt caching for multi-user scenarios"},
-		"OLLAMA_CONTEXT_LENGTH":    {"OLLAMA_CONTEXT_LENGTH", ContextLength(), "Context length to use unless otherwise specified (default 4096 or 2048 with low VRAM)"},
+		"OLLAMA_CONTEXT_LENGTH":    {"OLLAMA_CONTEXT_LENGTH", ContextLength(), "Context length to use unless otherwise specified (default: 4096)"},
 		"OLLAMA_NEW_ENGINE":        {"OLLAMA_NEW_ENGINE", NewEngine(), "Enable the new Ollama engine"},
 
 		// Informational

envconfig/config_test.go

@@ -278,9 +278,9 @@ func TestVar(t *testing.T) {
 }
 
 func TestContextLength(t *testing.T) {
-	cases := map[string]int64{
-		"":     -1,
-		"4096": 4096,
+	cases := map[string]uint{
+		"":     4096,
+		"2048": 2048,
 	}
 
 	for k, v := range cases {

server/routes_generate_test.go

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

@@ -81,6 +81,10 @@ 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 *api.Duration) (chan *runnerRef, chan error) {
+	if opts.NumCtx < 4 {
+		opts.NumCtx = 4
+	}
+
 	req := &LlmRequest{
 		ctx:             c,
 		model:           model,
@@ -110,11 +114,6 @@ func (s *Scheduler) Run(ctx context.Context) {
 	}()
 }
 
-const (
-	defaultContextLength  = 4096
-	smallGpuContextLength = 2048
-)
-
 func (s *Scheduler) processPending(ctx context.Context) {
 	for {
 		select {
@@ -167,17 +166,6 @@ 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

@@ -148,7 +148,6 @@ 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
 }