From e00becb643b3a99de16462b3a032717a93b983d9 Mon Sep 17 00:00:00 2001 From: Onur Gündüz Date: Wed, 17 Jun 2026 01:21:41 +0200 Subject: all: accept F16 safetensors weights (widen to F32 at load) --- src/late.c | 10 ++++++++-- 1 file changed, 8 insertions(+), 2 deletions(-) (limited to 'src/late.c') diff --git a/src/late.c b/src/late.c index 0829787..6ba52cc 100644 --- a/src/late.c +++ b/src/late.c @@ -96,8 +96,10 @@ embed_late_model_t *embed_late_model_load(const char *model_dir) { if (!tensor_has_supported_shape(late->dense_sf, t, "linear.weight", shape, 2, 1)) goto fail; - late->projection.dtype = t->dtype; - late->projection.data = safetensors_data(late->dense_sf, t); + /* Same F32/BF16-borrow, F16-widen contract as the backbone weights. */ + late->projection = weight_ref_from_tensor(late->dense_sf, t); + if (!late->projection.data) + goto fail; late->token_dim = (int)t->shape[0]; return late; @@ -110,6 +112,10 @@ void embed_late_model_free(embed_late_model_t *model) { if (!model) return; embed_model_free(model->base); + /* Free the projection before the mmap: only an F16-widened ref owns its + * buffer; a borrowed F32/BF16 ref points into dense_sf and must not be freed. */ + if (model->projection.owned) + free((void *)model->projection.data); safetensors_close(model->dense_sf); free(model); } -- cgit v1.3.1