api embed docs (#5282)

Refine error reporting for subprocess crash

cmd/cmd.go

@@ -1344,7 +1344,6 @@ func NewCLI() *cobra.Command {
 				envVars["OLLAMA_TMPDIR"],
 				envVars["OLLAMA_FLASH_ATTENTION"],
 				envVars["OLLAMA_LLM_LIBRARY"],
-				envVars["OLLAMA_MAX_VRAM"],
 			})
 		default:
 			appendEnvDocs(cmd, envs)

convert/mistral.go

@@ -71,6 +71,11 @@ func (m *MistralModel) WriteGGUF(ws io.WriteSeeker) error {
 		"tokenizer.ggml.unknown_token_id": uint32(0),
 	}
 
+	if m.Params.HeadDimension > 0 {
+		kv["llama.attention.key_length"] = uint32(m.Params.HeadDimension)
+		kv["llama.attention.value_length"] = uint32(m.Params.HeadDimension)
+	}
+
 	return llm.NewGGUFV3(m.Params.ByteOrder).Encode(ws, kv, m.Tensors)
 }
 

docs/api.md

@@ -1026,7 +1026,7 @@ If `stream` is set to `false`, then the response is a single JSON object:
 ## Generate Embeddings
 
 ```shell
-POST /api/embeddings
+POST /api/embed
 ```
 
 Generate embeddings from a model
@@ -1034,10 +1034,11 @@ Generate embeddings from a model
 ### Parameters
 
 - `model`: name of model to generate embeddings from
-- `prompt`: text to generate embeddings for
+- `input`: text or list of text to generate embeddings for
 
 Advanced parameters:
 
+- `truncate`: truncates the end of each input to fit within context length. Returns error if `false` and context length is exceeded. Defaults to `true`
 - `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
 - `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
 
@@ -1046,9 +1047,9 @@ Advanced parameters:
 #### Request
 
 ```shell
-curl http://localhost:11434/api/embeddings -d '{
+curl http://localhost:11434/api/embed -d '{
   "model": "all-minilm",
-  "prompt": "Here is an article about llamas..."
+  "input": "Why is the sky blue?"
 }'
 ```
 
@@ -1056,10 +1057,35 @@ curl http://localhost:11434/api/embeddings -d '{
 
 ```json
 {
-  "embedding": [
-    0.5670403838157654, 0.009260174818336964, 0.23178744316101074, -0.2916173040866852, -0.8924556970596313,
-    0.8785552978515625, -0.34576427936553955, 0.5742510557174683, -0.04222835972905159, -0.137906014919281
-  ]
+  "model": "all-minilm",
+  "embeddings": [[
+    0.010071029, -0.0017594862, 0.05007221, 0.04692972, 0.054916814,
+    0.008599704, 0.105441414, -0.025878139, 0.12958129, 0.031952348
+  ]]
+}
+```
+
+#### Request (Multiple input)
+
+```shell
+curl http://localhost:11434/api/embed -d '{
+  "model": "all-minilm",
+  "input": ["Why is the sky blue?", "Why is the grass green?"]
+}'
+```
+
+#### Response
+
+```json
+{
+  "model": "all-minilm",
+  "embeddings": [[
+    0.010071029, -0.0017594862, 0.05007221, 0.04692972, 0.054916814,
+    0.008599704, 0.105441414, -0.025878139, 0.12958129, 0.031952348
+  ],[
+    -0.0098027075, 0.06042469, 0.025257962, -0.006364387, 0.07272725,
+    0.017194884, 0.09032035, -0.051705178, 0.09951512, 0.09072481
+  ]]
 }
 ```
 
@@ -1106,3 +1132,45 @@ A single JSON object will be returned.
   ]
 }
 ```
+
+## Generate Embedding
+
+> Note: this endpoint has been superseded by `/api/embed`
+
+```shell
+POST /api/embeddings
+```
+
+Generate embeddings from a model
+
+### Parameters
+
+- `model`: name of model to generate embeddings from
+- `prompt`: text to generate embeddings for
+
+Advanced parameters:
+
+- `options`: additional model parameters listed in the documentation for the [Modelfile](./modelfile.md#valid-parameters-and-values) such as `temperature`
+- `keep_alive`: controls how long the model will stay loaded into memory following the request (default: `5m`)
+
+### Examples
+
+#### Request
+
+```shell
+curl http://localhost:11434/api/embeddings -d '{
+  "model": "all-minilm",
+  "prompt": "Here is an article about llamas..."
+}'
+```
+
+#### Response
+
+```json
+{
+  "embedding": [
+    0.5670403838157654, 0.009260174818336964, 0.23178744316101074, -0.2916173040866852, -0.8924556970596313,
+    0.8785552978515625, -0.34576427936553955, 0.5742510557174683, -0.04222835972905159, -0.137906014919281
+  ]
+}
+```

docs/gpu.md

@@ -46,13 +46,24 @@ sudo modprobe nvidia_uvm`
 
 ## AMD Radeon
 Ollama supports the following AMD GPUs:
+
+### Linux Support
 | Family         | Cards and accelerators                                                                                                               |
 | -------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- |
 | AMD Radeon RX  | `7900 XTX` `7900 XT` `7900 GRE` `7800 XT` `7700 XT` `7600 XT` `7600` `6950 XT` `6900 XTX` `6900XT` `6800 XT` `6800` `Vega 64` `Vega 56`    |
 | AMD Radeon PRO | `W7900` `W7800` `W7700` `W7600` `W7500` `W6900X` `W6800X Duo` `W6800X` `W6800` `V620` `V420` `V340` `V320` `Vega II Duo` `Vega II` `VII` `SSG` |
 | AMD Instinct   | `MI300X` `MI300A` `MI300` `MI250X` `MI250` `MI210` `MI200` `MI100` `MI60` `MI50`                                                               |
 
-### Overrides
+### Windows Support
+With ROCm v6.1, the following GPUs are supported on Windows.
+
+| Family         | Cards and accelerators                                                                                                               |
+| -------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- |
+| AMD Radeon RX  | `7900 XTX` `7900 XT` `7900 GRE` `7800 XT` `7700 XT` `7600 XT` `7600` `6950 XT` `6900 XTX` `6900XT` `6800 XT` `6800`    |
+| AMD Radeon PRO | `W7900` `W7800` `W7700` `W7600` `W7500` `W6900X` `W6800X Duo` `W6800X` `W6800` `V620` |
+
+
+### Overrides on Linux
 Ollama leverages the AMD ROCm library, which does not support all AMD GPUs. In
 some cases you can force the system to try to use a similar LLVM target that is
 close.  For example The Radeon RX 5400 is `gfx1034` (also known as 10.3.4)
@@ -63,7 +74,7 @@ would set `HSA_OVERRIDE_GFX_VERSION="10.3.0"` as an environment variable for the
 server.  If you have an unsupported AMD GPU you can experiment using the list of
 supported types below.
 
-At this time, the known supported GPU types are the following LLVM Targets.
+At this time, the known supported GPU types on linux are the following LLVM Targets.
 This table shows some example GPUs that map to these LLVM targets:
 | **LLVM Target** | **An Example GPU** |
 |-----------------|---------------------|

envconfig/config.go

@@ -43,8 +43,6 @@ var (
 	MaxRunners int
 	// Set via OLLAMA_MAX_QUEUE in the environment
 	MaxQueuedRequests int
-	// Set via OLLAMA_MAX_VRAM in the environment
-	MaxVRAM uint64
 	// Set via OLLAMA_MODELS in the environment
 	ModelsDir string
 	// Set via OLLAMA_NOHISTORY in the environment
@@ -89,7 +87,6 @@ func AsMap() map[string]EnvVar {
 		"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 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"},
@@ -194,16 +191,6 @@ func LoadConfig() {
 
 	TmpDir = clean("OLLAMA_TMPDIR")
 
-	userLimit := clean("OLLAMA_MAX_VRAM")
-	if userLimit != "" {
-		avail, err := strconv.ParseUint(userLimit, 10, 64)
-		if err != nil {
-			slog.Error("invalid setting, ignoring", "OLLAMA_MAX_VRAM", userLimit, "error", err)
-		} else {
-			MaxVRAM = avail
-		}
-	}
-
 	LLMLibrary = clean("OLLAMA_LLM_LIBRARY")
 
 	if onp := clean("OLLAMA_NUM_PARALLEL"); onp != "" {

gpu/amd_hip_windows.go

@@ -33,9 +33,10 @@ type HipLib struct {
 }
 
 func NewHipLib() (*HipLib, error) {
-	h, err := windows.LoadLibrary("amdhip64.dll")
+	// At runtime we depend on v6, so discover GPUs with the same library for a consistent set of GPUs
+	h, err := windows.LoadLibrary("amdhip64_6.dll")
 	if err != nil {
-		return nil, fmt.Errorf("unable to load amdhip64.dll: %w", err)
+		return nil, fmt.Errorf("unable to load amdhip64_6.dll, please make sure to upgrade to the latest amd driver: %w", err)
 	}
 	hl := &HipLib{}
 	hl.dll = h

gpu/amd_windows.go

@@ -92,7 +92,8 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 			continue
 		}
 		if gfxOverride == "" {
-			if !slices.Contains[[]string, string](supported, gfx) {
+			// Strip off Target Features when comparing
+			if !slices.Contains[[]string, string](supported, strings.Split(gfx, ":")[0]) {
 				slog.Warn("amdgpu is not supported", "gpu", i, "gpu_type", gfx, "library", libDir, "supported_types", supported)
 				// TODO - consider discrete markdown just for ROCM troubleshooting?
 				slog.Warn("See https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md for HSA_OVERRIDE_GFX_VERSION usage")

integration/concurrency_test.go

@@ -69,7 +69,7 @@ func TestIntegrationConcurrentPredictOrcaMini(t *testing.T) {
 	reqLimit := len(req)
 	iterLimit := 5
 
-	vram := os.Getenv("OLLAMA_MAX_VRAM")
+	vram := os.Getenv("OLLAMA_MAX_VRAM") // TODO - discover actual VRAM
 	if vram != "" {
 		max, err := strconv.ParseUint(vram, 10, 64)
 		require.NoError(t, err)
@@ -106,7 +106,7 @@ func TestIntegrationConcurrentPredictOrcaMini(t *testing.T) {
 
 // Stress the system if we know how much VRAM it has, and attempt to load more models than will fit
 func TestMultiModelStress(t *testing.T) {
-	vram := os.Getenv("OLLAMA_MAX_VRAM")
+	vram := os.Getenv("OLLAMA_MAX_VRAM") // TODO - discover actual VRAM
 	if vram == "" {
 		t.Skip("OLLAMA_MAX_VRAM not specified, can't pick the right models for the stress test")
 	}

integration/context_test.go

@@ -12,7 +12,7 @@ import (
 
 func TestContextExhaustion(t *testing.T) {
 	// Longer needed for small footprint GPUs
-	ctx, cancel := context.WithTimeout(context.Background(), 6*time.Minute)
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Minute)
 	defer cancel()
 	// Set up the test data
 	req := api.GenerateRequest{
@@ -25,5 +25,10 @@ func TestContextExhaustion(t *testing.T) {
 			"num_ctx":     128,
 		},
 	}
-	GenerateTestHelper(ctx, t, req, []string{"once", "upon", "lived"})
+	client, _, cleanup := InitServerConnection(ctx, t)
+	defer cleanup()
+	if err := PullIfMissing(ctx, client, req.Model); err != nil {
+		t.Fatalf("PullIfMissing failed: %v", err)
+	}
+	DoGenerate(ctx, t, client, req, []string{"once", "upon", "lived"}, 120*time.Second, 10*time.Second)
 }

llm/generate/gen_windows.ps1

@@ -7,8 +7,8 @@ function amdGPUs {
         return $env:AMDGPU_TARGETS
     }
     # Current supported rocblas list from ROCm v6.1.2 on windows
+    # https://rocm.docs.amd.com/projects/install-on-windows/en/latest/reference/system-requirements.html#windows-supported-gpus
     $GPU_LIST = @(
-        "gfx906:xnack-"
         "gfx1030"
         "gfx1100"
         "gfx1101"

llm/patches/05-default-pretokenizer.diff

@@ -1,8 +1,8 @@
 diff --git a/src/llama.cpp b/src/llama.cpp
-index 2b9ace28..172640e2 100644
+index 8fe51971..7113ba64 100644
 --- a/src/llama.cpp
 +++ b/src/llama.cpp
-@@ -5357,16 +5357,7 @@ static void llm_load_vocab(
+@@ -5433,16 +5433,7 @@ static void llm_load_vocab(
          if (vocab.type == LLAMA_VOCAB_TYPE_BPE) {
              vocab.tokenizer_add_space_prefix = false;
              vocab.tokenizer_clean_spaces = true;
@@ -20,9 +20,9 @@ index 2b9ace28..172640e2 100644
                  vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_DEFAULT;
              } else if (
                      tokenizer_pre == "llama3"   ||
-@@ -5439,7 +5430,8 @@ static void llm_load_vocab(
-                 tokenizer_pre == "jais") {
-                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_JAIS;
+@@ -5526,7 +5517,8 @@ static void llm_load_vocab(
+                 vocab.type_pre = LLAMA_VOCAB_PRE_TYPE_SMOLLM;
+                 vocab.tokenizer_clean_spaces = false;
              } else {
 -                throw std::runtime_error(format("unknown pre-tokenizer type: '%s'", tokenizer_pre.c_str()));
 +                LLAMA_LOG_WARN("%s: missing or unrecognized pre-tokenizer type, using: 'default'\n", __func__);

llm/patches/06-qwen2.diff

@@ -1,13 +0,0 @@
-diff --git a/src/llama.cpp b/src/llama.cpp
-index 40d2ec2c..f34eb79a 100644
---- a/src/llama.cpp
-+++ b/src/llama.cpp
-@@ -6943,7 +6943,7 @@ static struct ggml_tensor * llm_build_kqv(
-         struct ggml_tensor * kq = ggml_mul_mat(ctx, k, q);
-         cb(kq, "kq", il);
- 
--        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX) {
-+        if (model.arch == LLM_ARCH_PHI2 || model.arch == LLM_ARCH_PHI3 || model.arch == LLM_ARCH_GPTNEOX || model.arch == LLM_ARCH_QWEN2) {
-             // for this arch, we need to perform the KQ multiplication with F32 precision, otherwise we get NaNs
-             // ref: https://github.com/ggerganov/llama.cpp/pull/4490#issuecomment-1859055847
-             ggml_mul_mat_set_prec(kq, GGML_PREC_F32);

llm/patches/09-lora.diff

@@ -0,0 +1,360 @@
+diff --git a/common/common.cpp b/common/common.cpp
+index dbb724fb..c26fe6ee 100644
+--- a/common/common.cpp
++++ b/common/common.cpp
+@@ -2087,14 +2087,29 @@ std::tuple<struct llama_model *, struct llama_context *> llama_init_from_gpt_par
+     for (unsigned int i = 0; i < params.lora_adapter.size(); ++i) {
+         const std::string & lora_adapter = std::get<0>(params.lora_adapter[i]);
+         float lora_scale = std::get<1>(params.lora_adapter[i]);
++
++        // try to load as gguf
+         auto adapter = llama_lora_adapter_init(model, lora_adapter.c_str());
+         if (adapter == nullptr) {
+-            fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
+-            llama_free(lctx);
+-            llama_free_model(model);
+-            return std::make_tuple(nullptr, nullptr);
++            fprintf(stderr, "%s: error: failed to apply lora adapter, trying ggla\n", __func__);
++
++            // if that fails, try loading as ggla for compatibility
++            int err = llama_model_apply_lora_from_file(model,
++                                                    lora_adapter.c_str(),
++                                                    lora_scale,
++                                                    ((i > 0) || params.lora_base.empty())
++                                                        ? NULL
++                                                        : params.lora_base.c_str(),
++                                                    params.n_threads);
++            if (err != 0) {
++                fprintf(stderr, "%s: error: failed to apply lora adapter\n", __func__);
++                llama_free(lctx);
++                llama_free_model(model);
++                return std::make_tuple(nullptr, nullptr);
++            }
++        } else {
++            llama_lora_adapter_set(lctx, adapter, lora_scale);
+         }
+-        llama_lora_adapter_set(lctx, adapter, lora_scale);
+     }
+ 
+     if (params.ignore_eos) {
+diff --git a/include/llama.h b/include/llama.h
+index 93fd77ca..b0fb37a6 100644
+--- a/include/llama.h
++++ b/include/llama.h
+@@ -1160,6 +1160,20 @@ extern "C" {
+ 
+     LLAMA_API void llama_dump_timing_info_yaml(FILE * stream, const struct llama_context * ctx);
+ 
++    // Apply a LoRA adapter to a loaded model
++    // path_base_model is the path to a higher quality model to use as a base for
++    // the layers modified by the adapter. Can be NULL to use the current loaded model.
++    // The model needs to be reloaded before applying a new adapter, otherwise the adapter
++    // will be applied on top of the previous one
++    // Returns 0 on success
++    LLAMA_API int32_t llama_model_apply_lora_from_file(
++            const struct llama_model * model,
++                            const char * path_lora,
++                                float   scale,
++                            const char * path_base_model,
++                                int32_t   n_threads);
++
++
+ #ifdef __cplusplus
+ }
+ #endif
+diff --git a/src/llama.cpp b/src/llama.cpp
+index 80a0dd0f..9d7b0e17 100644
+--- a/src/llama.cpp
++++ b/src/llama.cpp
+@@ -21880,3 +21880,290 @@ static void llama_log_callback_default(ggml_log_level level, const char * text,
+     fputs(text, stderr);
+     fflush(stderr);
+ }
++
++static int llama_apply_lora_from_file_internal(
++    const struct llama_model & model, const char * path_lora, float scale, const char * path_base_model, int n_threads
++) {
++    LLAMA_LOG_INFO("%s: applying lora adapter from '%s' - please wait ...\n", __func__, path_lora);
++
++    const int64_t t_start_lora_us = ggml_time_us();
++
++    llama_file fin(path_lora, "rb");
++
++    // verify magic and version
++    {
++        uint32_t magic = fin.read_u32();
++        if (magic != LLAMA_FILE_MAGIC_GGLA) {
++            LLAMA_LOG_ERROR("%s: bad file magic\n", __func__);
++            return 1;
++        }
++
++        uint32_t format_version = fin.read_u32();
++        if (format_version != 1) {
++            LLAMA_LOG_ERROR("%s: unsupported file version\n", __func__ );
++            return 1;
++        }
++    }
++
++    int32_t lora_r = fin.read_u32();
++    int32_t lora_alpha = fin.read_u32();
++    float scaling = scale * (float)lora_alpha / (float)lora_r;
++
++    LLAMA_LOG_INFO("%s: r = %d, alpha = %d, scaling = %.2f\n", __func__, lora_r, lora_alpha, scaling);
++
++    // load base model
++    std::unique_ptr<llama_model_loader> ml;
++    if (path_base_model) {
++        LLAMA_LOG_INFO("%s: loading base model from '%s'\n", __func__, path_base_model);
++        ml.reset(new llama_model_loader(path_base_model, /*use_mmap*/ true, /*check_tensors*/ false, /*kv_overrides*/ nullptr));
++        ml->init_mappings(/*prefetch*/ false); // no prefetching
++    }
++
++    struct tensor_meta {
++        std::string name;
++        ggml_type type;
++        int32_t ne[2];
++        size_t offset;
++    };
++    std::map<std::string, tensor_meta> tensor_meta_map;
++
++    // load all tensor meta
++    while (true) {
++        if (fin.tell() == fin.size) {
++            // eof
++            break;
++        }
++
++        int32_t n_dims;
++        int32_t name_len;
++        int32_t ftype;
++
++        fin.read_raw(&n_dims, sizeof(n_dims));
++        fin.read_raw(&name_len, sizeof(name_len));
++        fin.read_raw(&ftype, sizeof(ftype));
++
++        if (n_dims != 1 && n_dims != 2) {
++            LLAMA_LOG_ERROR("%s: unsupported tensor dimension %d\n", __func__, n_dims);
++            return 1;
++        }
++
++        int32_t ne[2] = { 1, 1 };
++        for (int i = 0; i < n_dims; ++i) {
++            fin.read_raw(&ne[i], sizeof(ne[i]));
++        }
++
++        std::string name;
++        {
++            GGML_ASSERT(name_len < GGML_MAX_NAME);
++            char buf[GGML_MAX_NAME];
++            fin.read_raw(buf, name_len);
++            name = std::string(buf, name_len);
++        }
++
++        // check for lora suffix
++        std::string lora_suffix;
++        if (name.length() > 6) {
++            lora_suffix = name.substr(name.length() - 6);
++        }
++        if (lora_suffix != ".loraA" && lora_suffix != ".loraB") {
++            LLAMA_LOG_ERROR("%s: error: '%s' is not a lora tensor\n", __func__, name.c_str());
++            return 1;
++        }
++
++        // tensor type
++        ggml_type wtype;
++        switch (ftype) {
++            case 0: wtype = GGML_TYPE_F32;  break;
++            case 1: wtype = GGML_TYPE_F16;  break;
++            default:
++                    {
++                        LLAMA_LOG_ERROR("%s: invalid tensor data type '%d'\n",
++                                __func__, ftype);
++                        return 1;
++                    }
++        }
++
++        // data offset
++        size_t offset = fin.tell();
++        offset = (offset + 31) & -32;
++
++        // skip tensor data
++        fin.seek(offset + ggml_row_size(wtype, ne[0]) * ne[1], SEEK_SET);
++
++        tensor_meta_map.emplace(name, tensor_meta{ name, wtype, { ne[0], ne[1] }, offset });
++    }
++
++    bool warned = false;
++    int n_tensors = 0;
++
++    // apply
++    ggml_backend_t backend_cpu = ggml_backend_cpu_init();
++    if (backend_cpu == nullptr) {
++        LLAMA_LOG_ERROR("%s: error: failed to initialize cpu backend\n", __func__);
++        return 1;
++    }
++    ggml_backend_cpu_set_n_threads(backend_cpu, n_threads);
++
++    std::vector<no_init<uint8_t>> read_buf;
++    for (const auto & it : model.tensors_by_name) {
++        const std::string & base_name = it.first;
++        ggml_tensor * model_t = it.second;
++
++        if (tensor_meta_map.find(base_name + ".loraA") == tensor_meta_map.end() ||
++            tensor_meta_map.find(base_name + ".loraB") == tensor_meta_map.end()) {
++            continue;
++        }
++
++        tensor_meta & metaA = tensor_meta_map.at(base_name + ".loraA");
++        tensor_meta & metaB = tensor_meta_map.at(base_name + ".loraB");
++
++        ggml_init_params lora_init_params = {
++            /* .mem_size   */ ggml_tensor_overhead()*128 + ggml_graph_overhead(),
++            /* .mem_buffer */ nullptr,
++            /* .no_alloc   */ true,
++        };
++        ggml_context * lora_ctx = ggml_init(lora_init_params);
++        if (lora_ctx == nullptr) {
++            LLAMA_LOG_ERROR("%s: error: failed to initialize lora context\n", __func__);
++            ggml_backend_free(backend_cpu);
++            return 1;
++        }
++
++        // create tensors
++        ggml_tensor * loraA = ggml_new_tensor_2d(lora_ctx, metaA.type, metaA.ne[0], metaA.ne[1]);
++        ggml_tensor * loraB = ggml_new_tensor_2d(lora_ctx, metaB.type, metaB.ne[0], metaB.ne[1]);
++        ggml_set_name(loraA, metaA.name.c_str());
++        ggml_set_name(loraB, metaB.name.c_str());
++
++        ggml_tensor * base_t;
++        if (ml) {
++            if (!ml->get_tensor_meta(base_name.c_str())) {
++                LLAMA_LOG_ERROR("%s: error: tensor '%s' not found in base model\n", __func__, base_name.c_str());
++                return 1;
++            }
++            base_t = ggml_dup_tensor(lora_ctx, ml->get_tensor_meta(base_name.c_str()));
++        } else {
++            base_t = ggml_dup_tensor(lora_ctx, model_t);
++        }
++        ggml_set_name(base_t, base_name.c_str());
++
++        // allocate in backend buffer
++        ggml_backend_buffer_t lora_buf = ggml_backend_alloc_ctx_tensors_from_buft(lora_ctx, ggml_backend_cpu_buffer_type());
++        if (lora_buf == nullptr) {
++            LLAMA_LOG_ERROR("%s: error: failed to allocate lora tensors\n", __func__);
++            return 1;
++        }
++
++        // load tensor data
++        auto load_tensor = [&read_buf, &fin](const tensor_meta & tensor_meta, ggml_tensor * tensor) {
++            read_buf.resize(ggml_nbytes(tensor));
++            fin.seek(tensor_meta.offset, SEEK_SET);
++            fin.read_raw(read_buf.data(), ggml_nbytes(tensor));
++            ggml_backend_tensor_set(tensor, read_buf.data(), 0, read_buf.size());
++        };
++        load_tensor(metaA, loraA);
++        load_tensor(metaB, loraB);
++
++        // load base model tensor data
++        if (ml) {
++            ml->load_data_for(base_t);
++        } else {
++            ggml_backend_tensor_copy(model_t, base_t);
++        }
++
++        if (ggml_is_quantized(base_t->type) && !warned) {
++            LLAMA_LOG_WARN("%s: warning: using a lora adapter with a quantized model may result in poor quality, "
++                            "use a f16 or f32 base model with --lora-base\n", __func__);
++            warned = true;
++        }
++
++        if (base_t->ne[0] != loraA->ne[1] || base_t->ne[1] != loraB->ne[1]) {
++            LLAMA_LOG_ERROR("%s: incompatible tensor dimensions (%" PRId64 " and %" PRId64 ");"
++                            " are you sure that this adapter is for this model?\n", __func__, base_t->ne[0], loraA->ne[1]);
++            ggml_free(lora_ctx);
++            ggml_backend_buffer_free(lora_buf);
++            ggml_backend_free(backend_cpu);
++            return 1;
++        }
++
++        auto build_lora_graph = [&]() {
++            // w = w + BA*s
++            ggml_tensor * BA = ggml_mul_mat(lora_ctx, loraA, loraB);
++            ggml_set_name(BA, "BA");
++
++            if (scaling != 1.0f) {
++                BA = ggml_scale(lora_ctx, BA, scaling);
++                ggml_set_name(BA, "BA_scaled");
++            }
++
++            ggml_tensor * r;
++            r = ggml_add_inplace(lora_ctx, base_t, BA);
++            ggml_set_name(r, "r_add");
++
++            if (base_t->type != model_t->type) {
++                // convert the result to the model type
++                r = ggml_cast(lora_ctx, r, model_t->type);
++                ggml_set_name(r, "r_cast");
++            }
++
++            return r;
++        };
++
++        ggml_cgraph * gf = ggml_new_graph(lora_ctx);
++        ggml_tensor * r = build_lora_graph();
++        ggml_build_forward_expand(gf, r);
++
++        ggml_backend_buffer_t graph_buf = ggml_backend_alloc_ctx_tensors_from_buft(lora_ctx, ggml_backend_cpu_buffer_type());
++        if (graph_buf == nullptr) {
++            LLAMA_LOG_ERROR("%s: error: failed to allocate graph tensors\n", __func__);
++            ggml_free(lora_ctx);
++            ggml_backend_buffer_free(lora_buf);
++            ggml_backend_free(backend_cpu);
++            return 1;
++        }
++
++        ggml_backend_graph_compute(backend_cpu, gf);
++
++        ggml_backend_tensor_set(model_t, r->data, 0, ggml_nbytes(r));
++
++#if 0
++        // TODO: use scheduler with fallback to CPU for less copies between CPU and GPU
++        //ggml_backend_sched_t sched = ggml_backend_sched_new(backends.data(), backends.size(), GGML_DEFAULT_GRAPH_SIZE);
++
++        // sched compute
++        ggml_build_forward_expand(gf, build_graph());
++        ggml_backend_sched_init_measure(sched, gf);
++
++        // create the graph again, since the previous one was destroyed by the measure
++        ggml_graph_clear(gf);
++        ggml_build_forward_expand(gf, build_graph());
++        ggml_backend_sched_graph_compute(sched, gf);
++        ggml_backend_sched_free(sched);
++#endif
++
++        ggml_backend_buffer_free(lora_buf);
++        ggml_backend_buffer_free(graph_buf);
++        ggml_free(lora_ctx);
++
++        n_tensors++;
++        if (n_tensors % 4 == 0) {
++            LLAMA_LOG_INFO(".");
++        }
++    }
++
++    ggml_backend_free(backend_cpu);
++
++    const int64_t t_lora_us = ggml_time_us() - t_start_lora_us;
++    LLAMA_LOG_INFO(" done (%.2f ms)\n", t_lora_us / 1000.0);
++
++    return 0;
++}
++
++int32_t llama_model_apply_lora_from_file(const struct llama_model * model, const char * path_lora, float scale, const char * path_base_model, int32_t n_threads) {
++    try {
++        return llama_apply_lora_from_file_internal(*model, path_lora, scale, path_base_model, n_threads);
++    } catch (const std::exception & err) {
++        LLAMA_LOG_ERROR("%s: failed to apply lora adapter: %s\n", __func__, err.what());
++        return 1;
++    }
++}
+\ No newline at end of file

llm/server.go

@@ -385,8 +385,10 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 			filteredEnv := []string{}
 			for _, ev := range s.cmd.Env {
 				if strings.HasPrefix(ev, "CUDA_") ||
+					strings.HasPrefix(ev, "ROCR_") ||
 					strings.HasPrefix(ev, "ROCM_") ||
 					strings.HasPrefix(ev, "HIP_") ||
+					strings.HasPrefix(ev, "GPU_") ||
 					strings.HasPrefix(ev, "HSA_") ||
 					strings.HasPrefix(ev, "GGML_") ||
 					strings.HasPrefix(ev, "PATH=") ||
@@ -415,7 +417,17 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 
 		// reap subprocess when it exits
 		go func() {
-			s.done <- s.cmd.Wait()
+			err := s.cmd.Wait()
+			// Favor a more detailed message over the process exit status
+			if err != nil && s.status != nil && s.status.LastErrMsg != "" {
+				slog.Debug("llama runner terminated", "error", err)
+				if strings.Contains(s.status.LastErrMsg, "unknown model") {
+					s.status.LastErrMsg = "this model is not supported by your version of Ollama. You may need to upgrade"
+				}
+				s.done <- fmt.Errorf(s.status.LastErrMsg)
+			} else {
+				s.done <- err
+			}
 		}()
 
 		return s, nil
@@ -578,14 +590,7 @@ func (s *llmServer) WaitUntilRunning(ctx context.Context) error {
 			slog.Warn("client connection closed before server finished loading, aborting load")
 			return fmt.Errorf("timed out waiting for llama runner to start: %w", ctx.Err())
 		case err := <-s.done:
-			msg := ""
-			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)
+			return fmt.Errorf("llama runner process has terminated: %w", err)
 		default:
 		}
 		if time.Now().After(stallTimer) {

openai/openai.go

@@ -351,7 +351,6 @@ func fromChatRequest(r ChatCompletionRequest) (*api.ChatRequest, error) {
 		case string:
 			messages = append(messages, api.Message{Role: msg.Role, Content: content})
 		case []any:
-			message := api.Message{Role: msg.Role}
 			for _, c := range content {
 				data, ok := c.(map[string]any)
 				if !ok {
@@ -363,7 +362,7 @@ func fromChatRequest(r ChatCompletionRequest) (*api.ChatRequest, error) {
 					if !ok {
 						return nil, fmt.Errorf("invalid message format")
 					}
-					message.Content = text
+					messages = append(messages, api.Message{Role: msg.Role, Content: text})
 				case "image_url":
 					var url string
 					if urlMap, ok := data["image_url"].(map[string]any); ok {
@@ -395,12 +394,12 @@ func fromChatRequest(r ChatCompletionRequest) (*api.ChatRequest, error) {
 					if err != nil {
 						return nil, fmt.Errorf("invalid message format")
 					}
-					message.Images = append(message.Images, img)
+
+					messages = append(messages, api.Message{Role: msg.Role, Images: []api.ImageData{img}})
 				default:
 					return nil, fmt.Errorf("invalid message format")
 				}
 			}
-			messages = append(messages, message)
 		default:
 			if msg.ToolCalls == nil {
 				return nil, fmt.Errorf("invalid message content type: %T", content)
@@ -878,6 +877,7 @@ func ChatMiddleware() gin.HandlerFunc {
 		chatReq, err := fromChatRequest(req)
 		if err != nil {
 			c.AbortWithStatusJSON(http.StatusBadRequest, NewError(http.StatusBadRequest, err.Error()))
+			return
 		}
 
 		if err := json.NewEncoder(&b).Encode(chatReq); err != nil {

openai/openai_test.go

@@ -20,113 +20,59 @@ const prefix = `data:image/jpeg;base64,`
 const image = `iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAQAAAC1HAwCAAAAC0lEQVR42mNk+A8AAQUBAScY42YAAAAASUVORK5CYII=`
 const imageURL = prefix + image
 
-func TestMiddlewareRequests(t *testing.T) {
+func prepareRequest(req *http.Request, body any) {
+	bodyBytes, _ := json.Marshal(body)
+	req.Body = io.NopCloser(bytes.NewReader(bodyBytes))
+	req.Header.Set("Content-Type", "application/json")
+}
+
+func captureRequestMiddleware(capturedRequest any) gin.HandlerFunc {
+	return func(c *gin.Context) {
+		bodyBytes, _ := io.ReadAll(c.Request.Body)
+		c.Request.Body = io.NopCloser(bytes.NewReader(bodyBytes))
+		err := json.Unmarshal(bodyBytes, capturedRequest)
+		if err != nil {
+			c.AbortWithStatusJSON(http.StatusInternalServerError, "failed to unmarshal request")
+		}
+		c.Next()
+	}
+}
+
+func TestChatMiddleware(t *testing.T) {
 	type testCase struct {
 		Name     string
-		Method   string
-		Path     string
-		Handler  func() gin.HandlerFunc
 		Setup    func(t *testing.T, req *http.Request)
-		Expected func(t *testing.T, req *http.Request)
+		Expected func(t *testing.T, req *api.ChatRequest, resp *httptest.ResponseRecorder)
 	}
 
-	var capturedRequest *http.Request
-
-	captureRequestMiddleware := func() gin.HandlerFunc {
-		return func(c *gin.Context) {
-			bodyBytes, _ := io.ReadAll(c.Request.Body)
-			c.Request.Body = io.NopCloser(bytes.NewReader(bodyBytes))
-			capturedRequest = c.Request
-			c.Next()
-		}
-	}
+	var capturedRequest *api.ChatRequest
 
 	testCases := []testCase{
 		{
-			Name:    "chat handler",
-			Method:  http.MethodPost,
-			Path:    "/api/chat",
-			Handler: ChatMiddleware,
+			Name: "chat handler",
 			Setup: func(t *testing.T, req *http.Request) {
 				body := ChatCompletionRequest{
 					Model:    "test-model",
 					Messages: []Message{{Role: "user", Content: "Hello"}},
 				}
-
-				bodyBytes, _ := json.Marshal(body)
-
-				req.Body = io.NopCloser(bytes.NewReader(bodyBytes))
-				req.Header.Set("Content-Type", "application/json")
+				prepareRequest(req, body)
 			},
-			Expected: func(t *testing.T, req *http.Request) {
-				var chatReq api.ChatRequest
-				if err := json.NewDecoder(req.Body).Decode(&chatReq); err != nil {
-					t.Fatal(err)
+			Expected: func(t *testing.T, req *api.ChatRequest, resp *httptest.ResponseRecorder) {
+				if resp.Code != http.StatusOK {
+					t.Fatalf("expected 200, got %d", resp.Code)
 				}
 
-				if chatReq.Messages[0].Role != "user" {
-					t.Fatalf("expected 'user', got %s", chatReq.Messages[0].Role)
+				if req.Messages[0].Role != "user" {
+					t.Fatalf("expected 'user', got %s", req.Messages[0].Role)
 				}
 
-				if chatReq.Messages[0].Content != "Hello" {
-					t.Fatalf("expected 'Hello', got %s", chatReq.Messages[0].Content)
+				if req.Messages[0].Content != "Hello" {
+					t.Fatalf("expected 'Hello', got %s", req.Messages[0].Content)
 				}
 			},
 		},
 		{
-			Name:    "completions handler",
-			Method:  http.MethodPost,
-			Path:    "/api/generate",
-			Handler: CompletionsMiddleware,
-			Setup: func(t *testing.T, req *http.Request) {
-				temp := float32(0.8)
-				body := CompletionRequest{
-					Model:       "test-model",
-					Prompt:      "Hello",
-					Temperature: &temp,
-					Stop:        []string{"\n", "stop"},
-					Suffix:      "suffix",
-				}
-
-				bodyBytes, _ := json.Marshal(body)
-
-				req.Body = io.NopCloser(bytes.NewReader(bodyBytes))
-				req.Header.Set("Content-Type", "application/json")
-			},
-			Expected: func(t *testing.T, req *http.Request) {
-				var genReq api.GenerateRequest
-				if err := json.NewDecoder(req.Body).Decode(&genReq); err != nil {
-					t.Fatal(err)
-				}
-
-				if genReq.Prompt != "Hello" {
-					t.Fatalf("expected 'Hello', got %s", genReq.Prompt)
-				}
-
-				if genReq.Options["temperature"] != 1.6 {
-					t.Fatalf("expected 1.6, got %f", genReq.Options["temperature"])
-				}
-
-				stopTokens, ok := genReq.Options["stop"].([]any)
-
-				if !ok {
-					t.Fatalf("expected stop tokens to be a list")
-				}
-
-				if stopTokens[0] != "\n" || stopTokens[1] != "stop" {
-					t.Fatalf("expected ['\\n', 'stop'], got %v", stopTokens)
-				}
-
-				if genReq.Suffix != "suffix" {
-					t.Fatalf("expected 'suffix', got %s", genReq.Suffix)
-				}
-			},
-		},
-		{
-			Name:    "chat handler with image content",
-			Method:  http.MethodPost,
-			Path:    "/api/chat",
-			Handler: ChatMiddleware,
+			Name: "chat handler with image content",
 			Setup: func(t *testing.T, req *http.Request) {
 				body := ChatCompletionRequest{
 					Model: "test-model",
@@ -139,87 +85,254 @@ func TestMiddlewareRequests(t *testing.T) {
 						},
 					},
 				}
-
-				bodyBytes, _ := json.Marshal(body)
-
-				req.Body = io.NopCloser(bytes.NewReader(bodyBytes))
-				req.Header.Set("Content-Type", "application/json")
+				prepareRequest(req, body)
 			},
-			Expected: func(t *testing.T, req *http.Request) {
-				var chatReq api.ChatRequest
-				if err := json.NewDecoder(req.Body).Decode(&chatReq); err != nil {
-					t.Fatal(err)
+			Expected: func(t *testing.T, req *api.ChatRequest, resp *httptest.ResponseRecorder) {
+				if resp.Code != http.StatusOK {
+					t.Fatalf("expected 200, got %d", resp.Code)
 				}
 
-				if chatReq.Messages[0].Role != "user" {
-					t.Fatalf("expected 'user', got %s", chatReq.Messages[0].Role)
+				if req.Messages[0].Role != "user" {
+					t.Fatalf("expected 'user', got %s", req.Messages[0].Role)
 				}
 
-				if chatReq.Messages[0].Content != "Hello" {
-					t.Fatalf("expected 'Hello', got %s", chatReq.Messages[0].Content)
+				if req.Messages[0].Content != "Hello" {
+					t.Fatalf("expected 'Hello', got %s", req.Messages[0].Content)
 				}
 
 				img, _ := base64.StdEncoding.DecodeString(imageURL[len(prefix):])
 
-				if !bytes.Equal(chatReq.Messages[0].Images[0], img) {
-					t.Fatalf("expected image encoding, got %s", chatReq.Messages[0].Images[0])
+				if req.Messages[1].Role != "user" {
+					t.Fatalf("expected 'user', got %s", req.Messages[1].Role)
+				}
+
+				if !bytes.Equal(req.Messages[1].Images[0], img) {
+					t.Fatalf("expected image encoding, got %s", req.Messages[1].Images[0])
 				}
 			},
 		},
 		{
-			Name:    "embed handler single input",
-			Method:  http.MethodPost,
-			Path:    "/api/embed",
-			Handler: EmbeddingsMiddleware,
+			Name: "chat handler with tools",
+			Setup: func(t *testing.T, req *http.Request) {
+				body := ChatCompletionRequest{
+					Model: "test-model",
+					Messages: []Message{
+						{Role: "user", Content: "What's the weather like in Paris Today?"},
+						{Role: "assistant", ToolCalls: []ToolCall{{
+							ID:   "id",
+							Type: "function",
+							Function: struct {
+								Name      string `json:"name"`
+								Arguments string `json:"arguments"`
+							}{
+								Name:      "get_current_weather",
+								Arguments: "{\"location\": \"Paris, France\", \"format\": \"celsius\"}",
+							},
+						}}},
+					},
+				}
+				prepareRequest(req, body)
+			},
+			Expected: func(t *testing.T, req *api.ChatRequest, resp *httptest.ResponseRecorder) {
+				if resp.Code != 200 {
+					t.Fatalf("expected 200, got %d", resp.Code)
+				}
+
+				if req.Messages[0].Content != "What's the weather like in Paris Today?" {
+					t.Fatalf("expected What's the weather like in Paris Today?, got %s", req.Messages[0].Content)
+				}
+
+				if req.Messages[1].ToolCalls[0].Function.Arguments["location"] != "Paris, France" {
+					t.Fatalf("expected 'Paris, France', got %v", req.Messages[1].ToolCalls[0].Function.Arguments["location"])
+				}
+
+				if req.Messages[1].ToolCalls[0].Function.Arguments["format"] != "celsius" {
+					t.Fatalf("expected celsius, got %v", req.Messages[1].ToolCalls[0].Function.Arguments["format"])
+				}
+			},
+		},
+		{
+			Name: "chat handler error forwarding",
+			Setup: func(t *testing.T, req *http.Request) {
+				body := ChatCompletionRequest{
+					Model:    "test-model",
+					Messages: []Message{{Role: "user", Content: 2}},
+				}
+				prepareRequest(req, body)
+			},
+			Expected: func(t *testing.T, req *api.ChatRequest, resp *httptest.ResponseRecorder) {
+				if resp.Code != http.StatusBadRequest {
+					t.Fatalf("expected 400, got %d", resp.Code)
+				}
+
+				if !strings.Contains(resp.Body.String(), "invalid message content type") {
+					t.Fatalf("error was not forwarded")
+				}
+			},
+		},
+	}
+
+	endpoint := func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(ChatMiddleware(), captureRequestMiddleware(&capturedRequest))
+	router.Handle(http.MethodPost, "/api/chat", endpoint)
+
+	for _, tc := range testCases {
+		t.Run(tc.Name, func(t *testing.T) {
+			req, _ := http.NewRequest(http.MethodPost, "/api/chat", nil)
+
+			tc.Setup(t, req)
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			tc.Expected(t, capturedRequest, resp)
+
+			capturedRequest = nil
+		})
+	}
+}
+
+func TestCompletionsMiddleware(t *testing.T) {
+	type testCase struct {
+		Name     string
+		Setup    func(t *testing.T, req *http.Request)
+		Expected func(t *testing.T, req *api.GenerateRequest, resp *httptest.ResponseRecorder)
+	}
+
+	var capturedRequest *api.GenerateRequest
+
+	testCases := []testCase{
+		{
+			Name: "completions handler",
+			Setup: func(t *testing.T, req *http.Request) {
+				temp := float32(0.8)
+				body := CompletionRequest{
+					Model:       "test-model",
+					Prompt:      "Hello",
+					Temperature: &temp,
+					Stop:        []string{"\n", "stop"},
+					Suffix:      "suffix",
+				}
+				prepareRequest(req, body)
+			},
+			Expected: func(t *testing.T, req *api.GenerateRequest, resp *httptest.ResponseRecorder) {
+				if req.Prompt != "Hello" {
+					t.Fatalf("expected 'Hello', got %s", req.Prompt)
+				}
+
+				if req.Options["temperature"] != 1.6 {
+					t.Fatalf("expected 1.6, got %f", req.Options["temperature"])
+				}
+
+				stopTokens, ok := req.Options["stop"].([]any)
+
+				if !ok {
+					t.Fatalf("expected stop tokens to be a list")
+				}
+
+				if stopTokens[0] != "\n" || stopTokens[1] != "stop" {
+					t.Fatalf("expected ['\\n', 'stop'], got %v", stopTokens)
+				}
+
+				if req.Suffix != "suffix" {
+					t.Fatalf("expected 'suffix', got %s", req.Suffix)
+				}
+			},
+		},
+		{
+			Name: "completions handler error forwarding",
+			Setup: func(t *testing.T, req *http.Request) {
+				body := CompletionRequest{
+					Model:       "test-model",
+					Prompt:      "Hello",
+					Temperature: nil,
+					Stop:        []int{1, 2},
+					Suffix:      "suffix",
+				}
+				prepareRequest(req, body)
+			},
+			Expected: func(t *testing.T, req *api.GenerateRequest, resp *httptest.ResponseRecorder) {
+				if resp.Code != http.StatusBadRequest {
+					t.Fatalf("expected 400, got %d", resp.Code)
+				}
+
+				if !strings.Contains(resp.Body.String(), "invalid type for 'stop' field") {
+					t.Fatalf("error was not forwarded")
+				}
+			},
+		},
+	}
+
+	endpoint := func(c *gin.Context) {
+		c.Status(http.StatusOK)
+	}
+
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(CompletionsMiddleware(), captureRequestMiddleware(&capturedRequest))
+	router.Handle(http.MethodPost, "/api/generate", endpoint)
+
+	for _, tc := range testCases {
+		t.Run(tc.Name, func(t *testing.T) {
+			req, _ := http.NewRequest(http.MethodPost, "/api/generate", nil)
+
+			tc.Setup(t, req)
+
+			resp := httptest.NewRecorder()
+			router.ServeHTTP(resp, req)
+
+			tc.Expected(t, capturedRequest, resp)
+
+			capturedRequest = nil
+		})
+	}
+}
+
+func TestEmbeddingsMiddleware(t *testing.T) {
+	type testCase struct {
+		Name     string
+		Setup    func(t *testing.T, req *http.Request)
+		Expected func(t *testing.T, req *api.EmbedRequest, resp *httptest.ResponseRecorder)
+	}
+
+	var capturedRequest *api.EmbedRequest
+
+	testCases := []testCase{
+		{
+			Name: "embed handler single input",
 			Setup: func(t *testing.T, req *http.Request) {
 				body := EmbedRequest{
 					Input: "Hello",
 					Model: "test-model",
 				}
-
-				bodyBytes, _ := json.Marshal(body)
-
-				req.Body = io.NopCloser(bytes.NewReader(bodyBytes))
-				req.Header.Set("Content-Type", "application/json")
+				prepareRequest(req, body)
 			},
-			Expected: func(t *testing.T, req *http.Request) {
-				var embedReq api.EmbedRequest
-				if err := json.NewDecoder(req.Body).Decode(&embedReq); err != nil {
-					t.Fatal(err)
+			Expected: func(t *testing.T, req *api.EmbedRequest, resp *httptest.ResponseRecorder) {
+				if req.Input != "Hello" {
+					t.Fatalf("expected 'Hello', got %s", req.Input)
 				}
 
-				if embedReq.Input != "Hello" {
-					t.Fatalf("expected 'Hello', got %s", embedReq.Input)
-				}
-
-				if embedReq.Model != "test-model" {
-					t.Fatalf("expected 'test-model', got %s", embedReq.Model)
+				if req.Model != "test-model" {
+					t.Fatalf("expected 'test-model', got %s", req.Model)
 				}
 			},
 		},
 		{
-			Name:    "embed handler batch input",
-			Method:  http.MethodPost,
-			Path:    "/api/embed",
-			Handler: EmbeddingsMiddleware,
+			Name: "embed handler batch input",
 			Setup: func(t *testing.T, req *http.Request) {
 				body := EmbedRequest{
 					Input: []string{"Hello", "World"},
 					Model: "test-model",
 				}
-
-				bodyBytes, _ := json.Marshal(body)
-
-				req.Body = io.NopCloser(bytes.NewReader(bodyBytes))
-				req.Header.Set("Content-Type", "application/json")
+				prepareRequest(req, body)
 			},
-			Expected: func(t *testing.T, req *http.Request) {
-				var embedReq api.EmbedRequest
-				if err := json.NewDecoder(req.Body).Decode(&embedReq); err != nil {
-					t.Fatal(err)
-				}
-
-				input, ok := embedReq.Input.([]any)
+			Expected: func(t *testing.T, req *api.EmbedRequest, resp *httptest.ResponseRecorder) {
+				input, ok := req.Input.([]any)
 
 				if !ok {
 					t.Fatalf("expected input to be a list")
@@ -233,36 +346,52 @@ func TestMiddlewareRequests(t *testing.T) {
 					t.Fatalf("expected 'World', got %s", input[1])
 				}
 
-				if embedReq.Model != "test-model" {
-					t.Fatalf("expected 'test-model', got %s", embedReq.Model)
+				if req.Model != "test-model" {
+					t.Fatalf("expected 'test-model', got %s", req.Model)
+				}
+			},
+		},
+		{
+			Name: "embed handler error forwarding",
+			Setup: func(t *testing.T, req *http.Request) {
+				body := EmbedRequest{
+					Model: "test-model",
+				}
+				prepareRequest(req, body)
+			},
+			Expected: func(t *testing.T, req *api.EmbedRequest, resp *httptest.ResponseRecorder) {
+				if resp.Code != http.StatusBadRequest {
+					t.Fatalf("expected 400, got %d", resp.Code)
+				}
+
+				if !strings.Contains(resp.Body.String(), "invalid input") {
+					t.Fatalf("error was not forwarded")
 				}
 			},
 		},
 	}
 
-	gin.SetMode(gin.TestMode)
-	router := gin.New()
-
 	endpoint := func(c *gin.Context) {
 		c.Status(http.StatusOK)
 	}
 
+	gin.SetMode(gin.TestMode)
+	router := gin.New()
+	router.Use(EmbeddingsMiddleware(), captureRequestMiddleware(&capturedRequest))
+	router.Handle(http.MethodPost, "/api/embed", endpoint)
+
 	for _, tc := range testCases {
 		t.Run(tc.Name, func(t *testing.T) {
-			router = gin.New()
-			router.Use(captureRequestMiddleware())
-			router.Use(tc.Handler())
-			router.Handle(tc.Method, tc.Path, endpoint)
-			req, _ := http.NewRequest(tc.Method, tc.Path, nil)
+			req, _ := http.NewRequest(http.MethodPost, "/api/embed", nil)
 
-			if tc.Setup != nil {
-				tc.Setup(t, req)
-			}
+			tc.Setup(t, req)
 
 			resp := httptest.NewRecorder()
 			router.ServeHTTP(resp, req)
 
-			tc.Expected(t, capturedRequest)
+			tc.Expected(t, capturedRequest, resp)
+
+			capturedRequest = nil
 		})
 	}
 }
@@ -280,36 +409,6 @@ func TestMiddlewareResponses(t *testing.T) {
 	}
 
 	testCases := []testCase{
-		{
-			Name:     "completions handler error forwarding",
-			Method:   http.MethodPost,
-			Path:     "/api/generate",
-			TestPath: "/api/generate",
-			Handler:  CompletionsMiddleware,
-			Endpoint: func(c *gin.Context) {
-				c.JSON(http.StatusBadRequest, gin.H{"error": "invalid request"})
-			},
-			Setup: func(t *testing.T, req *http.Request) {
-				body := CompletionRequest{
-					Model:  "test-model",
-					Prompt: "Hello",
-				}
-
-				bodyBytes, _ := json.Marshal(body)
-
-				req.Body = io.NopCloser(bytes.NewReader(bodyBytes))
-				req.Header.Set("Content-Type", "application/json")
-			},
-			Expected: func(t *testing.T, resp *httptest.ResponseRecorder) {
-				if resp.Code != http.StatusBadRequest {
-					t.Fatalf("expected 400, got %d", resp.Code)
-				}
-
-				if !strings.Contains(resp.Body.String(), `"invalid request"`) {
-					t.Fatalf("error was not forwarded")
-				}
-			},
-		},
 		{
 			Name:     "list handler",
 			Method:   http.MethodGet,
@@ -326,8 +425,6 @@ func TestMiddlewareResponses(t *testing.T) {
 				})
 			},
 			Expected: func(t *testing.T, resp *httptest.ResponseRecorder) {
-				assert.Equal(t, http.StatusOK, resp.Code)
-
 				var listResp ListCompletion
 				if err := json.NewDecoder(resp.Body).Decode(&listResp); err != nil {
 					t.Fatal(err)
@@ -391,6 +488,8 @@ func TestMiddlewareResponses(t *testing.T) {
 			resp := httptest.NewRecorder()
 			router.ServeHTTP(resp, req)
 
+			assert.Equal(t, http.StatusOK, resp.Code)
+
 			tc.Expected(t, resp)
 		})
 	}

server/images.go

@@ -492,6 +492,12 @@ func CreateModel(ctx context.Context, name model.Name, modelFileDir, quantizatio
 				layers = append(layers, baseLayer.Layer)
 			}
 		case "license", "template", "system":
+			if c.Name == "template" {
+				if _, err := template.Parse(c.Args); err != nil {
+					return fmt.Errorf("%w: %s", errBadTemplate, err)
+				}
+			}
+
 			if c.Name != "license" {
 				// replace
 				layers = slices.DeleteFunc(layers, func(layer *Layer) bool {

server/model.go

@@ -344,6 +344,10 @@ func (m *Model) parseToolCalls(s string) ([]api.ToolCall, bool) {
 		}
 	}
 
+	if name == "" || arguments == "" {
+		return nil, false
+	}
+
 	var objs []map[string]any
 	for offset := 0; offset < len(s); {
 		var obj map[string]any
@@ -361,23 +365,40 @@ func (m *Model) parseToolCalls(s string) ([]api.ToolCall, bool) {
 			return nil, false
 		} else {
 			offset += int(decoder.InputOffset())
-			objs = append(objs, obj)
+
+			// collect all nested objects
+			var collect func(any) []map[string]any
+			collect = func(obj any) (all []map[string]any) {
+				switch o := obj.(type) {
+				case map[string]any:
+					all = append(all, o)
+					for _, v := range o {
+						all = append(all, collect(v)...)
+					}
+				case []any:
+					for _, v := range o {
+						all = append(all, collect(v)...)
+					}
+				}
+
+				return all
+			}
+			objs = append(objs, collect(obj)...)
 		}
 	}
 
 	var toolCalls []api.ToolCall
 	for _, kv := range objs {
-		var call api.ToolCall
-		for k, v := range kv {
-			switch k {
-			case name:
-				call.Function.Name = v.(string)
-			case arguments:
-				call.Function.Arguments = v.(map[string]any)
-			}
+		n, nok := kv[name].(string)
+		a, aok := kv[arguments].(map[string]any)
+		if nok && aok {
+			toolCalls = append(toolCalls, api.ToolCall{
+				Function: api.ToolCallFunction{
+					Name:      n,
+					Arguments: a,
+				},
+			})
 		}
-
-		toolCalls = append(toolCalls, call)
 	}
 
 	return toolCalls, len(toolCalls) > 0

server/model_test.go

@@ -166,6 +166,7 @@ The temperature in San Francisco, CA is 70°F and in Toronto, Canada is 20°C.`,
 {"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}}
 {"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}
 </tool_call>`, true},
+		{"xlam", `{"tool_calls": [{"name": "get_current_weather", "arguments": {"format":"fahrenheit","location":"San Francisco, CA"}},{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Toronto, Canada"}}]}`, true},
 	}
 
 	var tools []api.Tool

server/routes.go

@@ -56,6 +56,7 @@ func init() {
 }
 
 var errRequired = errors.New("is required")
+var errBadTemplate = errors.New("template error")
 
 func modelOptions(model *Model, requestOpts map[string]interface{}) (api.Options, error) {
 	opts := api.DefaultOptions()
@@ -609,6 +610,9 @@ func (s *Server) CreateModelHandler(c *gin.Context) {
 
 		quantization := cmp.Or(r.Quantize, r.Quantization)
 		if err := CreateModel(ctx, name, filepath.Dir(r.Path), strings.ToUpper(quantization), f, fn); err != nil {
+			if errors.Is(err, errBadTemplate) {
+				ch <- gin.H{"error": err.Error(), "status": http.StatusBadRequest}
+			}
 			ch <- gin.H{"error": err.Error()}
 		}
 	}()
@@ -1196,11 +1200,15 @@ func waitForStream(c *gin.Context, ch chan interface{}) {
 				return
 			}
 		case gin.H:
+			status, ok := r["status"].(int)
+			if !ok {
+				status = http.StatusInternalServerError
+			}
 			if errorMsg, ok := r["error"].(string); ok {
-				c.JSON(http.StatusInternalServerError, gin.H{"error": errorMsg})
+				c.JSON(status, gin.H{"error": errorMsg})
 				return
 			} else {
-				c.JSON(http.StatusInternalServerError, gin.H{"error": "unexpected error format in progress response"})
+				c.JSON(status, gin.H{"error": "unexpected error format in progress response"})
 				return
 			}
 		default:

server/routes_create_test.go

@@ -491,6 +491,42 @@ func TestCreateTemplateSystem(t *testing.T) {
 	if string(system) != "Say bye!" {
 		t.Errorf("expected \"Say bye!\", actual %s", system)
 	}
+
+	t.Run("incomplete template", func(t *testing.T) {
+		w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+			Name:      "test",
+			Modelfile: fmt.Sprintf("FROM %s\nTEMPLATE {{ .Prompt", createBinFile(t, nil, nil)),
+			Stream:    &stream,
+		})
+	
+		if w.Code != http.StatusBadRequest {
+			t.Fatalf("expected status code 400, actual %d", w.Code)
+		}
+	})
+
+	t.Run("template with unclosed if", func(t *testing.T) {
+		w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+			Name:      "test",
+			Modelfile: fmt.Sprintf("FROM %s\nTEMPLATE {{ if .Prompt }}", createBinFile(t, nil, nil)),
+			Stream:    &stream,
+		})
+	
+		if w.Code != http.StatusBadRequest {
+			t.Fatalf("expected status code 400, actual %d", w.Code)
+		}
+	})
+
+	t.Run("template with undefined function", func(t *testing.T) {
+		w := createRequest(t, s.CreateModelHandler, api.CreateRequest{
+			Name:      "test",
+			Modelfile: fmt.Sprintf("FROM %s\nTEMPLATE {{  Prompt }}", createBinFile(t, nil, nil)),
+			Stream:    &stream,
+		})
+	
+		if w.Code != http.StatusBadRequest {
+			t.Fatalf("expected status code 400, actual %d", w.Code)
+		}
+	})
 }
 
 func TestCreateLicenses(t *testing.T) {

server/routes_generate_test.go

@@ -73,8 +73,8 @@ func TestGenerateChat(t *testing.T) {
 			getCpuFn:      gpu.GetCPUInfo,
 			reschedDelay:  250 * time.Millisecond,
 			loadFn: func(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel int) {
-				// add 10ms delay to simulate loading
-				time.Sleep(10 * time.Millisecond)
+				// add small delay to simulate loading
+				time.Sleep(time.Millisecond)
 				req.successCh <- &runnerRef{
 					llama: &mock,
 				}
@@ -371,6 +371,8 @@ func TestGenerate(t *testing.T) {
 			getCpuFn:      gpu.GetCPUInfo,
 			reschedDelay:  250 * time.Millisecond,
 			loadFn: func(req *LlmRequest, ggml *llm.GGML, gpus gpu.GpuInfoList, numParallel int) {
+				// add small delay to simulate loading
+				time.Sleep(time.Millisecond)
 				req.successCh <- &runnerRef{
 					llama: &mock,
 				}

server/sched_test.go

@@ -94,7 +94,7 @@ func TestLoad(t *testing.T) {
 	require.Len(t, s.expiredCh, 1)
 }
 
-type bundle struct {
+type reqBundle struct {
 	ctx     context.Context //nolint:containedctx
 	ctxDone func()
 	srv     *mockLlm
@@ -102,13 +102,13 @@ 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, numParallel int) (llm.LlamaServer, error) {
+func (scenario *reqBundle) 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
 }
 
-func newScenario(t *testing.T, ctx context.Context, modelName string, estimatedVRAM uint64) *bundle {
-	scenario := &bundle{}
-	scenario.ctx, scenario.ctxDone = context.WithCancel(ctx)
+func newScenarioRequest(t *testing.T, ctx context.Context, modelName string, estimatedVRAM uint64, duration *api.Duration) *reqBundle {
+	b := &reqBundle{}
+	b.ctx, b.ctxDone = context.WithCancel(ctx)
 	t.Helper()
 
 	f, err := os.CreateTemp(t.TempDir(), modelName)
@@ -135,124 +135,154 @@ func newScenario(t *testing.T, ctx context.Context, modelName string, estimatedV
 
 	fname := f.Name()
 	model := &Model{Name: modelName, ModelPath: fname}
-	scenario.ggml, err = llm.LoadModel(model.ModelPath, 0)
+	b.ggml, err = llm.LoadModel(model.ModelPath, 0)
 	require.NoError(t, err)
 
-	scenario.req = &LlmRequest{
-		ctx:             scenario.ctx,
+	if duration == nil {
+		duration = &api.Duration{Duration: 5 * time.Millisecond}
+	}
+	b.req = &LlmRequest{
+		ctx:             b.ctx,
 		model:           model,
 		opts:            api.DefaultOptions(),
-		sessionDuration: &api.Duration{Duration: 5 * time.Millisecond},
+		sessionDuration: duration,
 		successCh:       make(chan *runnerRef, 1),
 		errCh:           make(chan error, 1),
 	}
-	scenario.srv = &mockLlm{estimatedVRAM: estimatedVRAM, estimatedVRAMByGPU: map[string]uint64{"": estimatedVRAM}}
-	return scenario
+	b.srv = &mockLlm{estimatedVRAM: estimatedVRAM, estimatedVRAMByGPU: map[string]uint64{"": estimatedVRAM}}
+	return b
 }
 
-func TestRequests(t *testing.T) {
-	ctx, done := context.WithTimeout(context.Background(), 10*time.Second)
+func getGpuFn() gpu.GpuInfoList {
+	g := gpu.GpuInfo{Library: "metal"}
+	g.TotalMemory = 24 * format.GigaByte
+	g.FreeMemory = 12 * format.GigaByte
+	return []gpu.GpuInfo{g}
+}
+
+func getCpuFn() gpu.GpuInfoList {
+	g := gpu.GpuInfo{Library: "cpu"}
+	g.TotalMemory = 32 * format.GigaByte
+	g.FreeMemory = 26 * format.GigaByte
+	return []gpu.GpuInfo{g}
+}
+
+func TestRequestsSameModelSameRequest(t *testing.T) {
+	ctx, done := context.WithTimeout(context.Background(), 500*time.Millisecond)
 	defer done()
-
-	// Same model, same request
-	scenario1a := newScenario(t, ctx, "ollama-model-1", 10)
-	scenario1a.req.sessionDuration = &api.Duration{Duration: 5 * time.Millisecond}
-	scenario1b := newScenario(t, ctx, "ollama-model-1", 11)
-	scenario1b.req.model = scenario1a.req.model
-	scenario1b.ggml = scenario1a.ggml
-	scenario1b.req.sessionDuration = &api.Duration{Duration: 0}
-
-	// simple reload of same model
-	scenario2a := newScenario(t, ctx, "ollama-model-1", 20)
-	tmpModel := *scenario1a.req.model
-	scenario2a.req.model = &tmpModel
-	scenario2a.ggml = scenario1a.ggml
-	scenario2a.req.sessionDuration = &api.Duration{Duration: 5 * time.Millisecond}
-
-	// Multiple loaded models
-	scenario3a := newScenario(t, ctx, "ollama-model-3a", 1*format.GigaByte)
-	scenario3b := newScenario(t, ctx, "ollama-model-3b", 24*format.GigaByte)
-	scenario3c := newScenario(t, ctx, "ollama-model-4a", 30)
-	scenario3c.req.opts.NumGPU = 0                           // CPU load, will be allowed
-	scenario3d := newScenario(t, ctx, "ollama-model-3c", 30) // Needs prior unloaded
-
 	s := InitScheduler(ctx)
-	s.getGpuFn = func() gpu.GpuInfoList {
-		g := gpu.GpuInfo{Library: "metal"}
-		g.TotalMemory = 24 * format.GigaByte
-		g.FreeMemory = 12 * format.GigaByte
-		return []gpu.GpuInfo{g}
-	}
-	s.getCpuFn = func() gpu.GpuInfoList {
-		g := gpu.GpuInfo{Library: "cpu"}
-		g.TotalMemory = 32 * format.GigaByte
-		g.FreeMemory = 26 * format.GigaByte
-		return []gpu.GpuInfo{g}
-	}
-	s.newServerFn = scenario1a.newServer
-	slog.Info("scenario1a")
-	s.pendingReqCh <- scenario1a.req
+	s.getGpuFn = getGpuFn
+	s.getCpuFn = getCpuFn
+	a := newScenarioRequest(t, ctx, "ollama-model-1", 10, &api.Duration{Duration: 5 * time.Millisecond})
+	b := newScenarioRequest(t, ctx, "ollama-model-1", 11, &api.Duration{Duration: 0})
+	b.req.model = a.req.model
+	b.ggml = a.ggml
+
+	s.newServerFn = a.newServer
+	slog.Info("a")
+	s.pendingReqCh <- a.req
 	require.Len(t, s.pendingReqCh, 1)
 	s.Run(ctx)
 	select {
-	case resp := <-scenario1a.req.successCh:
-		require.Equal(t, resp.llama, scenario1a.srv)
+	case resp := <-a.req.successCh:
+		require.Equal(t, resp.llama, a.srv)
 		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, scenario1a.req.errCh)
-	case err := <-scenario1a.req.errCh:
+		require.Empty(t, a.req.errCh)
+	case err := <-a.req.errCh:
 		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
 	}
 
 	// Same runner as first request due to not needing a reload
-	s.newServerFn = scenario1b.newServer
-	slog.Info("scenario1b")
-	s.pendingReqCh <- scenario1b.req
+	s.newServerFn = b.newServer
+	slog.Info("b")
+	s.pendingReqCh <- b.req
 	select {
-	case resp := <-scenario1b.req.successCh:
-		require.Equal(t, resp.llama, scenario1a.srv)
+	case resp := <-b.req.successCh:
+		require.Equal(t, resp.llama, a.srv)
 		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, scenario1b.req.errCh)
-	case err := <-scenario1b.req.errCh:
+		require.Empty(t, b.req.errCh)
+	case err := <-b.req.errCh:
+		t.Fatal(err.Error())
+	case <-ctx.Done():
+		t.Fatal("timeout")
+	}
+}
+
+func TestRequestsSimpleReloadSameModel(t *testing.T) {
+	ctx, done := context.WithTimeout(context.Background(), 500*time.Millisecond)
+	defer done()
+	s := InitScheduler(ctx)
+	s.getGpuFn = getGpuFn
+	s.getCpuFn = getCpuFn
+	a := newScenarioRequest(t, ctx, "ollama-model-1", 10, &api.Duration{Duration: 5 * time.Millisecond})
+	b := newScenarioRequest(t, ctx, "ollama-model-1", 20, &api.Duration{Duration: 5 * time.Millisecond})
+	tmpModel := *a.req.model
+	b.req.model = &tmpModel
+	b.ggml = a.ggml
+
+	s.newServerFn = a.newServer
+	slog.Info("a")
+	s.pendingReqCh <- a.req
+	require.Len(t, s.pendingReqCh, 1)
+	s.Run(ctx)
+	select {
+	case resp := <-a.req.successCh:
+		require.Equal(t, resp.llama, a.srv)
+		require.Empty(t, s.pendingReqCh)
+		require.Empty(t, a.req.errCh)
+	case err := <-a.req.errCh:
 		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
 	}
 
 	// Trigger a reload
-	s.newServerFn = scenario2a.newServer
-	scenario2a.req.model.AdapterPaths = []string{"new"}
-	slog.Info("scenario2a")
-	s.pendingReqCh <- scenario2a.req
+	s.newServerFn = b.newServer
+	b.req.model.AdapterPaths = []string{"new"}
+	slog.Info("b")
+	s.pendingReqCh <- b.req
 	// finish first two requests, so model can reload
 	time.Sleep(1 * time.Millisecond)
-	scenario1a.ctxDone()
-	scenario1b.ctxDone()
+	a.ctxDone()
 	select {
-	case resp := <-scenario2a.req.successCh:
-		require.Equal(t, resp.llama, scenario2a.srv)
+	case resp := <-b.req.successCh:
+		require.Equal(t, resp.llama, b.srv)
 		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, scenario2a.req.errCh)
-	case err := <-scenario2a.req.errCh:
+		require.Empty(t, b.req.errCh)
+	case err := <-b.req.errCh:
 		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
 	}
+}
+
+func TestRequestsMultipleLoadedModels(t *testing.T) {
+	ctx, done := context.WithTimeout(context.Background(), 500*time.Millisecond)
+	defer done()
+	s := InitScheduler(ctx)
+	s.getGpuFn = getGpuFn
+	s.getCpuFn = getCpuFn
+
+	// Multiple loaded models
+	a := newScenarioRequest(t, ctx, "ollama-model-3a", 1*format.GigaByte, nil)
+	b := newScenarioRequest(t, ctx, "ollama-model-3b", 24*format.GigaByte, nil)
+	c := newScenarioRequest(t, ctx, "ollama-model-4a", 30, nil)
+	c.req.opts.NumGPU = 0                                       // CPU load, will be allowed
+	d := newScenarioRequest(t, ctx, "ollama-model-3c", 30, nil) // Needs prior unloaded
 
 	envconfig.MaxRunners = 1
-	s.newServerFn = scenario3a.newServer
-	slog.Info("scenario3a")
-	s.pendingReqCh <- scenario3a.req
-	// finish prior request, so new model can load
-	time.Sleep(1 * time.Millisecond)
-	scenario2a.ctxDone()
+	s.newServerFn = a.newServer
+	slog.Info("a")
+	s.pendingReqCh <- a.req
+	s.Run(ctx)
 	select {
-	case resp := <-scenario3a.req.successCh:
-		require.Equal(t, resp.llama, scenario3a.srv)
+	case resp := <-a.req.successCh:
+		require.Equal(t, resp.llama, a.srv)
 		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, scenario3a.req.errCh)
-	case err := <-scenario3a.req.errCh:
+		require.Empty(t, a.req.errCh)
+	case err := <-a.req.errCh:
 		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
@@ -262,15 +292,15 @@ func TestRequests(t *testing.T) {
 	s.loadedMu.Unlock()
 
 	envconfig.MaxRunners = 0
-	s.newServerFn = scenario3b.newServer
-	slog.Info("scenario3b")
-	s.pendingReqCh <- scenario3b.req
+	s.newServerFn = b.newServer
+	slog.Info("b")
+	s.pendingReqCh <- b.req
 	select {
-	case resp := <-scenario3b.req.successCh:
-		require.Equal(t, resp.llama, scenario3b.srv)
+	case resp := <-b.req.successCh:
+		require.Equal(t, resp.llama, b.srv)
 		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, scenario3b.req.errCh)
-	case err := <-scenario3b.req.errCh:
+		require.Empty(t, b.req.errCh)
+	case err := <-b.req.errCh:
 		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
@@ -280,15 +310,15 @@ func TestRequests(t *testing.T) {
 	s.loadedMu.Unlock()
 
 	// This is a CPU load with NumGPU = 0 so it should load
-	s.newServerFn = scenario3c.newServer
-	slog.Info("scenario3c")
-	s.pendingReqCh <- scenario3c.req
+	s.newServerFn = c.newServer
+	slog.Info("c")
+	s.pendingReqCh <- c.req
 	select {
-	case resp := <-scenario3c.req.successCh:
-		require.Equal(t, resp.llama, scenario3c.srv)
+	case resp := <-c.req.successCh:
+		require.Equal(t, resp.llama, c.srv)
 		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, scenario3c.req.errCh)
-	case err := <-scenario3c.req.errCh:
+		require.Empty(t, c.req.errCh)
+	case err := <-c.req.errCh:
 		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
@@ -298,25 +328,25 @@ func TestRequests(t *testing.T) {
 	s.loadedMu.Unlock()
 
 	// Try to load a model that wont fit
-	s.newServerFn = scenario3d.newServer
-	slog.Info("scenario3d")
+	s.newServerFn = d.newServer
+	slog.Info("d")
 	s.loadedMu.Lock()
 	require.Len(t, s.loaded, 3)
 	s.loadedMu.Unlock()
-	scenario3a.ctxDone() // Won't help since this one isn't big enough to make room
+	a.ctxDone() // Won't help since this one isn't big enough to make room
 	time.Sleep(2 * time.Millisecond)
-	s.pendingReqCh <- scenario3d.req
+	s.pendingReqCh <- d.req
 	// finish prior request, so new model can load
 	time.Sleep(6 * time.Millisecond)
 	s.loadedMu.Lock()
 	require.Len(t, s.loaded, 2)
 	s.loadedMu.Unlock()
-	scenario3b.ctxDone()
+	b.ctxDone()
 	select {
-	case resp := <-scenario3d.req.successCh:
-		require.Equal(t, resp.llama, scenario3d.srv)
+	case resp := <-d.req.successCh:
+		require.Equal(t, resp.llama, d.srv)
 		require.Empty(t, s.pendingReqCh)
-		require.Empty(t, scenario3d.req.errCh)
+		require.Empty(t, d.req.errCh)
 	case <-ctx.Done():
 		t.Fatal("timeout")
 	}
@@ -329,26 +359,19 @@ func TestGetRunner(t *testing.T) {
 	ctx, done := context.WithTimeout(context.Background(), 100*time.Millisecond)
 	defer done()
 
-	scenario1a := newScenario(t, ctx, "ollama-model-1a", 10)
-	scenario1a.req.sessionDuration = &api.Duration{Duration: 0}
-	scenario1b := newScenario(t, ctx, "ollama-model-1b", 10)
-	scenario1b.req.sessionDuration = &api.Duration{Duration: 0}
-	scenario1c := newScenario(t, ctx, "ollama-model-1c", 10)
-	scenario1c.req.sessionDuration = &api.Duration{Duration: 0}
+	a := newScenarioRequest(t, ctx, "ollama-model-1a", 10, &api.Duration{Duration: 2 * time.Millisecond})
+	b := newScenarioRequest(t, ctx, "ollama-model-1b", 10, &api.Duration{Duration: 2 * time.Millisecond})
+	c := newScenarioRequest(t, ctx, "ollama-model-1c", 10, &api.Duration{Duration: 2 * time.Millisecond})
 	envconfig.MaxQueuedRequests = 1
 	s := InitScheduler(ctx)
-	s.getGpuFn = func() gpu.GpuInfoList {
-		g := gpu.GpuInfo{Library: "metal"}
-		g.TotalMemory = 24 * format.GigaByte
-		g.FreeMemory = 12 * format.GigaByte
-		return []gpu.GpuInfo{g}
-	}
-	s.newServerFn = scenario1a.newServer
-	slog.Info("scenario1a")
-	successCh1a, errCh1a := s.GetRunner(scenario1a.ctx, scenario1a.req.model, scenario1a.req.opts, scenario1a.req.sessionDuration)
+	s.getGpuFn = getGpuFn
+	s.getCpuFn = getCpuFn
+	s.newServerFn = a.newServer
+	slog.Info("a")
+	successCh1a, errCh1a := s.GetRunner(a.ctx, a.req.model, a.req.opts, a.req.sessionDuration)
 	require.Len(t, s.pendingReqCh, 1)
-	slog.Info("scenario1b")
-	successCh1b, errCh1b := s.GetRunner(scenario1b.ctx, scenario1b.req.model, scenario1b.req.opts, scenario1b.req.sessionDuration)
+	slog.Info("b")
+	successCh1b, errCh1b := s.GetRunner(b.ctx, b.req.model, b.req.opts, b.req.sessionDuration)
 	require.Len(t, s.pendingReqCh, 1)
 	require.Empty(t, successCh1b)
 	require.Len(t, errCh1b, 1)
@@ -357,22 +380,24 @@ func TestGetRunner(t *testing.T) {
 	s.Run(ctx)
 	select {
 	case resp := <-successCh1a:
-		require.Equal(t, resp.llama, scenario1a.srv)
+		require.Equal(t, resp.llama, a.srv)
 		require.Empty(t, s.pendingReqCh)
 		require.Empty(t, errCh1a)
+	case err := <-errCh1a:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
 	}
-	scenario1a.ctxDone()
+	a.ctxDone() // Set "a" model to idle so it can unload
 	s.loadedMu.Lock()
 	require.Len(t, s.loaded, 1)
 	s.loadedMu.Unlock()
 
-	scenario1c.req.model.ModelPath = "bad path"
-	slog.Info("scenario1c")
-	successCh1c, errCh1c := s.GetRunner(scenario1c.ctx, scenario1c.req.model, scenario1c.req.opts, scenario1c.req.sessionDuration)
+	c.req.model.ModelPath = "bad path"
+	slog.Info("c")
+	successCh1c, errCh1c := s.GetRunner(c.ctx, c.req.model, c.req.opts, c.req.sessionDuration)
 	// Starts in pending channel, then should be quickly processsed to return an error
-	time.Sleep(5 * time.Millisecond)
+	time.Sleep(20 * time.Millisecond) // Long enough for the "a" model to expire and unload
 	require.Empty(t, successCh1c)
 	s.loadedMu.Lock()
 	require.Empty(t, s.loaded)
@@ -380,7 +405,7 @@ func TestGetRunner(t *testing.T) {
 	require.Len(t, errCh1c, 1)
 	err = <-errCh1c
 	require.Contains(t, err.Error(), "bad path")
-	scenario1b.ctxDone()
+	b.ctxDone()
 }
 
 // TODO - add one scenario that triggers the bogus finished event with positive ref count
@@ -389,7 +414,7 @@ func TestPrematureExpired(t *testing.T) {
 	defer done()
 
 	// Same model, same request
-	scenario1a := newScenario(t, ctx, "ollama-model-1a", 10)
+	scenario1a := newScenarioRequest(t, ctx, "ollama-model-1a", 10, nil)
 	s := InitScheduler(ctx)
 	s.getGpuFn = func() gpu.GpuInfoList {
 		g := gpu.GpuInfo{Library: "metal"}
@@ -411,6 +436,8 @@ func TestPrematureExpired(t *testing.T) {
 		s.loadedMu.Unlock()
 		slog.Info("sending premature expired event now")
 		s.expiredCh <- resp // Shouldn't happen in real life, but make sure its safe
+	case err := <-errCh1a:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
 	}
@@ -446,6 +473,8 @@ func TestUseLoadedRunner(t *testing.T) {
 	select {
 	case success := <-req.successCh:
 		require.Equal(t, r1, success)
+	case err := <-req.errCh:
+		t.Fatal(err.Error())
 	case <-ctx.Done():
 		t.Fatal("timeout")
 	}
@@ -625,8 +654,7 @@ func TestAlreadyCanceled(t *testing.T) {
 	defer done()
 	dctx, done2 := context.WithCancel(ctx)
 	done2()
-	scenario1a := newScenario(t, dctx, "ollama-model-1", 10)
-	scenario1a.req.sessionDuration = &api.Duration{Duration: 0}
+	scenario1a := newScenarioRequest(t, dctx, "ollama-model-1", 10, &api.Duration{Duration: 0})
 	s := InitScheduler(ctx)
 	slog.Info("scenario1a")
 	s.pendingReqCh <- scenario1a.req

server/testdata/tools/xlam.gotmpl

@@ -0,0 +1,45 @@
+{{- if .System }}{{ .System }}
+{{ end }}
+{{- range $i, $_ := .Messages }}
+{{- if eq .Role "user" }}### Instruction:
+{{- if and $.Tools (le (len (slice $.Messages $i)) 2) }}
+[BEGIN OF TASK INSTRUCTION]
+You are an expert in composing functions. You are given a question and a set of possible functions. 
+Based on the question, you will need to make one or more function/tool calls to achieve the purpose. 
+If none of the functions can be used, point it out and refuse to answer. 
+If the given question lacks the parameters required by the function, also point it out.
+[END OF TASK INSTRUCTION]
+
+[BEGIN OF AVAILABLE TOOLS]
+{{ $.Tools }}
+[END OF AVAILABLE TOOLS]
+
+[BEGIN OF FORMAT INSTRUCTION]
+The output MUST strictly adhere to the following JSON format, and NO other text MUST be included.
+The example format is as follows. Please make sure the parameter type is correct. If no function call is needed, please make tool_calls an empty list '[]'.
+```
+{
+    "tool_calls": [
+    {"name": "func_name1", "arguments": {"argument1": "value1", "argument2": "value2"}},
+    ... (more tool calls as required)
+    ]
+}
+```
+[END OF FORMAT INSTRUCTION]
+
+[BEGIN OF QUERY]
+{{ .Content }}
+[END OF QUERY]
+
+
+{{ else }}
+{{ .Content }}
+{{ end }}
+{{- else if .ToolCalls }}### Response:
+{"tool_calls": [{{ range .ToolCalls }}{"name": "{{ .Function.Name }}", "arguments": {{ .Function.Arguments }}}{{ end }}]}
+<|EOT|>
+{{ else if eq .Role "assistant" }}### Response:
+{{ .Content }}
+<|EOT|>
+{{ end }}
+{{- end }}### Response:

server/testdata/tools/xlam.out

@@ -0,0 +1,40 @@
+You are a knowledgable assistant. You can answer questions and perform tasks.
+### Instruction:
+What's the weather like today in Paris?
+### Response:
+{"tool_calls": [{"name": "get_current_weather", "arguments": {"format":"celsius","location":"Paris, France"}}]}
+<|EOT|>
+### Response:
+The current temperature in Paris, France is 22 degrees Celsius.
+<|EOT|>
+### Instruction:
+[BEGIN OF TASK INSTRUCTION]
+You are an expert in composing functions. You are given a question and a set of possible functions. 
+Based on the question, you will need to make one or more function/tool calls to achieve the purpose. 
+If none of the functions can be used, point it out and refuse to answer. 
+If the given question lacks the parameters required by the function, also point it out.
+[END OF TASK INSTRUCTION]
+
+[BEGIN OF AVAILABLE TOOLS]
+[{"type":"function","function":{"name":"get_current_weather","description":"Get the current weather","parameters":{"type":"object","required":["location","format"],"properties":{"format":{"type":"string","description":"The temperature unit to use. Infer this from the users location.","enum":["celsius","fahrenheit"]},"location":{"type":"string","description":"The city and state, e.g. San Francisco, CA"}}}}}]
+[END OF AVAILABLE TOOLS]
+
+[BEGIN OF FORMAT INSTRUCTION]
+The output MUST strictly adhere to the following JSON format, and NO other text MUST be included.
+The example format is as follows. Please make sure the parameter type is correct. If no function call is needed, please make tool_calls an empty list '[]'.
+```
+{
+    "tool_calls": [
+    {"name": "func_name1", "arguments": {"argument1": "value1", "argument2": "value2"}},
+    ... (more tool calls as required)
+    ]
+}
+```
+[END OF FORMAT INSTRUCTION]
+
+[BEGIN OF QUERY]
+What's the weather like today in San Francisco and Toronto?
+[END OF QUERY]
+
+
+### Response:

template/template.go

@@ -264,6 +264,7 @@ func (t *Template) Execute(w io.Writer, v Values) error {
 	nodes := deleteNode(t.Template.Root.Copy(), func(n parse.Node) bool {
 		if field, ok := n.(*parse.FieldNode); ok && slices.Contains(field.Ident, "Response") {
 			cut = true
+			return false
 		}
 
 		return cut
@@ -273,7 +274,7 @@ func (t *Template) Execute(w io.Writer, v Values) error {
 	if err := template.Must(template.New("").AddParseTree("", &tree)).Execute(&b, map[string]any{
 		"System":   system,
 		"Prompt":   prompt,
-		"Response": "",
+		"Response": response,
 	}); err != nil {
 		return err
 	}

template/template_test.go

@@ -260,6 +260,26 @@ func TestExecuteWithMessages(t *testing.T) {
 
 Hello friend![/INST] Hello human![INST] What is your name?[/INST] `,
 		},
+		{
+			"mistral assistant",
+			[]template{
+				{"no response", `[INST] {{ .Prompt }}[/INST] `},
+				{"response", `[INST] {{ .Prompt }}[/INST] {{ .Response }}`},
+				{"messages", `
+{{- range $i, $m := .Messages }}
+{{- if eq .Role "user" }}[INST] {{ .Content }}[/INST] {{ else if eq .Role "assistant" }}{{ .Content }}{{ end }}
+{{- end }}`},
+			},
+			Values{
+				Messages: []api.Message{
+					{Role: "user", Content: "Hello friend!"},
+					{Role: "assistant", Content: "Hello human!"},
+					{Role: "user", Content: "What is your name?"},
+					{Role: "assistant", Content: "My name is Ollama and I"},
+				},
+			},
+			`[INST] Hello friend![/INST] Hello human![INST] What is your name?[/INST] My name is Ollama and I`,
+		},
 		{
 			"chatml",
 			[]template{