llama_cpp 0.11.1 → 0.12.1

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 705f8a7e2228a324f14fa23ca093e2ce4408eacb839d891284c40e319b07940a
-  data.tar.gz: fc04d232f2c7ecfa1402aa711eda63e36a03d287a6bc7c8e0d2c791194ad9e9a
+  metadata.gz: 13381408318e71cc1fc55c40ee9be6e62ad9e3ad6a8ce39279bb8040614e9b3b
+  data.tar.gz: 6456734b18865a7811f08d0d9d599771f574f4b59bd5b54a964ece7428115907
 SHA512:
-  metadata.gz: 888ba5bfa23ab51746d49c2cc071f8a220d3de39f6c3a34576f35bcb993fc0be841481dcbca9762504397ca6555571f43b4ba2c0ae3dae3fcd8d29bd2735ae16
-  data.tar.gz: e8940f8ab7a542569a71ad5c869ac781b8ad958ca02d2a5547b792008c228ae1e9ff23cc5b9552e83b0c631805ec7edfea6138c5ae9d922daca06e5ab3f1490d
+  metadata.gz: 1014349771d7aa3c318027de11603e96d5482e4bd5b1bcf0fd4874040245daf44c4cfb801077a698846459a7619ca9e01e0afc3507fc7bd519e7ba68a000a15d
+  data.tar.gz: 1315ca8954397edb0db93347a10762e35f829377ef3dba0ea9cf6c67f986972ac8e75b46c410a3ceceefc0474f2abbe6f441e56a60e789ef1d2617fc15cfb29e

data/CHANGELOG.md

@@ -1,3 +1,18 @@
+## [[0.12.1](https://github.com/yoshoku/llama_cpp.rb/compare/v0.12.0...v0.12.1)] - 2024-01-13
+
+- Bump bundled llama.cpp from b1768 to b1833.
+  - Add model file type constants.
+  - Add `kv_cache_seq_div` method to `Context`.
+
+## [[0.12.0](https://github.com/yoshoku/llama_cpp.rb/compare/v0.11.1...v0.12.0)] - 2024-01-11
+
+- Add `get_one` singleton method to `Batch`.
+
+**Breaking Changes**
+
+- Add deprecation warning to `eval`, `eval_embd`, and `sample_temperature` methods on `Context`.
+- Change to avoid using deprecated methods on `generate` method and example scripts.
+
 ## [[0.11.1](https://github.com/yoshoku/llama_cpp.rb/compare/v0.11.0...v0.11.1)] - 2024-01-08
 
 - Fix to set the values of `@n_tokens` and `@has_evaluated` instance variables in `decode` method of `Context`.

data/README.md

@@ -22,14 +22,14 @@ If bundler is not being used to manage dependencies, install the gem by executin
 $ gem install llama_cpp
 ```
 
-There are several installation options for improving execution performance:
+There are several installation options:
 
 ```sh
 # use OpenBLAS
 $ gem install llama_cpp -- --with-openblas
 
-# use Metal on macOS
-$ gem install llama_cpp -- --with-metal
+# use CUDA
+$ gem install llama_cpp -- --with-cuda
 ```
 
 Those options are defined in [extconf.rb](https://github.com/yoshoku/llama_cpp.rb/blob/main/ext/llama_cpp/extconf.rb) by with_config method.

data/examples/chat.rb

@@ -9,6 +9,7 @@
 require 'llama_cpp'
 require 'thor'
 require 'readline'
+require 'etc'
 
 class Chat < Thor # rubocop:disable Metrics/ClassLength, Style/Documentation
   default_command :main
@@ -30,12 +31,15 @@ class Chat < Thor # rubocop:disable Metrics/ClassLength, Style/Documentation
   option :typical_p, type: :numeric, desc: 'locally typical sampling, parameter p', default: 1.0
   option :temp, type: :numeric, desc: 'temperature', default: 0.8
   option :n_gpu_layers, type: :numeric, desc: 'number of layers on GPU', default: 0
+  option :n_threads, type: :numeric, desc: 'number of threads', default: Etc.nprocessors
   def main # rubocop:disable Metrics/AbcSize, Metrics/CyclomaticComplexity, Metrics/MethodLength, Metrics/PerceivedComplexity
     mdl_params = LLaMACpp::ModelParams.new
     mdl_params.n_gpu_layers = options[:n_gpu_layers]
     model = LLaMACpp::Model.new(model_path: options[:model], params: mdl_params)
     ctx_params = LLaMACpp::ContextParams.new
     ctx_params.seed = options[:seed] if options[:seed] != -1
+    ctx_params.n_threads = options[:n_threads]
+    ctx_params.n_threads_batch = options[:n_threads]
     context = LLaMACpp::Context.new(model: model, params: ctx_params)
 
     antiprompt = options[:reverse_prompt] || 'User:'
@@ -70,7 +74,7 @@ class Chat < Thor # rubocop:disable Metrics/ClassLength, Style/Documentation
 
         0.step(embd.size - 1, options[:batch_size]) do |i|
           n_eval = [options[:batch_size], embd.size - i].min
-          context.eval(tokens: embd[i...i + n_eval], n_past: n_past)
+          context.decode(LLaMACpp::Batch.get_one(tokens: embd[i...(i + n_eval)], n_tokens: n_eval, pos_zero: n_past, seq_id: 0))
           n_past += n_eval
         end
       end
@@ -95,7 +99,7 @@ class Chat < Thor # rubocop:disable Metrics/ClassLength, Style/Documentation
         context.sample_tail_free(candidates, z: options[:tfs_z])
         context.sample_typical(candidates, prob: options[:typical_p])
         context.sample_top_p(candidates, prob: options[:top_p])
-        context.sample_temperature(candidates, temperature: options[:temp])
+        context.sample_temp(candidates, temp: options[:temp])
         id = context.sample_token(candidates)
 
         last_n_tokens.shift

data/examples/embedding.rb

@@ -7,6 +7,7 @@
 
 require 'llama_cpp'
 require 'thor'
+require 'etc'
 
 class Embedding < Thor # rubocop:disable Style/Documentation
   default_command :main
@@ -15,6 +16,7 @@ class Embedding < Thor # rubocop:disable Style/Documentation
   option :model, type: :string, aliases: '-m', desc: 'path to model file', required: true
   option :prompt, type: :string, aliases: '-p', desc: 'prompt to generate embedding', required: true
   option :n_gpu_layers, type: :numeric, desc: 'number of layers on GPU', default: 0
+  option :n_threads, type: :numeric, desc: 'number of threads', default: Etc.nprocessors
   def main # rubocop:disable Metrics/AbcSize, Metrics/MethodLength
     mdl_params = LLaMACpp::ModelParams.new
     mdl_params.n_gpu_layers = options[:n_gpu_layers]
@@ -22,13 +24,15 @@ class Embedding < Thor # rubocop:disable Style/Documentation
     ctx_params = LLaMACpp::ContextParams.new
     ctx_params.embedding = true
     ctx_params.seed = options[:seed] if options[:seed] != -1
+    ctx_params.n_threads = options[:n_threads]
+    ctx_params.n_threads_batch = options[:n_threads]
     context = LLaMACpp::Context.new(model: model, params: ctx_params)
 
     embd_input = context.model.tokenize(text: options[:prompt], add_bos: true)
 
     return unless embd_input.size.positive?
 
-    context.eval(tokens: embd_input, n_past: 0)
+    context.decode(LLaMACpp::Batch.get_one(tokens: embd_input, n_tokens: embd_input.size, pos_zero: 0, seq_id: 0))
 
     context.embeddings.each { |val| print("#{val} ") }
     print("\n")

data/examples/simple.rb

@@ -7,12 +7,14 @@
 
 require 'llama_cpp'
 require 'thor'
+require 'etc'
 
 class Simple < Thor # rubocop:disable Style/Documentation
   default_command :main
   desc 'main', 'Simple completion'
   option :model, type: :string, aliases: '-m', desc: 'path to model file', required: true
   option :prompt, type: :string, aliases: '-p', desc: 'prompt to start with', default: 'Hello my name is'
+  option :n_threads, type: :numeric, desc: 'number of threads', default: Etc.nprocessors
   def main # rubocop:disable Metrics/AbcSize, Metrics/MethodLength
     n_len = 32
     model_params = LLaMACpp::ModelParams.new
@@ -21,7 +23,8 @@ class Simple < Thor # rubocop:disable Style/Documentation
     context_params.seed = 1234
     context_params.n_ctx = 2048
     context_params.logits_all = true
-    context_params.n_threads = 4
+    context_params.n_threads = options[:n_threads]
+    context_params.n_threads_batch = options[:n_threads]
     context = LLaMACpp::Context.new(model: model, params: context_params)
 
     tokens_list = context.model.tokenize(text: options[:prompt], add_bos: true)

data/ext/llama_cpp/llama_cpp.cpp

@@ -51,6 +51,7 @@ public:
   static void define_class(VALUE outer) {
     rb_cLLaMABatch = rb_define_class_under(outer, "Batch", rb_cObject);
     rb_define_alloc_func(rb_cLLaMABatch, llama_batch_alloc);
+    rb_define_singleton_method(rb_cLLaMABatch, "get_one", RUBY_METHOD_FUNC(_llama_batch_get_one), -1);
     rb_define_method(rb_cLLaMABatch, "initialize", RUBY_METHOD_FUNC(_llama_batch_initialize), -1);
     rb_define_method(rb_cLLaMABatch, "n_tokens=", RUBY_METHOD_FUNC(_llama_batch_set_n_tokens), 1);
     rb_define_method(rb_cLLaMABatch, "n_tokens", RUBY_METHOD_FUNC(_llama_batch_get_n_tokens), 0);
@@ -75,6 +76,48 @@ public:
 private:
   static const rb_data_type_t llama_batch_type;
 
+  static VALUE _llama_batch_get_one(int argc, VALUE* argv, VALUE klass) {
+    VALUE kw_args = Qnil;
+    ID kw_table[4] = { rb_intern("tokens"), rb_intern("n_tokens"), rb_intern("pos_zero"), rb_intern("seq_id") };
+    VALUE kw_values[4] = { Qundef, Qundef, Qundef, Qundef };
+    rb_scan_args(argc, argv, ":", &kw_args);
+    rb_get_kwargs(kw_args, kw_table, 4, 0, kw_values);
+
+    if (!RB_TYPE_P(kw_values[0], T_ARRAY)) {
+      rb_raise(rb_eArgError, "tokens must be an array");
+      return Qnil;
+    }
+    if (!RB_INTEGER_TYPE_P(kw_values[1])) {
+      rb_raise(rb_eArgError, "n_tokens must be an integer");
+      return Qnil;
+    }
+    if (!RB_INTEGER_TYPE_P(kw_values[2])) {
+      rb_raise(rb_eArgError, "pos_zero must be an integer");
+      return Qnil;
+    }
+    if (!RB_INTEGER_TYPE_P(kw_values[3])) {
+      rb_raise(rb_eArgError, "seq_id must be an integer");
+      return Qnil;
+    }
+
+    const size_t sz_array = RARRAY_LEN(kw_values[0]);
+    const int32_t n_tokens = NUM2INT(kw_values[1]);
+    const llama_pos pos_zero = NUM2INT(kw_values[2]);
+    const llama_seq_id seq_id = NUM2INT(kw_values[3]);
+
+    LLaMABatchWrapper* ptr = (LLaMABatchWrapper*)ruby_xmalloc(sizeof(LLaMABatchWrapper));
+    new (ptr) LLaMABatchWrapper();
+    ptr->batch = llama_batch_get_one(nullptr, n_tokens, pos_zero, seq_id);
+
+    ptr->batch.token = (llama_token*)malloc(sizeof(llama_token) * sz_array);
+    for (size_t i = 0; i < sz_array; i++) {
+      VALUE el = rb_ary_entry(kw_values[0], i);
+      ptr->batch.token[i] = NUM2INT(el);
+    }
+
+    return TypedData_Wrap_Struct(klass, &llama_batch_type, ptr);
+  }
+
   static VALUE _llama_batch_initialize(int argc, VALUE* argv, VALUE self) {
     VALUE kw_args = Qnil;
     ID kw_table[3] = { rb_intern("max_n_token"), rb_intern("n_embd"), rb_intern("max_n_seq") };
@@ -1983,6 +2026,7 @@ public:
     rb_define_method(rb_cLLaMAContext, "kv_cache_seq_cp", RUBY_METHOD_FUNC(_llama_context_kv_cache_seq_cp), 4);
     rb_define_method(rb_cLLaMAContext, "kv_cache_seq_keep", RUBY_METHOD_FUNC(_llama_context_kv_cache_seq_keep), 1);
     rb_define_method(rb_cLLaMAContext, "kv_cache_seq_shift", RUBY_METHOD_FUNC(_llama_context_kv_cache_seq_shift), 4);
+    rb_define_method(rb_cLLaMAContext, "kv_cache_seq_div", RUBY_METHOD_FUNC(_llama_context_kv_cache_seq_div), 4);
     rb_define_method(rb_cLLaMAContext, "set_rng_seed", RUBY_METHOD_FUNC(_llama_context_set_rng_seed), 1);
     rb_define_method(rb_cLLaMAContext, "load_session_file", RUBY_METHOD_FUNC(_llama_context_load_session_file), -1);
     rb_define_method(rb_cLLaMAContext, "save_session_file", RUBY_METHOD_FUNC(_llama_context_save_session_file), -1);
@@ -2054,6 +2098,8 @@ private:
     rb_scan_args(argc, argv, ":", &kw_args);
     rb_get_kwargs(kw_args, kw_table, 2, 2, kw_values);
 
+    rb_warn("eval is deprecated. Use decode instead.");
+
     if (!RB_TYPE_P(kw_values[0], T_ARRAY)) {
       rb_raise(rb_eArgError, "tokens must be an Array");
       return Qnil;
@@ -2104,6 +2150,8 @@ private:
     rb_scan_args(argc, argv, ":", &kw_args);
     rb_get_kwargs(kw_args, kw_table, 2, 2, kw_values);
 
+    rb_warn("eval_embd is deprecated. Use decode instead.");
+
     if (!RB_TYPE_P(kw_values[0], T_ARRAY)) {
       rb_raise(rb_eArgError, "tokens must be an Array");
       return Qnil;
@@ -2331,6 +2379,16 @@ private:
     return Qnil;
   }
 
+  static VALUE _llama_context_kv_cache_seq_div(VALUE self, VALUE seq_id, VALUE p0, VALUE p1, VALUE d) {
+    LLaMAContextWrapper* ptr = get_llama_context(self);
+    if (ptr->ctx == NULL) {
+      rb_raise(rb_eArgError, "LLaMA context is not initialized");
+      return Qnil;
+    }
+    llama_kv_cache_seq_div(ptr->ctx, NUM2INT(seq_id), NUM2INT(p0), NUM2INT(p1), NUM2INT(d));
+    return Qnil;
+  }
+
   static VALUE _llama_context_set_rng_seed(VALUE self, VALUE seed_) {
     LLaMAContextWrapper* ptr = get_llama_context(self);
     if (ptr->ctx == NULL) {
@@ -2794,6 +2852,8 @@ private:
     rb_scan_args(argc, argv, "1:", &candidates, &kw_args);
     rb_get_kwargs(kw_args, kw_table, 1, 0, kw_values);
 
+    rb_warn("sample_temperature is deprecated. Use sample_temp instead.");
+
     if (!rb_obj_is_kind_of(candidates, rb_cLLaMATokenDataArray)) {
       rb_raise(rb_eArgError, "1st argument must be a TokenDataArray");
       return Qnil;
@@ -3160,6 +3220,9 @@ extern "C" void Init_llama_cpp(void) {
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_Q5_K_S", INT2NUM(LLAMA_FTYPE_MOSTLY_Q5_K_S));
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_Q5_K_M", INT2NUM(LLAMA_FTYPE_MOSTLY_Q5_K_M));
   rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_Q6_K", INT2NUM(LLAMA_FTYPE_MOSTLY_Q6_K));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_IQ2_XXS", INT2NUM(LLAMA_FTYPE_MOSTLY_IQ2_XXS));
+  rb_define_const(rb_mLLaMACpp, "LLAMA_FTYPE_MOSTLY_IQ2_XS", INT2NUM(LLAMA_FTYPE_MOSTLY_IQ2_XS));
+  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_GUESSED", INT2NUM(LLAMA_FTYPE_GUESSED));
 

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 = '0.11.1'
+  VERSION = '0.12.1'
 
   # The version of llama.cpp bundled with llama_cpp.rb.
-  LLAMA_CPP_VERSION = 'b1768'
+  LLAMA_CPP_VERSION = 'b1833'
 end

data/lib/llama_cpp.rb

@@ -54,7 +54,7 @@ module LLaMACpp
           embd.insert(0, last_n_tokens[(n_ctx - (n_left / 2) - embd.size)...-embd.size])
         end
 
-        context.eval(tokens: embd, n_past: n_past)
+        context.decode(LLaMACpp::Batch.get_one(tokens: embd, n_tokens: embd.size, pos_zero: n_past, seq_id: 0))
       end
 
       n_past += embd.size
@@ -77,7 +77,7 @@ module LLaMACpp
         context.sample_tail_free(candidates, z: tfs_z)
         context.sample_typical(candidates, prob: typical_p)
         context.sample_top_p(candidates, prob: top_p)
-        context.sample_temperature(candidates, temperature: temperature)
+        context.sample_temp(candidates, temp: temperature)
         id = context.sample_token(candidates)
 
         last_n_tokens.shift

data/sig/llama_cpp.rbs

@@ -22,6 +22,9 @@ module LLaMACpp
   LLAMA_FTYPE_MOSTLY_Q5_K_S: Integer
   LLAMA_FTYPE_MOSTLY_Q5_K_M: Integer
   LLAMA_FTYPE_MOSTLY_Q6_K: Integer
+  LLAMA_FTYPE_MOSTLY_IQ2_XXS: Integer
+  LLAMA_FTYPE_MOSTLY_IQ2_XS: Integer
+  LLAMA_FTYPE_MOSTLY_Q2_K_S: Integer
 
   LLAMA_KV_OVERRIDE_INT: Integer
   LLAMA_KV_OVERRIDE_FLOAT: Integer
@@ -149,6 +152,7 @@ module LLaMACpp
   class Batch
     public
 
+    def self.get_one: (tokens: Array[Integer], n_tokens: Integer, pos_zero: Integer, seq_id: Integer) -> ::LLaMACpp::Batch
     def initialize: (max_n_token: Integer, n_embd: Integer, max_n_seq: Integer) -> void
     def n_tokens=: (Integer) -> Integer
     def n_tokens: () -> Integer
@@ -192,6 +196,7 @@ module LLaMACpp
     def kv_cache_seq_cp: (Integer, Integer,Integer, Integer) -> void
     def kv_cache_seq_keep: (Integer) -> void
     def kv_cache_seq_shift: (Integer, Integer, Integer, Integer) -> void
+    def kv_cache_seq_div: (Integer, Integer, Integer, Integer) -> void
     def set_rng_seed: (Integer) -> void
     def load_session_file: (session_path: String) -> void
     def save_session_file: (session_path: String, session_tokens: Array[Integer]) -> void

data/vendor/tmp/llama.cpp/Makefile

@@ -1,8 +1,8 @@
 # Define the default target now so that it is always the first target
 BUILD_TARGETS = \
-	main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
+	main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
 	simple batched batched-bench save-load-state server gguf llama-bench libllava.a llava-cli baby-llama beam-search  \
-	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup tests/test-c.o
+	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = \
@@ -620,6 +620,9 @@ quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.o ggml.
 perplexity: examples/perplexity/perplexity.cpp                ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
+imatrix: examples/imatrix/imatrix.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 embedding: examples/embedding/embedding.cpp                   ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
@@ -671,6 +674,9 @@ lookahead: examples/lookahead/lookahead.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS
 lookup: examples/lookup/lookup.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
+passkey: examples/passkey/passkey.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 ifdef LLAMA_METAL
 metal: examples/metal/metal.cpp ggml.o $(OBJS)
 	$(CXX) $(CXXFLAGS) $^ -o $@ $(LDFLAGS)

data/vendor/tmp/llama.cpp/ggml-backend-impl.h

@@ -90,7 +90,7 @@ extern "C" {
         void                      (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
 
         // compute graph without a plan
-        void (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
+        bool (*graph_compute)(ggml_backend_t backend, struct ggml_cgraph * cgraph);
 
         // check if the backend supports an operation
         bool (*supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);

data/vendor/tmp/llama.cpp/ggml-backend.c

@@ -195,11 +195,14 @@ void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_
     ggml_backend_synchronize(backend);
 }
 
-void ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
-    backend->iface.graph_compute(backend, cgraph);
+bool ggml_backend_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+    if (!backend->iface.graph_compute(backend, cgraph)) {
+        return false;
+    }
 
     // TODO: optional sync
     ggml_backend_synchronize(backend);
+    return true;
 }
 
 bool ggml_backend_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {
@@ -597,7 +600,7 @@ static void ggml_backend_cpu_graph_plan_compute(ggml_backend_t backend, ggml_bac
     GGML_UNUSED(backend);
 }
 
-static void ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
+static bool ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_cgraph * cgraph) {
     struct ggml_backend_cpu_context * cpu_ctx = (struct ggml_backend_cpu_context *)backend->context;
 
     struct ggml_cplan cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
@@ -611,6 +614,7 @@ static void ggml_backend_cpu_graph_compute(ggml_backend_t backend, struct ggml_c
     cplan.work_data = cpu_ctx->work_data;
 
     ggml_graph_compute(cgraph, &cplan);
+    return true;
 }
 
 static bool ggml_backend_cpu_supports_op(ggml_backend_t backend, const struct ggml_tensor * op) {

data/vendor/tmp/llama.cpp/ggml-backend.h

@@ -58,7 +58,7 @@ extern "C" {
 
     GGML_API void ggml_backend_graph_plan_free   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
     GGML_API void ggml_backend_graph_plan_compute(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
-    GGML_API void ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+    GGML_API bool ggml_backend_graph_compute     (ggml_backend_t backend, struct ggml_cgraph * cgraph);
     GGML_API bool ggml_backend_supports_op       (ggml_backend_t backend, const struct ggml_tensor * op);
 
     // tensor copy between different backends