From 021d3822eea5926727788f8090f28f4d790febae Mon Sep 17 00:00:00 2001 From: losewayy Date: Tue, 17 Nov 2026 16:28:32 +0800 Subject: [PATCH] perf(vae): batch spatial padding in Hunyuan causal convolutions --- src/model/vae/hunyuan_vae.hpp | 11 +++++++++-- 1 file changed, 9 insertions(+), 2 deletions(-) diff --git a/src/model/vae/hunyuan_vae.hpp b/src/model/vae/hunyuan_vae.hpp index ff2b2bb63..567a70431 100644 --- a/src/model/vae/hunyuan_vae.hpp +++ b/src/model/vae/hunyuan_vae.hpp @@ -140,8 +140,15 @@ namespace Hunyuan { } return input; }; - padded = replicate_pad(padded, 0, pad_w, pad_w); - padded = replicate_pad(padded, 1, pad_h, pad_h); + if (pad_w > 0 || pad_h > 0) { + const int64_t frames = padded->ne[2]; + const int64_t channels = padded->ne[3]; + // Keep spatial padding in one outer slice to avoid per-channel CUDA launches. + padded = ggml_reshape_3d(ctx->ggml_ctx, ggml_ext_cont(ctx, padded), padded->ne[0], padded->ne[1], frames * channels); + padded = replicate_pad(padded, 0, pad_w, pad_w); + padded = replicate_pad(padded, 1, pad_h, pad_h); + padded = ggml_reshape_4d(ctx->ggml_ctx, padded, padded->ne[0], padded->ne[1], frames, channels); + } return conv->forward(ctx, padded); } };