From 142df17c9c296c846131041283c69edd2db754d8 Mon Sep 17 00:00:00 2001 From: Jeff Bolz Date: Wed, 26 Nov 2025 23:32:30 -0600 Subject: [PATCH] vulkan: use a fixed 1KB buffer for the add_rms_fusion opt (#17514) --- ggml/src/ggml-vulkan/ggml-vulkan.cpp | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/ggml/src/ggml-vulkan/ggml-vulkan.cpp b/ggml/src/ggml-vulkan/ggml-vulkan.cpp index 7f2cf795c9..7c7ce1d8e7 100644 --- a/ggml/src/ggml-vulkan/ggml-vulkan.cpp +++ b/ggml/src/ggml-vulkan/ggml-vulkan.cpp @@ -5289,7 +5289,8 @@ static void ggml_vk_init(ggml_backend_vk_context * ctx, size_t idx) { ctx->prealloc_size_x = 0; ctx->prealloc_size_y = 0; ctx->prealloc_size_split_k = 0; - ctx->prealloc_size_add_rms_partials = 0; + // Fixed size of 1KB, for deterministic behavior + ctx->prealloc_size_add_rms_partials = 1024; ctx->fence = ctx->device->device.createFence({}); ctx->almost_ready_fence = ctx->device->device.createFence({}); @@ -13095,7 +13096,6 @@ static ggml_status ggml_backend_vk_graph_compute(ggml_backend_t backend, ggml_cg ctx->fused_ops_write_mask = 0; } - ctx->prealloc_size_add_rms_partials = std::max(ctx->prealloc_size_add_rms_partials, ctx->prealloc_size_add_rms_partials_offset); ctx->last_total_mul_mat_bytes = total_mul_mat_bytes; if (vk_perf_logger_enabled) {