llama_cpp 0.14.0 → 0.14.1

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: c7d855ccd32ae097f26a671751d6a2178361cf8d8a6c1b99af37859f2c47ca03
-  data.tar.gz: 3b17318424d08c65ad34da3fa14956c86db0a2ea05ac174323a9b8d2b9e69d59
+  metadata.gz: c2a192fa17c1d313a93306e415ec27dfb8fb6ce993b9fc78797ed6e1d38ca63f
+  data.tar.gz: f800e54961a8bea5de95373d15f0cda30f7e95edd655cc0504247dfefcff473a
 SHA512:
-  metadata.gz: 2d90bf9fdd8dbaf5e67b7fb8797a9412168ae6ce5fcfc4c6aca34e194d5beb5204184b5bb36d65dc507a7a618ac9e938987e8d8bf5871e4eb6304b5e6de06020
-  data.tar.gz: eab524367ace146eb6e20786bd530cead145e1651bcdb726afbb5364609d04b22ca8a515016bb0c2d154ea97fb62f19222c122bc9bb5efe7fc389a6f259da6f0
+  metadata.gz: 48cefba1491319f82d52a46e8be34b5f0115dbe80bd6a9fdbf4fe0e190581a6b1ff8c3e2b2dfdaefeaa0b7cb11c8b9f5a84bcb60354f64248abbee3d488378ee
+  data.tar.gz: 9c6d75d3818b61192bd5c93a8b091003e2342f28102de1fbc9a1a02955a7c89e2a144b82bbe83e805b3f741261e967469c3ad2f6d347b1b870fb51880b850d89

data/CHANGELOG.md

@@ -1,3 +1,10 @@
+## [[0.14.1](https://github.com/yoshoku/llama_cpp.rb/compare/v0.14.0...v0.14.1)] - 2024-03-16
+
+- Bump llama.cpp from b2361 to b2435.
+  - Add constants for vocaburary type: `LLAMA_VOCAB_TYPE_NONE`.
+  - Add `n_ubatch` and `n_seq_max` accessors to `ContextParams`.
+  - Add `n_ubatch`, `n_seq_max`, `set_causal_attn`, and `synchronize` methods to `Context`.
+
 ## [[0.14.0](https://github.com/yoshoku/llama_cpp.rb/compare/v0.13.0...v0.14.0)] - 2024-03-09
 
 **Breaking Changes**

data/ext/llama_cpp/llama_cpp.cpp

@@ -946,6 +946,10 @@ public:
     rb_define_method(rb_cLLaMAContextParams, "n_ctx", RUBY_METHOD_FUNC(_llama_context_params_get_n_ctx), 0);
     rb_define_method(rb_cLLaMAContextParams, "n_batch=", RUBY_METHOD_FUNC(_llama_context_params_set_n_batch), 1);
     rb_define_method(rb_cLLaMAContextParams, "n_batch", RUBY_METHOD_FUNC(_llama_context_params_get_n_batch), 0);
+    rb_define_method(rb_cLLaMAContextParams, "n_ubatch=", RUBY_METHOD_FUNC(_llama_context_params_set_n_ubatch), 1);
+    rb_define_method(rb_cLLaMAContextParams, "n_ubatch", RUBY_METHOD_FUNC(_llama_context_params_get_n_ubatch), 0);
+    rb_define_method(rb_cLLaMAContextParams, "n_seq_max=", RUBY_METHOD_FUNC(_llama_context_params_set_n_seq_max), 1);
+    rb_define_method(rb_cLLaMAContextParams, "n_seq_max", RUBY_METHOD_FUNC(_llama_context_params_get_n_seq_max), 0);
     rb_define_method(rb_cLLaMAContextParams, "n_threads=", RUBY_METHOD_FUNC(_llama_context_params_set_n_threads), 1);
     rb_define_method(rb_cLLaMAContextParams, "n_threads", RUBY_METHOD_FUNC(_llama_context_params_get_n_threads), 0);
     rb_define_method(rb_cLLaMAContextParams, "n_threads_batch=", RUBY_METHOD_FUNC(_llama_context_params_set_n_threads_batch), 1);
@@ -1031,6 +1035,30 @@ private:
     return INT2NUM(ptr->params.n_batch);
   }
 
+  // n_ubatch
+  static VALUE _llama_context_params_set_n_ubatch(VALUE self, VALUE n_ubatch) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    ptr->params.n_ubatch = NUM2INT(n_ubatch);
+    return INT2NUM(ptr->params.n_ubatch);
+  }
+
+  static VALUE _llama_context_params_get_n_ubatch(VALUE self) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    return INT2NUM(ptr->params.n_ubatch);
+  }
+
+  // n_seq_max
+  static VALUE _llama_context_params_set_n_seq_max(VALUE self, VALUE n_seq_max) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    ptr->params.n_seq_max = NUM2INT(n_seq_max);
+    return INT2NUM(ptr->params.n_seq_max);
+  }
+
+  static VALUE _llama_context_params_get_n_seq_max(VALUE self) {
+    LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
+    return INT2NUM(ptr->params.n_seq_max);
+  }
+
   // n_threads
   static VALUE _llama_context_params_set_n_threads(VALUE self, VALUE n_threads) {
     LLaMAContextParamsWrapper* ptr = get_llama_context_params(self);
@@ -2019,6 +2047,8 @@ public:
     rb_define_method(rb_cLLaMAContext, "embeddings_seq", RUBY_METHOD_FUNC(_llama_context_embeddings_seq), 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, "n_ubatch", RUBY_METHOD_FUNC(_llama_context_n_ubatch), 0);
+    rb_define_method(rb_cLLaMAContext, "n_seq_max", RUBY_METHOD_FUNC(_llama_context_n_seq_max), 0);
     rb_define_method(rb_cLLaMAContext, "timings", RUBY_METHOD_FUNC(_llama_context_get_timings), 0);
     rb_define_method(rb_cLLaMAContext, "print_timings", RUBY_METHOD_FUNC(_llama_context_print_timings), 0);
     rb_define_method(rb_cLLaMAContext, "reset_timings", RUBY_METHOD_FUNC(_llama_context_reset_timings), 0);
@@ -2033,6 +2063,8 @@ public:
     rb_define_method(rb_cLLaMAContext, "kv_cache_kv_cache_defrag", RUBY_METHOD_FUNC(_llama_context_kv_cache_defrag), 0);
     rb_define_method(rb_cLLaMAContext, "kv_cache_kv_cache_update", RUBY_METHOD_FUNC(_llama_context_kv_cache_update), 0);
     rb_define_method(rb_cLLaMAContext, "set_rng_seed", RUBY_METHOD_FUNC(_llama_context_set_rng_seed), 1);
+    rb_define_method(rb_cLLaMAContext, "set_causal_attn", RUBY_METHOD_FUNC(_llama_context_set_causal_attn), 1);
+    rb_define_method(rb_cLLaMAContext, "synchronize", RUBY_METHOD_FUNC(_llama_context_synchronize), 0);
     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);
     rb_define_method(rb_cLLaMAContext, "sample_repetition_penalties", RUBY_METHOD_FUNC(_llama_context_sample_repetition_penalties), -1);
@@ -2250,6 +2282,24 @@ private:
     return UINT2NUM(llama_n_batch(ptr->ctx));
   }
 
+  static VALUE _llama_context_n_ubatch(VALUE self) {
+    LLaMAContextWrapper* ptr = get_llama_context(self);
+    if (ptr->ctx == NULL) {
+      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
+      return Qnil;
+    }
+    return UINT2NUM(llama_n_ubatch(ptr->ctx));
+  }
+
+  static VALUE _llama_context_n_seq_max(VALUE self) {
+    LLaMAContextWrapper* ptr = get_llama_context(self);
+    if (ptr->ctx == NULL) {
+      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
+      return Qnil;
+    }
+    return UINT2NUM(llama_n_seq_max(ptr->ctx));
+  }
+
   static VALUE _llama_context_get_timings(VALUE self) {
     LLaMAContextWrapper* ptr = get_llama_context(self);
     if (ptr->ctx == NULL) {
@@ -2395,6 +2445,26 @@ private:
     return Qnil;
   }
 
+  static VALUE _llama_context_set_causal_attn(VALUE self, VALUE causal_attn) {
+    LLaMAContextWrapper* ptr = get_llama_context(self);
+    if (ptr->ctx == NULL) {
+      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
+      return Qnil;
+    }
+    llama_set_causal_attn(ptr->ctx, RTEST(causal_attn) ? true : false);
+    return Qnil;
+  }
+
+  static VALUE _llama_context_synchronize(VALUE self) {
+    LLaMAContextWrapper* ptr = get_llama_context(self);
+    if (ptr->ctx == NULL) {
+      rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
+      return Qnil;
+    }
+    llama_synchronize(ptr->ctx);
+    return Qnil;
+  }
+
   static VALUE _llama_context_load_session_file(int argc, VALUE* argv, VALUE self) {
     VALUE kw_args = Qnil;
     ID kw_table[1] = { rb_intern("session_path") };
@@ -3204,6 +3274,7 @@ extern "C" void Init_llama_cpp(void) {
   rb_define_module_function(rb_mLLaMACpp, "supports_mlock?", rb_llama_supports_mlock, 0);
   rb_define_module_function(rb_mLLaMACpp, "supports_gpu_offload?", rb_llama_supports_gpu_offload, 0);
 
+  rb_define_const(rb_mLLaMACpp, "LLAMA_VOCAB_TYPE_NONE", INT2NUM(LLAMA_VOCAB_TYPE_NONE));
   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));

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 = '0.14.0'
+  VERSION = '0.14.1'
 
   # The version of llama.cpp bundled with llama_cpp.rb.
-  LLAMA_CPP_VERSION = 'b2361'
+  LLAMA_CPP_VERSION = 'b2435'
 end

data/sig/llama_cpp.rbs

@@ -3,6 +3,7 @@ module LLaMACpp
   LLAMA_CPP_VERSION: String
   LLAMA_DEFALUT_SEED: String
 
+  LLAMA_VOCAB_TYPE_NONE: Integer
   LLAMA_VOCAB_TYPE_SPM: Integer
   LLAMA_VOCAB_TYPE_BPE: Integer
   LLAMA_VOCAB_TYPE_WPM: Integer
@@ -207,6 +208,8 @@ module LLaMACpp
     def logits: () -> Array[Float]
     def n_ctx: () -> Integer
     def n_batch: () -> Integer
+    def n_ubatch: () -> Integer
+    def n_seq_max: () -> Integer
     def timings: () -> ::LLaMACpp::Timings
     def print_timings: () -> void
     def reset_timings: () -> void
@@ -221,6 +224,8 @@ module LLaMACpp
     def kv_cache_defrag: () -> void
     def kv_cache_update: () -> void
     def set_rng_seed: (Integer) -> void
+    def set_causal_attn: (bool) -> void
+    def synchronize: () -> void
     def load_session_file: (session_path: String) -> void
     def save_session_file: (session_path: String, session_tokens: Array[Integer]) -> void
     def sample_repetition_penalties: (::LLaMACpp::TokenDataArray, Array[Integer], penalty_repeat: Float, penalty_freq: Float, penalty_present: Float) -> void
@@ -250,6 +255,10 @@ module LLaMACpp
     def n_ctx=: (Integer) -> Integer
     def n_batch: () -> Integer
     def n_batch=: (Integer) -> Integer
+    def n_ubatch: () -> Integer
+    def n_ubatch=: (Integer) -> Integer
+    def n_seq_max: () -> Integer
+    def n_seq_max=: (Integer) -> Integer
     def n_threads: () -> Integer
     def n_threads=: (Integer) -> Integer
     def n_threads_batch: () -> Integer

data/vendor/tmp/llama.cpp/Makefile

@@ -2,7 +2,7 @@
 BUILD_TARGETS = \
 	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 passkey tests/test-c.o
+	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey gritlm tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = \
@@ -167,6 +167,10 @@ ifeq ($(UNAME_S),OpenBSD)
 	MK_CPPFLAGS += -D_BSD_SOURCE
 endif
 
+ifdef LLAMA_SCHED_MAX_COPIES
+	MK_CPPFLAGS += -DGGML_SCHED_MAX_COPIES=$(LLAMA_SCHED_MAX_COPIES)
+endif
+
 ifdef LLAMA_DEBUG
 	MK_CFLAGS   += -O0 -g
 	MK_CXXFLAGS += -O0 -g
@@ -201,6 +205,10 @@ ifdef LLAMA_SERVER_VERBOSE
 	MK_CPPFLAGS += -DSERVER_VERBOSE=$(LLAMA_SERVER_VERBOSE)
 endif
 
+ifdef LLAMA_SERVER_SSL
+	MK_CPPFLAGS += -DCPPHTTPLIB_OPENSSL_SUPPORT
+	MK_LDFLAGS += -lssl -lcrypto
+endif
 
 ifdef LLAMA_CODE_COVERAGE
 	MK_CXXFLAGS += -fprofile-arcs -ftest-coverage -dumpbase ''
@@ -451,7 +459,7 @@ endif # LLAMA_CUDA_PEER_MAX_BATCH_SIZE
 ifdef LLAMA_CUDA_CCBIN
 	MK_NVCCFLAGS += -ccbin $(LLAMA_CUDA_CCBIN)
 endif
-ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
+ggml-cuda.o: ggml-cuda.cu ggml-cuda.h ggml-common.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) $(CPPFLAGS) -Xcompiler "$(CUDA_CXXFLAGS)" -c $< -o $@
 else
@@ -551,15 +559,16 @@ 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
+ggml-metal-embed.o: ggml-metal.metal ggml-common.h
 	@echo "Embedding Metal library"
+	@sed -e '/#include "ggml-common.h"/r ggml-common.h' -e '/#include "ggml-common.h"/d' < ggml-metal.metal > ggml-metal-embed.metal
 	$(eval TEMP_ASSEMBLY=$(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)
+	@echo ".section __DATA, __ggml_metallib"   >  $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_start"        >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_start:"              >> $(TEMP_ASSEMBLY)
+	@echo ".incbin \"ggml-metal-embed.metal\"" >> $(TEMP_ASSEMBLY)
+	@echo ".globl _ggml_metallib_end"          >> $(TEMP_ASSEMBLY)
+	@echo "_ggml_metallib_end:"                >> $(TEMP_ASSEMBLY)
 	@$(AS) $(TEMP_ASSEMBLY) -o $@
 	@rm -f ${TEMP_ASSEMBLY}
 endif
@@ -628,12 +637,15 @@ ggml-alloc.o: ggml-alloc.c ggml.h ggml-alloc.h
 ggml-backend.o: ggml-backend.c ggml.h ggml-backend.h
 	$(CC)  $(CFLAGS)   -c $< -o $@
 
-ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h
+ggml-quants.o: ggml-quants.c ggml.h ggml-quants.h ggml-common.h
 	$(CC) $(CFLAGS)    -c $< -o $@
 
-OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o
+unicode.o: unicode.cpp unicode.h
+	$(CXX) $(CXXFLAGS) -c $< -o $@
 
-llama.o: llama.cpp ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
+OBJS += ggml-alloc.o ggml-backend.o ggml-quants.o unicode.o
+
+llama.o: llama.cpp unicode.h ggml.h ggml-alloc.h ggml-backend.h ggml-cuda.h ggml-metal.h llama.h
 	$(CXX) $(CXXFLAGS) -c $< -o $@
 
 COMMON_H_DEPS = common/common.h common/sampling.h common/log.h
@@ -725,6 +737,10 @@ embedding: examples/embedding/embedding.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)
 
+gritlm: examples/gritlm/gritlm.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)
+
 save-load-state: examples/save-load-state/save-load-state.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)

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

@@ -61,7 +61,6 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
     }
 }
 
-// TODO: GGML_PAD ?
 static size_t aligned_offset(const void * buffer, size_t offset, size_t alignment) {
     assert(alignment && !(alignment & (alignment - 1))); // power of 2
     size_t align = (alignment - (((uintptr_t)buffer + offset) % alignment)) % alignment;
@@ -69,25 +68,14 @@ static size_t aligned_offset(const void * buffer, size_t offset, size_t alignmen
 }
 
 // tallocr
-struct ggml_tallocr {
-    ggml_backend_buffer_t buffer;
-    void * base;
-    size_t alignment;
-    size_t offset;
-};
-
-ggml_tallocr_t ggml_tallocr_new(ggml_backend_buffer_t buffer) {
-    ggml_tallocr_t talloc = malloc(sizeof(struct ggml_tallocr));
-    if (talloc == NULL) {
-        return NULL;
-    }
 
+struct ggml_tallocr ggml_tallocr_new(ggml_backend_buffer_t buffer) {
     void * base = ggml_backend_buffer_get_base(buffer);
     size_t align = ggml_backend_buffer_get_alignment(buffer);
 
     assert(align && !(align & (align - 1))); // power of 2
 
-    *talloc = (struct ggml_tallocr) {
+    struct ggml_tallocr talloc = (struct ggml_tallocr) {
         /*.buffer    = */ buffer,
         /*.base      = */ base,
         /*.alignment = */ align,
@@ -96,11 +84,7 @@ ggml_tallocr_t ggml_tallocr_new(ggml_backend_buffer_t buffer) {
     return talloc;
 }
 
-void ggml_tallocr_free(ggml_tallocr_t talloc) {
-    free(talloc);
-}
-
-void ggml_tallocr_alloc(ggml_tallocr_t talloc, struct ggml_tensor * tensor) {
+void ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tensor) {
     size_t size = ggml_backend_buffer_get_alloc_size(talloc->buffer, tensor);
     size = GGML_PAD(size, talloc->alignment);
 
@@ -354,12 +338,16 @@ struct hash_node {
     bool allocated;
 };
 
-//
 struct tensor_alloc {
     size_t offset;
     size_t size_max; // 0 = pre-allocated, unused, or view
 };
 
+struct leaf_alloc {
+    int buffer_id;
+    struct tensor_alloc leaf;
+};
+
 struct node_alloc {
     int buffer_id;
     struct tensor_alloc dst;
@@ -378,7 +366,7 @@ struct ggml_gallocr {
     struct node_alloc * node_allocs; // [n_nodes]
     int n_nodes;
 
-    struct tensor_alloc * leaf_allocs; // [n_leafs]
+    struct leaf_alloc * leaf_allocs; // [n_leafs]
     int n_leafs;
 };
 
@@ -543,13 +531,20 @@ static int get_node_buffer_id(const int * node_buffer_ids, int i) {
     return node_buffer_ids ? node_buffer_ids[i] : 0;
 }
 
-static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids) {
+static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
     // clear hash tables
     memset(galloc->hash_set.keys, 0, galloc->hash_set.size * sizeof(struct ggml_tensor *));
     memset(galloc->hash_values,   0, galloc->hash_set.size * sizeof(struct hash_node));
 
+    // allocate leafs
+    // these may be tensors that the application is not using in the graph, but may still want to allocate for other purposes
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        ggml_gallocr_allocate_node(galloc, leaf, get_node_buffer_id(leaf_buffer_ids, i));
+    }
+
     // count number of children and views
-    // allocate all graph inputs and leafs first to avoid overwriting them
+    // allocate other graph inputs and leafs first to avoid overwriting them
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
 
@@ -577,19 +572,6 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
         }
     }
 
-    // allocate the remaining leafs that are unused on the graph
-    // these are effectively static tensors that the application is not using in the graph, but may still want to allocate for other purposes
-    for (int i = 0; i < graph->n_leafs; i++) {
-        struct ggml_tensor * leaf = graph->leafs[i];
-        struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
-
-        if (hn->n_children == 0) {
-            assert(!hn->allocated);
-            // since buffer ids are only given for nodes, these leafs are always allocated in the first buffer
-            ggml_gallocr_allocate_node(galloc, leaf, 0);
-        }
-    }
-
     // allocate tensors
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
@@ -652,7 +634,7 @@ static void ggml_gallocr_alloc_graph_impl(ggml_gallocr_t galloc, struct ggml_cgr
     }
 }
 
-bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids) {
+bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
     size_t hash_size = graph->visited_hash_table.size;
 
     // initialize hash table
@@ -676,7 +658,7 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
     }
 
     // allocate in hash table
-    ggml_gallocr_alloc_graph_impl(galloc, graph, node_buffer_ids);
+    ggml_gallocr_alloc_graph_impl(galloc, graph, node_buffer_ids, leaf_buffer_ids);
 
     // set the node_allocs from the hash table
     if (galloc->n_nodes < graph->n_nodes) {
@@ -711,15 +693,16 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
     }
     if (galloc->n_leafs < graph->n_leafs) {
         free(galloc->leaf_allocs);
-        galloc->leaf_allocs = calloc(sizeof(struct tensor_alloc), graph->n_leafs);
+        galloc->leaf_allocs = calloc(sizeof(galloc->leaf_allocs[0]), graph->n_leafs);
         GGML_ASSERT(galloc->leaf_allocs != NULL);
     }
     galloc->n_leafs = graph->n_leafs;
     for (int i = 0; i < graph->n_leafs; i++) {
         struct ggml_tensor * leaf = graph->leafs[i];
         struct hash_node * hn = ggml_gallocr_hash_get(galloc, leaf);
-        galloc->leaf_allocs[i].offset = hn->offset;
-        galloc->leaf_allocs[i].size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
+        galloc->leaf_allocs[i].buffer_id = hn->buffer_id;
+        galloc->leaf_allocs[i].leaf.offset = hn->offset;
+        galloc->leaf_allocs[i].leaf.size_max = ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
     }
 
     // reallocate buffers if needed
@@ -727,7 +710,8 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
         size_t cur_size = galloc->buffers[i] ? ggml_backend_buffer_get_size(galloc->buffers[i]) : 0;
         size_t new_size = ggml_dyn_tallocr_max_size(galloc->buf_tallocs[i]);
 
-        if (new_size > cur_size) {
+        // even if there are no tensors allocated in this buffer, we still need to allocate it to initialize views
+        if (new_size > cur_size || galloc->buffers[i] == NULL) {
 #ifndef NDEBUG
             fprintf(stderr, "%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
 #endif
@@ -744,30 +728,30 @@ bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, c
 }
 
 bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph *graph) {
-    return ggml_gallocr_reserve_n(galloc, graph, NULL);
+    return ggml_gallocr_reserve_n(galloc, graph, NULL, NULL);
 }
 
-static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * node, int buffer_id, struct tensor_alloc * tensor_alloc) {
-    assert(node->data || node->view_src || ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], node) <= tensor_alloc->size_max);
+static void ggml_gallocr_init_tensor(ggml_gallocr_t galloc, struct ggml_tensor * tensor, int buffer_id, struct tensor_alloc * tensor_alloc) {
+    assert(tensor->data || tensor->view_src || ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], tensor) <= tensor_alloc->size_max);
 
-    if (node->view_src != NULL) {
-        if (node->buffer == NULL) {
+    if (tensor->view_src != NULL) {
+        if (tensor->buffer == NULL) {
             assert(tensor_alloc->offset == SIZE_MAX);
-            if (node->view_src->buffer == NULL) {
+            if (tensor->view_src->buffer == NULL) {
                 // this tensor was allocated without ggml-backend
                 return;
             }
-            ggml_backend_view_init(galloc->buffers[buffer_id], node);
+            ggml_backend_view_init(galloc->buffers[buffer_id], tensor);
         }
     } else {
-        if (node->data == NULL) {
+        if (tensor->data == NULL) {
             assert(tensor_alloc->offset != SIZE_MAX);
-            assert(ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], node) <= tensor_alloc->size_max);
+            assert(ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], tensor) <= tensor_alloc->size_max);
             void * base = ggml_backend_buffer_get_base(galloc->buffers[buffer_id]);
             void * addr = (char *)base + tensor_alloc->offset;
-            ggml_backend_tensor_alloc(galloc->buffers[buffer_id], node, addr);
+            ggml_backend_tensor_alloc(galloc->buffers[buffer_id], tensor, addr);
         } else {
-            if (node->buffer == NULL) {
+            if (tensor->buffer == NULL) {
                 // this tensor was allocated without ggml-backend
                 return;
             }
@@ -843,13 +827,18 @@ bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph)
 
     // reset buffers
     for (int i = 0; i < galloc->n_buffers; i++) {
-        // zero size buffers are not allocated
         if (galloc->buffers[i] != NULL) {
             ggml_backend_buffer_reset(galloc->buffers[i]);
         }
     }
 
     // allocate the graph tensors from the previous assignments
+    // leafs
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct ggml_tensor * leaf = graph->leafs[i];
+        struct leaf_alloc * leaf_alloc = &galloc->leaf_allocs[i];
+        ggml_gallocr_init_tensor(galloc, leaf, leaf_alloc->buffer_id, &leaf_alloc->leaf);
+    }
     // nodes
     for (int i = 0; i < graph->n_nodes; i++) {
         struct ggml_tensor * node = graph->nodes[i];
@@ -863,12 +852,6 @@ bool ggml_gallocr_alloc_graph(ggml_gallocr_t galloc, struct ggml_cgraph * graph)
         }
         ggml_gallocr_init_tensor(galloc, node, node_alloc->buffer_id, &node_alloc->dst);
     }
-    // leafs
-    for (int i = 0; i < graph->n_leafs; i++) {
-        struct ggml_tensor * leaf = graph->leafs[i];
-        struct tensor_alloc * leaf_alloc = &galloc->leaf_allocs[i];
-        ggml_gallocr_init_tensor(galloc, leaf, 0, leaf_alloc);
-    }
 
     return true;
 }
@@ -900,12 +883,12 @@ static bool alloc_tensor_range(struct ggml_context * ctx,
         return false;
     }
 
-    struct ggml_tallocr * tallocr = ggml_tallocr_new(buffer);
+    struct ggml_tallocr tallocr = ggml_tallocr_new(buffer);
 
     for (struct ggml_tensor * t = first; t != last; t = ggml_get_next_tensor(ctx, t)) {
         if (t->data == NULL) {
             if (t->view_src == NULL) {
-                ggml_tallocr_alloc(tallocr, t);
+                ggml_tallocr_alloc(&tallocr, t);
             } else if (t->buffer == NULL) {
                 ggml_backend_view_init(buffer, t);
             }
@@ -917,8 +900,6 @@ static bool alloc_tensor_range(struct ggml_context * ctx,
         }
     }
 
-    ggml_tallocr_free(tallocr);
-
     *buffers = realloc(*buffers, sizeof(ggml_backend_buffer_t) * (*n_buffers + 1));
     (*buffers)[(*n_buffers)++] = buffer;
 

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

@@ -11,11 +11,15 @@ typedef struct ggml_backend_buffer * ggml_backend_buffer_t;
 typedef struct ggml_backend * ggml_backend_t;
 
 // Tensor allocator
-typedef struct ggml_tallocr * ggml_tallocr_t;
+struct ggml_tallocr {
+    ggml_backend_buffer_t buffer;
+    void * base;
+    size_t alignment;
+    size_t offset;
+};
 
-GGML_API ggml_tallocr_t ggml_tallocr_new(ggml_backend_buffer_t buffer);
-GGML_API void           ggml_tallocr_free(ggml_tallocr_t talloc);
-GGML_API void           ggml_tallocr_alloc(ggml_tallocr_t talloc, struct ggml_tensor * tensor);
+GGML_API struct ggml_tallocr ggml_tallocr_new(ggml_backend_buffer_t buffer);
+GGML_API void                ggml_tallocr_alloc(struct ggml_tallocr * talloc, struct ggml_tensor * tensor);
 
 // Graph allocator
 /*
@@ -50,7 +54,11 @@ GGML_API void           ggml_gallocr_free(ggml_gallocr_t galloc);
 // not strictly required for single buffer usage: ggml_gallocr_alloc_graph will reallocate the buffers automatically if needed
 // returns false if the buffer allocation failed
 GGML_API bool ggml_gallocr_reserve(ggml_gallocr_t galloc, struct ggml_cgraph * graph);
-GGML_API bool ggml_gallocr_reserve_n(ggml_gallocr_t galloc, struct ggml_cgraph * graph, const int * node_buffer_ids);
+GGML_API bool ggml_gallocr_reserve_n(
+    ggml_gallocr_t galloc,
+    struct ggml_cgraph * graph,
+    const int * node_buffer_ids,
+    const int * leaf_buffer_ids);
 
 // automatic reallocation if the topology changes when using a single buffer
 // returns false if using multiple buffers and a re-allocation is needed (call ggml_gallocr_reserve_n first to set the node buffers)

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

@@ -86,12 +86,12 @@ extern "C" {
         // (optional) asynchronous tensor data access
         void (*GGML_CALL set_tensor_async)(ggml_backend_t backend,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
         void (*GGML_CALL get_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
-        bool (*GGML_CALL cpy_tensor_async)(ggml_backend_t backend, const struct ggml_tensor * src, struct ggml_tensor * dst);
+        bool (*GGML_CALL cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
 
         // (optional) complete all pending operations
         void (*GGML_CALL synchronize)(ggml_backend_t backend);
 
-        // create a plan for ggml_cgraph and free it
+        // compute graph with a plan (not used currently)
         ggml_backend_graph_plan_t (*GGML_CALL graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
         void                      (*GGML_CALL graph_plan_free)   (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
 
@@ -102,16 +102,27 @@ extern "C" {
 
         // check if the backend supports an operation
         bool (*GGML_CALL supports_op)(ggml_backend_t backend, const struct ggml_tensor * op);
+
+        // (optional) event synchronization
+        ggml_backend_event_t (*GGML_CALL event_new)         (ggml_backend_t backend);
+        void                 (*GGML_CALL event_free)        (ggml_backend_event_t event);
+        void                 (*GGML_CALL event_record)      (ggml_backend_event_t event);
+        void                 (*GGML_CALL event_wait)        (ggml_backend_t backend, ggml_backend_event_t event);
+        void                 (*GGML_CALL event_synchronize) (ggml_backend_event_t event);
     };
 
     struct ggml_backend {
         ggml_guid_t guid;
 
         struct ggml_backend_i iface;
-
         ggml_backend_context_t context;
     };
 
+    struct ggml_backend_event {
+        ggml_backend_t backend;
+        void * context;
+    };
+
     //
     // Backend registry
     //