baalajimaestro/ollama
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

Merge https://github.com/ollama/ollama

Browse source 
Signed-off-by: baalajimaestro <me@baalajimaestro.me>
This commit is contained in:
baalajimaestro 2024-07-06 23:37:14 +05:30
commit 415d9f0f15
Signed by: baalajimaestro
GPG key ID: F93C394FE9BBAFD5
79 changed files with 1964 additions and 980 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

4
.github/workflows/release.yaml vendored
View file

@ -304,6 +304,10 @@ jobs:
write-host "Installing plugin"
& "${env:RUNNER_TEMP}\plugin\*\kmscng.msi" /quiet
write-host "plugin installed"
- name: remove unwanted mingw dll.a files
run: |
Remove-Item "C:\mingw64\x86_64-w64-mingw32\lib\libpthread.dll.a"
Remove-Item "C:\mingw64\x86_64-w64-mingw32\lib\libwinpthread.dll.a"
- uses: actions/setup-go@v5
with:
go-version-file: go.mod

2
.github/workflows/test.yaml vendored
View file

@ -58,6 +58,7 @@ jobs:
runs-on: ${{ matrix.os }}
env:
GOARCH: ${{ matrix.arch }}
CGO_ENABLED: '1'
steps:
- uses: actions/checkout@v4
- uses: actions/setup-go@v5
@ -79,6 +80,7 @@ jobs:
- run: go generate -x ./...
if: ${{ ! startsWith(matrix.os, 'windows-') }}
name: 'Unix Go Generate'
- run: go build .
- uses: actions/upload-artifact@v4
with:
name: ${{ matrix.os }}-${{ matrix.arch }}-libraries

112
api/types.go
View file

@ -159,49 +159,18 @@ type Options struct {
// Runner options which must be set when the model is loaded into memory
type Runner struct {
UseNUMA bool `json:"numa,omitempty"`
NumCtx int `json:"num_ctx,omitempty"`
NumBatch int `json:"num_batch,omitempty"`
NumGPU int `json:"num_gpu,omitempty"`
MainGPU int `json:"main_gpu,omitempty"`
LowVRAM bool `json:"low_vram,omitempty"`
F16KV bool `json:"f16_kv,omitempty"`
LogitsAll bool `json:"logits_all,omitempty"`
VocabOnly bool `json:"vocab_only,omitempty"`
UseMMap TriState `json:"use_mmap,omitempty"`
UseMLock bool `json:"use_mlock,omitempty"`
NumThread int `json:"num_thread,omitempty"`
}
type TriState int
const (
TriStateUndefined TriState = -1
TriStateFalse TriState = 0
TriStateTrue TriState = 1
)
func (b *TriState) UnmarshalJSON(data []byte) error {
var v bool
if err := json.Unmarshal(data, &v); err != nil {
return err
}
if v {
*b = TriStateTrue
}
*b = TriStateFalse
return nil
}
func (b *TriState) MarshalJSON() ([]byte, error) {
if *b == TriStateUndefined {
return nil, nil
}
var v bool
if *b == TriStateTrue {
v = true
}
return json.Marshal(v)
UseNUMA bool `json:"numa,omitempty"`
NumCtx int `json:"num_ctx,omitempty"`
NumBatch int `json:"num_batch,omitempty"`
NumGPU int `json:"num_gpu,omitempty"`
MainGPU int `json:"main_gpu,omitempty"`
LowVRAM bool `json:"low_vram,omitempty"`
F16KV bool `json:"f16_kv,omitempty"`
LogitsAll bool `json:"logits_all,omitempty"`
VocabOnly bool `json:"vocab_only,omitempty"`
UseMMap *bool `json:"use_mmap,omitempty"`
UseMLock bool `json:"use_mlock,omitempty"`
NumThread int `json:"num_thread,omitempty"`
}
// EmbeddingRequest is the request passed to [Client.Embeddings].
@ -345,6 +314,13 @@ type ProcessModelResponse struct {
SizeVRAM int64 `json:"size_vram"`
}
type RetrieveModelResponse struct {
Id string `json:"id"`
Object string `json:"object"`
Created int64 `json:"created"`
OwnedBy string `json:"owned_by"`
}
type TokenResponse struct {
Token string `json:"token"`
}
@ -437,19 +413,6 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
continue
}
if reflect.PointerTo(field.Type()) == reflect.TypeOf((*TriState)(nil)) {
val, ok := val.(bool)
if !ok {
return fmt.Errorf("option %q must be of type boolean", key)
}
if val {
field.SetInt(int64(TriStateTrue))
} else {
field.SetInt(int64(TriStateFalse))
}
continue
}
switch field.Kind() {
case reflect.Int:
switch t := val.(type) {
@ -496,6 +459,17 @@ func (opts *Options) FromMap(m map[string]interface{}) error {
slice[i] = str
}
field.Set(reflect.ValueOf(slice))
case reflect.Pointer:
var b bool
if field.Type() == reflect.TypeOf(&b) {
val, ok := val.(bool)
if !ok {
return fmt.Errorf("option %q must be of type boolean", key)
}
field.Set(reflect.ValueOf(&val))
} else {
return fmt.Errorf("unknown type loading config params: %v %v", field.Kind(), field.Type())
}
default:
return fmt.Errorf("unknown type loading config params: %v", field.Kind())
}
@ -538,7 +512,7 @@ func DefaultOptions() Options {
LowVRAM: false,
F16KV: true,
UseMLock: false,
UseMMap: TriStateUndefined,
UseMMap: nil,
UseNUMA: false,
},
}
@ -608,19 +582,6 @@ func FormatParams(params map[string][]string) (map[string]interface{}, error) {
} else {
field := valueOpts.FieldByName(opt.Name)
if field.IsValid() && field.CanSet() {
if reflect.PointerTo(field.Type()) == reflect.TypeOf((*TriState)(nil)) {
boolVal, err := strconv.ParseBool(vals[0])
if err != nil {
return nil, fmt.Errorf("invalid bool value %s", vals)
}
if boolVal {
out[key] = TriStateTrue
} else {
out[key] = TriStateFalse
}
continue
}
switch field.Kind() {
case reflect.Float32:
floatVal, err := strconv.ParseFloat(vals[0], 32)
@ -648,6 +609,17 @@ func FormatParams(params map[string][]string) (map[string]interface{}, error) {
case reflect.Slice:
// TODO: only string slices are supported right now
out[key] = vals
case reflect.Pointer:
var b bool
if field.Type() == reflect.TypeOf(&b) {
boolVal, err := strconv.ParseBool(vals[0])
if err != nil {
return nil, fmt.Errorf("invalid bool value %s", vals)
}
out[key] = &boolVal
} else {
return nil, fmt.Errorf("unknown type %s for %s", field.Kind(), key)
}
default:
return nil, fmt.Errorf("unknown type %s for %s", field.Kind(), key)
}

40
api/types_test.go
View file

@ -108,25 +108,27 @@ func TestDurationMarshalUnmarshal(t *testing.T) {
}
func TestUseMmapParsingFromJSON(t *testing.T) {
tr := true
fa := false
tests := []struct {
name string
req string
exp TriState
exp *bool
}{
{
name: "Undefined",
req: `{ }`,
exp: TriStateUndefined,
exp: nil,
},
{
name: "True",
req: `{ "use_mmap": true }`,
exp: TriStateTrue,
exp: &tr,
},
{
name: "False",
req: `{ "use_mmap": false }`,
exp: TriStateFalse,
exp: &fa,
},
}
@ -144,50 +146,52 @@ func TestUseMmapParsingFromJSON(t *testing.T) {
}
func TestUseMmapFormatParams(t *testing.T) {
tr := true
fa := false
tests := []struct {
name string
req map[string][]string
exp TriState
exp *bool
err error
}{
{
name: "True",
req: map[string][]string{
"use_mmap": []string{"true"},
"use_mmap": {"true"},
},
exp: TriStateTrue,
exp: &tr,
err: nil,
},
{
name: "False",
req: map[string][]string{
"use_mmap": []string{"false"},
"use_mmap": {"false"},
},
exp: TriStateFalse,
exp: &fa,
err: nil,
},
{
name: "Numeric True",
req: map[string][]string{
"use_mmap": []string{"1"},
"use_mmap": {"1"},
},
exp: TriStateTrue,
exp: &tr,
err: nil,
},
{
name: "Numeric False",
req: map[string][]string{
"use_mmap": []string{"0"},
"use_mmap": {"0"},
},
exp: TriStateFalse,
exp: &fa,
err: nil,
},
{
name: "invalid string",
req: map[string][]string{
"use_mmap": []string{"foo"},
"use_mmap": {"foo"},
},
exp: TriStateUndefined,
exp: nil,
err: fmt.Errorf("invalid bool value [foo]"),
},
}
@ -195,11 +199,11 @@ func TestUseMmapFormatParams(t *testing.T) {
for _, test := range tests {
t.Run(test.name, func(t *testing.T) {
resp, err := FormatParams(test.req)
require.Equal(t, err, test.err)
require.Equal(t, test.err, err)
respVal, ok := resp["use_mmap"]
if test.exp != TriStateUndefined {
if test.exp != nil {
assert.True(t, ok, "resp: %v", resp)
assert.Equal(t, test.exp, respVal)
assert.Equal(t, *test.exp, *respVal.(*bool))
}
})
}

2
docs/development.md
View file

@ -104,7 +104,7 @@ like to use. For example, to compile an optimized binary for an Intel i9-9880H,
you might use:
```
OLLAMA_CUSTOM_CPU_DEFS="-DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_F16C=on -DLLAMA_FMA=on" go generate ./...
OLLAMA_CUSTOM_CPU_DEFS="-DGGML_AVX=on -DGGML_AVX2=on -DGGML_F16C=on -DGGML_FMA=on" go generate ./...
go build .
```

16
docs/faq.md
View file

@ -257,3 +257,19 @@ If you wish to override the `OLLAMA_KEEP_ALIVE` setting, use the `keep_alive` AP
## How do I manage the maximum number of requests the Ollama server can queue?
If too many requests are sent to the server, it will respond with a 503 error indicating the server is overloaded. You can adjust how many requests may be queue by setting `OLLAMA_MAX_QUEUE`.
## How does Ollama handle concurrent requests?
Ollama supports two levels of concurrent processing. If your system has sufficient available memory (system memory when using CPU inference, or VRAM for GPU inference) then multiple models can be loaded at the same time. For a given model, if there is sufficient available memory when the model is loaded, it is configured to allow parallel request processing.
If there is insufficient available memory to load a new model request while one or more models are already loaded, all new requests will be queued until the new model can be loaded. As prior models become idle, one or more will be unloaded to make room for the new model. Queued requests will be processed in order. When using GPU inference new models must be able to completely fit in VRAM to allow concurrent model loads.
Parallel request processing for a given model results in increasing the context size by the number of parallel requests. For example, a 2K context with 4 parallel requests will result in an 8K context and additional memory allocation.
The following server settings may be used to adjust how Ollama handles concurrent requests on most platforms:
- `OLLAMA_MAX_LOADED_MODELS` - The maximum number of models that can be loaded concurrently provided they fit in available memory. The default is 3 * the number of GPUs or 3 for CPU inference.
- `OLLAMA_NUM_PARALLEL` - The maximum number of parallel requests each model will process at the same time. The default will auto-select either 4 or 1 based on available memory.
- `OLLAMA_MAX_QUEUE` - The maximum number of requests Ollama will queue when busy before rejecting additional requests. The default is 512
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.

1
docs/openai.md
View file

@ -65,6 +65,7 @@ curl http://localhost:11434/v1/chat/completions \
}
]
}'
```
## Endpoints

19
docs/troubleshooting.md
View file

@ -70,14 +70,18 @@ curl -fsSL https://ollama.com/install.sh | OLLAMA_VERSION="0.1.29" sh
If your system is configured with the "noexec" flag where Ollama stores its temporary executable files, you can specify an alternate location by setting OLLAMA_TMPDIR to a location writable by the user ollama runs as. For example OLLAMA_TMPDIR=/usr/share/ollama/
## Container fails to run on NVIDIA GPU
## NVIDIA GPU Discovery
Make sure you've set up the container runtime first as described in [docker.md](./docker.md)
When Ollama starts up, it takes inventory of the GPUs present in the system to determine compatibility and how much VRAM is available. Sometimes this discovery can fail to find your GPUs. In general, running the latest driver will yield the best results.
Sometimes the container runtime can have difficulties initializing the GPU. When you check the server logs, this can show up as various error codes, such as "3" (not initialized), "46" (device unavailable), "100" (no device), "999" (unknown), or others. The following troubleshooting techniques may help resolve the problem
### Linux NVIDIA Troubleshooting
- Is the container runtime working? Try `docker run --gpus all ubuntu nvidia-smi` - if this doesn't work, Ollama wont be able to see your NVIDIA GPU.
- Is the uvm driver not loaded? `sudo nvidia-modprobe -u`
If you are using a container to run Ollama, make sure you've set up the container runtime first as described in [docker.md](./docker.md)
Sometimes the Ollama can have difficulties initializing the GPU. When you check the server logs, this can show up as various error codes, such as "3" (not initialized), "46" (device unavailable), "100" (no device), "999" (unknown), or others. The following troubleshooting techniques may help resolve the problem
- If you are using a container, is the container runtime working? Try `docker run --gpus all ubuntu nvidia-smi` - if this doesn't work, Ollama wont be able to see your NVIDIA GPU.
- Is the uvm driver loaded? `sudo nvidia-modprobe -u`
- Try reloading the nvidia_uvm driver - `sudo rmmod nvidia_uvm` then `sudo modprobe nvidia_uvm`
- Try rebooting
- Make sure you're running the latest nvidia drivers
@ -85,3 +89,8 @@ Sometimes the container runtime can have difficulties initializing the GPU. When
If none of those resolve the problem, gather additional information and file an issue:
- Set `CUDA_ERROR_LEVEL=50` and try again to get more diagnostic logs
- Check dmesg for any errors `sudo dmesg | grep -i nvrm` and `sudo dmesg | grep -i nvidia`
## Windows Terminal Errors
Older versions of Windows 10 (e.g., 21H1) are known to have a bug where the standard terminal program does not display control characters correctly. This can result in a long string of strings like `←[?25h←[?25l` being displayed, sometimes erroring with `The parameter is incorrect` To resolve this problem, please update to Win 10 22H1 or newer.

2
docs/windows.md
View file

@ -19,7 +19,7 @@ Logs will often be helpful in diagnosing the problem (see
## System Requirements
* Windows 10 or newer, Home or Pro
* Windows 10 22H2 or newer, Home or Pro
* NVIDIA 452.39 or newer Drivers if you have an NVIDIA card
* AMD Radeon Driver https://www.amd.com/en/support if you have a Radeon card

51
envconfig/config.go
View file

@ -4,12 +4,14 @@ import (
"errors"
"fmt"
"log/slog"
"math"
"net"
"os"
"path/filepath"
"runtime"
"strconv"
"strings"
"time"
)
type OllamaHost struct {
@ -34,17 +36,17 @@ var (
// Set via OLLAMA_HOST in the environment
Host *OllamaHost
// Set via OLLAMA_KEEP_ALIVE in the environment
KeepAlive string
KeepAlive time.Duration
// Set via OLLAMA_LLM_LIBRARY in the environment
LLMLibrary string
// Set via OLLAMA_MAX_LOADED_MODELS in the environment
MaxRunners int
// Set via OLLAMA_MAX_QUEUE in the environment
MaxQueuedRequests int
// Set via OLLAMA_MODELS in the environment
ModelsDir string
// 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
NoHistory bool
// Set via OLLAMA_NOPRUNE in the environment
@ -85,13 +87,13 @@ func AsMap() map[string]EnvVar {
"OLLAMA_HOST": {"OLLAMA_HOST", Host, "IP Address for the ollama server (default 127.0.0.1:11434)"},
"OLLAMA_KEEP_ALIVE": {"OLLAMA_KEEP_ALIVE", KeepAlive, "The duration that models stay loaded in memory (default \"5m\")"},
"OLLAMA_LLM_LIBRARY": {"OLLAMA_LLM_LIBRARY", LLMLibrary, "Set LLM library to bypass autodetection"},
"OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners, "Maximum number of loaded models (default 1)"},
"OLLAMA_MAX_LOADED_MODELS": {"OLLAMA_MAX_LOADED_MODELS", MaxRunners, "Maximum number of loaded models per GPU"},
"OLLAMA_MAX_QUEUE": {"OLLAMA_MAX_QUEUE", MaxQueuedRequests, "Maximum number of queued requests"},
"OLLAMA_MAX_VRAM": {"OLLAMA_MAX_VRAM", MaxVRAM, "Maximum VRAM"},
"OLLAMA_MODELS": {"OLLAMA_MODELS", ModelsDir, "The path to the models directory"},
"OLLAMA_NOHISTORY": {"OLLAMA_NOHISTORY", NoHistory, "Do not preserve readline history"},
"OLLAMA_NOPRUNE": {"OLLAMA_NOPRUNE", NoPrune, "Do not prune model blobs on startup"},
"OLLAMA_NUM_PARALLEL": {"OLLAMA_NUM_PARALLEL", NumParallel, "Maximum number of parallel requests (default 1)"},
"OLLAMA_NUM_PARALLEL": {"OLLAMA_NUM_PARALLEL", NumParallel, "Maximum number of parallel requests"},
"OLLAMA_ORIGINS": {"OLLAMA_ORIGINS", AllowOrigins, "A comma separated list of allowed origins"},
"OLLAMA_RUNNERS_DIR": {"OLLAMA_RUNNERS_DIR", RunnersDir, "Location for runners"},
"OLLAMA_SCHED_SPREAD": {"OLLAMA_SCHED_SPREAD", SchedSpread, "Always schedule model across all GPUs"},
@ -129,9 +131,10 @@ func clean(key string) string {
func init() {
// default values
NumParallel = 1
MaxRunners = 1
NumParallel = 0 // Autoselect
MaxRunners = 0 // Autoselect
MaxQueuedRequests = 512
KeepAlive = 5 * time.Minute
LoadConfig()
}
@ -205,8 +208,8 @@ func LoadConfig() {
if onp := clean("OLLAMA_NUM_PARALLEL"); onp != "" {
val, err := strconv.Atoi(onp)
if err != nil || val <= 0 {
slog.Error("invalid setting must be greater than zero", "OLLAMA_NUM_PARALLEL", onp, "error", err)
if err != nil {
slog.Error("invalid setting, ignoring", "OLLAMA_NUM_PARALLEL", onp, "error", err)
} else {
NumParallel = val
}
@ -251,7 +254,7 @@ func LoadConfig() {
if maxRunners != "" {
m, err := strconv.Atoi(maxRunners)
if err != nil {
slog.Error("invalid setting", "OLLAMA_MAX_LOADED_MODELS", maxRunners, "error", err)
slog.Error("invalid setting, ignoring", "OLLAMA_MAX_LOADED_MODELS", maxRunners, "error", err)
} else {
MaxRunners = m
}
@ -260,13 +263,16 @@ func LoadConfig() {
if onp := os.Getenv("OLLAMA_MAX_QUEUE"); onp != "" {
p, err := strconv.Atoi(onp)
if err != nil || p <= 0 {
slog.Error("invalid setting", "OLLAMA_MAX_QUEUE", onp, "error", err)
slog.Error("invalid setting, ignoring", "OLLAMA_MAX_QUEUE", onp, "error", err)
} else {
MaxQueuedRequests = p
}
}
KeepAlive = clean("OLLAMA_KEEP_ALIVE")
ka := clean("OLLAMA_KEEP_ALIVE")
if ka != "" {
loadKeepAlive(ka)
}
var err error
ModelsDir, err = getModelsDir()
@ -344,3 +350,24 @@ func getOllamaHost() (*OllamaHost, error) {
Port: port,
}, nil
}
func loadKeepAlive(ka string) {
v, err := strconv.Atoi(ka)
if err != nil {
d, err := time.ParseDuration(ka)
if err == nil {
if d < 0 {
KeepAlive = time.Duration(math.MaxInt64)
} else {
KeepAlive = d
}
}
} else {
d := time.Duration(v) * time.Second
if d < 0 {
KeepAlive = time.Duration(math.MaxInt64)
} else {
KeepAlive = d
}
}
}

17
envconfig/config_test.go
View file

@ -2,8 +2,10 @@ package envconfig
import (
"fmt"
"math"
"net"
"testing"
"time"
"github.com/stretchr/testify/assert"
"github.com/stretchr/testify/require"
@ -23,6 +25,21 @@ func TestConfig(t *testing.T) {
t.Setenv("OLLAMA_FLASH_ATTENTION", "1")
LoadConfig()
require.True(t, FlashAttention)
t.Setenv("OLLAMA_KEEP_ALIVE", "")
LoadConfig()
require.Equal(t, 5*time.Minute, KeepAlive)
t.Setenv("OLLAMA_KEEP_ALIVE", "3")
LoadConfig()
require.Equal(t, 3*time.Second, KeepAlive)
t.Setenv("OLLAMA_KEEP_ALIVE", "1h")
LoadConfig()
require.Equal(t, 1*time.Hour, KeepAlive)
t.Setenv("OLLAMA_KEEP_ALIVE", "-1s")
LoadConfig()
require.Equal(t, time.Duration(math.MaxInt64), KeepAlive)
t.Setenv("OLLAMA_KEEP_ALIVE", "-1")
LoadConfig()
require.Equal(t, time.Duration(math.MaxInt64), KeepAlive)
}
func TestClientFromEnvironment(t *testing.T) {

5
gpu/amd_windows.go
View file

@ -115,8 +115,6 @@ func AMDGetGPUInfo() []RocmGPUInfo {
continue
}
// TODO revisit this once ROCm v6 is available on windows.
// v5.7 only reports VRAM used by this process, so it's completely wrong and unusable
slog.Debug("amdgpu memory", "gpu", i, "total", format.HumanBytes2(totalMemory))
slog.Debug("amdgpu memory", "gpu", i, "available", format.HumanBytes2(freeMemory))
gpuInfo := RocmGPUInfo{
@ -126,6 +124,9 @@ func AMDGetGPUInfo() []RocmGPUInfo {
TotalMemory: totalMemory,
FreeMemory: freeMemory,
},
// Free memory reporting on Windows is not reliable until we bump to ROCm v6.2
UnreliableFreeMemory: true,
ID: strconv.Itoa(i), // TODO this is probably wrong if we specify visible devices
DependencyPath: libDir,
MinimumMemory: rocmMinimumMemory,

23
gpu/gpu.go
View file

@ -202,7 +202,7 @@ func GetGPUInfo() GpuInfoList {
}()
if !bootstrapped {
slog.Debug("Detecting GPUs")
slog.Info("looking for compatible GPUs")
needRefresh = false
cpuCapability = GetCPUCapability()
var memInfo C.mem_info_t
@ -320,6 +320,9 @@ func GetGPUInfo() GpuInfoList {
rocmGPUs = AMDGetGPUInfo()
bootstrapped = true
if len(cudaGPUs) == 0 && len(rocmGPUs) == 0 && len(oneapiGPUs) == 0 {
slog.Info("no compatible GPUs were discovered")
}
}
// For detected GPUs, load library if not loaded
@ -514,7 +517,23 @@ func LoadNVCUDAMgmt(nvcudaLibPaths []string) (int, *C.nvcuda_handle_t, string) {
defer C.free(unsafe.Pointer(lib))
C.nvcuda_init(lib, &resp)
if resp.err != nil {
slog.Debug("Unable to load nvcuda", "library", libPath, "error", C.GoString(resp.err))
// Decide what log level based on the type of error message to help users understand why
msg := C.GoString(resp.err)
switch resp.cudaErr {
case C.CUDA_ERROR_INSUFFICIENT_DRIVER, C.CUDA_ERROR_SYSTEM_DRIVER_MISMATCH:
slog.Warn("version mismatch between driver and cuda driver library - reboot or upgrade may be required", "library", libPath, "error", msg)
case C.CUDA_ERROR_NO_DEVICE:
slog.Info("no nvidia devices detected", "library", libPath)
case C.CUDA_ERROR_UNKNOWN:
slog.Warn("unknown error initializing cuda driver library", "library", libPath, "error", msg)
slog.Warn("see https://github.com/ollama/ollama/blob/main/docs/troubleshooting.md for more information")
default:
if strings.Contains(msg, "wrong ELF class") {
slog.Debug("skipping 32bit library", "library", libPath)
} else {
slog.Info("unable to load cuda driver library", "library", libPath, "error", msg)
}
}
C.free(unsafe.Pointer(resp.err))
} else {
return int(resp.num_devices), &resp.ch, libPath

31
gpu/gpu_info_nvcuda.c
View file

@ -7,6 +7,7 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
CUresult ret;
resp->err = NULL;
resp->num_devices = 0;
resp->cudaErr = CUDA_SUCCESS;
const int buflen = 256;
char buf[buflen + 1];
int i;
@ -38,6 +39,7 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
nvcuda_lib_path, msg);
free(msg);
resp->err = strdup(buf);
resp->cudaErr = -1;
return;
}
@ -52,6 +54,7 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
msg);
free(msg);
resp->err = strdup(buf);
resp->cudaErr = -1;
return;
}
}
@ -61,12 +64,9 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
LOG(resp->ch.verbose, "cuInit err: %d\n", ret);
UNLOAD_LIBRARY(resp->ch.handle);
resp->ch.handle = NULL;
if (ret == CUDA_ERROR_INSUFFICIENT_DRIVER) {
resp->err = strdup("your nvidia driver is too old or missing. If you have a CUDA GPU please upgrade to run ollama");
return;
}
snprintf(buf, buflen, "nvcuda init failure: %d", ret);
snprintf(buf, buflen, "cuda driver library init failure: %d", ret);
resp->err = strdup(buf);
resp->cudaErr = ret;
return;
}
@ -91,6 +91,7 @@ void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp) {
resp->ch.handle = NULL;
snprintf(buf, buflen, "unable to get device count: %d", ret);
resp->err = strdup(buf);
resp->cudaErr = ret;
return;
}
}
@ -106,13 +107,13 @@ void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {
CUuuid uuid = {0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0};
if (h.handle == NULL) {
resp->err = strdup("nvcuda handle isn't initialized");
resp->err = strdup("cuda driver library handle isn't initialized");
return;
}
ret = (*h.cuDeviceGet)(&device, i);
if (ret != CUDA_SUCCESS) {
snprintf(buf, buflen, "nvcuda device failed to initialize");
snprintf(buf, buflen, "cuda driver library device failed to initialize");
resp->err = strdup(buf);
return;
}
@ -168,14 +169,14 @@ void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {
// To get memory we have to set (and release) a context
ret = (*h.cuCtxCreate_v3)(&ctx, NULL, 0, 0, device);
if (ret != CUDA_SUCCESS) {
snprintf(buf, buflen, "nvcuda failed to get device context %d", ret);
snprintf(buf, buflen, "cuda driver library failed to get device context %d", ret);
resp->err = strdup(buf);
return;
}
ret = (*h.cuMemGetInfo_v2)(&memInfo.free, &memInfo.total);
if (ret != CUDA_SUCCESS) {
snprintf(buf, buflen, "nvcuda device memory info lookup failure %d", ret);
snprintf(buf, buflen, "cuda driver library device memory info lookup failure %d", ret);
resp->err = strdup(buf);
// Best effort on failure...
(*h.cuCtxDestroy)(ctx);
@ -193,7 +194,7 @@ void nvcuda_bootstrap(nvcuda_handle_t h, int i, mem_info_t *resp) {
ret = (*h.cuCtxDestroy)(ctx);
if (ret != CUDA_SUCCESS) {
LOG(1, "nvcuda failed to release device context %d", ret);
LOG(1, "cuda driver library failed to release device context %d", ret);
}
}
@ -206,7 +207,7 @@ void nvcuda_get_free(nvcuda_handle_t h, int i, uint64_t *free, uint64_t *total)
ret = (*h.cuDeviceGet)(&device, i);
if (ret != CUDA_SUCCESS) {
LOG(1, "nvcuda device failed to initialize");
LOG(1, "cuda driver library device failed to initialize");
return;
}
@ -214,13 +215,13 @@ void nvcuda_get_free(nvcuda_handle_t h, int i, uint64_t *free, uint64_t *total)
// To get memory we have to set (and release) a context
ret = (*h.cuCtxCreate_v3)(&ctx, NULL, 0, 0, device);
if (ret != CUDA_SUCCESS) {
LOG(1, "nvcuda failed to get device context %d", ret);
LOG(1, "cuda driver library failed to get device context %d", ret);
return;
}
ret = (*h.cuMemGetInfo_v2)(free, total);
if (ret != CUDA_SUCCESS) {
LOG(1, "nvcuda device memory info lookup failure %d", ret);
LOG(1, "cuda driver library device memory info lookup failure %d", ret);
// Best effort on failure...
(*h.cuCtxDestroy)(ctx);
return;
@ -228,12 +229,12 @@ void nvcuda_get_free(nvcuda_handle_t h, int i, uint64_t *free, uint64_t *total)
ret = (*h.cuCtxDestroy)(ctx);
if (ret != CUDA_SUCCESS) {
LOG(1, "nvcuda failed to release device context %d", ret);
LOG(1, "cuda driver library failed to release device context %d", ret);
}
}
void nvcuda_release(nvcuda_handle_t h) {
LOG(h.verbose, "releasing nvcuda library\n");
LOG(h.verbose, "releasing cuda driver library\n");
UNLOAD_LIBRARY(h.handle);
// TODO and other context release logic?
h.handle = NULL;

6
gpu/gpu_info_nvcuda.h
View file

@ -7,9 +7,12 @@
typedef enum cudaError_enum {
CUDA_SUCCESS = 0,
CUDA_ERROR_INVALID_VALUE = 1,
CUDA_ERROR_MEMORY_ALLOCATION = 2,
CUDA_ERROR_OUT_OF_MEMORY = 2,
CUDA_ERROR_NOT_INITIALIZED = 3,
CUDA_ERROR_INSUFFICIENT_DRIVER = 35,
CUDA_ERROR_NO_DEVICE = 100,
CUDA_ERROR_SYSTEM_DRIVER_MISMATCH = 803,
CUDA_ERROR_UNKNOWN = 999,
// Other values omitted for now...
} CUresult;
@ -64,6 +67,7 @@ typedef struct nvcuda_init_resp {
char *err; // If err is non-null handle is invalid
nvcuda_handle_t ch;
int num_devices;
CUresult cudaErr;
} nvcuda_init_resp_t;
void nvcuda_init(char *nvcuda_lib_path, nvcuda_init_resp_t *resp);

5
gpu/types.go
View file

@ -29,6 +29,11 @@ type GpuInfo struct {
// Extra environment variables specific to the GPU as list of [key,value]
EnvWorkarounds [][2]string `json:"envs,omitempty"`
// Set to true if we can NOT reliably discover FreeMemory. A value of true indicates
// the FreeMemory is best effort, and may over or under report actual memory usage
// False indicates FreeMemory can generally be trusted on this GPU
UnreliableFreeMemory bool
// GPU information
ID string `json:"gpu_id"` // string to use for selection of this specific GPU
Name string `json:"name"` // user friendly name if available

25
llm/ext_server/CMakeLists.txt vendored
View file

@ -1,14 +1,13 @@
set(TARGET ollama_llama_server)
option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
include_directories(${CMAKE_CURRENT_SOURCE_DIR})
add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h)
install(TARGETS ${TARGET} RUNTIME)
target_compile_definitions(${TARGET} PRIVATE
SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
)
target_link_libraries(${TARGET} PRIVATE common llava ${CMAKE_THREAD_LIBS_INIT})
if (WIN32)
TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
endif()
set(TARGET ollama_llama_server)
option(LLAMA_SERVER_VERBOSE "Build verbose logging option for Server" ON)
include_directories(${CMAKE_CURRENT_SOURCE_DIR})
add_executable(${TARGET} server.cpp utils.hpp json.hpp httplib.h)
install(TARGETS ${TARGET} RUNTIME)
target_compile_definitions(${TARGET} PRIVATE
SERVER_VERBOSE=$<BOOL:${LLAMA_SERVER_VERBOSE}>
)
target_link_libraries(${TARGET} PRIVATE ggml llama common llava ${CMAKE_THREAD_LIBS_INIT})
if (WIN32)
TARGET_LINK_LIBRARIES(${TARGET} PRIVATE ws2_32)
endif()
target_compile_features(${TARGET} PRIVATE cxx_std_11)

42
llm/ext_server/server.cpp vendored
View file

@ -1382,12 +1382,50 @@ struct llama_server_context
}
}
std::string common_prefix(const std::string& str1, const std::string& str2) {
auto mismatch_pair = std::mismatch(str1.begin(), str1.end(), str2.begin());
return std::string(str1.begin(), mismatch_pair.first);
}
// Find the slot that has the greatest common prefix
server_slot *prefix_slot(const json &prompt) {
if (!prompt.is_string()) {
return nullptr;
}
std::string prompt_str = prompt.get<std::string>();
server_slot *slot = nullptr;
size_t longest = 0;
for (server_slot &s : slots) {
if (s.available() && s.prompt.is_string()) {
std::string s_prompt = s.prompt.get<std::string>();
std::string prefix = common_prefix(s_prompt, prompt_str);
if (prefix.size() > longest) {
slot = &s;
longest = prefix.size();
}
}
}
if (!slot) {
return get_slot(-1);
}
LOG_INFO("slot with common prefix found", {{
"slot_id", slot->id,
"characters", longest
}});
return slot;
}
void process_single_task(task_server& task)
{
switch (task.type)
{
case TASK_TYPE_COMPLETION: {
server_slot *slot = get_slot(json_value(task.data, "slot_id", -1));
server_slot *slot = prefix_slot(task.data["prompt"]);
if (slot == nullptr)
{
// if no slot is available, we defer this task for processing later
@ -1732,7 +1770,7 @@ struct llama_server_context
slot.n_past -= 1;
}
slot.n_prompt_tokens_processed = slot.n_prompt_tokens - slot.n_past;
slot.n_prompt_tokens_processed = slot.n_prompt_tokens;
if (slot.ga_n != 1)
{

16
llm/generate/gen_darwin.sh
View file

@ -18,16 +18,16 @@ sign() {
fi
}
COMMON_DARWIN_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DLLAMA_METAL_MACOSX_VERSION_MIN=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DLLAMA_METAL_EMBED_LIBRARY=on -DLLAMA_OPENMP=off"
COMMON_DARWIN_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DLLAMA_METAL_MACOSX_VERSION_MIN=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DGGML_METAL_EMBED_LIBRARY=on -DGGML_OPENMP=off"
case "${GOARCH}" in
"amd64")
COMMON_CPU_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=off -DLLAMA_NATIVE=off"
COMMON_CPU_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DGGML_METAL=off -DGGML_NATIVE=off"
# Static build for linking into the Go binary
init_vars
CMAKE_TARGETS="--target llama --target ggml"
CMAKE_DEFS="${COMMON_CPU_DEFS} -DBUILD_SHARED_LIBS=off -DLLAMA_BLAS=off -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_BLAS=off -DGGML_ACCELERATE=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}_static"
echo "Building static library"
build
@ -37,7 +37,7 @@ case "${GOARCH}" in
# CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_BLAS=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=off -DGGML_BLAS=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}/cpu"
echo "Building LCD CPU"
build
@ -49,7 +49,7 @@ case "${GOARCH}" in
# Approximately 400% faster than LCD on same CPU
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=off -DLLAMA_BLAS=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=off -DGGML_BLAS=off -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}/cpu_avx"
echo "Building AVX CPU"
build
@ -61,7 +61,7 @@ case "${GOARCH}" in
# Approximately 10% faster than AVX on same CPU
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_ACCELERATE=on -DLLAMA_BLAS=off -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_ACCELERATE=on -DGGML_BLAS=off -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}/cpu_avx2"
echo "Building AVX2 CPU"
EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation"
@ -75,14 +75,14 @@ case "${GOARCH}" in
# Static build for linking into the Go binary
init_vars
CMAKE_TARGETS="--target llama --target ggml"
CMAKE_DEFS="-DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DLLAMA_BLAS=off -DCMAKE_SYSTEM_NAME=Darwin -DBUILD_SHARED_LIBS=off -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=off -DLLAMA_ACCELERATE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_OSX_DEPLOYMENT_TARGET=11.3 -DCMAKE_SYSTEM_NAME=Darwin -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}_static"
echo "Building static library"
build
if [ -z "$OLLAMA_SKIP_METAL_GENERATE" ]; then
init_vars
CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DLLAMA_ACCELERATE=on -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} -DLLAMA_METAL=on ${CMAKE_DEFS}"
CMAKE_DEFS="${COMMON_DARWIN_DEFS} -DCMAKE_SYSTEM_PROCESSOR=${ARCH} -DCMAKE_OSX_ARCHITECTURES=${ARCH} ${CMAKE_DEFS}"
BUILD_DIR="../build/darwin/${ARCH}/metal"
EXTRA_LIBS="${EXTRA_LIBS} -framework Accelerate -framework Foundation -framework Metal -framework MetalKit -framework MetalPerformanceShaders"
build

36
llm/generate/gen_linux.sh
View file

@ -51,7 +51,7 @@ if [ -z "${CUDACXX}" ]; then
export CUDACXX=$(command -v nvcc)
fi
fi
COMMON_CMAKE_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off -DLLAMA_OPENMP=off"
COMMON_CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_OPENMP=off"
source $(dirname $0)/gen_common.sh
init_vars
git_module_setup
@ -64,7 +64,7 @@ if [ -z "${OLLAMA_SKIP_STATIC_GENERATE}" -o "${OLLAMA_CPU_TARGET}" = "static" ];
# Static build for linking into the Go binary
init_vars
CMAKE_TARGETS="--target llama --target ggml"
CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DLLAMA_NATIVE=off -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off -DLLAMA_OPENMP=off ${CMAKE_DEFS}"
CMAKE_DEFS="-DBUILD_SHARED_LIBS=off -DGGML_NATIVE=off -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off -DGGML_OPENMP=off ${CMAKE_DEFS}"
BUILD_DIR="../build/linux/${ARCH}_static"
echo "Building static library"
build
@ -84,22 +84,22 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
compress
else
# Darwin Rosetta x86 emulation does NOT support AVX, AVX2, AVX512
# -DLLAMA_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
# -DLLAMA_F16C -- 2012 Intel Ivy Bridge & AMD 2011 Bulldozer (No significant improvement over just AVX)
# -DLLAMA_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
# -DLLAMA_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
# -DGGML_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
# -DGGML_F16C -- 2012 Intel Ivy Bridge & AMD 2011 Bulldozer (No significant improvement over just AVX)
# -DGGML_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
# -DGGML_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
# Note: the following seem to yield slower results than AVX2 - ymmv
# -DLLAMA_AVX512 -- 2017 Intel Skylake and High End DeskTop (HEDT)
# -DLLAMA_AVX512_VBMI -- 2018 Intel Cannon Lake
# -DLLAMA_AVX512_VNNI -- 2021 Intel Alder Lake
# -DGGML_AVX512 -- 2017 Intel Skylake and High End DeskTop (HEDT)
# -DGGML_AVX512_VBMI -- 2018 Intel Cannon Lake
# -DGGML_AVX512_VNNI -- 2021 Intel Alder Lake
COMMON_CPU_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DLLAMA_NATIVE=off -DLLAMA_OPENMP=off"
COMMON_CPU_DEFS="-DCMAKE_POSITION_INDEPENDENT_CODE=on -DGGML_NATIVE=off -DGGML_OPENMP=off"
if [ -z "${OLLAMA_CPU_TARGET}" -o "${OLLAMA_CPU_TARGET}" = "cpu" ]; then
#
# CPU first for the default library, set up as lowest common denominator for maximum compatibility (including Rosetta)
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
BUILD_DIR="../build/linux/${ARCH}/cpu"
echo "Building LCD CPU"
build
@ -116,7 +116,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
# Approximately 400% faster than LCD on same CPU
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=on -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_FMA=off -DLLAMA_F16C=off ${CMAKE_DEFS}"
CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_FMA=off -DGGML_F16C=off ${CMAKE_DEFS}"
BUILD_DIR="../build/linux/${ARCH}/cpu_avx"
echo "Building AVX CPU"
build
@ -129,7 +129,7 @@ if [ -z "${OLLAMA_SKIP_CPU_GENERATE}" ]; then
# Approximately 10% faster than AVX on same CPU
#
init_vars
CMAKE_DEFS="${COMMON_CPU_DEFS} -DLLAMA_AVX=on -DLLAMA_AVX2=on -DLLAMA_AVX512=off -DLLAMA_FMA=on -DLLAMA_F16C=on ${CMAKE_DEFS}"
CMAKE_DEFS="${COMMON_CPU_DEFS} -DGGML_AVX=on -DGGML_AVX2=on -DGGML_AVX512=off -DGGML_FMA=on -DGGML_F16C=on ${CMAKE_DEFS}"
BUILD_DIR="../build/linux/${ARCH}/cpu_avx2"
echo "Building AVX2 CPU"
build
@ -170,15 +170,15 @@ if [ -z "${OLLAMA_SKIP_CUDA_GENERATE}" -a -d "${CUDA_LIB_DIR}" ]; then
#
# CUDA compute < 6.0 lacks proper FP16 support on ARM.
# Disabling has minimal performance effect while maintaining compatibility.
ARM64_DEFS="-DLLAMA_AVX=off -DLLAMA_AVX2=off -DLLAMA_AVX512=off -DLLAMA_CUDA_F16=off"
ARM64_DEFS="-DGGML_AVX=off -DGGML_AVX2=off -DGGML_AVX512=off -DGGML_CUDA_F16=off"
fi
# Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
if [ -n "${OLLAMA_CUSTOM_CUDA_DEFS}" ]; then
echo "OLLAMA_CUSTOM_CUDA_DEFS=\"${OLLAMA_CUSTOM_CUDA_DEFS}\""
CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} ${OLLAMA_CUSTOM_CUDA_DEFS}"
echo "Building custom CUDA GPU"
else
CMAKE_CUDA_DEFS="-DLLAMA_CUDA=on -DCMAKE_CUDA_FLAGS=-t8 -DLLAMA_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES}"
CMAKE_CUDA_DEFS="-DGGML_CUDA=on -DCMAKE_CUDA_FLAGS=-t8 -DGGML_CUDA_FORCE_MMQ=on -DCMAKE_CUDA_ARCHITECTURES=${CMAKE_CUDA_ARCHITECTURES} -DCMAKE_LIBRARY_PATH=/usr/local/cuda/compat"
fi
CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} ${ARM64_DEFS} ${CMAKE_CUDA_DEFS}"
BUILD_DIR="../build/linux/${ARCH}/cuda${CUDA_VARIANT}"
@ -216,7 +216,7 @@ if [ -z "${OLLAMA_SKIP_ONEAPI_GENERATE}" -a -d "${ONEAPI_ROOT}" ]; then
init_vars
source ${ONEAPI_ROOT}/setvars.sh --force # set up environment variables for oneAPI
CC=icx
CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DLLAMA_SYCL=ON -DLLAMA_SYCL_F16=OFF"
CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DCMAKE_C_COMPILER=icx -DCMAKE_CXX_COMPILER=icpx -DGGML_SYCL=ON -DGGML_SYCL_F16=OFF"
BUILD_DIR="../build/linux/${ARCH}/oneapi"
EXTRA_LIBS="-fsycl -Wl,-rpath,${ONEAPI_ROOT}/compiler/latest/lib,-rpath,${ONEAPI_ROOT}/mkl/latest/lib,-rpath,${ONEAPI_ROOT}/tbb/latest/lib,-rpath,${ONEAPI_ROOT}/compiler/latest/opt/oclfpga/linux64/lib -lOpenCL -lmkl_core -lmkl_sycl_blas -lmkl_intel_ilp64 -lmkl_tbb_thread -ltbb"
DEBUG_FLAGS="" # icx compiles with -O0 if we pass -g, so we must remove it
@ -254,7 +254,7 @@ if [ -z "${OLLAMA_SKIP_ROCM_GENERATE}" -a -d "${ROCM_PATH}" ]; then
ROCM_VARIANT=_v$(ls ${ROCM_PATH}/lib/librocblas.so.*.*.????? | cut -f5 -d. || true)
fi
init_vars
CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DLLAMA_HIPBLAS=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
CMAKE_DEFS="${COMMON_CMAKE_DEFS} ${CMAKE_DEFS} -DGGML_HIPBLAS=on -DCMAKE_C_COMPILER=$ROCM_PATH/llvm/bin/clang -DCMAKE_CXX_COMPILER=$ROCM_PATH/llvm/bin/clang++ -DAMDGPU_TARGETS=$(amdGPUs) -DGPU_TARGETS=$(amdGPUs)"
# Users building from source can tune the exact flags we pass to cmake for configuring llama.cpp
if [ -n "${OLLAMA_CUSTOM_ROCM_DEFS}" ]; then
echo "OLLAMA_CUSTOM_ROCM_DEFS=\"${OLLAMA_CUSTOM_ROCM_DEFS}\""

44
llm/generate/gen_windows.ps1
View file

@ -39,8 +39,8 @@ function init_vars {
}
$script:cmakeDefs = @(
"-DBUILD_SHARED_LIBS=on",
"-DLLAMA_NATIVE=off",
"-DLLAMA_OPENMP=off"
"-DGGML_NATIVE=off",
"-DGGML_OPENMP=off"
)
$script:commonCpuDefs = @("-DCMAKE_POSITION_INDEPENDENT_CODE=on")
$script:ARCH = $Env:PROCESSOR_ARCHITECTURE.ToLower()
@ -182,9 +182,9 @@ function cleanup {
}
# -DLLAMA_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
# -DLLAMA_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
# -DLLAMA_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
# -DGGML_AVX -- 2011 Intel Sandy Bridge & AMD Bulldozer
# -DGGML_AVX2 -- 2013 Intel Haswell & 2015 AMD Excavator / 2017 AMD Zen
# -DGGML_FMA (FMA3) -- 2013 Intel Haswell & 2012 AMD Piledriver
function build_static() {
@ -204,13 +204,13 @@ function build_static() {
"-DCMAKE_C_COMPILER=gcc.exe",
"-DCMAKE_CXX_COMPILER=g++.exe",
"-DBUILD_SHARED_LIBS=off",
"-DLLAMA_NATIVE=off",
"-DLLAMA_AVX=off",
"-DLLAMA_AVX2=off",
"-DLLAMA_AVX512=off",
"-DLLAMA_F16C=off",
"-DLLAMA_FMA=off",
"-DLLAMA_OPENMP=off")
"-DGGML_NATIVE=off",
"-DGGML_AVX=off",
"-DGGML_AVX2=off",
"-DGGML_AVX512=off",
"-DGGML_F16C=off",
"-DGGML_FMA=off",
"-DGGML_OPENMP=off")
$script:buildDir="../build/windows/${script:ARCH}_static"
write-host "Building static library"
build
@ -224,7 +224,7 @@ function build_cpu($gen_arch) {
if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu"))) {
# remaining llama.cpp builds use MSVC
init_vars
$script:cmakeDefs = $script:commonCpuDefs + @("-A", $gen_arch, "-DLLAMA_AVX=off", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
$script:cmakeDefs = $script:commonCpuDefs + @("-A", $gen_arch, "-DGGML_AVX=off", "-DGGML_AVX2=off", "-DGGML_AVX512=off", "-DGGML_FMA=off", "-DGGML_F16C=off") + $script:cmakeDefs
$script:buildDir="../build/windows/${script:ARCH}/cpu"
$script:distDir="$script:DIST_BASE\cpu"
write-host "Building LCD CPU"
@ -239,7 +239,7 @@ function build_cpu($gen_arch) {
function build_cpu_avx() {
if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu_avx"))) {
init_vars
$script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=off", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=off", "-DLLAMA_F16C=off") + $script:cmakeDefs
$script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DGGML_AVX=on", "-DGGML_AVX2=off", "-DGGML_AVX512=off", "-DGGML_FMA=off", "-DGGML_F16C=off") + $script:cmakeDefs
$script:buildDir="../build/windows/${script:ARCH}/cpu_avx"
$script:distDir="$script:DIST_BASE\cpu_avx"
write-host "Building AVX CPU"
@ -254,7 +254,7 @@ function build_cpu_avx() {
function build_cpu_avx2() {
if ((-not "${env:OLLAMA_SKIP_CPU_GENERATE}" ) -and ((-not "${env:OLLAMA_CPU_TARGET}") -or ("${env:OLLAMA_CPU_TARGET}" -eq "cpu_avx2"))) {
init_vars
$script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DLLAMA_AVX=on", "-DLLAMA_AVX2=on", "-DLLAMA_AVX512=off", "-DLLAMA_FMA=on", "-DLLAMA_F16C=on") + $script:cmakeDefs
$script:cmakeDefs = $script:commonCpuDefs + @("-A", "x64", "-DGGML_AVX=on", "-DGGML_AVX2=on", "-DGGML_AVX512=off", "-DGGML_FMA=on", "-DGGML_F16C=on") + $script:cmakeDefs
$script:buildDir="../build/windows/${script:ARCH}/cpu_avx2"
$script:distDir="$script:DIST_BASE\cpu_avx2"
write-host "Building AVX2 CPU"
@ -279,9 +279,9 @@ function build_cuda() {
$script:distDir="$script:DIST_BASE\cuda$script:CUDA_VARIANT"
$script:cmakeDefs += @(
"-A", "x64",
"-DLLAMA_CUDA=ON",
"-DLLAMA_AVX=on",
"-DLLAMA_AVX2=off",
"-DGGML_CUDA=ON",
"-DGGML_AVX=on",
"-DGGML_AVX2=off",
"-DCUDAToolkit_INCLUDE_DIR=$script:CUDA_INCLUDE_DIR",
"-DCMAKE_CUDA_FLAGS=-t8",
"-DCMAKE_CUDA_ARCHITECTURES=${script:CMAKE_CUDA_ARCHITECTURES}"
@ -319,7 +319,7 @@ function build_oneapi() {
$script:distDir ="$script:DIST_BASE\oneapi$script:ONEAPI_VARIANT"
$script:cmakeDefs += @(
"-G", "MinGW Makefiles",
"-DLLAMA_SYCL=ON",
"-DGGML_SYCL=ON",
"-DCMAKE_C_COMPILER=icx",
"-DCMAKE_CXX_COMPILER=icx",
"-DCMAKE_BUILD_TYPE=Release"
@ -365,10 +365,10 @@ function build_rocm() {
"-G", "Ninja",
"-DCMAKE_C_COMPILER=clang.exe",
"-DCMAKE_CXX_COMPILER=clang++.exe",
"-DLLAMA_HIPBLAS=on",
"-DGGML_HIPBLAS=on",
"-DHIP_PLATFORM=amd",
"-DLLAMA_AVX=on",
"-DLLAMA_AVX2=off",
"-DGGML_AVX=on",
"-DGGML_AVX2=off",
"-DCMAKE_POSITION_INDEPENDENT_CODE=on",
"-DAMDGPU_TARGETS=$(amdGPUs)",
"-DGPU_TARGETS=$(amdGPUs)"

2
llm/llama.cpp

@ -1 +1 @@
Subproject commit 7c26775adb579e92b59c82e8084c07a1d0f75e9c
Subproject commit d7fd29fff16456ce9c3a23fd2d09a66256b05aff

16
llm/llm.go
View file

@ -1,12 +1,14 @@
package llm
// #cgo CFLAGS: -Illama.cpp
// #cgo darwin,arm64 LDFLAGS: ${SRCDIR}/build/darwin/arm64_static/libllama.a -lstdc++
// #cgo darwin,amd64 LDFLAGS: ${SRCDIR}/build/darwin/x86_64_static/libllama.a -lstdc++
// #cgo windows,amd64 LDFLAGS: ${SRCDIR}/build/windows/amd64_static/libllama.a -static -lstdc++
// #cgo windows,arm64 LDFLAGS: ${SRCDIR}/build/windows/arm64_static/libllama.a -static -lstdc++
// #cgo linux,amd64 LDFLAGS: ${SRCDIR}/build/linux/x86_64_static/libllama.a -lstdc++
// #cgo linux,arm64 LDFLAGS: ${SRCDIR}/build/linux/arm64_static/libllama.a -lstdc++
// #cgo CFLAGS: -Illama.cpp -Illama.cpp/include -Illama.cpp/ggml/include
// #cgo windows LDFLAGS: -static-libstdc++
// #cgo LDFLAGS: -lllama -lggml -lstdc++ -lpthread
// #cgo darwin,arm64 LDFLAGS: -L${SRCDIR}/build/darwin/arm64_static -L${SRCDIR}/build/darwin/arm64_static/src -L${SRCDIR}/build/darwin/arm64_static/ggml/src -framework Accelerate -framework Metal
// #cgo darwin,amd64 LDFLAGS: -L${SRCDIR}/build/darwin/x86_64_static -L${SRCDIR}/build/darwin/x86_64_static/src -L${SRCDIR}/build/darwin/x86_64_static/ggml/src
// #cgo windows,amd64 LDFLAGS: -L${SRCDIR}/build/windows/amd64_static -L${SRCDIR}/build/windows/amd64_static/src -L${SRCDIR}/build/windows/amd64_static/ggml/src
// #cgo windows,arm64 LDFLAGS: -L${SRCDIR}/build/windows/arm64_static -L${SRCDIR}/build/windows/arm64_static/src -L${SRCDIR}/build/windows/arm64_static/ggml/src
// #cgo linux,amd64 LDFLAGS: -L${SRCDIR}/build/linux/x86_64_static -L${SRCDIR}/build/linux/x86_64_static/src -L${SRCDIR}/build/linux/x86_64_static/ggml/src
// #cgo linux,arm64 LDFLAGS: -L${SRCDIR}/build/linux/arm64_static -L${SRCDIR}/build/linux/arm64_static/src -L${SRCDIR}/build/linux/arm64_static/ggml/src
// #include <stdlib.h>
// #include "llama.h"
import "C"

14
llm/patches/01-load-progress.diff
View file

@ -1,8 +1,8 @@
diff --git a/common/common.cpp b/common/common.cpp
index 73ff0e85..6adb1a92 100644
index 2c05a4d4..927f0e3d 100644
--- a/common/common.cpp
+++ b/common/common.cpp
@@ -2447,6 +2447,8 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
@@ -2093,6 +2093,8 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
mparams.use_mmap = params.use_mmap;
mparams.use_mlock = params.use_mlock;
mparams.check_tensors = params.check_tensors;
@ -12,10 +12,10 @@ index 73ff0e85..6adb1a92 100644
mparams.kv_overrides = NULL;
} else {
diff --git a/common/common.h b/common/common.h
index 58ed72f4..0bb2605e 100644
index 65c0ef81..ebca2c77 100644
--- a/common/common.h
+++ b/common/common.h
@@ -180,6 +180,13 @@ struct gpt_params {
@@ -184,6 +184,13 @@ struct gpt_params {
std::string mmproj = ""; // path to multimodal projector
std::vector<std::string> image; // path to image file(s)