llama_cpp 0.10.0 → 0.10.2

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: 7f406c15621a7c247adaacf1d588ddf278225e6846466afd1184c00f1ee61768
-  data.tar.gz: df73657c75a80cb44f41d34a3c1054676cf59a5d7d56cb1c2ce8a94264002293
+  metadata.gz: b4e94e20f142572fb46cff141e109025e1c5b91b9cd6cabfbaeac163a920bb82
+  data.tar.gz: 10fcd2922c057a0c960cbde83a0fa22a56eae6e21a73e20931cfe082108dec26
 SHA512:
-  metadata.gz: acd08d5099f14bf2bd4c8f9bf016253f0e316179b79d72fbe7066b0d645ca31e9bab427fcc53d93874f8df74cb1746731e2cd21864bfecdecff91f9778919b42
-  data.tar.gz: 5014a1bd545be90c56bebd48119a198cf7276513cb6c5f00d8322aa6eaa9a27442bc51bf06953a11c2fc04145f797c630cefee17b36589fe38f9226003416a09
+  metadata.gz: 570f5f257d947bad8489fcafe6c7cc0dc342d593132111d3479c0202e0998e444ca75c2140883dc8cf6a22f0f4f74fecd264a577cee6aabde4c168f5003b4c98
+  data.tar.gz: d105e41aac3bb39b2ee53c43eb320272ffa5290caa7e14e2f08bebbade94417aed065e01d23b8def9cce08fd6789576d33ba4082750647c0bbf9ff4c776db9ac

data/CHANGELOG.md

@@ -1,3 +1,15 @@
+## [[0.10.2](https://github.com/yoshoku/llama_cpp.rb/compare/v0.10.1...v0.10.2)] - 2023-12-23
+
+- Bump bundled llama.cpp from b1641 to b1686.
+  - Add `LLAMA_FILE_MAGIC_GGLA` constant.
+  - Add `n_batch` method to `Context`.
+
+## [[0.10.1](https://github.com/yoshoku/llama_cpp.rb/compare/v0.10.0...v0.10.1)] - 2023-12-16
+
+- Bump bundled llama.cpp from b1620 to b1641.
+- Add attribute reader for `params` to `Model`.
+- Add `Batch` class, this class was not published because the author forgot to write `rb_define_class`.
+
 ## [[0.10.0](https://github.com/yoshoku/llama_cpp.rb/compare/v0.9.5...v0.10.0)] - 2023-12-09
 
 - Bump bundled llama.cpp from b1593 to b1620.

data/ext/llama_cpp/llama_cpp.cpp

@@ -1333,6 +1333,7 @@ public:
   static void define_class(VALUE outer) {
     rb_cLLaMAModel = rb_define_class_under(outer, "Model", rb_cObject);
     rb_define_alloc_func(rb_cLLaMAModel, llama_model_alloc);
+    rb_define_attr(rb_cLLaMAModel, "params", 1, 0);
     rb_define_method(rb_cLLaMAModel, "initialize", RUBY_METHOD_FUNC(_llama_model_initialize), -1);
     rb_define_method(rb_cLLaMAModel, "empty?", RUBY_METHOD_FUNC(_llama_model_empty), 0);
     rb_define_method(rb_cLLaMAModel, "free", RUBY_METHOD_FUNC(_llama_model_free), 0);
@@ -1948,6 +1949,7 @@ public:
     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, "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);
     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);
@@ -2200,7 +2202,16 @@ private:
       rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
       return Qnil;
     }
-    return INT2NUM(llama_n_ctx(ptr->ctx));
+    return UINT2NUM(llama_n_ctx(ptr->ctx));
+  }
+
+  static VALUE _llama_context_n_batch(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_batch(ptr->ctx));
   }
 
   static VALUE _llama_context_get_timings(VALUE self) {
@@ -3071,6 +3082,7 @@ static VALUE rb_llama_max_devices(VALUE self) {
 extern "C" void Init_llama_cpp(void) {
   rb_mLLaMACpp = rb_define_module("LLaMACpp");
 
+  RbLLaMABatch::define_class(rb_mLLaMACpp);
   RbLLaMATokenData::define_class(rb_mLLaMACpp);
   RbLLaMATokenDataArray::define_class(rb_mLLaMACpp);
   RbLLaMAModel::define_class(rb_mLLaMACpp);
@@ -3144,6 +3156,11 @@ extern "C" void Init_llama_cpp(void) {
   rb_define_const(rb_mLLaMACpp, "LLAMA_ROPE_SCALING_MAX_VALUE", INT2NUM(LLAMA_ROPE_SCALING_MAX_VALUE));
 
   std::stringstream ss_magic;
+  ss_magic << std::showbase << std::hex << LLAMA_FILE_MAGIC_GGLA;
+  rb_define_const(rb_mLLaMACpp, "LLAMA_FILE_MAGIC_GGLA", rb_str_new2(ss_magic.str().c_str()));
+
+  ss_magic.str("");
+  ss_magic.clear(std::stringstream::goodbit);
   ss_magic << std::showbase << std::hex << LLAMA_FILE_MAGIC_GGSN;
   rb_define_const(rb_mLLaMACpp, "LLAMA_FILE_MAGIC_GGSN", rb_str_new2(ss_magic.str().c_str()));
 

data/ext/llama_cpp/src/ggml-alloc.c

@@ -449,11 +449,10 @@ static void init_view(ggml_gallocr_t galloc, struct ggml_tensor * view, bool upd
     if (update_backend) {
         view->backend = view->view_src->backend;
     }
-    view->buffer  = view->view_src->buffer;
+    // views are initialized in the alloc buffer rather than the view_src buffer
+    view->buffer  = alloc->buffer;
     view->data    = (char *)view->view_src->data + view->view_offs;
 
-    // FIXME: the view should be initialized by the owning buffer, but currently this breaks the CUDA backend
-    // due to the ggml_tensor_extra_gpu ring buffer overwriting the KV cache extras
     assert(ggml_tallocr_is_measure(alloc) || !view->buffer || view->buffer->buft == alloc->buffer->buft);
 
     if (!alloc->measure) {
@@ -736,6 +735,10 @@ void ggml_allocr_set_parse_seq(ggml_allocr_t alloc, const int * list, int n) {
 }
 
 void ggml_allocr_free(ggml_allocr_t alloc) {
+    if (alloc == NULL) {
+        return;
+    }
+
     ggml_gallocr_free(alloc->galloc);
     ggml_tallocr_free(alloc->talloc);
     free(alloc);
@@ -775,7 +778,7 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
     }
 
     if (nbytes == 0) {
-        fprintf(stderr, "%s: no tensors to allocate\n", __func__);
+        // all the tensors in the context are already allocated
         return NULL;
     }
 
@@ -789,6 +792,11 @@ ggml_backend_buffer_t ggml_backend_alloc_ctx_tensors_from_buft(struct ggml_conte
             } else {
                 ggml_backend_view_init(buffer, t);
             }
+        } else {
+            if (t->view_src != NULL) {
+                // view of a pre-allocated tensor
+                ggml_backend_view_init(buffer, t);
+            }
         }
     }
 

data/ext/llama_cpp/src/ggml-alloc.h

@@ -43,7 +43,7 @@ GGML_API size_t ggml_allocr_alloc_graph(ggml_allocr_t alloc, struct ggml_cgraph
 // ggml-backend v2 API
 //
 
-// Seperate tensor and graph allocator objects
+// Separate tensor and graph allocator objects
 // This is necessary for multi-backend allocation because the graph allocator needs to use multiple tensor allocators
 // The original API is kept as a wrapper around the new API
 

data/ext/llama_cpp/src/ggml-backend-impl.h

@@ -20,6 +20,9 @@ extern "C" {
         size_t                (*get_alignment)   (ggml_backend_buffer_type_t buft); // tensor alignment
         size_t                (*get_alloc_size)  (ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor); // data size needed to allocate the tensor, including padding
         bool                  (*supports_backend)(ggml_backend_buffer_type_t buft, ggml_backend_t backend); // check if the buffer type is usable by the backend
+        // check if tensor data is in host memory
+        // should be equivalent to supports_backend(buft, ggml_backend_cpu_init())
+        bool                  (*is_host)         (ggml_backend_buffer_type_t buft);
     };
 
     struct ggml_backend_buffer_type {
@@ -31,15 +34,16 @@ extern "C" {
     typedef void * ggml_backend_buffer_context_t;
 
     struct ggml_backend_buffer_i {
-        void     (*free_buffer)(ggml_backend_buffer_t buffer);
+        void   (*free_buffer)    (ggml_backend_buffer_t buffer);
         //void     (*reset)      (ggml_backend_buffer_t buffer); // reset any internal state due to tensor initialization, such as tensor extras
-        void *   (*get_base)   (ggml_backend_buffer_t buffer);
-        void     (*init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
-        void     (*set_tensor) (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
-        void     (*get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
+        void * (*get_base)       (ggml_backend_buffer_t buffer);
+        void   (*init_tensor)    (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+        void   (*set_tensor)     (ggml_backend_buffer_t buffer,       struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+        void   (*get_tensor)     (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
         // (optional) copy tensor between different buffer-type, allow for single-copy tranfers
-        void (*cpy_tensor_from)(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
-        void (*cpy_tensor_to)  (ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void   (*cpy_tensor_from)(ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void   (*cpy_tensor_to)  (ggml_backend_buffer_t buffer, struct ggml_tensor * src, struct ggml_tensor * dst);
+        void   (*clear)          (ggml_backend_buffer_t buffer, uint8_t value);
     };
 
     struct ggml_backend_buffer {
@@ -78,7 +82,7 @@ extern "C" {
         void (*cpy_tensor_from_async)(ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
         void (*cpy_tensor_to_async)  (ggml_backend_t backend, struct ggml_tensor * src, struct ggml_tensor * dst);
 
-        void (*synchronize)     (ggml_backend_t backend);
+        void (*synchronize)(ggml_backend_t backend);
 
         // compute graph with a plan
         ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, struct ggml_cgraph * cgraph);

data/ext/llama_cpp/src/ggml-backend.c

@@ -35,6 +35,13 @@ bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_ba
     return buft->iface.supports_backend(buft, backend);
 }
 
+bool ggml_backend_buft_is_host(ggml_backend_buffer_type_t buft) {
+    if (buft->iface.is_host) {
+        return buft->iface.is_host(buft);
+    }
+    return false;
+}
+
 // backend buffer
 
 ggml_backend_buffer_t ggml_backend_buffer_init(
@@ -94,6 +101,14 @@ size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct g
     return ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type(buffer), tensor);
 }
 
+void ggml_backend_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    buffer->iface.clear(buffer, value);
+}
+
+bool ggml_backend_buffer_is_host(ggml_backend_buffer_t buffer) {
+    return ggml_backend_buft_is_host(ggml_backend_buffer_type(buffer));
+}
+
 ggml_backend_buffer_type_t ggml_backend_buffer_type(ggml_backend_buffer_t buffer) {
     return buffer->buft;
 }
@@ -378,7 +393,6 @@ static void * ggml_backend_cpu_buffer_get_base(ggml_backend_buffer_t buffer) {
 
 static void ggml_backend_cpu_buffer_free_buffer(ggml_backend_buffer_t buffer) {
     free(buffer->context);
-    GGML_UNUSED(buffer);
 }
 
 static void ggml_backend_cpu_buffer_set_tensor(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
@@ -411,6 +425,10 @@ static void ggml_backend_cpu_buffer_cpy_tensor_to(ggml_backend_buffer_t buffer,
     GGML_UNUSED(buffer);
 }
 
+static void ggml_backend_cpu_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
+    memset(buffer->context, value, buffer->size);
+}
+
 static struct ggml_backend_buffer_i cpu_backend_buffer_i = {
     /* .free_buffer     = */ ggml_backend_cpu_buffer_free_buffer,
     /* .get_base        = */ ggml_backend_cpu_buffer_get_base,
@@ -419,6 +437,7 @@ static struct ggml_backend_buffer_i cpu_backend_buffer_i = {
     /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
     /* .cpy_tensor_from = */ ggml_backend_cpu_buffer_cpy_tensor_from,
     /* .cpy_tensor_to   = */ ggml_backend_cpu_buffer_cpy_tensor_to,
+    /* .clear           = */ ggml_backend_cpu_buffer_clear,
 };
 
 // for buffers from ptr, free is not called
@@ -430,6 +449,7 @@ static struct ggml_backend_buffer_i cpu_backend_buffer_i_from_ptr = {
     /* .get_tensor      = */ ggml_backend_cpu_buffer_get_tensor,
     /* .cpy_tensor_from = */ ggml_backend_cpu_buffer_cpy_tensor_from,
     /* .cpy_tensor_to   = */ ggml_backend_cpu_buffer_cpy_tensor_to,
+    /* .clear           = */ ggml_backend_cpu_buffer_clear,
 };
 
 static const size_t TENSOR_ALIGNMENT = 64; // should be enough for AVX 512
@@ -455,20 +475,70 @@ static bool ggml_backend_cpu_buffer_type_supports_backend(ggml_backend_buffer_ty
     GGML_UNUSED(buft);
 }
 
+static bool ggml_backend_cpu_buffer_type_is_host(ggml_backend_buffer_type_t buft) {
+    return true;
+
+    GGML_UNUSED(buft);
+}
+
 ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void) {
-    static struct ggml_backend_buffer_type ggml_backend_buffer_type_cpu = {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type = {
         /* .iface = */ {
             /* .alloc_buffer     = */ ggml_backend_cpu_buffer_type_alloc_buffer,
             /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
             /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
             /* .supports_backend = */ ggml_backend_cpu_buffer_type_supports_backend,
+            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
         },
         /* .context = */ NULL,
     };
 
-    return &ggml_backend_buffer_type_cpu;
+    return &ggml_backend_cpu_buffer_type;
 }
 
+#ifdef GGML_USE_CPU_HBM
+
+// buffer type HBM
+
+#include <hbwmalloc.h>
+
+static void ggml_backend_cpu_hbm_buffer_free_buffer(ggml_backend_buffer_t buffer) {
+    hbw_free(buffer->context);
+}
+
+static ggml_backend_buffer_t ggml_backend_cpu_hbm_buffer_type_alloc_buffer(ggml_backend_buffer_type_t buft, size_t size) {
+    //void * ptr = hbw_malloc(size);
+    void * ptr;
+    int result = hbw_posix_memalign(&ptr, ggml_backend_cpu_buffer_type_get_alignment(buft), size);
+    if (result != 0) {
+        fprintf(stderr, "failed to allocate HBM buffer of size %zu\n", size);
+        return NULL;
+    }
+
+    // FIXME: this is a hack to avoid having to implement a new buffer type
+    ggml_backend_buffer_t buffer = ggml_backend_cpu_buffer_from_ptr(ptr, size);
+    buffer->buft = buft;
+    buffer->iface.free_buffer = ggml_backend_cpu_hbm_buffer_free_buffer;
+
+    return buffer;
+}
+
+ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type() {
+    static struct ggml_backend_buffer_type ggml_backend_cpu_buffer_type_hbm = {
+        /* .iface    = */ {
+            /* .alloc_buffer     = */ ggml_backend_cpu_hbm_buffer_type_alloc_buffer,
+            /* .get_alignment    = */ ggml_backend_cpu_buffer_type_get_alignment,
+            /* .get_alloc_size   = */ NULL, // defaults to ggml_nbytes
+            /* .supports_backend = */ ggml_backend_cpu_buffer_type_supports_backend,
+            /* .is_host          = */ ggml_backend_cpu_buffer_type_is_host,
+        },
+        /* .context  = */ NULL,
+    };
+
+    return &ggml_backend_cpu_buffer_type_hbm;
+}
+#endif
+
 struct ggml_backend_cpu_context {
     int n_threads;
     void * work_data;
@@ -505,7 +575,7 @@ static ggml_backend_graph_plan_t ggml_backend_cpu_graph_plan_create(ggml_backend
     struct ggml_backend_plan_cpu * cpu_plan = malloc(sizeof(struct ggml_backend_plan_cpu));
 
     cpu_plan->cplan = ggml_graph_plan(cgraph, cpu_ctx->n_threads);
-    cpu_plan->cgraph = *cgraph;
+    cpu_plan->cgraph = *cgraph; // FIXME: deep copy
 
     if (cpu_plan->cplan.work_size > 0) {
         cpu_plan->cplan.work_data = malloc(cpu_plan->cplan.work_size);
@@ -1180,7 +1250,7 @@ void ggml_backend_sched_set_node_backend(ggml_backend_sched_t sched, struct ggml
 // utils
 void ggml_backend_view_init(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor) {
     GGML_ASSERT(tensor->buffer == NULL);
-    GGML_ASSERT(tensor->data == NULL);
+    //GGML_ASSERT(tensor->data == NULL); // views of pre-allocted tensors may have the data set, but still need to be initialized
     GGML_ASSERT(tensor->view_src != NULL);
     GGML_ASSERT(tensor->view_src->buffer != NULL);
     GGML_ASSERT(tensor->view_src->data != NULL);

data/ext/llama_cpp/src/ggml-backend.h

@@ -21,6 +21,7 @@ extern "C" {
     GGML_API size_t ggml_backend_buft_get_alignment (ggml_backend_buffer_type_t buft);
     GGML_API size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, struct ggml_tensor * tensor);
     GGML_API bool ggml_backend_buft_supports_backend(ggml_backend_buffer_type_t buft, ggml_backend_t backend);
+    GGML_API bool ggml_backend_buft_is_host         (ggml_backend_buffer_type_t buft);
 
     // buffer
     GGML_API void   ggml_backend_buffer_free          (ggml_backend_buffer_t buffer);
@@ -29,6 +30,8 @@ extern "C" {
     GGML_API void   ggml_backend_buffer_init_tensor   (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
     GGML_API size_t ggml_backend_buffer_get_alignment (ggml_backend_buffer_t buffer);
     GGML_API size_t ggml_backend_buffer_get_alloc_size(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+    GGML_API void   ggml_backend_buffer_clear         (ggml_backend_buffer_t buffer, uint8_t value);
+    GGML_API bool   ggml_backend_buffer_is_host       (ggml_backend_buffer_t buffer);
     GGML_API ggml_backend_buffer_type_t ggml_backend_buffer_type(ggml_backend_buffer_t buffer);
 
     //
@@ -76,6 +79,10 @@ extern "C" {
 
     GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_buffer_type(void);
 
+#ifdef GGML_USE_CPU_HBM
+    GGML_API ggml_backend_buffer_type_t ggml_backend_cpu_hbm_buffer_type(void);
+#endif
+
     //
     // Backend registry
     //