@@ -46,9 +46,11 @@ are detected from their weights.
 Omitting `--audio-vae` still runs the joint diffusion model but produces video without a
 decoded audio track.
 
-`--diffusion-fa` also enables flash attention in the video VAE decoder (a ViT), which removes
-most of its attention cost without touching the text encoder. Set `SD_H3_VAE_FLASH_ATTN=0`
-to decode with the previous mul_mat + softmax attention.
+On CUDA and ROCm, `--diffusion-fa` also enables flash attention in the video VAE decoder (a ViT),
+which removes most of its attention cost without touching the text encoder. Set
+`SD_H3_VAE_FLASH_ATTN=0` to decode with the previous mul_mat + softmax attention. Other backends
+keep that attention by default (on Vulkan the flash attention decode was slower);
+`SD_H3_VAE_FLASH_ATTN=1` turns it on there.
 
 On NVIDIA GPUs from Ampere on, the unmasked DiT and VAE attention runs a pipelined
 long-sequence variant of the ggml-cuda flash attention kernel (up to 4x faster at H3
 
@@ -1178,12 +1178,15 @@ bool StableDiffusionGGML::validate_and_load_runners() {
             high_noise_diffusion_model->set_flash_attention_enabled(true);
         }
     }
-    // The H3 video VAE decoder is a ViT whose f32 L x L attention scores dominate decode time, so --diffusion-fa
-    // covers it too, without switching the text encoder. SD_H3_VAE_FLASH_ATTN=0 restores the old path.
+    // The H3 video VAE decoder is a ViT whose f32 L x L attention scores dominate decode time, so on CUDA and ROCm
+    // --diffusion-fa covers it too, without switching the text encoder. It decoded slower on Vulkan, so other
+    // backends keep the old path unless SD_H3_VAE_FLASH_ATTN=1; SD_H3_VAE_FLASH_ATTN=0 restores it everywhere.
     if (!sd_ctx_params->flash_attn && sd_ctx_params->diffusion_flash_attn && first_stage_model &&
         sd_version_is_minimax_h3(version)) {
-        const char* env = getenv("SD_H3_VAE_FLASH_ATTN");
-        if (env == nullptr || strcmp(env, "0") != 0) {
+        const char* env                 = getenv("SD_H3_VAE_FLASH_ATTN");
+        ggml_backend_t vae_backend      = backend_for(SDBackendModule::VAE);
+        const bool fa_faster_by_default = sd_backend_is(vae_backend, "CUDA") || sd_backend_is(vae_backend, "ROCm");
+        if (env != nullptr ? strcmp(env, "0") != 0 : fa_faster_by_default) {
             LOG_INFO("Using flash attention in the MiniMax-H3 video VAE");
             first_stage_model->set_flash_attention_enabled(true);
         }