Carry ggml-metal.metal as payload
This commit is contained in:
parent
5646826a79
commit
54dbfa4c4a
44
llm/llama.go
44
llm/llama.go
|
@ -6,8 +6,12 @@ import (
|
|||
_ "embed"
|
||||
"errors"
|
||||
"fmt"
|
||||
"io"
|
||||
"io/fs"
|
||||
"log"
|
||||
"os"
|
||||
"os/exec"
|
||||
"path/filepath"
|
||||
"sync"
|
||||
"time"
|
||||
|
||||
|
@ -116,6 +120,7 @@ type ImageData struct {
|
|||
var (
|
||||
errNvidiaSMI = errors.New("warning: gpu support may not be enabled, check that you have installed GPU drivers: nvidia-smi command failed")
|
||||
errAvailableVRAM = errors.New("not enough VRAM available, falling back to CPU only")
|
||||
payloadMissing = fmt.Errorf("expected payload not included in this build of ollama")
|
||||
)
|
||||
|
||||
// StatusWriter is a writer that captures error messages from the llama runner process
|
||||
|
@ -202,3 +207,42 @@ type EmbeddingRequest struct {
|
|||
type EmbeddingResponse struct {
|
||||
Embedding []float64 `json:"embedding"`
|
||||
}
|
||||
|
||||
func extractLib(workDir, glob string) error {
|
||||
files, err := fs.Glob(libEmbed, glob)
|
||||
if err != nil || len(files) == 0 {
|
||||
return payloadMissing
|
||||
}
|
||||
|
||||
if len(files) != 1 {
|
||||
// Shouldn't happen, but just use the first one we find
|
||||
log.Printf("WARNING: multiple payloads detected - using %s", files[0])
|
||||
}
|
||||
|
||||
srcFile, err := libEmbed.Open(files[0])
|
||||
if err != nil {
|
||||
return fmt.Errorf("read payload %s: %v", files[0], err)
|
||||
}
|
||||
defer srcFile.Close()
|
||||
if err := os.MkdirAll(workDir, 0o755); err != nil {
|
||||
return fmt.Errorf("create payload temp dir %s: %v", workDir, err)
|
||||
}
|
||||
|
||||
destFile := filepath.Join(workDir, filepath.Base(files[0]))
|
||||
|
||||
_, err = os.Stat(destFile)
|
||||
switch {
|
||||
case errors.Is(err, os.ErrNotExist):
|
||||
destFile, err := os.OpenFile(destFile, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
|
||||
if err != nil {
|
||||
return fmt.Errorf("write payload %s: %v", files[0], err)
|
||||
}
|
||||
defer destFile.Close()
|
||||
if _, err := io.Copy(destFile, srcFile); err != nil {
|
||||
return fmt.Errorf("copy payload %s: %v", files[0], err)
|
||||
}
|
||||
case err != nil:
|
||||
return fmt.Errorf("stat payload %s: %v", files[0], err)
|
||||
}
|
||||
return nil
|
||||
}
|
||||
|
|
|
@ -1,18 +1,32 @@
|
|||
package llm
|
||||
|
||||
import (
|
||||
"embed"
|
||||
"fmt"
|
||||
"log"
|
||||
"os"
|
||||
|
||||
"github.com/jmorganca/ollama/api"
|
||||
)
|
||||
|
||||
// no-op stubs for mac
|
||||
//go:embed llama.cpp/gguf/build/*/bin/ggml-metal.metal
|
||||
var libEmbed embed.FS
|
||||
|
||||
func newRocmShimExtServer(model string, adapters, projectors []string, numLayers int64, opts api.Options) (extServer, error) {
|
||||
// should never happen...
|
||||
return nil, fmt.Errorf("ROCM GPUs not supported on Mac")
|
||||
}
|
||||
|
||||
func nativeInit(workDir string) error {
|
||||
func nativeInit(workdir string) error {
|
||||
err := extractLib(workdir, "llama.cpp/gguf/build/*/bin/ggml-metal.metal")
|
||||
if err != nil {
|
||||
if err == payloadMissing {
|
||||
// TODO perhaps consider this a hard failure on arm macs?
|
||||
log.Printf("ggml-meta.metal payload missing")
|
||||
return nil
|
||||
}
|
||||
return err
|
||||
}
|
||||
os.Setenv("GGML_METAL_PATH_RESOURCES", workdir)
|
||||
return nil
|
||||
}
|
||||
|
|
|
@ -14,7 +14,6 @@ import (
|
|||
"embed"
|
||||
"errors"
|
||||
"fmt"
|
||||
"io"
|
||||
"io/fs"
|
||||
"log"
|
||||
"os"
|
||||
|
@ -109,13 +108,15 @@ func (llm *shimExtServer) Close() {
|
|||
}
|
||||
|
||||
func nativeInit(workdir string) error {
|
||||
err := extractLib(workdir)
|
||||
err := extractLib(workdir, "llama.cpp/gguf/build/*/lib/*rocm_server*")
|
||||
if err != nil {
|
||||
if err == RocmShimMissing {
|
||||
log.Printf("%s", err)
|
||||
if err == payloadMissing {
|
||||
log.Printf("%s", RocmShimMissing)
|
||||
return nil
|
||||
}
|
||||
return err
|
||||
} else {
|
||||
ShimPresent = true
|
||||
}
|
||||
|
||||
// Verify we have permissions - either running as root, or we have group access to the driver
|
||||
|
@ -168,44 +169,3 @@ func nativeInit(workdir string) error {
|
|||
}
|
||||
return nil
|
||||
}
|
||||
|
||||
func extractLib(workDir string) error {
|
||||
files, err := fs.Glob(libEmbed, "llama.cpp/gguf/build/*/lib/*rocm_server*")
|
||||
if err != nil || len(files) == 0 {
|
||||
// this is expected, ollama may be compiled without shim library packed in
|
||||
return RocmShimMissing
|
||||
}
|
||||
|
||||
if len(files) != 1 {
|
||||
// Shouldn't happen, but just use the first one we find
|
||||
log.Printf("WARNING: multiple rocm libraries detected - using %s", files[0])
|
||||
}
|
||||
|
||||
srcFile, err := libEmbed.Open(files[0])
|
||||
if err != nil {
|
||||
return fmt.Errorf("read ROCm shim %s: %v", files[0], err)
|
||||
}
|
||||
defer srcFile.Close()
|
||||
if err := os.MkdirAll(workDir, 0o755); err != nil {
|
||||
return fmt.Errorf("create ROCm shim temp dir %s: %v", workDir, err)
|
||||
}
|
||||
|
||||
destFile := filepath.Join(workDir, filepath.Base(files[0]))
|
||||
|
||||
_, err = os.Stat(destFile)
|
||||
switch {
|
||||
case errors.Is(err, os.ErrNotExist):
|
||||
destFile, err := os.OpenFile(destFile, os.O_WRONLY|os.O_CREATE|os.O_TRUNC, 0o755)
|
||||
if err != nil {
|
||||
return fmt.Errorf("write ROCm shim %s: %v", files[0], err)
|
||||
}
|
||||
defer destFile.Close()
|
||||
if _, err := io.Copy(destFile, srcFile); err != nil {
|
||||
return fmt.Errorf("copy ROCm shim %s: %v", files[0], err)
|
||||
}
|
||||
case err != nil:
|
||||
return fmt.Errorf("stat ROCm shim %s: %v", files[0], err)
|
||||
}
|
||||
ShimPresent = true
|
||||
return nil
|
||||
}
|
||||
|
|
Loading…
Reference in a new issue