From fd098afc6dc2f284f1b6fda39c912df71854fdaf Mon Sep 17 00:00:00 2001
From: bssrdf <bssrdf@gmail.com>
Date: Tue, 27 Feb 2024 14:49:39 -0500
Subject: [PATCH] bring back freeing pmid lora params buffer; simply pooled
 output of CLIPvision

---
 clip.hpp             |  8 ++++----
 stable-diffusion.cpp | 12 ++++++------
 2 files changed, 10 insertions(+), 10 deletions(-)

diff --git a/clip.hpp b/clip.hpp
index fc171aec..75d6de28 100644
--- a/clip.hpp
+++ b/clip.hpp
@@ -904,10 +904,10 @@ public:
         GGML_ASSERT(x->ne[3] == 1);
         // int64_t max_token_idx  = 0;
         // ggml_tensor* pooled    = ggml_view_1d(ctx, x, x->ne[0], x->nb[1] * max_token_idx);  // assert N == 1
-        x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));
-        ggml_tensor* pooled    = ggml_view_2d(ctx, x, x->ne[0], x->ne[1], x->nb[0], 0);
-        pooled    = ggml_scale(ctx, ggml_cont(ctx, pooled), 1.f);
-        // print_ggml_tensor(pooled, true, "pooled");
+        // x = ggml_cont(ctx, ggml_permute(ctx, x, 0, 2, 1, 3));
+        // ggml_tensor* pooled    = ggml_view_2d(ctx, x, x->ne[0], x->ne[1], x->nb[0], 0);
+        // pooled  = ggml_cont(ctx, pooled);
+        ggml_tensor* pooled    = ggml_cont(ctx, ggml_view_2d(ctx, x, x->ne[0], x->ne[2], x->nb[2], 0));
         return pooled;  // [N, projection_dim]
     }
 
diff --git a/stable-diffusion.cpp b/stable-diffusion.cpp
index 9f2bbead..5708af44 100644
--- a/stable-diffusion.cpp
+++ b/stable-diffusion.cpp
@@ -1641,13 +1641,13 @@ sd_image_t* txt2img(sd_ctx_t* sd_ctx,
     LOG_INFO("apply_loras completed, taking %.2fs", (t1 - t0) * 1.0f / 1000);
 
     if(sd_ctx->sd->stacked_id){
-        // sd_ctx->sd->pmid_lora.apply(sd_ctx->sd->tensors, sd_ctx->sd->n_threads);
+        t0 = ggml_time_ms();
         sd_ctx->sd->pmid_lora->apply(sd_ctx->sd->tensors, sd_ctx->sd->n_threads);
-        // sd_ctx->sd->pmid_lora.free_compute_buffer();
-        // if (sd_ctx->sd->free_params_immediately) {
-        //     sd_ctx->sd->pmid_lora.free_params_buffer();
-        // }
-        LOG_INFO("pmid_lora apply completed");
+        t1 = ggml_time_ms();
+        LOG_INFO("pmid_lora apply completed, taking %.2fs", (t1 - t0) * 1.0f / 1000);
+        if (sd_ctx->sd->free_params_immediately) {
+            sd_ctx->sd->pmid_lora->free_params_buffer();
+        }
     }
 
     struct ggml_init_params params;