Update MLX and MLX-C with threading fixes (#15845)

* Update MLX and MLX-C

* Run MLX CGO work on a locked OS thread

MLX now relies on OS-thread-local execution state for streams, encoders, and caches. Add an mlxthread executor backed by runtime.LockOSThread and route runner initialization, model load, inference, status memory reads, and cleanup through the worker so Go goroutine migration cannot split MLX state across native threads.

Also stop caching default MLX streams before the runner owns the thread and add worker/threaded MLX regression tests.

* mlx: use common status writer

* mlx: bundle missing libjaccl on arm64

Inspired by #15793

* review comments
This commit is contained in:
Daniel Hiltgen
2026-05-03 10:03:14 -07:00
committed by GitHub
parent 9ba5a04914
commit 534342e7e2
29 changed files with 2122 additions and 279 deletions

View File

@@ -65,7 +65,12 @@ func (w *StatusWriter) AppendError(msg string) {
// logs, add a small rolling buffer here to capture those fragments.
var errorPrefixes = []string{
"mlx:",
"MLX:",
"panic:",
"fatal error:",
"error:",
"Error:",
"CUDA error",
"ROCm error",
"cudaMalloc failed",
@@ -79,15 +84,21 @@ var errorPrefixes = []string{
func (w *StatusWriter) Write(b []byte) (int, error) {
var errMsg string
errStart := -1
var errPrefix string
for _, prefix := range errorPrefixes {
if _, after, ok := bytes.Cut(b, []byte(prefix)); ok {
line := after
if j := bytes.IndexByte(line, '\n'); j >= 0 {
line = line[:j]
}
errMsg = prefix + string(bytes.TrimRight(line, " \t\r"))
if i := bytes.Index(b, []byte(prefix)); i >= 0 && (errStart < 0 || i < errStart) {
errStart = i
errPrefix = prefix
}
}
if errStart >= 0 {
line := b[errStart+len(errPrefix):]
if j := bytes.IndexByte(line, '\n'); j >= 0 {
line = line[:j]
}
errMsg = errPrefix + string(bytes.TrimRight(line, " \t\r"))
}
if errMsg != "" {
w.AppendError(errMsg)
}

View File

@@ -1,35 +1,59 @@
package llm
import (
"os"
"io"
"testing"
)
func TestStatusWriterCapturesErrorLine(t *testing.T) {
f, err := os.CreateTemp(t.TempDir(), "status-writer")
if err != nil {
t.Fatal(err)
}
defer f.Close()
w := NewStatusWriter(f)
if _, err := w.Write([]byte("llama_init_from_model: failed to initialize the context: failed to initialize Metal backend\n")); err != nil {
t.Fatal(err)
tests := []struct {
name string
log string
want string
}{
{
name: "llama init",
log: "llama_init_from_model: failed to initialize the context: failed to initialize Metal backend\n",
want: "llama_init_from_model: failed to initialize the context: failed to initialize Metal backend",
},
{
name: "cobra error",
log: "Error: foo baz bar\n",
want: "Error: foo baz bar",
},
{
name: "uppercase mlx",
log: "MLX: there was an error\n",
want: "MLX: there was an error",
},
{
name: "panic header",
log: "time=2026-05-01T15:36:45.053Z level=INFO source=pipeline.go:71 msg=\"peak memory\" size=\"8.26 GiB\"\n" +
"panic: mlx: Failed to compile kernel: nvrtc: error: invalid value for --gpu-architecture (-arch)\n" +
"\t. at /go/src/github.com/ollama/ollama/build/_deps/mlx-c-src/mlx/c/transforms.cpp:15\n\n" +
"goroutine 31 [running]:\n" +
"golang.org/x/sync/errgroup.(*Group).Go.func1()\n" +
"\tgolang.org/x/sync@v0.17.0/errgroup/errgroup.go:93 +0x50\n",
want: "panic: mlx: Failed to compile kernel: nvrtc: error: invalid value for --gpu-architecture (-arch)",
},
}
if got, want := w.LastError(), "llama_init_from_model: failed to initialize the context: failed to initialize Metal backend"; got != want {
t.Fatalf("LastError = %q, want %q", got, want)
for _, tt := range tests {
t.Run(tt.name, func(t *testing.T) {
w := NewStatusWriter(io.Discard)
if _, err := w.Write([]byte(tt.log)); err != nil {
t.Fatal(err)
}
if got := w.LastError(); got != tt.want {
t.Fatalf("LastError = %q, want %q", got, tt.want)
}
})
}
}
func TestStatusWriterAccumulatesErrorLines(t *testing.T) {
f, err := os.CreateTemp(t.TempDir(), "status-writer")
if err != nil {
t.Fatal(err)
}
defer f.Close()
w := NewStatusWriter(f)
w := NewStatusWriter(io.Discard)
if _, err := w.Write([]byte("error: failed to initialize the Metal library\n")); err != nil {
t.Fatal(err)
}