llama_cpp 0.12.5 → 0.12.7

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 143fb1bb108c9cc679ed6eddaaca4cb8a52a5321ee4ffd965440a2c92aeeb99e
-  data.tar.gz: f522cbf943f82143d1a4eae679473468a9920a6ef6fe6cf88147b82bc6a1f279
+  metadata.gz: 350a80cc8b804b23ee7b0f4e90604110b09664892d3d7c4217c4cd48c77cf775
+  data.tar.gz: 7a127d3b83cb680969589368eb741c6a2ac6a9765adf9f57dd23c0c1b54ca13d
 SHA512:
-  metadata.gz: 1646833e8e1ffd6dd22d809ce2c4f2b0f3de78d84504713da4e8d5ab1c2b466c5cbc47a3c787297753f6d56656635e12cf522acffbe37253bdae0c57f8cc51c9
-  data.tar.gz: fbbf0372d52ba8862dcc4ff61f590f634cdcde039dc31f09a93ac6cd8e112c34a1c6d567d54a9ec2d0679e1c4ec8c2e8153071c6952f67af34fa0c4ccf49ac76
+  metadata.gz: dbf25eb8f0fd60332eb8452ea400294d5b9b2b09127d0f3c5ef347135f30f565b161123d0f76a8553bcabf9e35db9fac3fff6cdd9df407fb830ab124d0d85d47
+  data.tar.gz: 2bbefd5b502150f052ab556c372c4f37b9cf2de2e22e34f4b2153a3b7ff93d7fca768eec5572d5514d7c46dc2a9c03121487907adc5ede612ecb6cea72de682d

data/CHANGELOG.md

@@ -1,3 +1,20 @@
+## [[0.12.7](https://github.com/yoshoku/llama_cpp.rb/compare/v0.12.6...v0.12.7)] - 2024-02-24
+
+- Bump bundled llama.cpp from b2106 to b2143.
+  - Add constants for file type: `LLAMA_FTYPE_MOSTLY_IQ1_S` and `LLAMA_FTYPE_MOSTLY_IQ4_NL`.
+  - Add constants for pooling type: `LLAMA_POOLING_NONE`, `LLAMA_POOLING_MEAN`, and `LLAMA_POOLING_CLS`.
+  - Add `numa_init` module function to `LLaMACpp`.
+  - Remove unnecessary argument from `backend_init`.
+
+Implementation of llama_chat_apply_template binding has been postponed for the time being.
+
+## [[0.12.6](https://github.com/yoshoku/llama_cpp.rb/compare/v0.12.5...v0.12.6)] - 2024-02-17
+
+- Bump bundled llama.cpp from b2106 to b2143.
+  - Add constant: `LLAMA_VOCAB_TYPE_WPM`.
+  - Add `do_pooling` accessors to ContextParams.
+  - Add `embeddings_ith` method to Context.
+
 ## [[0.12.5](https://github.com/yoshoku/llama_cpp.rb/compare/v0.12.4...v0.12.5)] - 2024-02-09
 
 - Bump bundled llama.cpp from b2047 to b2106.

data/ext/llama_cpp/llama_cpp.cpp

@@ -978,6 +978,8 @@ public:
     rb_define_method(rb_cLLaMAContextParams, "embedding", RUBY_METHOD_FUNC(_llama_context_params_get_embedding), 0);
     rb_define_method(rb_cLLaMAContextParams, "offload_kqv=", RUBY_METHOD_FUNC(_llama_context_params_set_offload_kqv), 1);
     rb_define_method(rb_cLLaMAContextParams, "offload_kqv", RUBY_METHOD_FUNC(_llama_context_params_get_offload_kqv), 0);
+    rb_define_method(rb_cLLaMAContextParams, "do_pooling=", RUBY_METHOD_FUNC(_llama_context_params_set_do_pooling), 1);
+    rb_define_method(rb_cLLaMAContextParams, "do_pooling", RUBY_METHOD_FUNC(_llama_context_params_get_do_pooling), 0);
   }
 
 private:
@@ -1220,6 +1222,18 @@ private:
     LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
     return ptr->params.offload_kqv ? Qtrue : Qfalse;
   }
+
+  // do_pooling
+  static VALUE _llama_context_params_set_do_pooling(VALUE self, VALUE do_pooling) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    ptr->params.do_pooling = RTEST(do_pooling) ? true : false;
+    return ptr->params.do_pooling ? Qtrue : Qfalse;
+  }
+
+  static VALUE _llama_context_params_get_do_pooling(VALUE self) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    return ptr->params.do_pooling ? Qtrue : Qfalse;
+  }
 };
 
 const rb_data_type_t RbLLaMAContextParams::llama_context_params_type = {
@@ -2029,6 +2043,7 @@ public:
     rb_define_method(rb_cLLaMAContext, "decode", RUBY_METHOD_FUNC(_llama_context_decode), 1);
     rb_define_method(rb_cLLaMAContext, "logits", RUBY_METHOD_FUNC(_llama_context_logits), 0);
     rb_define_method(rb_cLLaMAContext, "embeddings", RUBY_METHOD_FUNC(_llama_context_embeddings), 0);
+    rb_define_method(rb_cLLaMAContext, "embeddings_ith", RUBY_METHOD_FUNC(_llama_context_embeddings_ith), 1);
     rb_define_method(rb_cLLaMAContext, "n_ctx", RUBY_METHOD_FUNC(_llama_context_n_ctx), 0);
     rb_define_method(rb_cLLaMAContext, "n_batch", RUBY_METHOD_FUNC(_llama_context_n_batch), 0);
     rb_define_method(rb_cLLaMAContext, "timings", RUBY_METHOD_FUNC(_llama_context_get_timings), 0);
@@ -2286,6 +2301,36 @@ private:
     return output;
   }
 
+  static VALUE _llama_context_embeddings_ith(VALUE self, VALUE ith) {
+    if (!RB_INTEGER_TYPE_P(ith)) {
+      rb_raise(rb_eArgError, "ith must be an integer");
+      return Qnil;
+    }
+    LLaMAContextWrapper* ptr = get_llama_context(self);
+    if (ptr->ctx == NULL) {
+      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
+      return Qnil;
+    }
+    VALUE params = rb_iv_get(self, "@params");
+    LLaMAContextParamsWrapper* prms_ptr = RbLLaMAContextParams::get_llama_context_params(params);
+    if (!prms_ptr->params.embedding) {
+      rb_raise(rb_eRuntimeError, "embedding parameter is false");
+      return Qnil;
+    }
+
+    VALUE model = rb_iv_get(self, "@model");
+    LLaMAModelWrapper* model_ptr = RbLLaMAModel::get_llama_model(model);
+    const int n_embd = llama_n_embd(model_ptr->model);
+
+    VALUE output = rb_ary_new();
+    const float* embd = llama_get_embeddings_ith(ptr->ctx, NUM2INT(ith));
+    for (int i = 0; i < n_embd; i++) {
+      rb_ary_push(output, DBL2NUM((double)(embd[i])));
+    }
+
+    return output;
+  }
+
   static VALUE _llama_context_n_ctx(VALUE self) {
     LLaMAContextWrapper* ptr = get_llama_context(self);
     if (ptr->ctx == NULL) {
@@ -3198,15 +3243,8 @@ const rb_data_type_t RbLLaMAContext::llama_context_type = {
 
 // module functions
 
-static VALUE rb_llama_llama_backend_init(int argc, VALUE* argv, VALUE self) {
-  VALUE kw_args = Qnil;
-  ID kw_table[1] = { rb_intern("numa") };
-  VALUE kw_values[1] = { Qundef };
-  rb_scan_args(argc, argv, ":", &kw_args);
-  rb_get_kwargs(kw_args, kw_table, 0, 1, kw_values);
-
-  const bool numa = kw_values[0] == Qundef ? false : (RTEST(kw_values[0]) ? true : false);
-  llama_backend_init(numa);
+static VALUE rb_llama_llama_backend_init(VALUE self) {
+  llama_backend_init();
 
   return Qnil;
 }
@@ -3217,6 +3255,17 @@ static VALUE rb_llama_llama_backend_free(VALUE self) {
   return Qnil;
 }
 
+static VALUE rb_llama_llama_numa_init(VALUE self, VALUE strategy) {
+  if (!RB_INTEGER_TYPE_P(strategy)) {
+    rb_raise(rb_eArgError, "strategy must be an integer");
+    return Qnil;
+  }
+
+  llama_numa_init(static_cast<enum ggml_numa_strategy>(NUM2INT(strategy)));
+
+  return Qnil;
+}
+
 static VALUE rb_llama_model_quantize(int argc, VALUE* argv, VALUE self) {
   VALUE kw_args = Qnil;
   ID kw_table[3] = { rb_intern("input_path"), rb_intern("output_path"), rb_intern("params") };
@@ -3300,8 +3349,9 @@ extern "C" void Init_llama_cpp(void) {
   RbLLaMAGrammarElement::define_class(rb_mLLaMACpp);
   RbLLaMAGrammar::define_class(rb_mLLaMACpp);
 
-  rb_define_module_function(rb_mLLaMACpp, "backend_init", rb_llama_llama_backend_init, -1);
+  rb_define_module_function(rb_mLLaMACpp, "backend_init", rb_llama_llama_backend_init, 0);
   rb_define_module_function(rb_mLLaMACpp, "backend_free", rb_llama_llama_backend_free, 0);
+  rb_define_module_function(rb_mLLaMACpp, "numa_init", rb_llama_llama_numa_init, 1);
   rb_define_module_function(rb_mLLaMACpp, "model_quantize", rb_llama_model_quantize, -1);
   rb_define_module_function(rb_mLLaMACpp, "print_system_info", rb_llama_print_system_info, 0);
   rb_define_module_function(rb_mLLaMACpp, "time_us", rb_llama_time_us, 0);
@@ -3314,6 +3364,7 @@ extern "C" void Init_llama_cpp(void) {
 
   rb_define_const(rb_mLLaMACpp, "LLAMA_VOCAB_TYPE_SPM", INT2NUM(LLAMA_VOCAB_TYPE_SPM));
   rb_define_const(rb_mLLaMACpp, "LLAMA_VOCAB_TYPE_BPE", INT2NUM(LLAMA_VOCAB_TYPE_BPE));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_VOCAB_TYPE_WPM", INT2NUM(LLAMA_VOCAB_TYPE_WPM));
 
   rb_define_const(rb_mLLaMACpp, "LLAMA_TOKEN_TYPE_UNDEFINED", INT2NUM(LLAMA_TOKEN_TYPE_UNDEFINED));
   rb_define_const(rb_mLLaMACpp, "LLAMA_TOKEN_TYPE_NORMAL", INT2NUM(LLAMA_TOKEN_TYPE_NORMAL));
@@ -3345,6 +3396,8 @@ extern "C" void Init_llama_cpp(void) {
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_Q2_K_S", INT2NUM(LLAMA_FTYPE_MOSTLY_Q2_K_S));
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_Q3_K_XS", INT2NUM(LLAMA_FTYPE_MOSTLY_Q3_K_XS));
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_IQ3_XXS", INT2NUM(LLAMA_FTYPE_MOSTLY_IQ3_XXS));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_IQ1_S", INT2NUM(LLAMA_FTYPE_MOSTLY_IQ1_S));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_IQ4_NL", INT2NUM(LLAMA_FTYPE_MOSTLY_IQ4_NL));
 
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_GUESSED", INT2NUM(LLAMA_FTYPE_GUESSED));
 
@@ -3366,6 +3419,10 @@ extern "C" void Init_llama_cpp(void) {
   rb_define_const(rb_mLLaMACpp, "LLAMA_ROPE_SCALING_YARN", INT2NUM(LLAMA_ROPE_SCALING_YARN));
   rb_define_const(rb_mLLaMACpp, "LLAMA_ROPE_SCALING_MAX_VALUE", INT2NUM(LLAMA_ROPE_SCALING_MAX_VALUE));
 
+  rb_define_const(rb_mLLaMACpp, "LLAMA_POOLING_NONE", INT2NUM(LLAMA_POOLING_NONE));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_POOLING_MEAN", INT2NUM(LLAMA_POOLING_MEAN));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_POOLING_CLS", INT2NUM(LLAMA_POOLING_CLS));
+
   rb_define_const(rb_mLLaMACpp, "LLAMA_SPLIT_NONE", INT2NUM(LLAMA_SPLIT_NONE));
   rb_define_const(rb_mLLaMACpp, "LLAMA_SPLIT_LAYER", INT2NUM(LLAMA_SPLIT_LAYER));
   rb_define_const(rb_mLLaMACpp, "LLAMA_SPLIT_ROW", INT2NUM(LLAMA_SPLIT_ROW));

data/lib/llama_cpp/version.rb

@@ -3,8 +3,8 @@
 # llama_cpp.rb provides Ruby bindings for the llama.cpp.
 module LLaMACpp
   # The version of llama_cpp.rb you install.
-  VERSION = '0.12.5'
+  VERSION = '0.12.7'
 
   # The version of llama.cpp bundled with llama_cpp.rb.
-  LLAMA_CPP_VERSION = 'b2106'
+  LLAMA_CPP_VERSION = 'b2249'
 end

data/sig/llama_cpp.rbs

@@ -3,6 +3,10 @@ module LLaMACpp
   LLAMA_CPP_VERSION: String
   LLAMA_DEFALUT_SEED: String
 
+  LLAMA_VOCAB_TYPE_SPM: Integer
+  LLAMA_VOCAB_TYPE_BPE: Integer
+  LLAMA_VOCAB_TYPE_WPM: Integer
+
   LLAMA_FTYPE_ALL_F32: Integer
   LLAMA_FTYPE_MOSTLY_F16: Integer
   LLAMA_FTYPE_MOSTLY_Q4_0: Integer
@@ -25,6 +29,8 @@ module LLaMACpp
   LLAMA_FTYPE_MOSTLY_Q2_K_S: Integer
   LLAMA_FTYPE_MOSTLY_Q3_K_XS: Integer
   LLAMA_FTYPE_MOSTLY_IQ3_XXS: Integer
+  LLAMA_FTYPE_MOSTLY_IQ1_S: Integer
+  LLAMA_FTYPE_MOSTLY_IQ4_NL: Integer
 
   LLAMA_KV_OVERRIDE_INT: Integer
   LLAMA_KV_OVERRIDE_FLOAT: Integer
@@ -44,12 +50,17 @@ module LLaMACpp
   LLAMA_ROPE_SCALING_YARN: Integer
   LLAMA_ROPE_SCALING_MAX_VALUE: Integer
 
+  LLAMA_POOLING_NONE: Integer
+  LLAMA_POOLING_MEAN: Integer
+  LLAMA_POOLING_CLS: Integer
+
   LLAMA_SPLIT_NONE: Integer
   LLAMA_SPLIT_LAYER: Integer
   LLAMA_SPLIT_ROW: Integer
 
-  def self?.backend_init: (?numa: bool) -> void
+  def self?.backend_init: () -> void
   def self?.backend_free: () -> void
+  def self?.numa_init: (Integer) -> void
   def self?.model_quantize: (input_path: String, output_path: String, params: ModelQuantizeParams) -> void
   def self?.generate: (::LLaMACpp::Context, String,
     ?n_predict: Integer, ?n_threads: Integer, ?n_keep: Integer, ?n_batch: Integer,
@@ -190,6 +201,7 @@ module LLaMACpp
 
     def initialize: (model: ::LLaMACpp::Model, params: ::LLaMACpp::ContextParams) -> void
     def embeddings: () -> Array[Float]
+    def embeddings_ith: (Integer) -> Array[Float]
     def eval: (tokens: Array[Integer], n_past: Integer, ?n_tokens: Integer) -> void
     def eval_embd: (tokens: Array[Float], n_past: Integer, ?n_tokens: Integer) -> void
     def decode: (::LLaMACpp::Batch) -> void
@@ -270,6 +282,8 @@ module LLaMACpp
     def embedding=: (bool) -> bool
     def offload_kqv: () -> bool
     def offload_kqv=: (bool) -> bool
+    def do_pooling: () -> bool
+    def do_pooling=: (bool) -> bool
   end
 
   class ModelQuantizeParams

data/vendor/tmp/llama.cpp/Makefile

@@ -97,9 +97,10 @@ endif
 #
 
 # keep standard at C11 and C++11
-MK_CPPFLAGS = -I. -Icommon
-MK_CFLAGS   = -std=c11   -fPIC
-MK_CXXFLAGS = -std=c++11 -fPIC
+MK_CPPFLAGS  = -I. -Icommon
+MK_CFLAGS    = -std=c11   -fPIC
+MK_CXXFLAGS  = -std=c++11 -fPIC
+MK_NVCCFLAGS = -std=c++11
 
 # -Ofast tends to produce faster code, but may not be available for some compilers.
 ifdef LLAMA_FAST
@@ -172,7 +173,7 @@ ifdef LLAMA_DEBUG
 	MK_LDFLAGS  += -g
 
 	ifeq ($(UNAME_S),Linux)
-		MK_CXXFLAGS += -Wp,-D_GLIBCXX_ASSERTIONS
+		MK_CPPFLAGS += -D_GLIBCXX_ASSERTIONS
 	endif
 else
 	MK_CPPFLAGS += -DNDEBUG
@@ -215,6 +216,11 @@ MK_CFLAGS    += $(WARN_FLAGS) -Wshadow -Wstrict-prototypes -Wpointer-arith -Wmis
 				-Werror=implicit-function-declaration
 MK_CXXFLAGS  += $(WARN_FLAGS) -Wmissing-declarations -Wmissing-noreturn
 
+ifeq ($(LLAMA_FATAL_WARNINGS),1)
+	MK_CFLAGS   += -Werror
+	MK_CXXFLAGS += -Werror
+endif
+
 # this version of Apple ld64 is buggy
 ifneq '' '$(findstring dyld-1015.7,$(shell $(CC) $(LDFLAGS) -Wl,-v 2>&1))'
 	MK_CPPFLAGS += -DHAVE_BUGGY_APPLE_LINKER
@@ -381,6 +387,9 @@ ifdef LLAMA_CUBLAS
 	MK_LDFLAGS   += -lcuda -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64 -L/opt/cuda/lib64 -L$(CUDA_PATH)/targets/x86_64-linux/lib -L/usr/local/cuda/targets/aarch64-linux/lib -L/usr/lib/wsl/lib
 	OBJS         += ggml-cuda.o
 	MK_NVCCFLAGS += -use_fast_math
+ifdef LLAMA_FATAL_WARNINGS
+	MK_NVCCFLAGS += -Werror all-warnings
+endif # LLAMA_FATAL_WARNINGS
 ifndef JETSON_EOL_MODULE_DETECT
 	MK_NVCCFLAGS += --forward-unknown-to-host-compiler
 endif # JETSON_EOL_MODULE_DETECT
@@ -439,9 +448,9 @@ ifdef LLAMA_CUDA_CCBIN
 endif
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 ifdef JETSON_EOL_MODULE_DETECT
-	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+	$(NVCC) -I. -Icommon -D_XOPEN_SOURCE=600 -D_GNU_SOURCE -DNDEBUG -DGGML_USE_CUBLAS -I/usr/local/cuda/include -I/opt/cuda/include -I/usr/local/cuda/targets/aarch64-linux/include -std=c++11 -O3 $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 else
-	$(NVCC) $(BASE_CXXFLAGS) $(NVCCFLAGS) -Wno-pedantic -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
+	$(NVCC) $(NVCCFLAGS) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 endif # JETSON_EOL_MODULE_DETECT
 endif # LLAMA_CUBLAS
 
@@ -526,11 +535,29 @@ ifdef LLAMA_METAL
 ifdef LLAMA_METAL_NDEBUG
 	MK_CPPFLAGS += -DGGML_METAL_NDEBUG
 endif
+ifdef LLAMA_METAL_EMBED_LIBRARY
+	MK_CPPFLAGS += -DGGML_METAL_EMBED_LIBRARY
+	OBJS        += ggml-metal-embed.o
+endif
 endif # LLAMA_METAL
 
 ifdef LLAMA_METAL
 ggml-metal.o: ggml-metal.m ggml-metal.h
 	$(CC) $(CFLAGS) -c $< -o $@
+
+ifdef LLAMA_METAL_EMBED_LIBRARY
+ggml-metal-embed.o: ggml-metal.metal
+	@echo "Embedding Metal library"
+	$(eval TEMP_ASSEMBLY=$(shell mktemp))
+	@echo ".section __DATA, __ggml_metallib" > $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_start" >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_start:" >> $(TEMP_ASSEMBLY)
+	@echo ".incbin \"$<\"" >> $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_end" >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_end:" >> $(TEMP_ASSEMBLY)
+	@$(AS) $(TEMP_ASSEMBLY) -o $@
+	@rm -f ${TEMP_ASSEMBLY}
+endif
 endif # LLAMA_METAL
 
 ifdef LLAMA_MPI
@@ -542,9 +569,10 @@ GF_CC := $(CC)
 include scripts/get-flags.mk
 
 # combine build flags with cmdline overrides
-override CFLAGS    := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CFLAGS) $(GF_CFLAGS) $(CFLAGS)
-BASE_CXXFLAGS      := $(MK_CPPFLAGS) $(CPPFLAGS) $(MK_CXXFLAGS) $(CXXFLAGS)
-override CXXFLAGS  := $(BASE_CXXFLAGS) $(HOST_CXXFLAGS) $(GF_CXXFLAGS)
+override CPPFLAGS  := $(MK_CPPFLAGS) $(CPPFLAGS)
+override CFLAGS    := $(CPPFLAGS) $(MK_CFLAGS) $(GF_CFLAGS) $(CFLAGS)
+BASE_CXXFLAGS      := $(MK_CXXFLAGS) $(CXXFLAGS)
+override CXXFLAGS  := $(BASE_CXXFLAGS) $(HOST_CXXFLAGS) $(GF_CXXFLAGS) $(CPPFLAGS)
 override NVCCFLAGS := $(MK_NVCCFLAGS) $(NVCCFLAGS)
 override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
 
@@ -552,7 +580,7 @@ override LDFLAGS   := $(MK_LDFLAGS) $(LDFLAGS)
 ifdef LLAMA_CUBLAS
 GF_CC := $(NVCC) $(NVCCFLAGS) 2>/dev/null .c -Xcompiler
 include scripts/get-flags.mk
-CUDA_CXXFLAGS := $(GF_CXXFLAGS)
+CUDA_CXXFLAGS := $(BASE_CXXFLAGS) $(GF_CXXFLAGS) -Wno-pedantic
 endif
 
 #
@@ -571,6 +599,14 @@ $(info I CC:        $(shell $(CC)   --version | head -n 1))
 $(info I CXX:       $(shell $(CXX)  --version | head -n 1))
 ifdef LLAMA_CUBLAS
 $(info I NVCC:      $(shell $(NVCC) --version | tail -n 1))
+CUDA_VERSION := $(shell nvcc --version | grep -oP 'release (\K[0-9]+\.[0-9])')
+ifeq ($(shell awk -v "v=$(CUDA_VERSION)" 'BEGIN { print (v < 11.7) }'),1)
+ifndef CUDA_DOCKER_ARCH
+ifndef CUDA_POWER_ARCH
+$(error I ERROR: For CUDA versions < 11.7 a target CUDA architecture must be explicitly provided via CUDA_DOCKER_ARCH)
+endif # CUDA_POWER_ARCH
+endif # CUDA_DOCKER_ARCH
+endif # eq ($(shell echo "$(CUDA_VERSION) < 11.7" | bc),1)
 endif # LLAMA_CUBLAS
 $(info )
 
@@ -625,7 +661,6 @@ lib: llama.o ggml.o $(OBJS)
 
 clean:
 	rm -vrf *.o tests/*.o *.so *.a *.dll *.dylib benchmark-matmult common/build-info.cpp *.dot $(COV_TARGETS) $(BUILD_TARGETS) $(TEST_TARGETS)
-	find examples pocs -type f -name "*.o" -delete
 
 #
 # Examples
@@ -689,7 +724,7 @@ save-load-state: examples/save-load-state/save-load-state.cpp ggml.o llama.o $(C
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
 
-server: examples/server/server.cpp examples/server/oai.hpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
+server: examples/server/server.cpp examples/server/oai.hpp examples/server/utils.hpp examples/server/httplib.h examples/server/json.hpp examples/server/index.html.hpp examples/server/index.js.hpp examples/server/completion.js.hpp examples/llava/clip.cpp examples/llava/clip.h examples/llava/llava.h examples/llava/llava.cpp common/stb_image.h ggml.o llama.o $(COMMON_DEPS) grammar-parser.o $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) -c examples/llava/clip.cpp -o $(call GET_OBJ_FILE, examples/llava/clip.cpp) -Wno-cast-qual
 	$(CXX) $(CXXFLAGS) -Iexamples/server $(filter-out %.h %.hpp $< examples/llava/clip.cpp,$^) $(call GET_OBJ_FILE, $<) $(call GET_OBJ_FILE, examples/llava/clip.cpp) -o $@ $(LDFLAGS) $(LWINSOCK2)
@@ -860,3 +895,7 @@ tests/test-model-load-cancel: tests/test-model-load-cancel.cpp ggml.o llama.o te
 tests/test-autorelease: tests/test-autorelease.cpp ggml.o llama.o tests/get-model.cpp $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)
+
+tests/test-chat-template: tests/test-chat-template.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) -c $< -o $(call GET_OBJ_FILE, $<)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h $<,$^) $(call GET_OBJ_FILE, $<) -o $@ $(LDFLAGS)