llama_cpp 0.12.5 → 0.12.6

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 143fb1bb108c9cc679ed6eddaaca4cb8a52a5321ee4ffd965440a2c92aeeb99e
-  data.tar.gz: f522cbf943f82143d1a4eae679473468a9920a6ef6fe6cf88147b82bc6a1f279
+  metadata.gz: 296b29b7d20c7bfd66f69749ccd41e63d6998589af0d3514db8f6c08011d545f
+  data.tar.gz: 48f8787a63759a95049bbc515f4b35c74d07b356f1635d751d8d9d852e386c5a
 SHA512:
-  metadata.gz: 1646833e8e1ffd6dd22d809ce2c4f2b0f3de78d84504713da4e8d5ab1c2b466c5cbc47a3c787297753f6d56656635e12cf522acffbe37253bdae0c57f8cc51c9
-  data.tar.gz: fbbf0372d52ba8862dcc4ff61f590f634cdcde039dc31f09a93ac6cd8e112c34a1c6d567d54a9ec2d0679e1c4ec8c2e8153071c6952f67af34fa0c4ccf49ac76
+  metadata.gz: 5cd4c284a31fcdd36565b481c2456545eaf3fe19fda3778121f26f529ca01d18a894ba73739d966dc29f5aa239f8784ed56801bac5db3d21ae13e5b5aa2b4012
+  data.tar.gz: 7d03f1d081d097913fe3489a0432a5869a13e0a0371458c6c4d6cdea7296422a5af51c13ae05ea0d752e068865cc99e52ee0c4f3d67de892003c76e9126d5940

data/CHANGELOG.md

@@ -1,3 +1,10 @@
+## [[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) {
@@ -3314,6 +3359,7 @@ 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));

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 = '0.12.6'
 
   # The version of llama.cpp bundled with llama_cpp.rb.
-  LLAMA_CPP_VERSION = 'b2106'
+  LLAMA_CPP_VERSION = 'b2143'
 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
@@ -190,6 +194,7 @@ 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 +275,8 @@ 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

@@ -571,6 +571,14 @@ $(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 +633,7 @@ 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
+# find examples pocs -type f -name "*.o" -delete
 
 #
 # Examples