Remove default auto from help message

This may confuse users thinking "auto" is an acceptable string - it must be numeric

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 2            | 9B         | 5.5GB | `ollama run gemma2`            |
-| Gemma              | 7B         | 4.8GB | `ollama run gemma:7b`          |
+| 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`       |
@@ -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
 

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 {
@@ -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
 	}
 
+	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]}
+	showInfo(resp)
-	resp, err := client.Show(cmd.Context(), &req)
+
-	if err != nil {
+	return nil
-		return err
 }
 
+func showInfo(resp *api.ShowResponse) {
 	arch := resp.ModelInfo["general.architecture"].(string)
 
 	modelData := [][]string{
@@ -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)},
@@ -705,8 +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:")
+					showInfo(resp)
-					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("")
 				case "license":
 					if resp.License == "" {
 						fmt.Println("No license was specified for this model.")

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
 

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

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`                                                                                 |

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

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"},
 		"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
+	NumParallel = 0 // Autoselect
-	MaxRunners = 1
+	MaxRunners = 0  // Autoselect
 	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 {
+		if err != nil {
-			slog.Error("invalid setting must be greater than zero", "OLLAMA_NUM_PARALLEL", onp, "error", err)
+			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
 		}

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,

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

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 n_shift = n_left / 2;
-                        const int erased_blocks = (slot.n_prompt_tokens - slot.params.n_keep - n_block_size) / n_block_size;
+                        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", {
+                        LOG_INFO("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())},
+                            {"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);

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":
+	case "gemma", "gemma2":
-		fullOffload = 4 * batch * (embedding + vocab)
+		fullOffload = max(
-		partialOffload = 4*batch*(2*embedding+vocab+1) + embedding*vocab*105/128
+			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),

llm/patches/07-gemma.diff

@@ -0,0 +1,305 @@
+From 5cadb45f39d001ffbad95b690d6cf0abcb4a6d96 Mon Sep 17 00:00:00 2001
+From: Ollama maintainers <hello@ollama.com>
+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, const char *> 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_arch, std::map<llm_tensor, std::string>> 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 << "<s>assistant\n";
+         }
+-    } else if (tmpl == "gemma" || tmpl.find("<start_of_turn>") != std::string::npos) {
++    } else if (tmpl == "gemma" || tmpl == "gemma2" || tmpl.find("<start_of_turn>") != std::string::npos) {
+         // google/gemma-7b-it
+         std::string system_prompt = "";
+         for (auto message : chat) {
+-- 
+2.45.2
+

llm/server.go

@@ -82,7 +82,7 @@ func LoadModel(model string, maxArraySize int) (*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
@@ -218,8 +218,10 @@ 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")
 	}
@@ -232,15 +234,6 @@ 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 != "" {
@@ -567,6 +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:
 		}

llm/status.go

@@ -25,6 +25,7 @@ var errorPrefixes = []string{
 	"CUDA error",
 	"cudaMalloc failed",
 	"\"ERR\"",
+	"architecture",
 }
 
 func (w *StatusWriter) Write(b []byte) (int, error) {

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
 	}

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))
+			}
+		})
+	}
+}

server/routes.go

@@ -1237,6 +1237,11 @@ 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})
 }
 

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
@@ -38,13 +39,23 @@ type Scheduler struct {
 	loaded   map[string]*runnerRef
 	loadedMu sync.Mutex
 
-	loadFn       func(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList)
+	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) (llm.LlamaServer, error)
+	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
 }
 
+// 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 {
@@ -65,13 +76,10 @@ 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,
@@ -110,11 +118,25 @@ 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
@@ -143,6 +165,26 @@ func (s *Scheduler) processPending(ctx context.Context) {
 						gpus = s.getGpuFn()
 					}
 
+					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, 0)
 					if err != nil {
@@ -152,26 +194,32 @@ 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 +234,10 @@ 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 +398,11 @@ func (pending *LlmRequest) useLoadedRunner(runner *runnerRef, finished chan *Llm
 	}()
 }
 
-func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList) {
+func (s *Scheduler) load(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel int) {
-	llama, err := s.newServerFn(gpus, req.model.ModelPath, ggml, req.model.AdapterPaths, req.model.ProjectorPaths, req.opts)
+	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 +426,7 @@ 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()
@@ -485,6 +537,7 @@ type runnerRef struct {
 
 	model       *Model
 	modelPath   string
+	numParallel int
 	*api.Options
 }
 
@@ -525,6 +578,9 @@ 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?
@@ -611,36 +667,56 @@ 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, defaultParallel, 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
+		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.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))
+						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}
 					}
 				}
 			}
+		}
 
 		// TODO future refinements
 		// - if multiple Libraries, see if any single GPU in any Library will fit
 		// - try subsets of GPUs instead of just falling back to 1 or all in a family
 
 		// Now try all the GPUs
+		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.Debug("new model will fit in available VRAM, loading", "model", req.model.ModelPath, "library", sgl[0].Library, "required", format.HumanBytes2(estimatedVRAM))
+				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
 }
 

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}
+	r1 := &runnerRef{llama: llm1, gpus: gpus, numParallel: 1}
-	r2 := &runnerRef{llama: llm2, gpus: gpus}
+	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}
+	r1 := &runnerRef{refCount: 1, sessionDuration: 1, numParallel: 1}
-	r2 := &runnerRef{sessionDuration: 2}
+	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"}},
+		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}
+	r1 := &runnerRef{llama: llm1, numParallel: 1}
-	r2 := &runnerRef{llama: llm2}
+	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}
+	r1 := &runnerRef{llama: llm1, numParallel: 1}
-	r2 := &runnerRef{model: &Model{AdapterPaths: []string{"A"}}}
+	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