aboutsummaryrefslogtreecommitdiff
path: root/embed_mlx.c
diff options
context:
space:
mode:
authorOnur Gündüz <ogunduz@gmail.com>2026-06-16 08:39:03 +0200
committerOnur Gündüz <ogunduz@gmail.com>2026-06-16 08:39:03 +0200
commit707002111e3c6366d6822a3bb0c646b96534dd31 (patch)
treeaf48baf096be578b48ba89de0a9fcd3b7510b929 /embed_mlx.c
parent187e5e8ad924cde6436ed799a0b8f3a7e0ed2efd (diff)
downloadffwd-707002111e3c6366d6822a3bb0c646b96534dd31.tar.gz
ffwd-707002111e3c6366d6822a3bb0c646b96534dd31.zip
mlx: upcast BERT norm/bias/embedding fields for BF16 models
Diffstat (limited to 'embed_mlx.c')
-rw-r--r--embed_mlx.c13
1 files changed, 13 insertions, 0 deletions
diff --git a/embed_mlx.c b/embed_mlx.c
index 8b615ee..8d9de39 100644
--- a/embed_mlx.c
+++ b/embed_mlx.c
@@ -621,8 +621,21 @@ loaded:
upcast_f32_inplace(&l->k_norm, ctx->stream);
upcast_f32_inplace(&l->input_norm, ctx->stream);
upcast_f32_inplace(&l->post_attn_norm, ctx->stream);
+ /* BERT block biases (no-ops for Qwen, whose arrays are unset); the
+ * two BERT LayerNorm weights reuse input_norm/post_attn_norm above. */
+ upcast_f32_inplace(&l->o_bias, ctx->stream);
+ upcast_f32_inplace(&l->attn_ln_bias, ctx->stream);
+ upcast_f32_inplace(&l->ffn_inter_bias, ctx->stream);
+ upcast_f32_inplace(&l->ffn_out_bias, ctx->stream);
+ upcast_f32_inplace(&l->ffn_ln_bias, ctx->stream);
}
upcast_f32_inplace(&ctx->norm, ctx->stream);
+ /* BERT embedding stage (no-ops for Qwen): learned position and token-type
+ * embeddings and the embedding LayerNorm all feed the F32 residual. */
+ upcast_f32_inplace(&ctx->position_embeddings, ctx->stream);
+ upcast_f32_inplace(&ctx->token_type_embeddings, ctx->stream);
+ upcast_f32_inplace(&ctx->embed_ln_w, ctx->stream);
+ upcast_f32_inplace(&ctx->embed_ln_b, ctx->stream);
}
if (ctx->quantize_bits) {
for (int i = layer_start; i < layer_end; i++) {