aboutsummaryrefslogtreecommitdiff
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-06-04 20:18:19 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-14 08:27:03 +0200
commitcdbbd670fb1917edd625ff1c19f7a28f2852a930 (patch)
tree7e2b982f73d1f44501ef7860471747b5ebe93ed7
parenta89fc2041de88b2b6fb326b3031a602333068052 (diff)
downloadffwd-cdbbd670fb1917edd625ff1c19f7a28f2852a930.tar.gz
ffwd-cdbbd670fb1917edd625ff1c19f7a28f2852a930.zip
mlx: prepare contiguous linear weights
-rw-r--r--pplx_embed_mlx.c25
1 files changed, 21 insertions, 4 deletions
diff --git a/pplx_embed_mlx.c b/pplx_embed_mlx.c
index bb948a8..e083cb8 100644
--- a/pplx_embed_mlx.c
+++ b/pplx_embed_mlx.c
@@ -173,7 +173,7 @@ static mlx_linear_t load_linear_tensor(multi_safetensors_t *ms,
static int linear_ok(mlx_linear_t l)
{
- return arr_ok(l.w);
+ return arr_ok(l.w) || arr_ok(l.wt);
}
static void free_linear(mlx_linear_t *l)
@@ -190,14 +190,27 @@ static int prepare_linear_transpose(mlx_linear_t *l, mlx_stream s)
if (!l || l->quantized || !arr_ok(l->w) || arr_ok(l->wt))
return 0;
+ mlx_array transposed = mlx_array_new();
+ mlx_transpose(&transposed, l->w, s);
+ if (!arr_ok(transposed)) {
+ mlx_array_free(transposed);
+ fprintf(stderr, "mlx: failed to prepare transposed linear weight\n");
+ return -1;
+ }
+
l->wt = mlx_array_new();
- mlx_transpose(&l->wt, l->w, s);
+ mlx_contiguous(&l->wt, transposed, false, s);
+ mlx_array_free(transposed);
if (!arr_ok(l->wt)) {
mlx_array_free(l->wt);
l->wt = (mlx_array){0};
- fprintf(stderr, "mlx: failed to prepare transposed linear weight\n");
+ fprintf(stderr, "mlx: failed to materialize transposed linear weight\n");
return -1;
}
+ mlx_array_eval(l->wt);
+ mlx_synchronize(s);
+ mlx_array_free(l->w);
+ l->w = (mlx_array){0};
return 0;
}
@@ -259,6 +272,7 @@ static int quantize_linear(mlx_linear_t *l, int bits, int group_size,
static mlx_array linear(mlx_array x, const mlx_linear_t *W, mlx_stream s)
{
if (W->quantized) {
+ if (!arr_ok(W->w)) return (mlx_array){0};
mlx_array res = mlx_array_new();
mlx_optional_int gs = {.value = W->group_size, .has_value = true};
mlx_optional_int bits = {.value = W->bits, .has_value = true};
@@ -271,11 +285,14 @@ static mlx_array linear(mlx_array x, const mlx_linear_t *W, mlx_stream s)
mlx_array res = mlx_array_new();
if (arr_ok(W->wt)) {
mlx_matmul(&res, x, W->wt, s);
- } else {
+ } else if (arr_ok(W->w)) {
mlx_array Wt = mlx_array_new();
mlx_transpose(&Wt, W->w, s);
mlx_matmul(&res, x, Wt, s);
mlx_array_free(Wt);
+ } else {
+ mlx_array_free(res);
+ return (mlx_array){0};
}
return res;
}