llama_cpp 0.12.0 → 0.12.1

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 87010edca1b352ae7bdd3a693451893b13dd75e9e109f9e2b42f6164cc186b08
-  data.tar.gz: ff34254b6377698903dcf771663b91c3c804111228888d96e91363bd0f29d3a6
+  metadata.gz: 13381408318e71cc1fc55c40ee9be6e62ad9e3ad6a8ce39279bb8040614e9b3b
+  data.tar.gz: 6456734b18865a7811f08d0d9d599771f574f4b59bd5b54a964ece7428115907
 SHA512:
-  metadata.gz: a23aa59fa4936940b28942398bfe98bdb09574162943ebaff31cdbda19394c7690f6c780f49da31eecc4b77427718a8b7ee58e62b2adb087100e1eee66310abc
-  data.tar.gz: 5cc105e69fc81d4616d93cd036af70f809be0c99b9155a6d3e386c9900ca012123353c23417ce56a5a64a1d805108b35de2d9feb5a6265c110d9341e5a2e242b
+  metadata.gz: 1014349771d7aa3c318027de11603e96d5482e4bd5b1bcf0fd4874040245daf44c4cfb801077a698846459a7619ca9e01e0afc3507fc7bd519e7ba68a000a15d
+  data.tar.gz: 1315ca8954397edb0db93347a10762e35f829377ef3dba0ea9cf6c67f986972ac8e75b46c410a3ceceefc0474f2abbe6f441e56a60e789ef1d2617fc15cfb29e

data/CHANGELOG.md

@@ -1,3 +1,9 @@
+## [[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`.

data/ext/llama_cpp/llama_cpp.cpp

@@ -2026,6 +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);
@@ -2378,6 +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) {
@@ -3209,6 +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 = '0.12.0'
+  VERSION = '0.12.1'
 
   # The version of llama.cpp bundled with llama_cpp.rb.
-  LLAMA_CPP_VERSION = 'b1768'
+  LLAMA_CPP_VERSION = 'b1833'
 end

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
@@ -193,6 +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