feat: reduce CLIP memory usage with no embeddings

The CLIP weights need to be converted to f32 for textual inversions
(fbd42b6), but that increases the
amount of allocated VRAM even when embeddings aren't being used.

Author: wbruna

clip.hpp

@@ -544,9 +544,15 @@ class CLIPEmbeddings : public GGMLBlock {
     int64_t embed_dim;
     int64_t vocab_size;
     int64_t num_positions;
+    bool force_clip_f32;
 
     void init_params(struct ggml_context* ctx, const String2GGMLType& tensor_types = {}, const std::string prefix = "") {
         enum ggml_type token_wtype    = GGML_TYPE_F32;
+        if (!force_clip_f32) {
+            auto tensor_type = tensor_types.find(prefix + "token_embedding.weight");
+            if (tensor_type != tensor_types.end())
+                token_wtype = tensor_type->second;
+        }
         enum ggml_type position_wtype = GGML_TYPE_F32;
 
         params["token_embedding.weight"]    = ggml_new_tensor_2d(ctx, token_wtype, embed_dim, vocab_size);
@@ -556,10 +562,12 @@ class CLIPEmbeddings : public GGMLBlock {
 public:
     CLIPEmbeddings(int64_t embed_dim,
                    int64_t vocab_size    = 49408,
-                   int64_t num_positions = 77)
+                   int64_t num_positions = 77,
+                   bool force_clip_f32   = false)
         : embed_dim(embed_dim),
           vocab_size(vocab_size),
-          num_positions(num_positions) {
+          num_positions(num_positions),
+          force_clip_f32(force_clip_f32) {
     }
 
     struct ggml_tensor* get_token_embed_weight() {
@@ -677,7 +685,8 @@ class CLIPTextModel : public GGMLBlock {
     bool with_final_ln        = true;
 
     CLIPTextModel(CLIPVersion version = OPENAI_CLIP_VIT_L_14,
-                  bool with_final_ln  = true)
+                  bool with_final_ln  = true,
+                  bool force_clip_f32 = false)
         : version(version), with_final_ln(with_final_ln) {
         if (version == OPEN_CLIP_VIT_H_14) {
             hidden_size       = 1024;
@@ -691,7 +700,7 @@ class CLIPTextModel : public GGMLBlock {
             n_layer           = 32;
         }
 
-        blocks["embeddings"]       = std::shared_ptr<GGMLBlock>(new CLIPEmbeddings(hidden_size, vocab_size, n_token));
+        blocks["embeddings"]       = std::shared_ptr<GGMLBlock>(new CLIPEmbeddings(hidden_size, vocab_size, n_token, force_clip_f32));
         blocks["encoder"]          = std::shared_ptr<GGMLBlock>(new CLIPEncoder(n_layer, hidden_size, n_head, intermediate_size));
         blocks["final_layer_norm"] = std::shared_ptr<GGMLBlock>(new LayerNorm(hidden_size));
     }
@@ -862,8 +871,9 @@ struct CLIPTextModelRunner : public GGMLRunner {
                         const String2GGMLType& tensor_types,
                         const std::string prefix,
                         CLIPVersion version = OPENAI_CLIP_VIT_L_14,
-                        bool with_final_ln  = true)
-        : GGMLRunner(backend), model(version, with_final_ln) {
+                        bool with_final_ln  = true,
+                        bool force_clip_f32 = false)
+        : GGMLRunner(backend), model(version, with_final_ln, force_clip_f32) {
         model.init(params_ctx, tensor_types, prefix);
     }
 

conditioner.hpp

@@ -62,13 +62,14 @@ struct FrozenCLIPEmbedderWithCustomWords : public Conditioner {
                                       SDVersion version = VERSION_SD1,
                                       PMVersion pv      = PM_VERSION_1)
         : version(version), pm_version(pv), tokenizer(sd_version_is_sd2(version) ? 0 : 49407), embd_dir(embd_dir) {
+        bool force_clip_f32 = embd_dir.size() > 0;
         if (sd_version_is_sd1(version)) {
-            text_model = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14);
+            text_model = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, true, force_clip_f32);
         } else if (sd_version_is_sd2(version)) {
-            text_model = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPEN_CLIP_VIT_H_14);
+            text_model = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPEN_CLIP_VIT_H_14, true, force_clip_f32);
         } else if (sd_version_is_sdxl(version)) {
-            text_model  = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, false);
-            text_model2 = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.1.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false);
+            text_model  = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.transformer.text_model", OPENAI_CLIP_VIT_L_14, false, force_clip_f32);
+            text_model2 = std::make_shared<CLIPTextModelRunner>(backend, tensor_types, "cond_stage_model.1.transformer.text_model", OPEN_CLIP_VIT_BIGG_14, false, force_clip_f32);
         }
     }