fix: sd3.5 large

thxCode · thxCode · commit c2889690df7c · 2024-11-12T13:38:55.000+08:00
Signed-off-by: thxCode &lt;thxcode0824@gmail.com&gt;
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -111,14 +111,13 @@ if(SD_SYCL)
     target_compile_options(${SD_LIB} PRIVATE ${SYCL_COMPILE_OPTIONS})
 endif()
 
-set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)
-
 # see https://github.com/ggerganov/ggml/pull/682
 add_definitions(-DGGML_MAX_NAME=128)
 
 # deps
 # Only add ggml if it hasn't been added yet
 if (NOT TARGET ggml)
+    set(CMAKE_POLICY_DEFAULT_CMP0077 NEW)
     add_subdirectory(ggml)
 endif()
 
diff --git a/examples/convert/main.cpp b/examples/convert/main.cpp
@@ -139,7 +139,7 @@ static bool convert_params_parse(int argc, char** argv, convert_params& params)
                 }
                 const char* outtype    = argv[i++];
                 params.vae_output_type = convert_str_to_ggml_type(outtype);
-                if (params.vae_output_type == GGML_TYPE_COUNT) {
+                if (params.vae_output_type >= GGML_TYPE_COUNT) {
                     invalid("--vae-outtype");
                 }
                 continue;
@@ -151,7 +151,7 @@ static bool convert_params_parse(int argc, char** argv, convert_params& params)
                 }
                 const char* outtype     = argv[i++];
                 params.clip_output_type = convert_str_to_ggml_type(outtype);
-                if (params.clip_output_type == GGML_TYPE_COUNT) {
+                if (params.clip_output_type >= GGML_TYPE_COUNT) {
                     invalid("--clip-outtype");
                 }
                 continue;
@@ -163,7 +163,7 @@ static bool convert_params_parse(int argc, char** argv, convert_params& params)
                 }
                 const char* outtype = argv[i++];
                 params.output_type  = convert_str_to_ggml_type(outtype);
-                if (params.output_type == GGML_TYPE_COUNT) {
+                if (params.output_type >= GGML_TYPE_COUNT) {
                     invalid("--outtype");
                 }
                 continue;
diff --git a/model.cpp b/model.cpp
@@ -1566,20 +1566,7 @@ SDVersion ModelLoader::get_sd_version() {
 }
 
 ggml_type ModelLoader::get_sd_wtype() {
-    for (auto& tensor_storage : tensor_storages) {
-        if (is_unused_tensor(tensor_storage.name)) {
-            continue;
-        }
-
-        if (ggml_is_quantized(tensor_storage.type)) {
-            return tensor_storage.type;
-        }
-
-        if (tensor_should_be_converted(tensor_storage, GGML_TYPE_Q4_K)) {
-            return tensor_storage.type;
-        }
-    }
-    return GGML_TYPE_COUNT;
+    return get_diffusion_model_wtype();
 }
 
 ggml_type ModelLoader::get_conditioner_wtype() {
@@ -1962,13 +1949,17 @@ bool ModelLoader::tensor_should_be_converted(const TensorStorage& tensor_storage
 }
 
 bool ModelLoader::save_to_gguf_file(const std::string& file_path, ggml_type outtype, ggml_type vae_outtype, ggml_type clip_outtype) {
-    if (vae_outtype == GGML_TYPE_COUNT) {
+    if (vae_outtype >= GGML_TYPE_COUNT) {
         vae_outtype = outtype;
     }
-    if (clip_outtype == GGML_TYPE_COUNT) {
+    if (clip_outtype >= GGML_TYPE_COUNT) {
         clip_outtype = outtype;
     }
 
+    LOG_INFO("save conditioner weight type: %s", ggml_type_name(clip_outtype));
+    LOG_INFO("save diffusion model weight type: %s", ggml_type_name(outtype));
+    LOG_INFO("save vae weight type: %s", ggml_type_name(vae_outtype));
+
     auto backend    = ggml_backend_cpu_init();
     size_t mem_size = 1 * 1024 * 1024;  // for padding
     mem_size += tensor_storages.size() * ggml_tensor_overhead();
diff --git a/stable-diffusion.cpp b/stable-diffusion.cpp
@@ -283,7 +283,7 @@ class StableDiffusionGGML {
             model_wtype = model_loader.get_sd_wtype();
             if (model_wtype == GGML_TYPE_COUNT) {
                 model_wtype = GGML_TYPE_F32;
-                LOG_WARN("can not get mode wtype frome weight, use f32");
+                LOG_WARN("can not get mode wtype from weight, use f32");
             }
             conditioner_wtype = model_loader.get_conditioner_wtype();
             if (conditioner_wtype == GGML_TYPE_COUNT) {

Original file line number	Diff line number	Diff line change
`@@ -139,7 +139,7 @@ static bool convert_params_parse(int argc, char** argv, convert_params& params)`
`139`	`139`	`}`
`140`	`140`	`const char* outtype = argv[i++];`
`141`	`141`	`params.vae_output_type = convert_str_to_ggml_type(outtype);`
`142`		`- if (params.vae_output_type == GGML_TYPE_COUNT) {`
	`142`	`+ if (params.vae_output_type >= GGML_TYPE_COUNT) {`
`143`	`143`	`invalid("--vae-outtype");`
`144`	`144`	`}`
`145`	`145`	`continue;`
`@@ -151,7 +151,7 @@ static bool convert_params_parse(int argc, char** argv, convert_params& params)`
`151`	`151`	`}`
`152`	`152`	`const char* outtype = argv[i++];`
`153`	`153`	`params.clip_output_type = convert_str_to_ggml_type(outtype);`
`154`		`- if (params.clip_output_type == GGML_TYPE_COUNT) {`
	`154`	`+ if (params.clip_output_type >= GGML_TYPE_COUNT) {`
`155`	`155`	`invalid("--clip-outtype");`
`156`	`156`	`}`
`157`	`157`	`continue;`
`@@ -163,7 +163,7 @@ static bool convert_params_parse(int argc, char** argv, convert_params& params)`
`163`	`163`	`}`
`164`	`164`	`const char* outtype = argv[i++];`
`165`	`165`	`params.output_type = convert_str_to_ggml_type(outtype);`
`166`		`- if (params.output_type == GGML_TYPE_COUNT) {`
	`166`	`+ if (params.output_type >= GGML_TYPE_COUNT) {`
`167`	`167`	`invalid("--outtype");`
`168`	`168`	`}`
`169`	`169`	`continue;`
Original file line number	Diff line number	Diff line change
`@@ -283,7 +283,7 @@ class StableDiffusionGGML {`
`283`	`283`	`model_wtype = model_loader.get_sd_wtype();`
`284`	`284`	`if (model_wtype == GGML_TYPE_COUNT) {`
`285`	`285`	`model_wtype = GGML_TYPE_F32;`
`286`		`- LOG_WARN("can not get mode wtype frome weight, use f32");`
	`286`	`+ LOG_WARN("can not get mode wtype from weight, use f32");`
`287`	`287`	`}`
`288`	`288`	`conditioner_wtype = model_loader.get_conditioner_wtype();`
`289`	`289`	`if (conditioner_wtype == GGML_TYPE_COUNT) {`