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

api/types.go

@@ -327,6 +327,7 @@ type ProcessResponse struct {
 // ListModelResponse is a single model description in [ListResponse].
 type ListModelResponse struct {
 	Name       string       `json:"name"`
+	Model      string       `json:"model"`
 	ModifiedAt time.Time    `json:"modified_at"`
 	Size       int64        `json:"size"`
 	Digest     string       `json:"digest"`
@@ -335,6 +336,7 @@ type ListModelResponse struct {
 
 // ProcessModelResponse is a single model description in [ProcessResponse].
 type ProcessModelResponse struct {
+	Name      string       `json:"name"`
 	Model     string       `json:"model"`
 	Size      int64        `json:"size"`
 	Digest    string       `json:"digest"`

cmd/cmd.go

@@ -526,7 +526,7 @@ func ListRunningHandler(cmd *cobra.Command, args []string) error {
 	var data [][]string
 
 	for _, m := range models.Models {
-		if len(args) == 0 || strings.HasPrefix(m.Model, args[0]) {
+		if len(args) == 0 || strings.HasPrefix(m.Name, args[0]) {
 			var procStr string
 			switch {
 			case m.SizeVRAM == 0:
@@ -540,7 +540,7 @@ func ListRunningHandler(cmd *cobra.Command, args []string) error {
 				cpuPercent := math.Round(float64(sizeCPU) / float64(m.Size) * 100)
 				procStr = fmt.Sprintf("%d%%/%d%% CPU/GPU", int(cpuPercent), int(100-cpuPercent))
 			}
-			data = append(data, []string{m.Model, m.Digest[:12], format.HumanBytes(m.Size), procStr, format.HumanTime(m.ExpiresAt, "Never")})
+			data = append(data, []string{m.Name, m.Digest[:12], format.HumanBytes(m.Size), procStr, format.HumanTime(m.ExpiresAt, "Never")})
 		}
 	}
 
@@ -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 {

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.")

docs/api.md

@@ -1086,6 +1086,7 @@ A single JSON object will be returned.
 {
   "models": [
     {
+      "name": "mistral:latest",
       "model": "mistral:latest",
       "size": 5137025024,
       "digest": "2ae6f6dd7a3dd734790bbbf58b8909a606e0e7e97e94b7604e0aa7ae4490e6d8",

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<llama_token> 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);

server/routes.go

@@ -800,6 +800,7 @@ func (s *Server) ListModelsHandler(c *gin.Context) {
 
 		// tag should never be masked
 		models = append(models, api.ListModelResponse{
+			Model:      n.DisplayShortest(),
 			Name:       n.DisplayShortest(),
 			Size:       m.Size(),
 			Digest:     m.digest,
@@ -1218,6 +1219,7 @@ func (s *Server) ProcessHandler(c *gin.Context) {
 
 		mr := api.ProcessModelResponse{
 			Model:     model.ShortName,
+			Name:      model.ShortName,
 			Size:      int64(v.estimatedTotal),
 			SizeVRAM:  int64(v.estimatedVRAM),
 			Digest:    model.Digest,