starred/ollama-ollama
1
0
Fork 0
mirror of https://github.com/ollama/ollama.git synced 2026-04-17 15:53:27 +02:00
Code Issues Packages Projects Releases Wiki Activity

launch: skip context length warning for MLX models and show model name (#15102)

Browse Source
This commit is contained in:
Parth Sareen
2026-03-27 15:01:33 -07:00
committed by GitHub
parent ac83ac20c4
commit b00bd1dfd4
2 changed files with 56 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

33
cmd/launch/launch_test.go
View File

@@ -1506,6 +1506,7 @@ func TestConfirmLowContextLength(t *testing.T) {
statusBody string
statusCode int
showParams string // Parameters field returned by /api/show
showBody string // full JSON body for /api/show (overrides showParams when set)
wantWarning bool
wantModelfile bool // true if warning should mention Modelfile
}{
@@ -1596,6 +1597,28 @@ func TestConfirmLowContextLength(t *testing.T) {
statusCode: http.StatusOK,
showParams: "SHOW_ERROR", // sentinel to make show return 500
},
{
name: "no warning for safetensors model with high context length",
models: []string{"qwen3.5"},
statusBody: `{"cloud":{},"context_length":32768}`,
statusCode: http.StatusOK,
showBody: `{"details":{"format":"safetensors"},"model_info":{"qwen3_5_moe.context_length":262144}}`,
},
{
name: "warns for safetensors model with low context length",
models: []string{"small-model"},
statusBody: `{"cloud":{},"context_length":32768}`,
statusCode: http.StatusOK,
showBody: `{"details":{"format":"safetensors"},"model_info":{"small.context_length":4096}}`,
wantWarning: true,
},
{
name: "no warning for safetensors model even when server context is low",
models: []string{"qwen3.5"},
statusBody: `{"cloud":{},"context_length":4096}`,
statusCode: http.StatusOK,
showBody: `{"details":{"format":"safetensors"},"model_info":{"qwen3_5_moe.context_length":262144}}`,
},
}
for _, tt := range tests {
@@ -1614,7 +1637,11 @@ func TestConfirmLowContextLength(t *testing.T) {
return
}
w.WriteHeader(http.StatusOK)
fmt.Fprintf(w, `{"parameters":%q}`, tt.showParams)
if tt.showBody != "" {
fmt.Fprint(w, tt.showBody)
} else {
fmt.Fprintf(w, `{"parameters":%q}`, tt.showParams)
}
return
}
http.NotFound(w, r)
@@ -1648,12 +1675,12 @@ func TestConfirmLowContextLength(t *testing.T) {
t.Fatalf("expected warning=%v, got output: %q", tt.wantWarning, output)
}
if tt.wantWarning && tt.wantModelfile {
if !strings.Contains(output, "Use the model:") {
if !strings.Contains(output, "Use the base model") {
t.Fatalf("expected parent model hint in output: %q", output)
}
}
if tt.wantWarning && !tt.wantModelfile {
if strings.Contains(output, "Use the model:") {
if strings.Contains(output, "Use the base model") {
t.Fatalf("expected server hint, not parent model hint in output: %q", output)
}
}

28
cmd/launch/models.go
View File

@@ -479,16 +479,24 @@ func lowContextLength(ctx context.Context, client *api.Client, models []string)
modelfileOverride := false
var info *api.ShowResponse
if info, err = client.Show(ctx, &api.ShowRequest{Model: m}); err == nil {
// Safetensors (MLX) models always load at their full max context
// length, so the server default num_ctx doesn't apply.
if info.Details.Format == "safetensors" {
// Context length check in case models with low context length are added
if modelCtx := modelInfoContextLength(info.ModelInfo); modelCtx >= recommendedContextLength {
continue
}
}
if numCtx := parseNumCtx(info.Parameters); numCtx > 0 {
effectiveCtx = numCtx
modelfileOverride = true
}
}
if effectiveCtx < recommendedContextLength {
fmt.Fprintf(os.Stderr, "\n%sWarning: context window is %d tokens (recommended: %d+)%s\n", ansiYellow, effectiveCtx, recommendedContextLength, ansiReset)
fmt.Fprintf(os.Stderr, "\n%sWarning: %s has a context length of %d tokens, which is below the recommended %d.%s\n", ansiYellow, m, effectiveCtx, recommendedContextLength, ansiReset)
if modelfileOverride {
parentModel := info.Details.ParentModel
fmt.Fprintf(os.Stderr, "%sUse the model: %s and increase the context length to at least %d in Ollama App Settings.%s\n\n", ansiYellow, parentModel, recommendedContextLength, ansiReset)
fmt.Fprintf(os.Stderr, "%sUse the base model %s and increase the context length in Ollama App Settings.%s\n\n", ansiYellow, parentModel, ansiReset)
} else {
if runtime.GOOS == "windows" {
fmt.Fprintf(os.Stderr, "%sIncrease it in Ollama App Settings or with $env:OLLAMA_CONTEXT_LENGTH=%d; ollama serve%s\n\n", ansiYellow, recommendedContextLength, ansiReset)
@@ -515,6 +523,22 @@ func parseNumCtx(parameters string) int {
return 0
}
// modelInfoContextLength extracts the model's architectural context length
// from the ModelInfo map (e.g. "qwen3_5_moe.context_length" → 262144).
func modelInfoContextLength(modelInfo map[string]any) int {
for k, v := range modelInfo {
if strings.HasSuffix(k, ".context_length") {
switch n := v.(type) {
case float64:
return int(n)
case int:
return n
}
}
}
return 0
}
// TODO(parthsareen): this duplicates the pull progress UI in cmd.PullHandler.
// Move the shared pull rendering to a small utility once the package boundary settles.
func pullModel(ctx context.Context, client *api.Client, model string, insecure bool) error {