revert to rocm 5.7

* server: fix `contet`, `load_duration` and `total_duration` fields

* Update server/routes.go

.github/workflows/release.yaml

@@ -147,7 +147,7 @@ jobs:
         run: |
           $ErrorActionPreference = "Stop"
           write-host "downloading AMD HIP Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
           write-host "Installing AMD HIP"
           Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
           write-host "Completed AMD HIP"

.github/workflows/test.yaml

@@ -169,7 +169,7 @@ jobs:
         run: |
           $ErrorActionPreference = "Stop"
           write-host "downloading AMD HIP Installer"
-          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-23.Q4-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
+          Invoke-WebRequest -Uri "https://download.amd.com/developer/eula/rocm-hub/AMD-Software-PRO-Edition-24.Q3-WinSvr2022-For-HIP.exe" -OutFile "${env:RUNNER_TEMP}\rocm-install.exe"
           write-host "Installing AMD HIP"
           Start-Process "${env:RUNNER_TEMP}\rocm-install.exe" -ArgumentList '-install' -NoNewWindow -Wait
           write-host "Completed AMD HIP"

.gitmodules

@@ -1,4 +1,4 @@
 [submodule "llama.cpp"]
 	path = llm/llama.cpp
-	url = https://github.com/ggerganov/llama.cpp.git
+	url = https://githubfast.com/ggerganov/llama.cpp.git
 	shallow = true

README.md

@@ -311,6 +311,7 @@ See the [API documentation](./docs/api.md) for all endpoints.
 - [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)
+- [Kerlig AI](https://www.kerlig.com/) (AI writing assistant for macOS)
 
 ### Terminal
 

app/ollama.iss

@@ -127,6 +127,10 @@ Type: filesandordirs; Name: "{%USERPROFILE}\.ollama\models"
 Type: filesandordirs; Name: "{%USERPROFILE}\.ollama\history"
 ; NOTE: if the user has a custom OLLAMA_MODELS it will be preserved
 
+[InstallDelete]
+Type: filesandordirs; Name: "{%TEMP}\ollama*"
+Type: filesandordirs; Name: "{%LOCALAPPDATA}\Programs\Ollama"
+
 [Messages]
 WizardReady=Ollama Windows Preview
 ReadyLabel1=%nLet's get you up and running with your own large language models.

docs/faq.md

@@ -272,4 +272,4 @@ The following server settings may be used to adjust how Ollama handles concurren
 - `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
 
-Note: Windows with Radeon GPUs currently default to 1 model maximum due to limitations in ROCm v5.7 for available VRAM reporting.  Once ROCm v6 is available, Windows Radeon will follow the defaults above.  You may enable concurrent model loads on Radeon on Windows, but ensure you don't load more models than will fit into your GPUs VRAM.
+Note: Windows with Radeon GPUs currently default to 1 model maximum due to limitations in ROCm v5.7 for available VRAM reporting.  Once ROCm v6.2 is available, Windows Radeon will follow the defaults above.  You may enable concurrent model loads on Radeon on Windows, but ensure you don't load more models than will fit into your GPUs VRAM.

gpu/amd_common.go

@@ -49,9 +49,17 @@ func rocmGetVisibleDevicesEnv(gpuInfo []GpuInfo) (string, string) {
 }
 
 func commonAMDValidateLibDir() (string, error) {
-	// We try to favor system paths first, so that we can wire up the subprocess to use
-	// the system version.  Only use our bundled version if the system version doesn't work
-	// This gives users a more recovery options if versions have subtle problems at runtime
+	// Favor our bundled version
+
+	// Installer payload location if we're running the installed binary
+	exe, err := os.Executable()
+	if err == nil {
+		rocmTargetDir := filepath.Join(filepath.Dir(exe), "rocm")
+		if rocmLibUsable(rocmTargetDir) {
+			slog.Debug("detected ROCM next to ollama executable " + rocmTargetDir)
+			return rocmTargetDir, nil
+		}
+	}
 
 	// Prefer explicit HIP env var
 	hipPath := os.Getenv("HIP_PATH")
@@ -87,14 +95,5 @@ func commonAMDValidateLibDir() (string, error) {
 		}
 	}
 
-	// Installer payload location if we're running the installed binary
-	exe, err := os.Executable()
-	if err == nil {
-		rocmTargetDir := filepath.Join(filepath.Dir(exe), "rocm")
-		if rocmLibUsable(rocmTargetDir) {
-			slog.Debug("detected ROCM next to ollama executable " + rocmTargetDir)
-			return rocmTargetDir, nil
-		}
-	}
 	return "", fmt.Errorf("no suitable rocm found, falling back to CPU")
 }

gpu/amd_hip_windows.go

@@ -84,9 +84,8 @@ func (hl *HipLib) AMDDriverVersion() (driverMajor, driverMinor int, err error) {
 	}
 
 	slog.Debug("hipDriverGetVersion", "version", version)
-	// TODO - this isn't actually right, but the docs claim hipDriverGetVersion isn't accurate anyway...
-	driverMajor = version / 1000
-	driverMinor = (version - (driverMajor * 1000)) / 10
+	driverMajor = version / 10000000
+	driverMinor = (version - (driverMajor * 10000000)) / 100000
 
 	return driverMajor, driverMinor, nil
 }

gpu/amd_windows.go

@@ -8,20 +8,21 @@ import (
 	"path/filepath"
 	"slices"
 	"strconv"
+	"strings"
 
 	"github.com/ollama/ollama/envconfig"
 	"github.com/ollama/ollama/format"
 )
 
-//const (
+const (
 
 	// TODO  We're lookinng for this exact name to detect iGPUs since hipGetDeviceProperties never reports integrated==true
-//	iGPUName = "AMD Radeon(TM) Graphics"
-//)
+	iGPUName = "AMD 2099 Graphics"
+)
 
 var (
 	// Used to validate if the given ROCm lib is usable
-	ROCmLibGlobs          = []string{"hipblas.dll", "rocblas"}                 // TODO - probably include more coverage of files here...
+	ROCmLibGlobs          = []string{"hipblas.dll", "rocblas"}                 // This is not sufficient to discern v5 vs v6
 	RocmStandardLocations = []string{"C:\\Program Files\\AMD\\ROCm\\5.7\\bin"} // TODO glob?
 )
 
@@ -34,12 +35,11 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 	}
 	defer hl.Release()
 
-	// TODO - this reports incorrect version information, so omitting for now
-	// driverMajor, driverMinor, err := hl.AMDDriverVersion()
-	// if err != nil {
-	// 	// For now this is benign, but we may eventually need to fail compatibility checks
-	// 	slog.Debug("error looking up amd driver version", "error", err)
-	// }
+	driverMajor, driverMinor, err := hl.AMDDriverVersion()
+	if err != nil {
+		// For now this is benign, but we may eventually need to fail compatibility checks
+		slog.Debug("error looking up amd driver version", "error", err)
+	}
 
 	// Note: the HIP library automatically handles subsetting to any HIP_VISIBLE_DEVICES the user specified
 	count := hl.HipGetDeviceCount()
@@ -87,10 +87,10 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 		slog.Debug("hip device", "id", i, "name", name, "gfx", gfx)
 		//slog.Info(fmt.Sprintf("[%d] Integrated: %d", i, props.iGPU)) // DOESN'T REPORT CORRECTLY!  Always 0
 		// TODO  Why isn't props.iGPU accurate!?
-		//if strings.EqualFold(name, iGPUName) {
-		//	slog.Info("unsupported Radeon iGPU detected skipping", "id", i, "name", name, "gfx", gfx)
-		//	continue
-		//}
+		if strings.EqualFold(name, iGPUName) {
+			slog.Info("unsupported Radeon iGPU detected skipping", "id", i, "name", name, "gfx", gfx)
+			continue
+		}
 		if gfxOverride == "" {
 			if !slices.Contains[[]string, string](supported, gfx) {
 				//slog.Warn("amdgpu is not supported", "gpu", i, "gpu_type", gfx, "library", libDir, "supported_types", supported)
@@ -109,10 +109,10 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 		}
 
 		// iGPU detection, remove this check once we can support an iGPU variant of the rocm library
-		//if totalMemory < IGPUMemLimit {
-		//	slog.Info("amdgpu appears to be an iGPU, skipping", "gpu", i, "total", format.HumanBytes2(totalMemory))
-		//	continue
-		//}
+		if totalMemory < IGPUMemLimit {
+			slog.Info("amdgpu appears to be an iGPU, skipping", "gpu", i, "total", format.HumanBytes2(totalMemory))
+			continue
+		}
 
 		slog.Debug("amdgpu memory", "gpu", i, "total", format.HumanBytes2(totalMemory))
 		slog.Debug("amdgpu memory", "gpu", i, "available", format.HumanBytes2(freeMemory))
@@ -131,10 +131,8 @@ func AMDGetGPUInfo() []RocmGPUInfo {
 				MinimumMemory:  rocmMinimumMemory,
 				Name:           name,
 				Compute:        gfx,
-
-				// TODO - this information isn't accurate on windows, so don't report it until we find the right way to retrieve
-				// DriverMajor:    driverMajor,
-				// DriverMinor:    driverMinor,
+				DriverMajor:    driverMajor,
+				DriverMinor:    driverMinor,
 			},
 			index: i,
 		}

gpu/gpu.go

@@ -274,6 +274,28 @@ func GetGPUInfo() GpuInfoList {
 				gpuInfo.DriverMajor = driverMajor
 				gpuInfo.DriverMinor = driverMinor
 
+				// query the management library as well so we can record any skew between the two
+				// which represents overhead on the GPU we must set aside on subsequent updates
+				if cHandles.nvml != nil {
+					C.nvml_get_free(*cHandles.nvml, C.int(gpuInfo.index), &memInfo.free, &memInfo.total, &memInfo.used)
+					if memInfo.err != nil {
+						slog.Warn("error looking up nvidia GPU memory", "error", C.GoString(memInfo.err))
+						C.free(unsafe.Pointer(memInfo.err))
+					} else {
+						if memInfo.free != 0 && uint64(memInfo.free) > gpuInfo.FreeMemory {
+							gpuInfo.OSOverhead = uint64(memInfo.free) - gpuInfo.FreeMemory
+							slog.Info("detected OS VRAM overhead",
+								"id", gpuInfo.ID,
+								"library", gpuInfo.Library,
+								"compute", gpuInfo.Compute,
+								"driver", fmt.Sprintf("%d.%d", gpuInfo.DriverMajor, gpuInfo.DriverMinor),
+								"name", gpuInfo.Name,
+								"overhead", format.HumanBytes2(gpuInfo.OSOverhead),
+							)
+						}
+					}
+				}
+
 				// TODO potentially sort on our own algorithm instead of what the underlying GPU library does...
 				cudaGPUs = append(cudaGPUs, gpuInfo)
 			}
@@ -338,14 +360,17 @@ func GetGPUInfo() GpuInfoList {
 					"before",
 					"total", format.HumanBytes2(cpus[0].TotalMemory),
 					"free", format.HumanBytes2(cpus[0].FreeMemory),
+					"free_swap", format.HumanBytes2(cpus[0].FreeSwap),
 				),
 				slog.Group(
 					"now",
 					"total", format.HumanBytes2(mem.TotalMemory),
 					"free", format.HumanBytes2(mem.FreeMemory),
+					"free_swap", format.HumanBytes2(mem.FreeSwap),
 				),
 			)
 			cpus[0].FreeMemory = mem.FreeMemory
+			cpus[0].FreeSwap = mem.FreeSwap
 		}
 
 		var memInfo C.mem_info_t
@@ -374,9 +399,14 @@ func GetGPUInfo() GpuInfoList {
 				slog.Warn("error looking up nvidia GPU memory")
 				continue
 			}
+			if cHandles.nvml != nil && gpu.OSOverhead > 0 {
+				// When using the management library update based on recorded overhead
+				memInfo.free -= C.uint64_t(gpu.OSOverhead)
+			}
 			slog.Debug("updating cuda memory data",
 				"gpu", gpu.ID,
 				"name", gpu.Name,
+				"overhead", format.HumanBytes2(gpu.OSOverhead),
 				slog.Group(
 					"before",
 					"total", format.HumanBytes2(gpu.TotalMemory),

gpu/gpu_darwin.go

@@ -57,6 +57,7 @@ func GetCPUMem() (memInfo, error) {
 	return memInfo{
 		TotalMemory: uint64(C.getPhysicalMemory()),
 		FreeMemory:  uint64(C.getFreeMemory()),
+		// FreeSwap omitted as Darwin uses dynamic paging
 	}, nil
 }
 

gpu/gpu_linux.go

@@ -50,7 +50,7 @@ var OneapiMgmtName = "libze_intel_gpu.so"
 
 func GetCPUMem() (memInfo, error) {
 	var mem memInfo
-	var total, available, free, buffers, cached uint64
+	var total, available, free, buffers, cached, freeSwap uint64
 	f, err := os.Open("/proc/meminfo")
 	if err != nil {
 		return mem, err
@@ -70,20 +70,21 @@ func GetCPUMem() (memInfo, error) {
 			_, err = fmt.Sscanf(line, "Buffers:%d", &buffers)
 		case strings.HasPrefix(line, "Cached:"):
 			_, err = fmt.Sscanf(line, "Cached:%d", &cached)
+		case strings.HasPrefix(line, "SwapFree:"):
+			_, err = fmt.Sscanf(line, "SwapFree:%d", &freeSwap)
 		default:
 			continue
 		}
 		if err != nil {
 			return mem, err
 		}
-
-		if total > 0 && available > 0 {
-			mem.TotalMemory = total * format.KibiByte
-			mem.FreeMemory = available * format.KibiByte
-			return mem, nil
-		}
 	}
 	mem.TotalMemory = total * format.KibiByte
-	mem.FreeMemory = (free + buffers + cached) * format.KibiByte
+	mem.FreeSwap = freeSwap * format.KibiByte
+	if available > 0 {
+		mem.FreeMemory = available * format.KibiByte
+	} else {
+		mem.FreeMemory = (free + buffers + cached) * format.KibiByte
+	}
 	return mem, nil
 }

gpu/gpu_windows.go

@@ -51,5 +51,5 @@ func GetCPUMem() (memInfo, error) {
 	if r1 == 0 {
 		return memInfo{}, fmt.Errorf("GlobalMemoryStatusEx failed: %w", err)
 	}
-	return memInfo{TotalMemory: memStatus.TotalPhys, FreeMemory: memStatus.AvailPhys}, nil
+	return memInfo{TotalMemory: memStatus.TotalPhys, FreeMemory: memStatus.AvailPhys, FreeSwap: memStatus.AvailPageFile}, nil
 }

gpu/types.go

@@ -10,6 +10,7 @@ import (
 type memInfo struct {
 	TotalMemory uint64 `json:"total_memory,omitempty"`
 	FreeMemory  uint64 `json:"free_memory,omitempty"`
+	FreeSwap    uint64 `json:"free_swap,omitempty"`
 }
 
 // Beginning of an `ollama info` command
@@ -52,7 +53,8 @@ type CPUInfo struct {
 
 type CudaGPUInfo struct {
 	GpuInfo
-	index int //nolint:unused,nolintlint
+	OSOverhead uint64 // Memory overhead between the driver library and management library
+	index      int    //nolint:unused,nolintlint
 }
 type CudaGPUInfoList []CudaGPUInfo
 

llm/generate/gen_linux.sh

@@ -187,7 +187,7 @@ if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
         CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
         echo "Building custom CUDA GPU"
     else
-        CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_FLAGS=-t8 -DGGML_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} -DCMAKE_LIBRARY_PATH=/usr/local/cuda/compat"
+        CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_FLAGS=-t8 -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
     fi
     CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS}"
     BUILD_DIR="../build/linux/${ARCH}/cuda${CUDA_VARIANT}"

llm/generate/gen_windows.ps1

@@ -6,12 +6,11 @@ function amdGPUs {
     if ($env:AMDGPU_TARGETS) {
         return $env:AMDGPU_TARGETS
     }
-    # TODO - load from some common data file for linux + windows build consistency
+    # Current supported rocblas list from ROCm v6.1.2 on windows
     $GPU_LIST = @(
         "gfx803"
         "gfx900"
         "gfx902"
-        "gfx904"
         "gfx90c:xnack-"
         "gfx906:xnack-"
         "gfx908:xnack-"
@@ -407,7 +406,6 @@ function build_rocm() {
         sign
         install
 
-        # Assumes v5.7, may need adjustments for v6
         rm -ea 0 -recurse -force -path "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"
         md "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\rocblas\library\" -ea 0 > $null
         cp "${env:HIP_PATH}\bin\hipblas.dll" "${script:SRC_DIR}\dist\windows-${script:ARCH}\rocm\"

llm/ggml.go

@@ -424,6 +424,32 @@ func (llm GGML) GraphSize(context, batch uint64) (partialOffload, fullOffload ui
 			4*batch*(3*embedding+vocab)+embedding*vocab*105/128,
 			4*batch*(2*embedding+1+2*embeddingHeadsK*headsKV+context+context*headsKV)+4*embeddingHeadsK*context*headsKV+embedding*embeddingHeadsK*headsKV*9/16,
 		)
+	case "chatglm":
+		fullOffload = 4 * batch * (embedding + vocab)
+		partialOffload = 4*batch*(embedding+vocab) + embedding*vocab*105/128
+		if qkvBias, ok := layers["blk.0"]["attn_qkv.bias"]; ok {
+			fullOffload = max(
+				fullOffload,
+				4*batch*(2+
+					2*embedding+
+					context+
+					context*heads+
+					embeddingHeadsK*heads+
+					qkvBias.Shape[0]),
+			)
+
+			partialOffload = max(
+				partialOffload,
+				4*batch*(1+
+					2*embedding+
+					embeddingHeadsK*heads+
+					context+
+					context*heads)+
+					4*embeddingHeadsK*context+
+					4*context*embeddingHeadsK+
+					4*qkvBias.Shape[0],
+			)
+		}
 	}
 
 	return

llm/llm.go

@@ -33,7 +33,7 @@ func Quantize(infile, outfile string, ftype fileType) error {
 	params.ftype = ftype.Value()
 
 	if rc := C.llama_model_quantize(cinfile, coutfile, &params); rc != 0 {
-		return fmt.Errorf("llama_model_quantize: %d", rc)
+		return fmt.Errorf("failed to quantize model. This model architecture may not be supported, or you may need to upgrade Ollama to the latest version")
 	}
 
 	return nil

llm/server.go

@@ -88,6 +88,7 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 	var estimate MemoryEstimate
 	var systemTotalMemory uint64
 	var systemFreeMemory uint64
+	var systemSwapFreeMemory uint64
 
 	systemMemInfo, err := gpu.GetCPUMem()
 	if err != nil {
@@ -95,7 +96,8 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 	} else {
 		systemTotalMemory = systemMemInfo.TotalMemory
 		systemFreeMemory = systemMemInfo.FreeMemory
-		slog.Debug("system memory", "total", format.HumanBytes2(systemTotalMemory), "free", systemFreeMemory)
+		systemSwapFreeMemory = systemMemInfo.FreeSwap
+		slog.Debug("system memory", "total", format.HumanBytes2(systemTotalMemory), "free", format.HumanBytes2(systemFreeMemory), "free_swap", format.HumanBytes2(systemSwapFreeMemory))
 	}
 
 	// If the user wants zero GPU layers, reset the gpu list to be CPU/system ram info
@@ -122,6 +124,16 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		}
 	}
 
+	// On linux, over-allocating CPU memory will almost always result in an error
+	if runtime.GOOS == "linux" {
+		systemMemoryRequired := estimate.TotalSize - estimate.VRAMSize
+		available := systemFreeMemory + systemSwapFreeMemory
+		if systemMemoryRequired > available {
+			slog.Warn("model request too large for system", "requested", format.HumanBytes2(systemMemoryRequired), "available", available, "total", format.HumanBytes2(systemTotalMemory), "free", format.HumanBytes2(systemFreeMemory), "swap", format.HumanBytes2(systemSwapFreeMemory))
+			return nil, fmt.Errorf("model requires more system memory (%s) than is available (%s)", format.HumanBytes2(systemMemoryRequired), format.HumanBytes2(available))
+		}
+	}
+
 	estimate.log()
 
 	// Loop through potential servers
@@ -254,10 +266,6 @@ func NewLlamaServer(gpus gpu.GpuInfoList, model string, ggml *GGML, adapters, pr
 		params = append(params, "--tensor-split", estimate.TensorSplit)
 	}
 
-	if estimate.TensorSplit != "" {
-		params = append(params, "--tensor-split", estimate.TensorSplit)
-	}
-
 	for i := range len(servers) {
 		dir := availableServers[servers[i]]
 		if dir == "" {

scripts/build_windows.ps1

@@ -107,9 +107,12 @@ function gatherDependencies() {
 
     # TODO - this varies based on host build system and MSVC version - drive from dumpbin output
     # currently works for Win11 + MSVC 2019 + Cuda V11
-    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\msvcp140.dll" "${script:DEPS_DIR}\ollama_runners\"
+    cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\msvcp140*.dll" "${script:DEPS_DIR}\ollama_runners\"
     cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\vcruntime140.dll" "${script:DEPS_DIR}\ollama_runners\"
     cp "${env:VCToolsRedistDir}\x64\Microsoft.VC*.CRT\vcruntime140_1.dll" "${script:DEPS_DIR}\ollama_runners\"
+    foreach ($part in $("runtime", "stdio", "filesystem", "math", "convert", "heap", "string", "time", "locale", "environment")) {
+        cp "$env:VCToolsRedistDir\..\..\..\Tools\Llvm\x64\bin\api-ms-win-crt-${part}*.dll" "${script:DEPS_DIR}\ollama_runners\"
+    }
 
 
     cp "${script:SRC_DIR}\app\ollama_welcome.ps1" "${script:SRC_DIR}\dist\"

server/prompt.go

@@ -4,7 +4,6 @@ import (
 	"bytes"
 	"context"
 	"log/slog"
-	"slices"
 
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/llm"
@@ -17,26 +16,18 @@ type tokenizeFunc func(context.Context, string) ([]int, error)
 // chatPrompt truncates any messages that exceed the context window of the model, making sure to always include 1) the
 // latest message and 2) system messages
 func chatPrompt(ctx context.Context, m *Model, tokenize tokenizeFunc, opts *api.Options, msgs []api.Message) (prompt string, images []llm.ImageData, _ error) {
-	// pull out any system messages which should always be included in the prompt
 	var system []api.Message
-	msgs = slices.DeleteFunc(msgs, func(m api.Message) bool {
-		if m.Role == "system" {
-			system = append(system, m)
-			return true
-		}
-
-		return false
-	})
-
-	if len(system) == 0 && m.System != "" {
-		// add model system prompt since it wasn't provided
-		system = append(system, api.Message{Role: "system", Content: m.System})
-	}
-
 	// always include the last message
 	n := len(msgs) - 1
 	// in reverse, find all messages that fit into context window
 	for i := n - 1; i >= 0; i-- {
+		system = make([]api.Message, 0)
+		for j := range i {
+			if msgs[j].Role == "system" {
+				system = append(system, msgs[j])
+			}
+		}
+
 		var b bytes.Buffer
 		if err := m.Template.Execute(&b, template.Values{Messages: append(system, msgs[i:]...)}); err != nil {
 			return "", nil, err

server/prompt_test.go

@@ -6,6 +6,7 @@ import (
 	"strings"
 	"testing"
 
+	"github.com/google/go-cmp/cmp"
 	"github.com/ollama/ollama/api"
 	"github.com/ollama/ollama/template"
 )
@@ -160,6 +161,19 @@ func TestChatPrompt(t *testing.T) {
 				{Role: "assistant", Content: "I-I'm a what?"},
 				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
 			},
+			expect: expect{
+				prompt: "You are the Test Who Lived. You're a test, Harry! I-I'm a what? A test. And a thumping good one at that, I'd wager. ",
+			},
+		},
+		{
+			name:  "out of order system",
+			limit: 2048,
+			msgs: []api.Message{
+				{Role: "user", Content: "You're a test, Harry!"},
+				{Role: "assistant", Content: "I-I'm a what?"},
+				{Role: "system", Content: "You are the Test Who Lived."},
+				{Role: "user", Content: "A test. And a thumping good one at that, I'd wager."},
+			},
 			expect: expect{
 				prompt: "You're a test, Harry! I-I'm a what? You are the Test Who Lived. A test. And a thumping good one at that, I'd wager. ",
 			},
@@ -187,6 +201,10 @@ func TestChatPrompt(t *testing.T) {
 				t.Errorf("expected %q, got %q", tt.prompt, prompt)
 			}
 
+			if diff := cmp.Diff(prompt, tt.prompt); diff != "" {
+				t.Errorf("mismatch (-got +want):\n%s", diff)
+			}
+
 			if len(images) != len(tt.images) {
 				t.Fatalf("expected %d images, got %d", len(tt.images), len(images))
 			}

server/routes.go

@@ -102,6 +102,7 @@ func (s *Server) scheduleRunner(ctx context.Context, name string, caps []Capabil
 }
 
 func (s *Server) GenerateHandler(c *gin.Context) {
+	checkpointStart := time.Now()
 	var req api.GenerateRequest
 	if err := c.ShouldBindJSON(&req); errors.Is(err, io.EOF) {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
@@ -129,6 +130,8 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 		return
 	}
 
+	checkpointLoaded := time.Now()
+
 	if req.Prompt == "" {
 		c.JSON(http.StatusOK, api.GenerateResponse{
 			Model:      req.Model,
@@ -191,26 +194,48 @@ func (s *Server) GenerateHandler(c *gin.Context) {
 
 	ch := make(chan any)
 	go func() {
+		// TODO (jmorganca): avoid building the response twice both here and below
+		var sb strings.Builder
 		defer close(ch)
 		if err := r.Completion(c.Request.Context(), llm.CompletionRequest{
 			Prompt:  prompt,
 			Images:  images,
 			Format:  req.Format,
 			Options: opts,
-		}, func(r llm.CompletionResponse) {
-			ch <- api.GenerateResponse{
+		}, func(cr llm.CompletionResponse) {
+			res := api.GenerateResponse{
 				Model:      req.Model,
 				CreatedAt:  time.Now().UTC(),
-				Response:   r.Content,
-				Done:       r.Done,
-				DoneReason: r.DoneReason,
+				Response:   cr.Content,
+				Done:       cr.Done,
+				DoneReason: cr.DoneReason,
 				Metrics: api.Metrics{
-					PromptEvalCount:    r.PromptEvalCount,
-					PromptEvalDuration: r.PromptEvalDuration,
-					EvalCount:          r.EvalCount,
-					EvalDuration:       r.EvalDuration,
+					PromptEvalCount:    cr.PromptEvalCount,
+					PromptEvalDuration: cr.PromptEvalDuration,
+					EvalCount:          cr.EvalCount,
+					EvalDuration:       cr.EvalDuration,
 				},
 			}
+
+			if _, err := sb.WriteString(cr.Content); err != nil {
+				ch <- gin.H{"error": err.Error()}
+			}
+
+			if cr.Done {
+				res.TotalDuration = time.Since(checkpointStart)
+				res.LoadDuration = checkpointLoaded.Sub(checkpointStart)
+
+				if !req.Raw {
+					tokens, err := r.Tokenize(c.Request.Context(), prompt+sb.String())
+					if err != nil {
+						ch <- gin.H{"error": err.Error()}
+						return
+					}
+					res.Context = append(req.Context, tokens...)
+				}
+			}
+
+			ch <- res
 		}); err != nil {
 			ch <- gin.H{"error": err.Error()}
 		}
@@ -1122,6 +1147,8 @@ func (s *Server) ProcessHandler(c *gin.Context) {
 }
 
 func (s *Server) ChatHandler(c *gin.Context) {
+	checkpointStart := time.Now()
+
 	var req api.ChatRequest
 	if err := c.ShouldBindJSON(&req); errors.Is(err, io.EOF) {
 		c.AbortWithStatusJSON(http.StatusBadRequest, gin.H{"error": "missing request body"})
@@ -1141,6 +1168,8 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		return
 	}
 
+	checkpointLoaded := time.Now()
+
 	if len(req.Messages) == 0 {
 		c.JSON(http.StatusOK, api.ChatResponse{
 			Model:      req.Model,
@@ -1152,6 +1181,10 @@ func (s *Server) ChatHandler(c *gin.Context) {
 		return
 	}
 
+	if req.Messages[0].Role != "system" {
+		req.Messages = append([]api.Message{{Role: "system", Content: m.System}}, req.Messages...)
+	}
+
 	prompt, images, err := chatPrompt(c.Request.Context(), m, r.Tokenize, opts, req.Messages)
 	if err != nil {
 		c.JSON(http.StatusInternalServerError, gin.H{"error": err.Error()})
@@ -1169,7 +1202,7 @@ func (s *Server) ChatHandler(c *gin.Context) {
 			Format:  req.Format,
 			Options: opts,
 		}, func(r llm.CompletionResponse) {
-			ch <- api.ChatResponse{
+			res := api.ChatResponse{
 				Model:      req.Model,
 				CreatedAt:  time.Now().UTC(),
 				Message:    api.Message{Role: "assistant", Content: r.Content},
@@ -1182,6 +1215,13 @@ func (s *Server) ChatHandler(c *gin.Context) {
 					EvalDuration:       r.EvalDuration,
 				},
 			}
+
+			if r.Done {
+				res.TotalDuration = time.Since(checkpointStart)
+				res.LoadDuration = checkpointLoaded.Sub(checkpointStart)
+			}
+
+			ch <- res
 		}); err != nil {
 			ch <- gin.H{"error": err.Error()}
 		}

server/routes_create_test.go

@@ -546,8 +546,8 @@ func TestCreateDetectTemplate(t *testing.T) {
 
 		checkFileExists(t, filepath.Join(p, "blobs", "*"), []string{
 			filepath.Join(p, "blobs", "sha256-553c4a3f747b3d22a4946875f1cc8ed011c2930d83f864a0c7265f9ec0a20413"),
-			filepath.Join(p, "blobs", "sha256-9512c372dfc7d84d6065b8dd2b601aeed8cc1a78e7a7aa784a42fff37f5524b7"),
-			filepath.Join(p, "blobs", "sha256-b8b78cb8c6eefd14c06f1af042e6161255bf87bbf2dd14fce57cdac893db8139"),
+			filepath.Join(p, "blobs", "sha256-c608dc615584cd20d9d830363dabf8a4783ae5d34245c3d8c115edb3bc7b28e4"),
+			filepath.Join(p, "blobs", "sha256-f836ee110db21567f826332e4cedd746c06d10664fd5a9ea3659e3683a944510"),
 		})
 	})
 

server/sched.go

@@ -135,11 +135,6 @@ func (s *Scheduler) processPending(ctx context.Context) {
 			}
 
 			for {
-				cpus := s.getCpuFn()
-				var systemMem gpu.GpuInfo
-				if len(cpus) > 0 {
-					systemMem = cpus[0]
-				}
 				var runnerToExpire *runnerRef
 				s.loadedMu.Lock()
 				runner := s.loaded[pending.model.ModelPath]
@@ -193,38 +188,6 @@ func (s *Scheduler) processPending(ctx context.Context) {
 						break
 					}
 
-					estimate := llm.EstimateGPULayers(gpus, ggml, pending.model.ProjectorPaths, pending.opts)
-					maxSize := systemMem.FreeMemory
-
-					// Add available GPU memory to the total pool
-					// macOS hardware has unified memory so don't double count
-					if runtime.GOOS != "darwin" {
-						for _, gpu := range gpus {
-							if gpu.Library == "cpu" {
-								continue
-							}
-							if loadedCount == 0 {
-								// If no other models are loaded, set the limit based on what's available
-								maxSize += gpu.FreeMemory
-							} else {
-								// Other models could be unloaded, favor total memory for limit
-								maxSize += gpu.TotalMemory
-							}
-						}
-					}
-
-					// Block attempting to load a model larger than system memory + GPU memory
-					if estimate.TotalSize > maxSize {
-						slog.Warn("model request too large for system", "requested", format.HumanBytes2(estimate.TotalSize), "system", format.HumanBytes2(maxSize))
-
-						// Linux will crash if over-allocating memory - return an error to the user.
-						// TODO (jmorganca): add reasonable upper limits for darwin and windows as well
-						if runtime.GOOS == "linux" {
-							pending.errCh <- fmt.Errorf("requested model (%s) is too large for this system (%s)", format.HumanBytes2(estimate.TotalSize), format.HumanBytes2(maxSize))
-							break
-						}
-					}
-
 					// 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

template/alfred.gotmpl

@@ -1,8 +1 @@
-{{- if .Messages }}
-{{- if .System }}<start_system>{{ .System }}<end_message>
-{{- end }}
-{{- range .Messages }}<start_{{ .Role }}>{{ .Content }}<end_message>
-{{- end }}<start_assistant>
-{{- else }}
 {{ if .System }}<start_system>{{ .System }}<end_message>{{ end }}{{ if .Prompt }}<start_user>{{ .Prompt }}<end_message>{{ end }}<start_assistant>{{ .Response }}<end_message>
-{{- end }}

template/alpaca.gotmpl

@@ -1,14 +1,3 @@
-{{- if .Messages }}
-{{- if .System }}{{ .System }}
-{{- end }}
-{{- range .Messages }}
-{{- if eq .Role "user" }}### Instruction:
-{{- else if eq .Role "assistant" }}### Response:
-{{- end }}
-{{ .Content }}
-
-{{ end }}### Response:
-{{ else }}
 {{ if .System }}{{ .System }}
 
 {{ end }}{{ if .Prompt }}### Instruction:
@@ -16,4 +5,4 @@
 
 {{ end }}### Response:
 {{ .Response }}
-{{- end }}
+

template/chatml.gotmpl

@@ -1,15 +1,6 @@
-{{- if .Messages }}
-{{- if .System }}<|im_start|>system
-{{ .System }}<|im_end|>
-{{ end }}
-{{- range .Messages }}<|im_start|>{{ .Role }}
-{{ .Content }}<|im_end|>
-{{ end }}<|im_start|>assistant
-{{ else }}
 {{ if .System }}<|im_start|>system
 {{ .System }}<|im_end|>
 {{ end }}{{ if .Prompt }}<|im_start|>user
 {{ .Prompt }}<|im_end|>
 {{ end }}<|im_start|>assistant
 {{ .Response }}<|im_end|>
-{{- end }}

template/chatqa.gotmpl

@@ -1,17 +1,6 @@
-{{- if .Messages }}
-{{- if .System }}System: {{ .System }}
-
-{{ end }}
-{{- range .Messages }}
-{{- if eq .Role "user" }}User:
-{{- else if eq .Role "assistant" }}Assistant:
-{{- end }} {{ .Content }}
-
-{{ end }}Assistant:
-{{- else }}
 {{ if .System }}System: {{ .System }}
 
 {{ end }}{{ if .Prompt }}User: {{ .Prompt }}
 
-{{ end }}Assistant: <|begin_of_text|>{{ .Response }}
-{{- end }}
+{{ end }}Assistant: {{ .Response }}
+

template/codellama-70b-instruct.gotmpl

@@ -1,19 +1,10 @@
-{{- if .Messages }}
-{{- if .System }}Source: system
+{{ if .System }}Source: system
 
- {{ .System }} <step> {{ end }}
-{{- range .Messages }}Source: {{ .Role }}
-
- {{ .Content }} <step> {{ end }}Source: assistant
-Destination: user
-
-{{ else }}
-{{ if .System }} Source: system
-
- {{ .System }} <step>{{ end }} Source: user
+ {{ .System }} <step> {{ end }}Source: user
 
  {{ .Prompt }} <step> Source: assistant
+{{- if not .Response }}
 Destination: user
-
- {{ .Response }}<step>
 {{- end }}
+
+ {{ .Response }} <step> 

template/falcon-instruct.gotmpl

@@ -1,13 +1,5 @@
-{{- if .Messages }}
-{{- if .System }}System: {{ .System }}
-{{ end }}
-{{- range .Messages }}
-{{- if eq .Role "user" }}User:
-{{ else if eq .Role "assistant" }}Falcon:
-{{ end }}{{ .Content }}
+{{ if .System }}System: {{ .System }}
+{{ end }}{{ if .Prompt }}User:
+{{ .Prompt }}
 {{ end }}Falcon:
-{{ else }}
-{{ if .System }}{{ .System }}
-{{ end }}{{ if .Prompt }}User: {{ .Prompt }}
-{{ end }}Assistant: {{ .Response }}
-{{- end }}
+{{ .Response }}

template/gemma-instruct.gotmpl

@@ -1,16 +1,5 @@
-{{- if .Messages }}
-{{- range $index, $_ := .Messages }}<start_of_turn>
-{{- if eq .Role "user" }}user
-{{- if and $.System (eq $index 0) }}
-{{ $.System }}
-{{- end }}
-{{- else if eq .Role "assistant" }}model
-{{- end }}
-{{ .Content }}<end_of_turn>
-{{ end }}<start_of_turn>model
-{{ else }}
 <start_of_turn>user
-{{ if .System }}{{ .System }} {{ end }}{{ .Prompt }}<end_of_turn>
+{{ if .System }}{{ .System }}
+{{ end }}{{ .Prompt }}<end_of_turn>
 <start_of_turn>model
 {{ .Response }}<end_of_turn>
-{{- end }}

template/granite-instruct.gotmpl

@@ -1,18 +1,4 @@
-{{- if .Messages }}
-{{- if .System }}System:
-{{ .System }}
-
-{{ end }}
-{{- range .Messages }}
-{{- if eq .Role "user" }}Question:
-{{- else if eq .Role "assistant" }}Answer:
-{{- end }}
-{{ .Content }}
-
-{{ end }}Answer:
-{{ else }}
-{{ if .System }}
-System:
+{{ if .System }}System:
 {{ .System }}
 
 {{ end }}{{ if .Prompt }}Question:
@@ -20,4 +6,4 @@ System:
 
 {{ end }}Answer:
 {{ .Response }}
-{{- end }}
+

template/llama2-chat.gotmpl

@@ -1,16 +1,6 @@
-{{- if .Messages }}
-{{- range $index, $_ := .Messages }}
-{{- if eq .Role "user" }}[INST] {{ if eq $index 0 }}<<SYS>>
-{{- if $.System }}
-{{ $.System }}
+[INST] <<SYS>>
+{{- if .System }}
+{{ .System }}
 {{ end }}<</SYS>>
 
-{{ end }}{{ .Content }}
-{{- else }} [/INST] {{ .Content }}</s><s>
-{{- end }}
-{{- end }} [/INST]
-{{- else }}
-[INST] <<SYS>>{{ .System }}<</SYS>>
-
-{{ .Prompt }} [/INST] {{ .Response }}
-{{- end }}
+{{ .Prompt }} [/INST] {{ .Response }}</s><s>

template/llama3-instruct.gotmpl

@@ -1,14 +1,3 @@
-{{- if .Messages }}
-{{- if .System }}<|start_header_id|>system<|end_header_id|>
-
-{{ .System }}<|eot_id|>
-{{- end }}
-{{- range .Messages }}<|start_header_id|>{{ .Role }}<|end_header_id|>
-
-{{ .Content }}<|eot_id|>
-{{- end }}<|start_header_id|>assistant<|end_header_id|>
-
-{{ else }}
 {{ if .System }}<|start_header_id|>system<|end_header_id|>
 
 {{ .System }}<|eot_id|>{{ end }}{{ if .Prompt }}<|start_header_id|>user<|end_header_id|>
@@ -16,4 +5,3 @@
 {{ .Prompt }}<|eot_id|>{{ end }}<|start_header_id|>assistant<|end_header_id|>
 
 {{ .Response }}<|eot_id|>
-{{- end }}

template/magicoder.gotmpl

@@ -1,15 +1,3 @@
-{{- if .Messages }}
-{{- if .System }}{{ .System }}
-
-{{ end }}
-{{- range .Messages }}
-{{- if eq .Role "user" }}@@ Instruction
-{{- else if eq .Role "assistant" }}@@ Response
-{{- end }}
-{{ .Content }}
-
-{{ end }}@@ Response
-{{ else }}
 {{ if .System }}{{ .System }}
 
 {{ end }}{{ if .Prompt }}@@ Instruction
@@ -17,4 +5,4 @@
 
 {{ end }}@@ Response
 {{ .Response }}
-{{- end }}
+

template/mistral-instruct.gotmpl

@@ -1,9 +1,3 @@
-{{- if .Messages }}
-{{- range $index, $_ := .Messages }}
-{{- if eq .Role "user" }}[INST] {{ if and $.System (eq (len (slice $.Messages $index)) 1) }}{{ $.System }}
-{{ end }}{{ .Content }}
-{{- else if eq .Role "assistant" }}[/INST] {{ .Content }}</s>
-{{- end }}
-{{- end }}[/INST]
-{{- else }}[INST] {{ if .System }}{{ .System }} {{ end }}{{ .Prompt }} [/INST] {{ .Response }}
-{{- end }}
+[INST] {{ if .System }}{{ .System }}
+
+{{ end }}{{ .Prompt }}[/INST] {{ .Response }}</s>

template/openchat.gotmpl

@@ -1,11 +1 @@
-{{- if .Messages }}
-{{- if .System }}GPT Correct System: {{ .System }}<|end_of_turn|>
-{{- end }}
-{{- range .Messages }}GPT Correct
-{{- if eq .Role "user" }} User:
-{{- else if eq .Role "assistant" }} Assistant:
-{{- end }} {{ .Content }}<|end_of_turn|>
-{{- end }}GPT Correct Assistant:
-{{- else }}
-{{ .System }}<|end_of_turn|>GPT4 Correct User: {{ .Prompt }}<|end_of_turn|>GPT4 Correct Assistant: {{ .Response }}<|end_of_turn|>
-{{- end }}
+{{ if .System }}GPT4 Correct System: {{ .System }}<|end_of_turn|>{{ end }}GPT4 Correct User: {{ .Prompt }}<|end_of_turn|>GPT4 Correct Assistant: {{ .Response }}<|end_of_turn|>

template/phi-3.gotmpl

@@ -1,15 +1,6 @@
-{{- if .Messages }}
-{{- if .System }}<|system|>
-{{ .System }}<|end|>
-{{ end }}
-{{- range .Messages }}<|{{ .Role }}|>
-{{ .Content }}<|end|>
-{{ end }}<|assistant|>
-{{ else }}
 {{ if .System }}<|system|>
 {{ .System }}<|end|>
 {{ end }}{{ if .Prompt }}<|user|>
 {{ .Prompt }}<|end|>
 {{ end }}<|assistant|>
 {{ .Response }}<|end|>
-{{- end }}

template/solar-instruct.gotmpl

@@ -1,16 +1,3 @@
-{{- if .Messages }}
-{{- if .System }}### System:
-{{ .System }}
-
-{{ end }}
-{{- range .Messages }}
-{{- if eq .Role "user" }}### User:
-{{ .Content }}
-{{ else if eq .Role "assistant" }}### Assistant:
-{{ .Content }}</s>
-{{ end }}
-{{ end }}### Assistant:
-{{ else }}
 {{ if .System }}### System:
 {{ .System }}
 
@@ -18,5 +5,5 @@
 {{ .Prompt }}
 
 {{ end }}### Assistant:
-{{ .Response }}
-{{- end }}
+{{ .Response }}</s>
+

template/starcoder2-instruct.gotmpl

@@ -1,24 +1,8 @@
-{{- if .Messages }}
-{{- if .System }}{{ .System }}
-
-{{ end }}
-{{- range .Messages }}
-{{- if eq .Role "user" }}### Instruction
-{{ .Content }}
-
-{{ else if eq .Role "assistant" }}### Response
-{{ .Content }}<|endoftext|>
-
-{{ end }}
-{{- end }}### Response
-{{ else }}
 {{ if .System }}{{ .System }}
 
 {{ end }}{{ if .Prompt }}### Instruction
 {{ .Prompt }}
 
-
 {{ end }}### Response
 {{ .Response }}<|endoftext|>
 
-{{- end }}

template/template.go

@@ -143,52 +143,74 @@ func (t *Template) Vars() []string {
 
 type Values struct {
 	Messages []api.Message
+
+	// forceLegacy is a flag used to test compatibility with legacy templates
+	forceLegacy bool
 }
 
 func (t *Template) Execute(w io.Writer, v Values) error {
-	system, collated := collate(v.Messages)
-	if slices.Contains(t.Vars(), "messages") {
+	system, messages := collate(v.Messages)
+	if !v.forceLegacy && slices.Contains(t.Vars(), "messages") {
 		return t.Template.Execute(w, map[string]any{
 			"System":   system,
-			"Messages": collated,
+			"Messages": messages,
 		})
 	}
 
+	system = ""
 	var b bytes.Buffer
 	var prompt, response string
-	for i, m := range collated {
-		if m.Role == "user" {
-			prompt = m.Content
-		} else {
-			response = m.Content
-		}
-
-		if i != len(collated)-1 && prompt != "" && response != "" {
+	for _, m := range messages {
+		execute := func () error {
 			if err := t.Template.Execute(&b, map[string]any{
-				"System":   "",
+				"System":   system,
 				"Prompt":   prompt,
 				"Response": response,
 			}); err != nil {
 				return err
 			}
 
+			system = ""
 			prompt = ""
 			response = ""
+			return nil
+		}
+
+		switch m.Role {
+		case "system":
+			if prompt != "" || response != "" {
+				if err := execute(); err != nil {
+					return err
+				}
+			}
+			system = m.Content
+		case "user":
+			if response != "" {
+				if err := execute(); err != nil {
+					return err
+				}
+			}
+			prompt = m.Content
+		case "assistant":
+			response = m.Content
 		}
 	}
 
 	var cut bool
-	tree := t.Template.Copy()
-	// for the last message, cut everything after "{{ .Response }}"
-	tree.Root.Nodes = slices.DeleteFunc(tree.Root.Nodes, func(n parse.Node) bool {
-		if slices.Contains(parseNode(n), "Response") {
-			cut = true
+	nodes := deleteNode(t.Template.Root.Copy(), func(n parse.Node) bool {
+		switch t := n.(type) {
+		case *parse.ActionNode:
+		case *parse.FieldNode:
+			if slices.Contains(t.Ident, "Response") {
+				cut = true
+			}
 		}
 
 		return cut
 	})
 
-	if err := template.Must(template.New("").AddParseTree("", tree)).Execute(&b, map[string]any{
+	tree := parse.Tree{Root: nodes.(*parse.ListNode)}
+	if err := template.Must(template.New("").AddParseTree("", &tree)).Execute(&b, map[string]any{
 		"System": system,
 		"Prompt": prompt,
 	}); err != nil {
@@ -199,25 +221,16 @@ func (t *Template) Execute(w io.Writer, v Values) error {
 	return err
 }
 
-type messages []*api.Message
-
 // collate messages based on role. consecutive messages of the same role are merged
-// into a single message. collate also pulls out and merges messages with Role == "system"
-// which are templated separately. As a side effect, it mangles message content adding image
-// tags ([img-%d]) as needed
-func collate(msgs []api.Message) (system string, collated messages) {
+// into a single message. collate also collects and returns all system messages.
+// collate mutates message content adding image tags ([img-%d]) as needed
+func collate(msgs []api.Message) (string, []*api.Message) {
 	var n int
+
+	var system []string
+	var collated []*api.Message
 	for i := range msgs {
 		msg := msgs[i]
-		if msg.Role == "system" {
-			if system != "" {
-				system += "\n\n"
-			}
-
-			system += msg.Content
-			continue
-		}
-
 		for range msg.Images {
 			imageTag := fmt.Sprintf("[img-%d]", n)
 			if !strings.Contains(msg.Content, "[img]") {
@@ -228,6 +241,10 @@ func collate(msgs []api.Message) (system string, collated messages) {
 			n++
 		}
 
+		if msg.Role == "system" {
+			system = append(system, msg.Content)
+		}
+
 		if len(collated) > 0 && collated[len(collated)-1].Role == msg.Role {
 			collated[len(collated)-1].Content += "\n\n" + msg.Content
 		} else {
@@ -235,7 +252,7 @@ func collate(msgs []api.Message) (system string, collated messages) {
 		}
 	}
 
-	return
+	return strings.Join(system, "\n\n"), collated
 }
 
 func parseNode(n parse.Node) []string {
@@ -286,3 +303,72 @@ func parseNode(n parse.Node) []string {
 
 	return nil
 }
+
+// deleteNode walks the node list and deletes nodes that match the predicate
+// this is currently to remove the {{ .Response }} node from templates
+func deleteNode(n parse.Node, fn func(parse.Node) bool) parse.Node {
+	var walk func(n parse.Node) parse.Node
+	walk = func(n parse.Node) parse.Node {
+		if fn(n) {
+			return nil
+		}
+
+		switch t := n.(type) {
+		case *parse.ListNode:
+			var nodes []parse.Node
+			for _, c := range t.Nodes {
+				if n := walk(c); n != nil {
+					nodes = append(nodes, n)
+				}
+			}
+
+			t.Nodes = nodes
+			return t
+		case *parse.IfNode:
+			t.BranchNode = *(walk(&t.BranchNode).(*parse.BranchNode))
+		case *parse.WithNode:
+			t.BranchNode = *(walk(&t.BranchNode).(*parse.BranchNode))
+		case *parse.RangeNode:
+			t.BranchNode = *(walk(&t.BranchNode).(*parse.BranchNode))
+		case *parse.BranchNode:
+			t.List = walk(t.List).(*parse.ListNode)
+			if t.ElseList != nil {
+				t.ElseList = walk(t.ElseList).(*parse.ListNode)
+			}
+		case *parse.ActionNode:
+			n := walk(t.Pipe)
+			if n == nil {
+				return nil
+			}
+
+			t.Pipe = n.(*parse.PipeNode)
+		case *parse.PipeNode:
+			var commands []*parse.CommandNode
+			for _, c := range t.Cmds {
+				var args []parse.Node
+				for _, a := range c.Args {
+					if n := walk(a); n != nil {
+						args = append(args, n)
+					}
+				}
+
+				if len(args) == 0 {
+					return nil
+				}
+
+				c.Args = args
+				commands = append(commands, c)
+			}
+
+			if len(commands) == 0 {
+				return nil
+			}
+
+			t.Cmds = commands
+		}
+
+		return n
+	}
+
+	return walk(n)
+}

template/template_test.go

@@ -105,8 +105,8 @@ func TestTemplate(t *testing.T) {
 			}
 
 			for n, tt := range cases {
+				var actual bytes.Buffer
 				t.Run(n, func(t *testing.T) {
-					var actual bytes.Buffer
 					if err := tmpl.Execute(&actual, Values{Messages: tt}); err != nil {
 						t.Fatal(err)
 					}
@@ -116,7 +116,34 @@ func TestTemplate(t *testing.T) {
 						t.Fatal(err)
 					}
 
-					if diff := cmp.Diff(actual.Bytes(), expect); diff != "" {
+					bts := actual.Bytes()
+
+					if slices.Contains([]string{"chatqa.gotmpl", "llama2-chat.gotmpl", "mistral-instruct.gotmpl", "openchat.gotmpl", "vicuna.gotmpl"}, match) && bts[len(bts)-1] == ' ' {
+						t.Log("removing trailing space from output")
+						bts = bts[:len(bts)-1]
+					}
+
+					if diff := cmp.Diff(bts, expect); diff != "" {
+						t.Errorf("mismatch (-got +want):\n%s", diff)
+					}
+				})
+
+				t.Run("legacy", func(t *testing.T) {
+					t.Skip("legacy outputs are currently default outputs")
+					var legacy bytes.Buffer
+					if err := tmpl.Execute(&legacy, Values{Messages: tt, forceLegacy: true}); err != nil {
+						t.Fatal(err)
+					}
+
+					legacyBytes := legacy.Bytes()
+					if slices.Contains([]string{"chatqa.gotmpl", "openchat.gotmpl", "vicuna.gotmpl"}, match) && legacyBytes[len(legacyBytes)-1] == ' ' {
+						t.Log("removing trailing space from legacy output")
+						legacyBytes = legacyBytes[:len(legacyBytes)-1]
+					} else if slices.Contains([]string{"codellama-70b-instruct.gotmpl", "llama2-chat.gotmpl", "mistral-instruct.gotmpl"}, match) {
+						t.Skip("legacy outputs cannot be compared to messages outputs")
+					}
+
+					if diff := cmp.Diff(legacyBytes, actual.Bytes()); diff != "" {
 						t.Errorf("mismatch (-got +want):\n%s", diff)
 					}
 				})
@@ -135,7 +162,24 @@ func TestParse(t *testing.T) {
 		{"{{ .System }} {{ .Prompt }} {{ .Response }}", []string{"prompt", "response", "system"}},
 		{"{{ with .Tools }}{{ . }}{{ end }} {{ .System }} {{ .Prompt }}", []string{"prompt", "response", "system", "tools"}},
 		{"{{ range .Messages }}{{ .Role }} {{ .Content }}{{ end }}", []string{"content", "messages", "role"}},
-		{"{{ range .Messages }}{{ if eq .Role \"system\" }}SYSTEM: {{ .Content }}{{ else if eq .Role \"user\" }}USER: {{ .Content }}{{ else if eq .Role \"assistant\" }}ASSISTANT: {{ .Content }}{{ end }}{{ end }}", []string{"content", "messages", "role"}},
+		{`{{- range .Messages }}
+{{- if eq .Role "system" }}SYSTEM:
+{{- else if eq .Role "user" }}USER:
+{{- else if eq .Role "assistant" }}ASSISTANT:
+{{- end }} {{ .Content }}
+{{- end }}`, []string{"content", "messages", "role"}},
+		{`{{- if .Messages }}
+{{- range .Messages }}<|im_start|>{{ .Role }}
+{{ .Content }}<|im_end|>
+{{ end }}<|im_start|>assistant
+{{ else -}}
+{{ if .System }}<|im_start|>system
+{{ .System }}<|im_end|>
+{{ end }}{{ if .Prompt }}<|im_start|>user
+{{ .Prompt }}<|im_end|>
+{{ end }}<|im_start|>assistant
+{{ .Response }}<|im_end|>
+{{- end -}}`, []string{"content", "messages", "prompt", "response", "role", "system"}},
 	}
 
 	for _, tt := range cases {
@@ -145,9 +189,8 @@ func TestParse(t *testing.T) {
 				t.Fatal(err)
 			}
 
-			vars := tmpl.Vars()
-			if !slices.Equal(tt.vars, vars) {
-				t.Errorf("expected %v, got %v", tt.vars, vars)
+			if diff := cmp.Diff(tmpl.Vars(), tt.vars); diff != "" {
+				t.Errorf("mismatch (-got +want):\n%s", diff)
 			}
 		})
 	}
@@ -167,12 +210,17 @@ func TestExecuteWithMessages(t *testing.T) {
 		{
 			"mistral",
 			[]template{
-				{"no response", `[INST] {{ if .System }}{{ .System }}{{ "\n\n" }}{{ end }}{{ .Prompt }}[/INST] `},
-				{"response", `[INST] {{ if .System }}{{ .System }}{{ "\n\n" }}{{ end }}{{ .Prompt }}[/INST] {{ .Response }}`},
-				{"messages", `{{- range $index, $_ := .Messages }}
-{{- if eq .Role "user" }}[INST] {{ if and (eq (len (slice $.Messages $index)) 1) $.System }}{{ $.System }}{{ "\n\n" }}
-{{- end }}{{ .Content }}[/INST] {{ else if eq .Role "assistant" }}{{ .Content }}
-{{- end }}
+				{"no response", `[INST] {{ if .System }}{{ .System }}
+
+{{ end }}{{ .Prompt }}[/INST] `},
+				{"response", `[INST] {{ if .System }}{{ .System }}
+
+{{ end }}{{ .Prompt }}[/INST] {{ .Response }}`},
+				{"messages", `[INST] {{ if .System }}{{ .System }}
+
+{{ end }}
+{{- range .Messages }}
+{{- if eq .Role "user" }}{{ .Content }}[/INST] {{ else if eq .Role "assistant" }}{{ .Content }}[INST] {{ end }}
 {{- end }}`},
 			},
 			Values{
@@ -187,13 +235,17 @@ func TestExecuteWithMessages(t *testing.T) {
 		{
 			"mistral system",
 			[]template{
-				{"no response", `[INST] {{ if .System }}{{ .System }}{{ "\n\n" }}{{ end }}{{ .Prompt }}[/INST] `},
-				{"response", `[INST] {{ if .System }}{{ .System }}{{ "\n\n" }}{{ end }}{{ .Prompt }}[/INST] {{ .Response }}`},
-				{"messages", `
-{{- range $index, $_ := .Messages }}
-{{- if eq .Role "user" }}[INST] {{ if and (eq (len (slice $.Messages $index)) 1) $.System }}{{ $.System }}{{ "\n\n" }}
-{{- end }}{{ .Content }}[/INST] {{ else if eq .Role "assistant" }}{{ .Content }}
-{{- end }}
+				{"no response", `[INST] {{ if .System }}{{ .System }}
+
+{{ end }}{{ .Prompt }}[/INST] `},
+				{"response", `[INST] {{ if .System }}{{ .System }}
+
+{{ end }}{{ .Prompt }}[/INST] {{ .Response }}`},
+				{"messages", `[INST] {{ if .System }}{{ .System }}
+
+{{ end }}
+{{- range .Messages }}
+{{- if eq .Role "user" }}{{ .Content }}[/INST] {{ else if eq .Role "assistant" }}{{ .Content }}[INST] {{ end }}
 {{- end }}`},
 			},
 			Values{
@@ -204,9 +256,9 @@ func TestExecuteWithMessages(t *testing.T) {
 					{Role: "user", Content: "What is your name?"},
 				},
 			},
-			`[INST] Hello friend![/INST] Hello human![INST] You are a helpful assistant!
+			`[INST] You are a helpful assistant!
 
-What is your name?[/INST] `,
+Hello friend![/INST] Hello human![INST] What is your name?[/INST] `,
 		},
 		{
 			"chatml",
@@ -220,12 +272,9 @@ What is your name?[/INST] `,
 {{ .Response }}<|im_end|>
 `},
 				{"messages", `
-{{- range $index, $_ := .Messages }}
-{{- if and (eq .Role "user") (eq (len (slice $.Messages $index)) 1) $.System }}<|im_start|>system
-{{ $.System }}<|im_end|>{{ "\n" }}
-{{- end }}<|im_start|>{{ .Role }}
-{{ .Content }}<|im_end|>{{ "\n" }}
-{{- end }}<|im_start|>assistant
+{{- range $index, $_ := .Messages }}<|im_start|>{{ .Role }}
+{{ .Content }}<|im_end|>
+{{ end }}<|im_start|>assistant
 `},
 			},
 			Values{
@@ -236,12 +285,12 @@ What is your name?[/INST] `,
 					{Role: "user", Content: "What is your name?"},
 				},
 			},
-			`<|im_start|>user
+			`<|im_start|>system
+You are a helpful assistant!<|im_end|>
+<|im_start|>user
 Hello friend!<|im_end|>
 <|im_start|>assistant
 Hello human!<|im_end|>
-<|im_start|>system
-You are a helpful assistant!<|im_end|>
 <|im_start|>user
 What is your name?<|im_end|>
 <|im_start|>assistant
@@ -258,9 +307,11 @@ What is your name?<|im_end|>
 `},
 				{"messages", `
 {{- range .Messages }}
-{{- if eq .Role "user" }}Question: {{ .Content }}{{ "\n\n" }}
-{{- else if eq .Role "assistant" }}Answer: {{ .Content }}{{ "\n\n" }}
-{{- end }}
+{{- if eq .Role "user" }}Question: {{ .Content }}
+
+{{ else if eq .Role "assistant" }}Answer: {{ .Content }}
+
+{{ end }}
 {{- end }}Answer: `},
 			},
 			Values{
@@ -300,8 +351,8 @@ Answer: `,
 						t.Fatal(err)
 					}
 
-					if b.String() != tt.expected {
-						t.Errorf("expected\n%s,\ngot\n%s", tt.expected, b.String())
+					if diff := cmp.Diff(b.String(), tt.expected); diff != "" {
+						t.Errorf("mismatch (-got +want):\n%s", diff)
 					}
 				})
 			}

template/testdata/alpaca.gotmpl/system-user-assistant-user

@@ -1,4 +1,6 @@
-You are a helpful assistant.### Instruction:
+You are a helpful assistant.
+
+### Instruction:
 Hello, how are you?
 
 ### Response:

template/testdata/codellama-70b-instruct.gotmpl/system-user-assistant-user

@@ -9,3 +9,4 @@ Source: system
  I'd like to show off how chat templating works! <step> Source: assistant
 Destination: user
 
+ 

template/testdata/codellama-70b-instruct.gotmpl/user

@@ -3,3 +3,4 @@ Source: user
  Hello, how are you? <step> Source: assistant
 Destination: user
 
+ 

template/testdata/codellama-70b-instruct.gotmpl/user-assistant-user

@@ -7,3 +7,4 @@ Source: user
  I'd like to show off how chat templating works! <step> Source: assistant
 Destination: user
 
+ 

template/testdata/llama2-chat.gotmpl/system-user-assistant-user

@@ -2,4 +2,6 @@
 You are a helpful assistant.
 <</SYS>>
 
-Hello, how are you? [/INST] I'm doing great. How can I help you today?</s><s>[INST] I'd like to show off how chat templating works! [/INST]
+Hello, how are you? [/INST] I'm doing great. How can I help you today?</s><s>[INST] <<SYS>><</SYS>>
+
+I'd like to show off how chat templating works! [/INST]

template/testdata/llama2-chat.gotmpl/user-assistant-user

@@ -1,3 +1,5 @@
 [INST] <<SYS>><</SYS>>
 
-Hello, how are you? [/INST] I'm doing great. How can I help you today?</s><s>[INST] I'd like to show off how chat templating works! [/INST]
+Hello, how are you? [/INST] I'm doing great. How can I help you today?</s><s>[INST] <<SYS>><</SYS>>
+
+I'd like to show off how chat templating works! [/INST]

template/testdata/mistral-instruct.gotmpl/system-user-assistant-user

@@ -1,2 +1,3 @@
-[INST] Hello, how are you?[/INST] I'm doing great. How can I help you today?</s>[INST] You are a helpful assistant.
-I'd like to show off how chat templating works![/INST]
+[INST] You are a helpful assistant.
+
+Hello, how are you?[/INST] I'm doing great. How can I help you today?</s>[INST] I'd like to show off how chat templating works![/INST]

template/testdata/openchat.gotmpl/system-user-assistant-user

@@ -1 +1 @@
-GPT Correct System: You are a helpful assistant.<|end_of_turn|>GPT Correct User: Hello, how are you?<|end_of_turn|>GPT Correct Assistant: I'm doing great. How can I help you today?<|end_of_turn|>GPT Correct User: I'd like to show off how chat templating works!<|end_of_turn|>GPT Correct Assistant:
+GPT4 Correct System: You are a helpful assistant.<|end_of_turn|>GPT4 Correct User: Hello, how are you?<|end_of_turn|>GPT4 Correct Assistant: I'm doing great. How can I help you today?<|end_of_turn|>GPT4 Correct User: I'd like to show off how chat templating works!<|end_of_turn|>GPT4 Correct Assistant:

template/testdata/openchat.gotmpl/user

@@ -1 +1 @@
-GPT Correct User: Hello, how are you?<|end_of_turn|>GPT Correct Assistant:
+GPT4 Correct User: Hello, how are you?<|end_of_turn|>GPT4 Correct Assistant:

template/testdata/openchat.gotmpl/user-assistant-user

@@ -1 +1 @@
-GPT Correct User: Hello, how are you?<|end_of_turn|>GPT Correct Assistant: I'm doing great. How can I help you today?<|end_of_turn|>GPT Correct User: I'd like to show off how chat templating works!<|end_of_turn|>GPT Correct Assistant:
+GPT4 Correct User: Hello, how are you?<|end_of_turn|>GPT4 Correct Assistant: I'm doing great. How can I help you today?<|end_of_turn|>GPT4 Correct User: I'd like to show off how chat templating works!<|end_of_turn|>GPT4 Correct Assistant:

template/vicuna.gotmpl

@@ -1,14 +1,4 @@
-{{- if .Messages }}
-{{- if .System }}{{ .System }}
-
-{{ end }}
-{{- range .Messages }}
-{{- if eq .Role "user" }}USER: {{ .Content }}
-{{ else if eq .Role "assistant" }}ASSISTANT: {{ .Content }}</s>
-{{ end }}
-{{- end }}ASSISTANT:
-{{- else }}
 {{ if .System }}{{ .System }}
+
 {{ end }}{{ if .Prompt }}USER: {{ .Prompt }}
-{{ end }}ASSISTANT: {{ .Response }}
-{{- end }}
+{{ end }}ASSISTANT: {{ .Response }}</s>

template/zephyr.gotmpl

@@ -1,15 +1,6 @@
-{{- if .Messages }}
-{{- if .System }}<|system|>
-{{ .System }}</s>
-{{ end }}
-{{- range .Messages }}<|{{ .Role }}|>
-{{ .Content }}</s>
-{{ end }}<|assistant|>
-{{ else }}
 {{ if .System }}<|system|>
 {{ .System }}</s>
 {{ end }}{{ if .Prompt }}<|user|>
 {{ .Prompt }}</s>
 {{ end }}<|assistant|>
 {{ .Response }}</s>
-{{- end }}