llama_cpp 0.4.0 → 0.5.1

checksums.yaml

@@ -1,7 +1,7 @@
 ---
 SHA256:
-  metadata.gz: af3a0e01bc9f3cfad4cee3f21144dd354640e1d4558125be36d4b499fa3b4c24
-  data.tar.gz: 042a3b0491d98fa6a093c684e6ab751152f37c8438a3b4a7b19cb2d8c7ab95a7
+  metadata.gz: fd67587510fff74b8b1d55e2e5861711709dfb5d8c44cf40b3bf762276e57d5b
+  data.tar.gz: 5cb5319136e538eb2ec9a6406caaaacdabdb2dceec5cade43769eda1b02de9c5
 SHA512:
-  metadata.gz: 7ed85bd8438ee3b3adab884795c4aecb5b0d72ad57b7e02bc281b62c3b1d669efab62a020e03b09defe3084ecd8afacc4220303e99167d04d668650768c7392b
-  data.tar.gz: b705a0ccd2c7c1e15aed6383acb9d5a3d79d0a0c882a74c42b9099df9a27aff88ba08a2f06aa4d195382e8f41c1b16c0014a2047d1923369f275ca481d52bb21
+  metadata.gz: c2ab28fe9bf5674976ff2e676ea4d76157bd2ebf24b92ca2f959a6cdf2c19de94fe95d76ab21ca313d9017f835387b0f9ad616cb3700024fc5394fa1e9984fda
+  data.tar.gz: 0ce0be3db250eb7d35f3784bd7a3bd54e7ab8833378745417da3504f69bc31910d4fec459d29ad28218fce2614e8321462e9873c96ed1c3793eb5f9bbe5a9eac

data/CHANGELOG.md

@@ -1,3 +1,16 @@
+## [[0.5.1](https://github.com/yoshoku/llama_cpp.rb/compare/v0.5.0...v0.5.1)] - 2023-09-08
+
+- Bump bundled llama.cpp from master-b1140 to master-b1198.
+
+## [[0.5.0](https://github.com/yoshoku/llama_cpp.rb/compare/v0.4.0...v0.5.0)] - 2023-09-02
+
+**Breaking Changes**
+- Bump bundled llama.cpp from master-b1060 to master-b1140.
+  - Rename `token_to_str` method on Context to `token_to_piece` method.
+  - Rename `token_to_str` method on Model to `token_to_piece` method.
+  - Rename `type` method on Model to `desc` method.
+  - Add `size` and `n_params` methods to Model.
+
 ## [[0.4.0](https://github.com/yoshoku/llama_cpp.rb/compare/v0.3.8...v0.4.0)] - 2023-08-26
 
 **Breaking Changes**

data/examples/chat.rb

@@ -122,7 +122,7 @@ class Chat < Thor # rubocop:disable Metrics/ClassLength, Style/Documentation
 
       if input_echo
         output = []
-        embd.each { |token| output << context.token_to_str(token) }
+        embd.each { |token| output << context.token_to_piece(token) }
         output_str = output.join
         output_str.chomp!(antiprompt) if first_input
         print(output_str)
@@ -131,7 +131,7 @@ class Chat < Thor # rubocop:disable Metrics/ClassLength, Style/Documentation
       if embd_input.size <= n_consumed
         if antiprompt.size.positive?
           last_output = []
-          last_n_tokens.each { |token| last_output << context.token_to_str(token) }
+          last_n_tokens.each { |token| last_output << context.token_to_piece(token) }
           last_output_str = last_output.join
 
           search_start_pos = last_output_str.size > antiprompt.size ? last_output_str.size - antiprompt.size : 0

data/ext/llama_cpp/extconf.rb

@@ -50,7 +50,7 @@ if with_config('accelerate')
 end
 
 if with_config('metal')
-  $CFLAGS << ' -DGGML_USE_METAL -DGGML_METAL_NDEBUG'
+  $CFLAGS << ' -DGGML_USE_METAL'
   $CXXFLAGS << ' -DGGML_USE_METAL'
   $LDFLAGS << ' -framework Foundation -framework Metal -framework MetalKit'
   $objs = %w[ggml.o ggml-alloc.o ggml-metal.o llama.o llama_cpp.o]

data/ext/llama_cpp/llama_cpp.cpp

@@ -811,9 +811,11 @@ public:
     rb_define_method(rb_cLLaMAModel, "n_vocab", RUBY_METHOD_FUNC(_llama_model_get_model_n_vocab), 0);
     rb_define_method(rb_cLLaMAModel, "n_ctx", RUBY_METHOD_FUNC(_llama_model_get_model_n_ctx), 0);
     rb_define_method(rb_cLLaMAModel, "n_embd", RUBY_METHOD_FUNC(_llama_model_get_model_n_embd), 0);
-    rb_define_method(rb_cLLaMAModel, "token_to_str", RUBY_METHOD_FUNC(_llama_model_token_to_str_with_model), 1);
+    rb_define_method(rb_cLLaMAModel, "token_to_piece", RUBY_METHOD_FUNC(_llama_model_token_to_piece_with_model), 1);
     rb_define_method(rb_cLLaMAModel, "tokenize", RUBY_METHOD_FUNC(_llama_model_tokenize_with_model), -1);
-    rb_define_method(rb_cLLaMAModel, "type", RUBY_METHOD_FUNC(_llama_model_get_model_type), 0);
+    rb_define_method(rb_cLLaMAModel, "desc", RUBY_METHOD_FUNC(_llama_model_get_model_desc), 0);
+    rb_define_method(rb_cLLaMAModel, "size", RUBY_METHOD_FUNC(_llama_model_get_model_size), 0);
+    rb_define_method(rb_cLLaMAModel, "n_params", RUBY_METHOD_FUNC(_llama_model_get_model_n_params), 0);
   }
 
 private:
@@ -974,7 +976,7 @@ private:
     return INT2NUM(llama_model_n_embd(ptr->model));
   }
 
-  static VALUE _llama_model_token_to_str_with_model(VALUE self, VALUE token_) {
+  static VALUE _llama_model_token_to_piece_with_model(VALUE self, VALUE token_) {
     if (!RB_INTEGER_TYPE_P(token_)) {
       rb_raise(rb_eArgError, "token must be an integer");
       return Qnil;
@@ -982,10 +984,10 @@ private:
     const llama_token token = NUM2INT(token_);
     LLaMAModelWrapper* ptr = get_llama_model(self);
     std::vector<char> result(8, 0);
-    const int n_tokens = llama_token_to_str_with_model(ptr->model, token, result.data(), result.size());
+    const int n_tokens = llama_token_to_piece_with_model(ptr->model, token, result.data(), result.size());
     if (n_tokens < 0) {
       result.resize(-n_tokens);
-      const int check = llama_token_to_str_with_model(ptr->model, token, result.data(), result.size());
+      const int check = llama_token_to_piece_with_model(ptr->model, token, result.data(), result.size());
       if (check != -n_tokens) {
         rb_raise(rb_eRuntimeError, "failed to convert");
         return Qnil;
@@ -1040,12 +1042,22 @@ private:
     return ret;
   }
 
-  static VALUE _llama_model_get_model_type(VALUE self) {
+  static VALUE _llama_model_get_model_desc(VALUE self) {
     LLaMAModelWrapper* ptr = get_llama_model(self);
     char buf[128];
-    ::llama_model_type(ptr->model, buf, sizeof(buf));
+    llama_model_desc(ptr->model, buf, sizeof(buf));
     return rb_str_new_cstr(buf);
   }
+
+  static VALUE _llama_model_get_model_size(VALUE self) {
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    return UINT2NUM(llama_model_size(ptr->model));
+  }
+
+  static VALUE _llama_model_get_model_n_params(VALUE self) {
+    LLaMAModelWrapper* ptr = get_llama_model(self);
+    return UINT2NUM(llama_model_n_params(ptr->model));
+  }
 };
 
 const rb_data_type_t RbLLaMAModel::llama_model_type = {
@@ -1326,7 +1338,7 @@ public:
     rb_define_method(rb_cLLaMAContext, "token_bos", RUBY_METHOD_FUNC(_llama_context_token_bos), 0);
     rb_define_method(rb_cLLaMAContext, "token_eos", RUBY_METHOD_FUNC(_llama_context_token_eos), 0);
     rb_define_method(rb_cLLaMAContext, "token_nl", RUBY_METHOD_FUNC(_llama_context_token_nl), 0);
-    rb_define_method(rb_cLLaMAContext, "token_to_str", RUBY_METHOD_FUNC(_llama_context_token_to_str), 1);
+    rb_define_method(rb_cLLaMAContext, "token_to_piece", RUBY_METHOD_FUNC(_llama_context_token_to_piece), 1);
     rb_define_method(rb_cLLaMAContext, "n_vocab", RUBY_METHOD_FUNC(_llama_context_n_vocab), 0);
     rb_define_method(rb_cLLaMAContext, "n_ctx", RUBY_METHOD_FUNC(_llama_context_n_ctx), 0);
     rb_define_method(rb_cLLaMAContext, "n_embd", RUBY_METHOD_FUNC(_llama_context_n_embd), 0);
@@ -1567,7 +1579,7 @@ private:
     return output;
   }
 
-  static VALUE _llama_context_token_to_str(VALUE self, VALUE token_) {
+  static VALUE _llama_context_token_to_piece(VALUE self, VALUE token_) {
     LLaMAContextWrapper* ptr = get_llama_context(self);
     if (ptr->ctx == NULL) {
       rb_raise(rb_eRuntimeError, "LLaMA context is not initialized");
@@ -1575,10 +1587,10 @@ private:
     }
     const llama_token token = NUM2INT(token_);
     std::vector<char> result(8, 0);
-    const int n_tokens = llama_token_to_str(ptr->ctx, token, result.data(), result.size());
+    const int n_tokens = llama_token_to_piece(ptr->ctx, token, result.data(), result.size());
     if (n_tokens < 0) {
       result.resize(-n_tokens);
-      const int check = llama_token_to_str(ptr->ctx, token, result.data(), result.size());
+      const int check = llama_token_to_piece(ptr->ctx, token, result.data(), result.size());
       if (check != -n_tokens) {
         rb_raise(rb_eRuntimeError, "failed to convert");
         return Qnil;

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

@@ -1,3 +1,8 @@
+// defines MAP_ANONYMOUS
+#ifndef _GNU_SOURCE
+#define _GNU_SOURCE
+#endif
+
 #include "ggml-alloc.h"
 #include "ggml.h"
 #include <assert.h>
@@ -6,6 +11,26 @@
 #include <stdlib.h>
 #include <string.h>
 
+#ifdef __has_include
+    #if __has_include(<unistd.h>)
+        #include <unistd.h>
+        #if defined(_POSIX_MAPPED_FILES)
+            #include <sys/types.h>
+            #include <sys/mman.h>
+        #endif
+    #endif
+#endif
+
+#if defined(_WIN32)
+    #define WIN32_LEAN_AND_MEAN
+    #ifndef NOMINMAX
+        #define NOMINMAX
+    #endif
+    #include <windows.h>
+    #include <memoryapi.h>
+#endif
+
+
 #define UNUSED(x) (void)(x)
 #define MAX(a, b) ((a) > (b) ? (a) : (b))
 #define GGML_MAX_CONCUR (2*GGML_MAX_NODES)
@@ -99,15 +124,24 @@ static void remove_allocated_tensor(struct ggml_allocr * alloc, struct ggml_tens
 }
 #endif
 
-
-static size_t ggml_allocator_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+static size_t ggml_allocr_get_alloc_size(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
     return ggml_nbytes(tensor);
 
     UNUSED(alloc);
 }
 
+// check if a tensor is allocated by this buffer
+static bool ggml_allocr_is_own(struct ggml_allocr * alloc, const struct ggml_tensor * tensor) {
+    void * ptr = tensor->data;
+    return ptr >= alloc->data && (char *)ptr < (char *)alloc->data + alloc->max_size;
+}
+
 void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
-    size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
+#ifdef GGML_ALLOCATOR_DEBUG
+    GGML_ASSERT(!ggml_is_view(tensor)); // views generally get data pointer from one of their sources
+    GGML_ASSERT(tensor->data == NULL); // avoid allocating tensor which already has memory allocated
+#endif
+    size_t size = ggml_allocr_get_alloc_size(alloc, tensor);
     size = aligned_offset(NULL, size, alloc->alignment);
 
     AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
@@ -131,14 +165,14 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor)
     if (best_fit_block == -1) {
         // the last block is our last resort
         struct free_block * block = &alloc->free_blocks[alloc->n_free_blocks - 1];
+        max_avail = MAX(max_avail, block->size);
         if (block->size >= size) {
             best_fit_block = alloc->n_free_blocks - 1;
-            max_avail = MAX(max_avail, block->size);
         } else {
             fprintf(stderr, "%s: not enough space in the buffer (needed %zu, largest block available %zu)\n",
                     __func__, size, max_avail);
             GGML_ASSERT(!"not enough space in the buffer");
-        return;
+            return;
         }
     }
     struct free_block * block = &alloc->free_blocks[best_fit_block];
@@ -173,17 +207,17 @@ void ggml_allocr_alloc(struct ggml_allocr * alloc, struct ggml_tensor * tensor)
 }
 
 // this is a very naive implementation, but for our case the number of free blocks should be very small
-static void ggml_allocator_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
+static void ggml_allocr_free_tensor(struct ggml_allocr * alloc, struct ggml_tensor * tensor) {
     void * ptr = tensor->data;
 
-    if (ptr < alloc->data || (char*)ptr >= (char*)alloc->data + alloc->max_size) {
+    if (ggml_allocr_is_own(alloc, tensor) == false) {
         // the tensor was not allocated in this buffer
         // this can happen because the graph allocator will try to free weights and other tensors from different buffers
         // the easiest way to deal with this is just to ignore it
         return;
     }
 
-    size_t size = ggml_allocator_get_alloc_size(alloc, tensor);
+    size_t size = ggml_allocr_get_alloc_size(alloc, tensor);
     size = aligned_offset(NULL, size, alloc->alignment);
     AT_PRINTF("%s: freeing %s (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, size, alloc->n_free_blocks);
 
@@ -268,7 +302,7 @@ struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment)
         /*.parse_seq     = */ {0},
         /*.parse_seq_len = */ 0,
 #ifdef GGML_ALLOCATOR_DEBUG
-        /*.allocated_tensors = */ = {0},
+        /*.allocated_tensors = */ {0},
 #endif
     };
 
@@ -277,17 +311,68 @@ struct ggml_allocr * ggml_allocr_new(void * data, size_t size, size_t alignment)
     return alloc;
 }
 
-// address and size of the buffer when measuring
-// it needs to be large enough to fit all the tensors, but it cannot overlap with other existing buffers
-static void * const MEASURE_BASE_ADDR = (void *) 0x1000;
-static const size_t MEASURE_MAX_SIZE  = 1ULL<<40; // 1 TB
+// OS specific functions to allocate and free uncommitted virtual memory
+static void * alloc_vmem(size_t size) {
+#if defined(_WIN32)
+    return VirtualAlloc(NULL, size, MEM_RESERVE, PAGE_NOACCESS);
+#elif defined(_POSIX_MAPPED_FILES)
+    void * ptr = mmap(NULL, size, PROT_NONE, MAP_PRIVATE | MAP_ANON, -1, 0);
+    if (ptr == MAP_FAILED) {
+        return NULL;
+    }
+    return ptr;
+#else
+    // use a fixed address for other platforms
+    uintptr_t base_addr = (uintptr_t)-size - 0x100;
+    return (void *)base_addr;
+#endif
+}
+
+static void free_vmem(void * base_addr, size_t size) {
+#if defined(_WIN32)
+    VirtualFree(base_addr, 0, MEM_RELEASE);
+    UNUSED(size);
+#elif defined(_POSIX_MAPPED_FILES)
+    munmap(base_addr, size);
+#else
+    // nothing to do
+    UNUSED(base_addr);
+    UNUSED(size);
+#endif
+}
+
+// allocate uncommitted virtual memory to measure the size of the graph
+static void alloc_measure_vmem(void ** base_addr, size_t * size) {
+    // 1TB for 64-bit, 1GB for 32-bit
+    *size = sizeof(void *) == 4 ? 1ULL<<30 : 1ULL<<40;
+    do {
+        *base_addr = alloc_vmem(*size);
+        if (*base_addr != NULL) {
+            AT_PRINTF("allocated %.2f GB of virtual memory for measure buffer at %p\n", *size / 1024.0 / 1024.0 / 1024.0, *base_addr);
+            return;
+        }
+        // try again with half the size
+        *size /= 2;
+    } while (*size > 0);
+
+    GGML_ASSERT(!"failed to allocate virtual memory for measure buffer");
+}
+
+static void free_measure_vmem(void * base_addr, size_t size) {
+    free_vmem(base_addr, size);
+}
 
 struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
     struct ggml_allocr * alloc = (struct ggml_allocr *)malloc(sizeof(struct ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
 
+    void * base_addr;
+    size_t size;
+
+    alloc_measure_vmem(&base_addr, &size);
+
     *alloc = (struct ggml_allocr){
-        /*.data          = */ MEASURE_BASE_ADDR,
-        /*.size          = */ MEASURE_MAX_SIZE,
+        /*.data          = */ base_addr,
+        /*.size          = */ size,
         /*.alignment     = */ alignment,
         /*.n_free_blocks = */ 0,
         /*.free_blocks   = */ {{0}},
@@ -297,7 +382,7 @@ struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
         /*.parse_seq     = */ {0},
         /*.parse_seq_len = */ 0,
 #ifdef GGML_ALLOCATOR_DEBUG
-        /*.allocated_tensors = */ = {0},
+        /*.allocated_tensors = */ {0},
 #endif
     };
 
@@ -307,6 +392,9 @@ struct ggml_allocr * ggml_allocr_new_measure(size_t alignment) {
 }
 
 void ggml_allocr_free(struct ggml_allocr * alloc) {
+    if (alloc->measure) {
+        free_measure_vmem(alloc->data, alloc->size);
+    }
     free(alloc);
 }
 
@@ -317,8 +405,7 @@ bool ggml_allocr_is_measure(struct ggml_allocr * alloc) {
 //////////// compute graph allocator
 
 static bool ggml_is_view(struct ggml_tensor * t) {
-    return t->op == GGML_OP_RESHAPE || t->op == GGML_OP_VIEW || t->op == GGML_OP_TRANSPOSE ||
-           t->op == GGML_OP_PERMUTE || t->op == GGML_OP_CPY;
+    return t->view_src != NULL;
 }
 
 static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml_tensor * b) {
@@ -336,28 +423,6 @@ static bool ggml_are_same_layout(const struct ggml_tensor * a, const struct ggml
     return true;
 }
 
-static struct ggml_tensor * get_view_parent(struct ggml_tensor * t) {
-    switch (t->op) {
-        case GGML_OP_PERMUTE:
-        case GGML_OP_RESHAPE:
-        case GGML_OP_TRANSPOSE:
-        case GGML_OP_VIEW:
-            return t->src[0];
-        case GGML_OP_CPY:
-            return t->src[1];
-        default:
-            return NULL;
-    }
-}
-
-static struct ggml_tensor * get_view_source(struct ggml_tensor * t) {
-    struct ggml_tensor * parent = t;
-    do {
-        parent = get_view_parent(parent);
-    } while (ggml_is_view(parent));
-    return parent;
-}
-
 static bool ggml_op_can_inplace(enum ggml_op op) {
     switch (op) {
         case GGML_OP_SCALE:
@@ -365,7 +430,6 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
         case GGML_OP_DIAG_MASK_INF:
         case GGML_OP_ADD:
         case GGML_OP_ADD1:
-        case GGML_OP_ACC:
         case GGML_OP_SUB:
         case GGML_OP_MUL:
         case GGML_OP_DIV:
@@ -375,7 +439,6 @@ static bool ggml_op_can_inplace(enum ggml_op op) {
         case GGML_OP_UNARY:
         case GGML_OP_ROPE:
         case GGML_OP_RMS_NORM:
-        case GGML_OP_SET:
         case GGML_OP_SOFT_MAX:
         case GGML_OP_CONT:
             return true;
@@ -389,24 +452,8 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
     struct hash_node * ht = alloc->hash_table;
     if (node->data == NULL) {
         if (ggml_is_view(node)) {
-            size_t offset;
-            switch(node->op) {
-                case GGML_OP_VIEW:
-                    memcpy(&offset, node->op_params, sizeof(size_t));
-                    node->data = (char *) node->src[0]->data + offset;
-                    break;
-                case GGML_OP_PERMUTE:
-                case GGML_OP_RESHAPE:
-                case GGML_OP_TRANSPOSE:
-                    node->data = node->src[0]->data;
-                    break;
-                case GGML_OP_CPY:
-                    node->data = node->src[1]->data;
-                    break;
-                default:
-                    GGML_ASSERT(!"unknown view op");
-                    break;
-            }
+            assert(node->view_src->data != NULL);
+            node->data = (char *)node->view_src->data + node->view_offs;
         } else {
             // see if we can reuse a parent's buffer (inplace)
             if (ggml_op_can_inplace(node->op)) {
@@ -417,8 +464,7 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
                     }
 
                     // if the node's data is external, then we cannot re-use it
-                    if ((char *) parent->data < (char *) alloc->data ||
-                        (char *) parent->data >= ((char *) alloc->data + alloc->size)) {
+                    if (ggml_allocr_is_own(alloc, parent) == false) {
                         AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data);
                         continue;
                     }
@@ -426,7 +472,7 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
                     struct hash_node * p_hn = hash_get(ht, parent);
                     if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && ggml_are_same_layout(node, parent)) {
                         if (ggml_is_view(parent)) {
-                            struct ggml_tensor * view_src = get_view_source(parent);
+                            struct ggml_tensor * view_src = parent->view_src;
                             struct hash_node * view_src_hn = hash_get(ht, view_src);
                             if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
                                 // TODO: the offset of the view parent must be kept to ensure that the op doesn't overwrite
@@ -452,7 +498,7 @@ static void allocate_node(struct ggml_allocr * alloc, struct ggml_tensor * node)
     }
 }
 
-static size_t ggml_allocator_alloc_graph_tensors_n(
+static size_t ggml_allocr_alloc_graph_tensors_n(
     struct ggml_allocr * alloc,
     struct ggml_cgraph ** graphs, int n_graphs,
     struct ggml_tensor *** inputs, struct ggml_tensor *** outputs) {
@@ -468,7 +514,7 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
             struct ggml_tensor * node = gf->nodes[i];
 
             if (ggml_is_view(node)) {
-                struct ggml_tensor * view_src = get_view_source(node);
+                struct ggml_tensor * view_src = node->view_src;
                 hash_get(ht, view_src)->n_views += 1;
             }
 
@@ -530,11 +576,10 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
                 AT_PRINTF("\n");
             }
 
-
             // update parents
             // update immediately if there is no parse_seq
             // update only at barriers if there is parse_seq
-            if ((alloc->parse_seq_len==0) || alloc->parse_seq[ind] == -1) {
+            if ((alloc->parse_seq_len == 0) || alloc->parse_seq[ind] == -1) {
                 int update_start = alloc->parse_seq_len ? last_barrier_pos : ind;
                 int update_end   = alloc->parse_seq_len ? ind              : ind + 1;
                 for (int i = update_start; i < update_end; i++) {
@@ -553,17 +598,17 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
 
                         if (p_hn->n_children == 0 && p_hn->n_views == 0) {
                             if (ggml_is_view(parent)) {
-                                struct ggml_tensor * view_src = get_view_source(parent);
+                                struct ggml_tensor * view_src = parent->view_src;
                                 struct hash_node * view_src_hn = hash_get(ht, view_src);
                                 view_src_hn->n_views -= 1;
-                                AT_PRINTF("view_src %s\n", view_src->name);
+                                AT_PRINTF("view_src %s: %d children, %d views\n", view_src->name, view_src_hn->n_children, view_src_hn->n_views);
                                 if (view_src_hn->n_views == 0 && view_src_hn->n_children == 0 && view_src->data != node->data) {
-                                    ggml_allocator_free_tensor(alloc, view_src);
+                                    ggml_allocr_free_tensor(alloc, view_src);
                                 }
                             }
                             else {
                                 if (parent->data != node->data) {
-                                    ggml_allocator_free_tensor(alloc, parent);
+                                    ggml_allocr_free_tensor(alloc, parent);
                                 }
                             }
                         }
@@ -580,7 +625,7 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
             for (int i = 0; outputs[g][i] != NULL; i++) {
                 struct ggml_tensor * output = outputs[g][i];
                 AT_PRINTF("output: %s\n", output->name);
-                ggml_allocator_free_tensor(alloc, output);
+                ggml_allocr_free_tensor(alloc, output);
             }
         }
     }
@@ -589,5 +634,5 @@ static size_t ggml_allocator_alloc_graph_tensors_n(
 }
 
 size_t ggml_allocr_alloc_graph(struct ggml_allocr * alloc, struct ggml_cgraph * graph) {
-    return ggml_allocator_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
+    return ggml_allocr_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
 }