From f463018cd700690336898d109226d489deca036d Mon Sep 17 00:00:00 2001 From: Onur Gündüz Date: Sun, 14 Jun 2026 15:19:12 +0200 Subject: all: add support for Qwen3-Embedding + causal GQA attention --- embed_mlx.h | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'embed_mlx.h') diff --git a/embed_mlx.h b/embed_mlx.h index 267c454..5f6e2c9 100644 --- a/embed_mlx.h +++ b/embed_mlx.h @@ -108,7 +108,7 @@ EMBED_API int embed_mlx_encode_into(embed_mlx_ctx_t *ctx, const int *token_ids, * Compute a true padded dense batch on MLX. * * out_embeddings is caller-provided [batch, hidden_size]. - * Padded tokens are masked out of attention keys and mean pooling. + * Pooling and attention masking follow the loaded model configuration. * Returns 0 on success, -1 on error. */ EMBED_API int embed_mlx_encode_batch(embed_mlx_ctx_t *ctx, const embed_input_t *inputs, -- cgit v1.3.1