whisper.rn 0.4.0-rc.1 → 0.4.0-rc.10

package/cpp/ggml-alloc.c

@@ -1,146 +1,151 @@
 #include "ggml-alloc.h"
+#include "ggml-backend-impl.h"
 #include "ggml.h"
+#include "ggml-impl.h"
 #include <assert.h>
+#include <limits.h>
 #include <stdarg.h>
 #include <stdio.h>
 #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 WSP_GGML_MAX_CONCUR (2*WSP_GGML_MAX_NODES)
+#define MAX_FREE_BLOCKS 256
 
 //#define WSP_GGML_ALLOCATOR_DEBUG
 
-//#define AT_PRINTF printf
-#define AT_PRINTF(...) ((void)0)
+//#define AT_PRINTF(...) WSP_GGML_LOG_DEBUG(__VA_ARGS__)
+#define AT_PRINTF(...)
 
-struct hash_node {
-    struct wsp_ggml_tensor * t;
-    int n_children;
-    int n_views;
-};
 
-static size_t hash(void * p) {
-    return (size_t)p % WSP_GGML_GRAPH_HASHTABLE_SIZE;
+static bool wsp_ggml_is_view(const struct wsp_ggml_tensor * t) {
+    return t->view_src != NULL;
 }
 
-static struct hash_node * hash_get(struct hash_node hash_table[], struct wsp_ggml_tensor * t) {
-    size_t h = hash(t);
-
-    // linear probing
-    size_t i = h;
-    while (hash_table[i].t != NULL) {
-        if (hash_table[i].t == t) {
-            return &hash_table[i];
+static bool wsp_ggml_are_same_layout(const struct wsp_ggml_tensor * a, const struct wsp_ggml_tensor * b) {
+    if (a->type != b->type) {
+        return false;
+    }
+    for (int i = 0; i < WSP_GGML_MAX_DIMS; i++) {
+        if (a->ne[i] != b->ne[i]) {
+            return false;
         }
-        i = (i + 1) % WSP_GGML_GRAPH_HASHTABLE_SIZE;
-        if (i == h) {
-            // hash table is full
-            WSP_GGML_ASSERT(false);
+        if (a->nb[i] != b->nb[i]) {
+            return false;
         }
     }
+    return true;
+}
 
-    hash_table[i].t = t;
-    return &hash_table[i];
+static bool wsp_ggml_op_can_inplace(enum wsp_ggml_op op) {
+    switch (op) {
+        case WSP_GGML_OP_SCALE:
+        case WSP_GGML_OP_DIAG_MASK_ZERO:
+        case WSP_GGML_OP_DIAG_MASK_INF:
+        case WSP_GGML_OP_ADD:
+        case WSP_GGML_OP_ADD1:
+        case WSP_GGML_OP_SUB:
+        case WSP_GGML_OP_MUL:
+        case WSP_GGML_OP_DIV:
+        case WSP_GGML_OP_SQR:
+        case WSP_GGML_OP_SQRT:
+        case WSP_GGML_OP_LOG:
+        case WSP_GGML_OP_UNARY:
+        case WSP_GGML_OP_ROPE:
+        case WSP_GGML_OP_RMS_NORM:
+        case WSP_GGML_OP_SOFT_MAX:
+            return true;
+
+        default:
+            return false;
+    }
 }
 
-// TODO: WSP_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;
     return offset + align;
 }
 
+// tallocr
+
+struct wsp_ggml_tallocr wsp_ggml_tallocr_new(wsp_ggml_backend_buffer_t buffer) {
+    void * base = wsp_ggml_backend_buffer_get_base(buffer);
+    size_t align = wsp_ggml_backend_buffer_get_alignment(buffer);
+
+    assert(align && !(align & (align - 1))); // power of 2
+
+    struct wsp_ggml_tallocr talloc = (struct wsp_ggml_tallocr) {
+        /*.buffer    = */ buffer,
+        /*.base      = */ base,
+        /*.alignment = */ align,
+        /*.offset    = */ aligned_offset(base, 0, align),
+    };
+    return talloc;
+}
+
+void wsp_ggml_tallocr_alloc(struct wsp_ggml_tallocr * talloc, struct wsp_ggml_tensor * tensor) {
+    size_t size = wsp_ggml_backend_buffer_get_alloc_size(talloc->buffer, tensor);
+    size = WSP_GGML_PAD(size, talloc->alignment);
+
+    if (talloc->offset + size > wsp_ggml_backend_buffer_get_size(talloc->buffer)) {
+        WSP_GGML_LOG_ERROR("%s: not enough space in the buffer to allocate %s (needed %zu, available %zu)\n",
+                __func__, tensor->name, size, wsp_ggml_backend_buffer_get_size(talloc->buffer) - talloc->offset);
+        WSP_GGML_ABORT("not enough space in the buffer");
+    }
+
+    void * addr = (char *)wsp_ggml_backend_buffer_get_base(talloc->buffer) + talloc->offset;
+    talloc->offset += size;
+
+    assert(((uintptr_t)addr % talloc->alignment) == 0);
+
+    wsp_ggml_backend_tensor_alloc(talloc->buffer, tensor, addr);
+}
+
+// dynamic tensor allocator
+
 struct free_block {
-    void * addr;
+    size_t offset;
     size_t size;
 };
 
-#define MAX_FREE_BLOCKS 128
-
-struct wsp_ggml_allocr {
-    void * data;
-    size_t size;
+struct wsp_ggml_dyn_tallocr {
     size_t alignment;
     int n_free_blocks;
     struct free_block free_blocks[MAX_FREE_BLOCKS];
-    struct hash_node hash_table[WSP_GGML_GRAPH_HASHTABLE_SIZE];
     size_t max_size;
-    bool measure;
-    int parse_seq[WSP_GGML_MAX_CONCUR];
-    int parse_seq_len;
 
 #ifdef WSP_GGML_ALLOCATOR_DEBUG
-    struct wsp_ggml_tensor * allocated_tensors[1024];
+    struct {
+        const struct wsp_ggml_tensor * tensor;
+        size_t offset;
+    } allocated_tensors[1024];
 #endif
 };
 
 #ifdef WSP_GGML_ALLOCATOR_DEBUG
-static void add_allocated_tensor(struct wsp_ggml_allocr * alloc, struct wsp_ggml_tensor * tensor) {
+static void add_allocated_tensor(struct wsp_ggml_dyn_tallocr * alloc, size_t offset, const struct wsp_ggml_tensor * tensor) {
     for (int i = 0; i < 1024; i++) {
-        if (alloc->allocated_tensors[i] == NULL) {
-            alloc->allocated_tensors[i] = tensor;
+        if (alloc->allocated_tensors[i].tensor == NULL) {
+            alloc->allocated_tensors[i].tensor = tensor;
+            alloc->allocated_tensors[i].offset = offset;
             return;
         }
     }
-    WSP_GGML_ASSERT(!"out of allocated_tensors");
+    WSP_GGML_ABORT("out of allocated_tensors");
 }
-static void remove_allocated_tensor(struct wsp_ggml_allocr * alloc, struct wsp_ggml_tensor * tensor) {
+static void remove_allocated_tensor(struct wsp_ggml_dyn_tallocr * alloc, size_t offset, const struct wsp_ggml_tensor * tensor) {
     for (int i = 0; i < 1024; i++) {
-        if (alloc->allocated_tensors[i] == tensor ||
-            (alloc->allocated_tensors[i] != NULL && alloc->allocated_tensors[i]->data == tensor->data)) {
-            alloc->allocated_tensors[i] = NULL;
+        if (alloc->allocated_tensors[i].offset == offset) {
+            alloc->allocated_tensors[i].tensor = NULL;
             return;
         }
     }
-    printf("tried to free tensor %s not found\n", tensor->name);
-    WSP_GGML_ASSERT(!"tensor not found");
+    WSP_GGML_ABORT("tried to free tensor %s not found\n", tensor->name);
 }
 #endif
 
-static size_t wsp_ggml_allocr_get_alloc_size(struct wsp_ggml_allocr * alloc, struct wsp_ggml_tensor * tensor) {
-    return wsp_ggml_nbytes(tensor);
-
-    UNUSED(alloc);
-}
-
-// check if a tensor is allocated by this buffer
-static bool wsp_ggml_allocr_is_own(struct wsp_ggml_allocr * alloc, const struct wsp_ggml_tensor * tensor) {
-    void * ptr = tensor->data;
-    return ptr >= alloc->data && (char *)ptr < (char *)alloc->data + alloc->max_size;
-}
-
-static bool wsp_ggml_is_view(struct wsp_ggml_tensor * t) {
-    return t->view_src != NULL;
-}
-
-void wsp_ggml_allocr_alloc(struct wsp_ggml_allocr * alloc, struct wsp_ggml_tensor * tensor) {
-#ifdef WSP_GGML_ALLOCATOR_DEBUG
-    WSP_GGML_ASSERT(!wsp_ggml_is_view(tensor)); // views generally get data pointer from one of their sources
-    WSP_GGML_ASSERT(tensor->data == NULL); // avoid allocating tensor which already has memory allocated
-#endif
-    size_t size = wsp_ggml_allocr_get_alloc_size(alloc, tensor);
+static size_t wsp_ggml_dyn_tallocr_alloc(struct wsp_ggml_dyn_tallocr * alloc, size_t size, const struct wsp_ggml_tensor * tensor) {
     size = aligned_offset(NULL, size, alloc->alignment);
 
     AT_PRINTF("%s: allocating %s (%zu bytes) - ", __func__, tensor->name, size);
@@ -159,8 +164,6 @@ void wsp_ggml_allocr_alloc(struct wsp_ggml_allocr * alloc, struct wsp_ggml_tenso
         }
     }
 
-    AT_PRINTF("block %d\n", best_fit_block);
-
     if (best_fit_block == -1) {
         // the last block is our last resort
         struct free_block * block = &alloc->free_blocks[alloc->n_free_blocks - 1];
@@ -168,15 +171,16 @@ void wsp_ggml_allocr_alloc(struct wsp_ggml_allocr * alloc, struct wsp_ggml_tenso
         if (block->size >= size) {
             best_fit_block = alloc->n_free_blocks - 1;
         } else {
-            fprintf(stderr, "%s: not enough space in the buffer (needed %zu, largest block available %zu)\n",
+            // this should never happen
+            WSP_GGML_LOG_ERROR("%s: not enough space in the buffer to allocate %zu bytes, largest block available %zu bytes\n",
                     __func__, size, max_avail);
-            WSP_GGML_ASSERT(!"not enough space in the buffer");
-            return;
+            WSP_GGML_ABORT("not enough space in the buffer");
         }
     }
+
     struct free_block * block = &alloc->free_blocks[best_fit_block];
-    void * addr = block->addr;
-    block->addr = (char*)block->addr + size;
+    size_t offset = block->offset;
+    block->offset = offset + size;
     block->size -= size;
     if (block->size == 0) {
         // remove block if empty
@@ -186,52 +190,63 @@ void wsp_ggml_allocr_alloc(struct wsp_ggml_allocr * alloc, struct wsp_ggml_tenso
         }
     }
 
-    tensor->data = addr;
+    AT_PRINTF("block %d, offset %zu\n", best_fit_block, offset);
 
 #ifdef WSP_GGML_ALLOCATOR_DEBUG
-    add_allocated_tensor(alloc, tensor);
-    size_t cur_max = (char*)addr - (char*)alloc->data + size;
+    add_allocated_tensor(alloc, offset, tensor);
+    size_t cur_max = offset + size;
     if (cur_max > alloc->max_size) {
-        printf("max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
+        // sort allocated_tensors by offset
+        for (int i = 0; i < 1024; i++) {
+            for (int j = i + 1; j < 1024; j++) {
+                if (alloc->allocated_tensors[i].offset > alloc->allocated_tensors[j].offset) {
+                    const struct wsp_ggml_tensor * tmp_tensor = alloc->allocated_tensors[i].tensor;
+                    size_t tmp_offset = alloc->allocated_tensors[i].offset;
+                    alloc->allocated_tensors[i].tensor = alloc->allocated_tensors[j].tensor;
+                    alloc->allocated_tensors[i].offset = alloc->allocated_tensors[j].offset;
+                    alloc->allocated_tensors[j].tensor = tmp_tensor;
+                    alloc->allocated_tensors[j].offset = tmp_offset;
+                }
+            }
+        }
+        WSP_GGML_LOG_DEBUG("max_size = %.2f MB: tensors: ", cur_max / 1024.0 / 1024.0);
         for (int i = 0; i < 1024; i++) {
-            if (alloc->allocated_tensors[i]) {
-                printf("%s (%.2f MB) ", alloc->allocated_tensors[i]->name, wsp_ggml_nbytes(alloc->allocated_tensors[i]) / 1024.0 / 1024.0);
+            if (alloc->allocated_tensors[i].tensor) {
+                WSP_GGML_LOG_DEBUG("%s [%zx-%zx] (%.2f MB) ", alloc->allocated_tensors[i].tensor->name,
+                    alloc->allocated_tensors[i].offset,
+                    alloc->allocated_tensors[i].offset + wsp_ggml_nbytes(alloc->allocated_tensors[i].tensor),
+                    wsp_ggml_nbytes(alloc->allocated_tensors[i].tensor) / 1024.0 / 1024.0);
             }
         }
-        printf("\n");
+        WSP_GGML_LOG_DEBUG("\n");
     }
 #endif
 
-    alloc->max_size = MAX(alloc->max_size, (char*)addr - (char*)alloc->data + size);
-}
+    alloc->max_size = MAX(alloc->max_size, offset + size);
 
-// this is a very naive implementation, but for our case the number of free blocks should be very small
-static void wsp_ggml_allocr_free_tensor(struct wsp_ggml_allocr * alloc, struct wsp_ggml_tensor * tensor) {
-    void * ptr = tensor->data;
+    return offset;
 
-    if (wsp_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;
-    }
+    WSP_GGML_UNUSED(tensor);
+}
 
-    size_t size = wsp_ggml_allocr_get_alloc_size(alloc, tensor);
+// this is a very naive implementation, but for our case the number of free blocks should be very small
+static void wsp_ggml_dyn_tallocr_free_tensor(struct wsp_ggml_dyn_tallocr * alloc, size_t offset, size_t size, const struct wsp_ggml_tensor * 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);
+
+    AT_PRINTF("%s: freeing %s at %zu (%zu bytes) - n_free_blocks = %d\n", __func__, tensor->name, offset, size, alloc->n_free_blocks);
 
 #ifdef WSP_GGML_ALLOCATOR_DEBUG
-    remove_allocated_tensor(alloc, tensor);
+    remove_allocated_tensor(alloc, offset, tensor);
 #endif
 
     // see if we can merge with an existing block
     for (int i = 0; i < alloc->n_free_blocks; i++) {
         struct free_block * block = &alloc->free_blocks[i];
         // check if ptr is at the end of the block
-        if ((char*)block->addr + block->size == ptr) {
+        if (block->offset + block->size == offset) {
             block->size += size;
             // check if we can merge with the next block
-            if (i < alloc->n_free_blocks - 1 && (char*)block->addr + block->size == alloc->free_blocks[i+1].addr) {
+            if (i < alloc->n_free_blocks - 1 && block->offset + block->size == alloc->free_blocks[i+1].offset) {
                 block->size += alloc->free_blocks[i+1].size;
                 alloc->n_free_blocks--;
                 for (int j = i+1; j < alloc->n_free_blocks; j++) {
@@ -241,11 +256,11 @@ static void wsp_ggml_allocr_free_tensor(struct wsp_ggml_allocr * alloc, struct w
             return;
         }
         // check if ptr is at the beginning of the block
-        if ((char*)ptr + size == block->addr) {
-            block->addr = ptr;
+        if (offset + size == block->offset) {
+            block->offset = offset;
             block->size += size;
             // check if we can merge with the previous block
-            if (i > 0 && (char*)alloc->free_blocks[i-1].addr + alloc->free_blocks[i-1].size == block->addr) {
+            if (i > 0 && alloc->free_blocks[i-1].offset + alloc->free_blocks[i-1].size == block->offset) {
                 alloc->free_blocks[i-1].size += block->size;
                 alloc->n_free_blocks--;
                 for (int j = i; j < alloc->n_free_blocks; j++) {
@@ -259,7 +274,7 @@ static void wsp_ggml_allocr_free_tensor(struct wsp_ggml_allocr * alloc, struct w
     WSP_GGML_ASSERT(alloc->n_free_blocks < MAX_FREE_BLOCKS && "out of free blocks");
     // insert the new block in the correct position to keep the array sorted by address (to make merging blocks faster)
     int insert_pos = 0;
-    while (insert_pos < alloc->n_free_blocks && alloc->free_blocks[insert_pos].addr < ptr) {
+    while (insert_pos < alloc->n_free_blocks && alloc->free_blocks[insert_pos].offset < offset) {
         insert_pos++;
     }
     // shift all blocks from insert_pos onward to make room for the new block
@@ -267,367 +282,757 @@ static void wsp_ggml_allocr_free_tensor(struct wsp_ggml_allocr * alloc, struct w
         alloc->free_blocks[i] = alloc->free_blocks[i-1];
     }
     // insert the new block
-    alloc->free_blocks[insert_pos].addr = ptr;
+    alloc->free_blocks[insert_pos].offset = offset;
     alloc->free_blocks[insert_pos].size = size;
     alloc->n_free_blocks++;
-}
 
-void wsp_ggml_allocr_set_parse_seq(struct wsp_ggml_allocr * alloc, const int * list, int n) {
-    for (int i = 0; i < n; i++) {
-        alloc->parse_seq[i] = list[i];
-    }
-    alloc->parse_seq_len = n;
+    WSP_GGML_UNUSED(tensor);
 }
 
-void wsp_ggml_allocr_reset(struct wsp_ggml_allocr * alloc) {
+static void wsp_ggml_dyn_tallocr_reset(struct wsp_ggml_dyn_tallocr * alloc) {
     alloc->n_free_blocks = 1;
-    size_t align_offset = aligned_offset(alloc->data, 0, alloc->alignment);
-    alloc->free_blocks[0].addr = (char *)alloc->data + align_offset;
-    alloc->free_blocks[0].size = alloc->size - align_offset;
+    alloc->free_blocks[0].offset = 0;
+    alloc->free_blocks[0].size = SIZE_MAX/2; // restrict maximum size of a measure allocator to half size_t max to avoid overflows
+    alloc->max_size = 0;
+
+#ifdef WSP_GGML_ALLOCATOR_DEBUG
+    for (int i = 0; i < 1024; i++) {
+        alloc->allocated_tensors[i].tensor = NULL;
+    }
+#endif
 }
 
-struct wsp_ggml_allocr * wsp_ggml_allocr_new(void * data, size_t size, size_t alignment) {
-    struct wsp_ggml_allocr * alloc = (struct wsp_ggml_allocr *)malloc(sizeof(struct wsp_ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
+static struct wsp_ggml_dyn_tallocr * wsp_ggml_dyn_tallocr_new(size_t alignment) {
+    struct wsp_ggml_dyn_tallocr * alloc = (struct wsp_ggml_dyn_tallocr *)malloc(sizeof(struct wsp_ggml_dyn_tallocr));
 
-    *alloc = (struct wsp_ggml_allocr){
-        /*.data          = */ data,
-        /*.size          = */ size,
+    *alloc = (struct wsp_ggml_dyn_tallocr) {
         /*.alignment     = */ alignment,
         /*.n_free_blocks = */ 0,
         /*.free_blocks   = */ {{0}},
-        /*.hash_table    = */ {{0}},
         /*.max_size      = */ 0,
-        /*.measure       = */ false,
-        /*.parse_seq     = */ {0},
-        /*.parse_seq_len = */ 0,
 #ifdef WSP_GGML_ALLOCATOR_DEBUG
-        /*.allocated_tensors = */ {0},
+        /*.allocated_tensors = */ {{0}},
 #endif
     };
 
-    wsp_ggml_allocr_reset(alloc);
+    wsp_ggml_dyn_tallocr_reset(alloc);
 
     return alloc;
 }
 
-// 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 wsp_ggml_dyn_tallocr_free(struct wsp_ggml_dyn_tallocr * alloc) {
+    free(alloc);
 }
 
-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
+static size_t wsp_ggml_dyn_tallocr_max_size(struct wsp_ggml_dyn_tallocr * alloc) {
+    return alloc->max_size;
 }
 
-// allocate uncommitted virtual memory to measure the size of the graph
-static void alloc_measure_vmem(void ** base_addr, size_t * size) {
-    // 128GB for 64-bit, 1GB for 32-bit
-    *size = sizeof(void *) == 4 ? 1ULL<<30 : 1ULL<<37;
-    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);
 
-    WSP_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);
-}
+// graph allocator
 
-struct wsp_ggml_allocr * wsp_ggml_allocr_new_measure(size_t alignment) {
-    struct wsp_ggml_allocr * alloc = (struct wsp_ggml_allocr *)malloc(sizeof(struct wsp_ggml_allocr) /* + n_free_blocks * sizeof(struct free_block) */);
+struct hash_node {
+    int n_children;
+    int n_views;
+    int buffer_id;
+    size_t offset; // offset within the buffer
+    bool allocated;
+};
 
-    void * base_addr;
-    size_t size;
+struct tensor_alloc {
+    int buffer_id;
+    size_t offset;
+    size_t size_max; // 0 = pre-allocated, unused, or view
+};
 
-    alloc_measure_vmem(&base_addr, &size);
+struct leaf_alloc {
+    struct tensor_alloc leaf;
+};
 
-    *alloc = (struct wsp_ggml_allocr){
-        /*.data          = */ base_addr,
-        /*.size          = */ size,
-        /*.alignment     = */ alignment,
-        /*.n_free_blocks = */ 0,
-        /*.free_blocks   = */ {{0}},
-        /*.hash_table    = */ {{0}},
-        /*.max_size      = */ 0,
-        /*.measure       = */ true,
-        /*.parse_seq     = */ {0},
-        /*.parse_seq_len = */ 0,
-#ifdef WSP_GGML_ALLOCATOR_DEBUG
-        /*.allocated_tensors = */ {0},
-#endif
-    };
+struct node_alloc {
+    struct tensor_alloc dst;
+    struct tensor_alloc src[WSP_GGML_MAX_SRC];
+};
 
-    wsp_ggml_allocr_reset(alloc);
+struct wsp_ggml_gallocr {
+    wsp_ggml_backend_buffer_type_t * bufts; // [n_buffers]
+    wsp_ggml_backend_buffer_t * buffers; // [n_buffers]
+    struct wsp_ggml_dyn_tallocr ** buf_tallocs; // [n_buffers]
+    int n_buffers;
 
-    return alloc;
-}
+    struct wsp_ggml_hash_set hash_set;
+    struct hash_node * hash_values; // [hash_set.size]
+
+    struct node_alloc * node_allocs; // [n_nodes]
+    int n_nodes;
+
+    struct leaf_alloc * leaf_allocs; // [n_leafs]
+    int n_leafs;
+};
+
+wsp_ggml_gallocr_t wsp_ggml_gallocr_new_n(wsp_ggml_backend_buffer_type_t * bufts, int n_bufs) {
+    wsp_ggml_gallocr_t galloc = (wsp_ggml_gallocr_t)calloc(1, sizeof(struct wsp_ggml_gallocr));
+    WSP_GGML_ASSERT(galloc != NULL);
+
+    galloc->bufts = calloc(n_bufs, sizeof(wsp_ggml_backend_buffer_type_t));
+    WSP_GGML_ASSERT(galloc->bufts != NULL);
+
+    galloc->buffers = calloc(n_bufs, sizeof(wsp_ggml_backend_buffer_t));
+    WSP_GGML_ASSERT(galloc->buffers != NULL);
+
+    galloc->buf_tallocs = calloc(n_bufs, sizeof(struct wsp_ggml_dyn_tallocr *));
+    WSP_GGML_ASSERT(galloc->buf_tallocs != NULL);
+
+    for (int i = 0; i < n_bufs; i++) {
+        galloc->bufts[i] = bufts[i];
+        galloc->buffers[i] = NULL;
 
-void wsp_ggml_allocr_free(struct wsp_ggml_allocr * alloc) {
-    if (alloc->measure) {
-        free_measure_vmem(alloc->data, alloc->size);
+        // check if the same buffer type is used multiple times and reuse the same allocator
+        for (int j = 0; j < i; j++) {
+            if (bufts[i] == bufts[j]) {
+                galloc->buf_tallocs[i] = galloc->buf_tallocs[j];
+                break;
+            }
+        }
+
+        if (galloc->buf_tallocs[i] == NULL) {
+            size_t alignment = wsp_ggml_backend_buft_get_alignment(bufts[i]);
+            galloc->buf_tallocs[i] = wsp_ggml_dyn_tallocr_new(alignment);
+        }
     }
-    free(alloc);
-}
+    galloc->n_buffers = n_bufs;
 
-bool wsp_ggml_allocr_is_measure(struct wsp_ggml_allocr * alloc) {
-    return alloc->measure;
+    return galloc;
 }
 
-//////////// compute graph allocator
+wsp_ggml_gallocr_t wsp_ggml_gallocr_new(wsp_ggml_backend_buffer_type_t buft) {
+    return wsp_ggml_gallocr_new_n(&buft, 1);
+}
 
-static bool wsp_ggml_are_same_layout(const struct wsp_ggml_tensor * a, const struct wsp_ggml_tensor * b) {
-    if (a->type != b->type) {
-        return false;
+void wsp_ggml_gallocr_free(wsp_ggml_gallocr_t galloc) {
+    if (galloc == NULL) {
+        return;
     }
-    for (int i = 0; i < WSP_GGML_MAX_DIMS; i++) {
-        if (a->ne[i] != b->ne[i]) {
-            return false;
+
+    for (int i = 0; i < galloc->n_buffers; i++) {
+        if (galloc->buffers != NULL) {
+            // skip if already freed
+            bool freed = false;
+            for (int j = 0; j < i; j++) {
+                if (galloc->buffers[j] == galloc->buffers[i]) {
+                    freed = true;
+                    break;
+                }
+            }
+            if (!freed) {
+                wsp_ggml_backend_buffer_free(galloc->buffers[i]);
+            }
         }
-        if (a->nb[i] != b->nb[i]) {
-            return false;
+        if (galloc->buf_tallocs != NULL) {
+            // skip if already freed
+            bool freed = false;
+            for (int j = 0; j < i; j++) {
+                if (galloc->buf_tallocs[j] == galloc->buf_tallocs[i]) {
+                    freed = true;
+                    break;
+                }
+            }
+            if (!freed) {
+                wsp_ggml_dyn_tallocr_free(galloc->buf_tallocs[i]);
+            }
         }
     }
-    return true;
+
+    wsp_ggml_hash_set_free(&galloc->hash_set);
+    free(galloc->hash_values);
+    free(galloc->bufts);
+    free(galloc->buffers);
+    free(galloc->buf_tallocs);
+    free(galloc->node_allocs);
+    free(galloc->leaf_allocs);
+    free(galloc);
 }
 
-static bool wsp_ggml_op_can_inplace(enum wsp_ggml_op op) {
-    switch (op) {
-        case WSP_GGML_OP_SCALE:
-        case WSP_GGML_OP_DIAG_MASK_ZERO:
-        case WSP_GGML_OP_DIAG_MASK_INF:
-        case WSP_GGML_OP_ADD:
-        case WSP_GGML_OP_ADD1:
-        case WSP_GGML_OP_SUB:
-        case WSP_GGML_OP_MUL:
-        case WSP_GGML_OP_DIV:
-        case WSP_GGML_OP_SQR:
-        case WSP_GGML_OP_SQRT:
-        case WSP_GGML_OP_LOG:
-        case WSP_GGML_OP_UNARY:
-        case WSP_GGML_OP_ROPE:
-        case WSP_GGML_OP_RMS_NORM:
-        case WSP_GGML_OP_SOFT_MAX:
-        case WSP_GGML_OP_CONT:
-            return true;
+typedef struct wsp_ggml_gallocr * wsp_ggml_gallocr_t;
 
-        default:
-            return false;
-    }
+static struct hash_node * wsp_ggml_gallocr_hash_get(wsp_ggml_gallocr_t galloc, struct wsp_ggml_tensor * t) {
+    size_t i = wsp_ggml_hash_find_or_insert(&galloc->hash_set, t);
+    return &galloc->hash_values[i];
 }
 
-static void allocate_node(struct wsp_ggml_allocr * alloc, struct wsp_ggml_tensor * node) {
-    struct hash_node * ht = alloc->hash_table;
-    if (node->data == NULL) {
-        if (wsp_ggml_is_view(node)) {
-            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 (wsp_ggml_op_can_inplace(node->op)) {
-                for (int i = 0; i < WSP_GGML_MAX_SRC; i++) {
-                    struct wsp_ggml_tensor * parent = node->src[i];
-                    if (parent == NULL) {
-                        break;
-                    }
+static bool wsp_ggml_gallocr_is_own(wsp_ggml_gallocr_t galloc, struct wsp_ggml_tensor * t) {
+    return wsp_ggml_gallocr_hash_get(galloc, t)->allocated;
+}
 
-                    // if the node's data is external, then we cannot re-use it
-                    if (wsp_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;
-                    }
+static bool wsp_ggml_gallocr_is_allocated(wsp_ggml_gallocr_t galloc, struct wsp_ggml_tensor * t) {
+    return t->data != NULL || wsp_ggml_gallocr_hash_get(galloc, t)->allocated;
+}
 
-                    struct hash_node * p_hn = hash_get(ht, parent);
-                    if (parent->data != NULL && p_hn->n_children == 1 && p_hn->n_views == 0 && wsp_ggml_are_same_layout(node, parent)) {
-                        if (wsp_ggml_is_view(parent)) {
-                            struct wsp_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
-                                // the parent's data that it will need later (same layout requirement). the problem is that then
-                                // we cannot free the tensor because the original address of the allocation is lost.
-                                // adding a view_src pointer to the tensor would solve this and simplify the code dealing with views
-                                // for now, we only reuse the parent's data if the offset is zero (view_src->data == parent->data)
-                                AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name);
-                                node->data = parent->data;
-                                return;
-                            }
-                        }
-                        else {
-                            AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
-                            node->data = parent->data;
+static void wsp_ggml_gallocr_allocate_node(wsp_ggml_gallocr_t galloc, struct wsp_ggml_tensor * node, int buffer_id) {
+    WSP_GGML_ASSERT(buffer_id >= 0);
+    struct hash_node * hn = wsp_ggml_gallocr_hash_get(galloc, node);
+
+    if (!wsp_ggml_gallocr_is_allocated(galloc, node) && !wsp_ggml_is_view(node)) {
+        hn->allocated = true;
+        assert(hn->offset == 0);
+
+        // try to reuse a parent's buffer (inplace)
+        if (wsp_ggml_op_can_inplace(node->op)) {
+            for (int i = 0; i < WSP_GGML_MAX_SRC; i++) {
+                struct wsp_ggml_tensor * parent = node->src[i];
+                if (parent == NULL) {
+                    continue;
+                }
+
+                // if the node's data is external, then we cannot re-use it
+                if (!wsp_ggml_gallocr_is_own(galloc, parent)) {
+                    AT_PRINTF("not reusing parent %s for %s as %p is external\n", parent->name, node->name, parent->data);
+                    continue;
+                }
+
+                // outputs cannot be reused
+                if (parent->flags & WSP_GGML_TENSOR_FLAG_OUTPUT || (parent->view_src != NULL && parent->view_src->flags & WSP_GGML_TENSOR_FLAG_OUTPUT)) {
+                    AT_PRINTF("not reusing parent %s for %s as it is an output\n", parent->name, node->name);
+                    continue;
+                }
+
+                if (!wsp_ggml_are_same_layout(node, parent)) {
+                    AT_PRINTF("not reusing parent %s for %s as layouts are different\n", parent->name, node->name);
+                    continue;
+                }
+
+                struct hash_node * p_hn = wsp_ggml_gallocr_hash_get(galloc, parent);
+                if (p_hn->n_children == 1 && p_hn->n_views == 0) {
+                    if (wsp_ggml_is_view(parent)) {
+                        struct wsp_ggml_tensor * view_src = parent->view_src;
+                        struct hash_node * view_src_hn = wsp_ggml_gallocr_hash_get(galloc, view_src);
+                        if (view_src_hn->n_views == 1 && view_src_hn->n_children == 0 && view_src->data == parent->data) {
+                            AT_PRINTF("reusing view parent %s (%s) for %s\n", parent->name, view_src->name, node->name);
+                            assert(view_src_hn->offset == p_hn->offset);
+                            hn->buffer_id = p_hn->buffer_id;
+                            hn->offset = p_hn->offset;
+                            p_hn->allocated = false; // avoid freeing the parent
+                            view_src_hn->allocated = false;
                             return;
                         }
+                    } else {
+                        AT_PRINTF("reusing parent %s for %s\n", parent->name, node->name);
+                        hn->buffer_id = p_hn->buffer_id;
+                        hn->offset = p_hn->offset;
+                        p_hn->allocated = false; // avoid freeing the parent
+                        return;
                     }
                 }
             }
-            wsp_ggml_allocr_alloc(alloc, node);
         }
+        // allocate tensor from the buffer
+        struct wsp_ggml_dyn_tallocr * alloc = galloc->buf_tallocs[buffer_id];
+        wsp_ggml_backend_buffer_type_t buft = galloc->bufts[buffer_id];
+        size_t size = wsp_ggml_backend_buft_get_alloc_size(buft, node);
+        size_t offset = wsp_ggml_dyn_tallocr_alloc(alloc, size, node);
+        hn->buffer_id = buffer_id;
+        hn->offset = offset;
+        return;
+    }
+}
+
+static void wsp_ggml_gallocr_free_node(wsp_ggml_gallocr_t galloc, struct wsp_ggml_tensor * node) {
+    // graph outputs are never freed
+    if (node->flags & WSP_GGML_TENSOR_FLAG_OUTPUT) {
+        AT_PRINTF("not freeing output %s\n", node->name);
+        return;
     }
+
+    struct hash_node * hn = wsp_ggml_gallocr_hash_get(galloc, node);
+    size_t offset = hn->offset;
+    int buffer_id = hn->buffer_id;
+    struct wsp_ggml_dyn_tallocr * alloc = galloc->buf_tallocs[buffer_id];
+    wsp_ggml_backend_buffer_type_t buft = galloc->bufts[buffer_id];
+    size_t size = wsp_ggml_backend_buft_get_alloc_size(buft, node);
+    wsp_ggml_dyn_tallocr_free_tensor(alloc, offset, size, node);
+    hn->allocated = false;
+}
+
+static int get_node_buffer_id(const int * node_buffer_ids, int i) {
+    return node_buffer_ids ? node_buffer_ids[i] : 0;
 }
 
-static size_t wsp_ggml_allocr_alloc_graph_tensors_n(
-    struct wsp_ggml_allocr * alloc,
-    struct wsp_ggml_cgraph ** graphs, int n_graphs,
-    struct wsp_ggml_tensor *** inputs, struct wsp_ggml_tensor *** outputs) {
+static void wsp_ggml_gallocr_alloc_graph_impl(wsp_ggml_gallocr_t galloc, struct wsp_ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
+    // clear hash tables
+    wsp_ggml_hash_set_reset(&galloc->hash_set);
+    memset(galloc->hash_values, 0, sizeof(struct hash_node) * galloc->hash_set.size);
 
-    // reset hash table
-    struct hash_node * ht = alloc->hash_table;
-    memset(ht, 0, sizeof(struct hash_node) * WSP_GGML_GRAPH_HASHTABLE_SIZE);
+    // 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 wsp_ggml_tensor * leaf = graph->leafs[i];
+        wsp_ggml_gallocr_allocate_node(galloc, leaf, get_node_buffer_id(leaf_buffer_ids, i));
+    }
 
     // count number of children and views
-    for (int g = 0; g < n_graphs; g++) {
-        struct wsp_ggml_cgraph * gf = graphs[g];
-        for (int i = 0; i < gf->n_nodes; i++) {
-            struct wsp_ggml_tensor * node = gf->nodes[i];
-
-            if (wsp_ggml_is_view(node)) {
-                struct wsp_ggml_tensor * view_src = node->view_src;
-                hash_get(ht, view_src)->n_views += 1;
+    // allocate other graph inputs and leafs first to avoid overwriting them
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct wsp_ggml_tensor * node = graph->nodes[i];
+
+        // TODO: better way to add external dependencies
+        // WSP_GGML_OP_NONE does not appear normally in the graph nodes, but is used by ggml-backend to add dependencies to
+        // control when some tensors are allocated and freed. in this case, the dependencies are in `src`, but the node
+        // itself is never used and should not be considered a dependency
+        if (wsp_ggml_is_view(node) && node->op != WSP_GGML_OP_NONE) {
+            struct wsp_ggml_tensor * view_src = node->view_src;
+            wsp_ggml_gallocr_hash_get(galloc, view_src)->n_views += 1;
+        }
+
+        if (node->flags & WSP_GGML_TENSOR_FLAG_INPUT) {
+            wsp_ggml_gallocr_allocate_node(galloc, graph->nodes[i], get_node_buffer_id(node_buffer_ids, i));
+        }
+
+        for (int j = 0; j < WSP_GGML_MAX_SRC; j++) {
+            struct wsp_ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                continue;
             }
 
-            for (int j = 0; j < WSP_GGML_MAX_SRC; j++) {
-                struct wsp_ggml_tensor * parent = node->src[j];
-                if (parent == NULL) {
-                    break;
-                }
-                hash_get(ht, parent)->n_children += 1;
+            wsp_ggml_gallocr_hash_get(galloc, src)->n_children += 1;
+
+            // allocate explicit inputs
+            if (src->flags & WSP_GGML_TENSOR_FLAG_INPUT) {
+                wsp_ggml_gallocr_allocate_node(galloc, src, get_node_buffer_id(node_buffer_ids, i));
             }
         }
     }
 
     // allocate tensors
-    for (int g = 0; g < n_graphs; g++) {
-        struct wsp_ggml_cgraph * gf = graphs[g];
-        AT_PRINTF("####### graph %d/%d\n", g, n_graphs);
-        // graph inputs are allocated first to ensure that they are not overwritten by each other
-        if (inputs != NULL && inputs[g] != NULL) {
-            for (int i = 0; inputs[g][i] != NULL; i++) {
-                struct wsp_ggml_tensor * input = inputs[g][i];
-                AT_PRINTF("input: %s\n", input->name);
-                allocate_node(alloc, input);
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct wsp_ggml_tensor * node = graph->nodes[i];
+        int buffer_id = get_node_buffer_id(node_buffer_ids, i);
+
+        // allocate parents (only leafs need to be allocated at this point)
+        for (int j = 0; j < WSP_GGML_MAX_SRC; j++) {
+            struct wsp_ggml_tensor * parent = node->src[j];
+            if (parent == NULL) {
+                continue;
             }
+            wsp_ggml_gallocr_allocate_node(galloc, parent, buffer_id);
         }
-        // if we have parse_seq then we allocate nodes following the list, and we only free nodes at barriers
-        int last_barrier_pos = 0;
-        int n_nodes = alloc->parse_seq_len ? alloc->parse_seq_len : gf->n_nodes;
-
-        for (int ind = 0; ind < n_nodes; ind++) {
-            // allocate a node if there is no parse_seq or this is not a barrier
-            if ((alloc->parse_seq_len==0) || alloc->parse_seq[ind] != -1) {
-                int i = alloc->parse_seq_len ? alloc->parse_seq[ind] : ind;
-                struct wsp_ggml_tensor * node = gf->nodes[i];
-
-                // allocate parents (leafs)
-                for (int j = 0; j < WSP_GGML_MAX_SRC; j++) {
-                    struct wsp_ggml_tensor * parent = node->src[j];
-                    if (parent == NULL) {
-                        break;
-                    }
-                    allocate_node(alloc, parent);
-                }
 
-                // allocate node
-                allocate_node(alloc, node);
+        // allocate node
+        wsp_ggml_gallocr_allocate_node(galloc, node, buffer_id);
 
-                AT_PRINTF("exec: %s (%s) <= ", wsp_ggml_op_name(node->op), node->name);
-                for (int j = 0; j < WSP_GGML_MAX_SRC; j++) {
-                    struct wsp_ggml_tensor * parent = node->src[j];
-                    if (parent == NULL) {
-                        break;
-                    }
-                    AT_PRINTF("%s", parent->name);
-                    if (j < WSP_GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
-                        AT_PRINTF(", ");
-                    }
-                }
-                AT_PRINTF("\n");
+        AT_PRINTF("exec: %s (%s) <= ", wsp_ggml_op_desc(node), node->name);
+        for (int j = 0; j < WSP_GGML_MAX_SRC; j++) {
+            struct wsp_ggml_tensor * parent = node->src[j];
+            if (parent == NULL) {
+                continue;
             }
+            AT_PRINTF("%s", parent->name);
+            if (j < WSP_GGML_MAX_SRC - 1 && node->src[j + 1] != NULL) {
+                AT_PRINTF(", ");
+            }
+        }
+        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) {
-                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++) {
-                    int node_i = alloc->parse_seq_len ? alloc->parse_seq[i] : i;
-                    struct wsp_ggml_tensor * node = gf->nodes[node_i];
-
-                    for (int j = 0; j < WSP_GGML_MAX_SRC; j++) {
-                        struct wsp_ggml_tensor * parent = node->src[j];
-                        if (parent == NULL) {
-                            break;
-                        }
-                        struct hash_node * p_hn = hash_get(ht, parent);
-                        p_hn->n_children -= 1;
-
-                        //AT_PRINTF("parent %s: %d children, %d views\n", parent->name, parent->n_children, parent->n_views);
-
-                        if (p_hn->n_children == 0 && p_hn->n_views == 0) {
-                            if (wsp_ggml_is_view(parent)) {
-                                struct wsp_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: %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) {
-                                    wsp_ggml_allocr_free_tensor(alloc, view_src);
-                                }
-                            }
-                            else {
-                                if (parent->data != node->data) {
-                                    wsp_ggml_allocr_free_tensor(alloc, parent);
-                                }
-                            }
-                        }
+        // update parents
+        for (int j = 0; j < WSP_GGML_MAX_SRC; j++) {
+            struct wsp_ggml_tensor * parent = node->src[j];
+            if (parent == NULL) {
+                continue;
+            }
+            struct hash_node * p_hn = wsp_ggml_gallocr_hash_get(galloc, parent);
+            p_hn->n_children -= 1;
+
+            AT_PRINTF("parent %s: %d children, %d views, allocated: %d\n",
+                parent->name, p_hn->n_children, p_hn->n_views, p_hn->allocated);
+
+            if (p_hn->n_children == 0 && p_hn->n_views == 0) {
+                if (wsp_ggml_is_view(parent)) {
+                    struct wsp_ggml_tensor * view_src = parent->view_src;
+                    struct hash_node * view_src_hn = wsp_ggml_gallocr_hash_get(galloc, view_src);
+                    view_src_hn->n_views -= 1;
+                    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_hn->allocated) {
+                        wsp_ggml_gallocr_free_node(galloc, view_src);
                     }
                 }
-                AT_PRINTF("\n");
-                if (alloc->parse_seq_len) {
-                    last_barrier_pos = ind + 1;
+                else if (p_hn->allocated) {
+                    wsp_ggml_gallocr_free_node(galloc, parent);
                 }
             }
+            AT_PRINTF("\n");
         }
-        // free graph outputs here that wouldn't be freed otherwise because they have no children
-        if (outputs != NULL && outputs[g] != NULL) {
-            for (int i = 0; outputs[g][i] != NULL; i++) {
-                struct wsp_ggml_tensor * output = outputs[g][i];
-                AT_PRINTF("output: %s\n", output->name);
-                wsp_ggml_allocr_free_tensor(alloc, output);
+    }
+}
+
+bool wsp_ggml_gallocr_reserve_n(wsp_ggml_gallocr_t galloc, struct wsp_ggml_cgraph * graph, const int * node_buffer_ids, const int * leaf_buffer_ids) {
+    size_t min_hash_size = graph->n_nodes + graph->n_leafs;
+    // add 25% margin to avoid hash collisions
+    min_hash_size += min_hash_size / 4;
+
+    // initialize hash table
+    if (galloc->hash_set.size < min_hash_size) {
+        wsp_ggml_hash_set_free(&galloc->hash_set);
+        galloc->hash_set = wsp_ggml_hash_set_new(min_hash_size);
+        WSP_GGML_ASSERT(galloc->hash_set.keys != NULL);
+
+        free(galloc->hash_values);
+        galloc->hash_values = malloc(sizeof(struct hash_node) * galloc->hash_set.size);
+        WSP_GGML_ASSERT(galloc->hash_values != NULL);
+    }
+
+    // reset allocators
+    for (int i = 0; i < galloc->n_buffers; i++) {
+        wsp_ggml_dyn_tallocr_reset(galloc->buf_tallocs[i]);
+    }
+
+    // allocate in hash table
+    wsp_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) {
+        free(galloc->node_allocs);
+        galloc->node_allocs = calloc(graph->n_nodes, sizeof(struct node_alloc));
+        WSP_GGML_ASSERT(galloc->node_allocs != NULL);
+    }
+    galloc->n_nodes = graph->n_nodes;
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct wsp_ggml_tensor * node = graph->nodes[i];
+        struct node_alloc * node_alloc = &galloc->node_allocs[i];
+        if (node->view_src || node->data) {
+            node_alloc->dst.buffer_id = -1;
+            node_alloc->dst.offset = SIZE_MAX;
+            node_alloc->dst.size_max = 0;
+        } else {
+            struct hash_node * hn = wsp_ggml_gallocr_hash_get(galloc, node);
+            node_alloc->dst.buffer_id = hn->buffer_id;
+            node_alloc->dst.offset    = hn->offset;
+            node_alloc->dst.size_max  = wsp_ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], node);
+        }
+        for (int j = 0; j < WSP_GGML_MAX_SRC; j++) {
+            struct wsp_ggml_tensor * src = node->src[j];
+            if (!src || src->view_src || src->data) {
+                node_alloc->src[j].buffer_id = -1;
+                node_alloc->src[j].offset = SIZE_MAX;
+                node_alloc->src[j].size_max = 0;
+            } else {
+                struct hash_node * hn = wsp_ggml_gallocr_hash_get(galloc, src);
+                node_alloc->src[j].buffer_id = hn->buffer_id;
+                node_alloc->src[j].offset   = hn->offset;
+                node_alloc->src[j].size_max = wsp_ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], src);
             }
         }
     }
+    if (galloc->n_leafs < graph->n_leafs) {
+        free(galloc->leaf_allocs);
+        galloc->leaf_allocs = calloc(graph->n_leafs, sizeof(galloc->leaf_allocs[0]));
+        WSP_GGML_ASSERT(galloc->leaf_allocs != NULL);
+    }
+    galloc->n_leafs = graph->n_leafs;
+    for (int i = 0; i < graph->n_leafs; i++) {
+        struct wsp_ggml_tensor * leaf = graph->leafs[i];
+        struct hash_node * hn = wsp_ggml_gallocr_hash_get(galloc, leaf);
+        if (leaf->view_src || leaf->data) {
+            galloc->leaf_allocs[i].leaf.buffer_id = -1;
+            galloc->leaf_allocs[i].leaf.offset = SIZE_MAX;
+            galloc->leaf_allocs[i].leaf.size_max = 0;
+        } else {
+            galloc->leaf_allocs[i].leaf.buffer_id = hn->buffer_id;
+            galloc->leaf_allocs[i].leaf.offset = hn->offset;
+            galloc->leaf_allocs[i].leaf.size_max = wsp_ggml_backend_buft_get_alloc_size(galloc->bufts[hn->buffer_id], leaf);
+        }
+    }
 
-    return alloc->max_size;
+    // reallocate buffers if needed
+    for (int i = 0; i < galloc->n_buffers; i++) {
+        // if the buffer type is used multiple times, we reuse the same buffer
+        for (int j = 0; j < i; j++) {
+            if (galloc->buf_tallocs[j] == galloc->buf_tallocs[i]) {
+                galloc->buffers[i] = galloc->buffers[j];
+                break;
+            }
+        }
+
+        size_t cur_size = galloc->buffers[i] ? wsp_ggml_backend_buffer_get_size(galloc->buffers[i]) : 0;
+        size_t new_size = wsp_ggml_dyn_tallocr_max_size(galloc->buf_tallocs[i]);
+
+        // 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
+            WSP_GGML_LOG_DEBUG("%s: reallocating %s buffer from size %.02f MiB to %.02f MiB\n", __func__, wsp_ggml_backend_buft_name(galloc->bufts[i]), cur_size / 1024.0 / 1024.0, new_size / 1024.0 / 1024.0);
+#endif
+
+            wsp_ggml_backend_buffer_free(galloc->buffers[i]);
+            galloc->buffers[i] = wsp_ggml_backend_buft_alloc_buffer(galloc->bufts[i], new_size);
+            if (galloc->buffers[i] == NULL) {
+                WSP_GGML_LOG_ERROR("%s: failed to allocate %s buffer of size %zu\n", __func__, wsp_ggml_backend_buft_name(galloc->bufts[i]), new_size);
+                return false;
+            }
+            wsp_ggml_backend_buffer_set_usage(galloc->buffers[i], WSP_GGML_BACKEND_BUFFER_USAGE_COMPUTE);
+        }
+    }
+
+    return true;
+}
+
+bool wsp_ggml_gallocr_reserve(wsp_ggml_gallocr_t galloc, struct wsp_ggml_cgraph *graph) {
+    return wsp_ggml_gallocr_reserve_n(galloc, graph, NULL, NULL);
+}
+
+static void wsp_ggml_gallocr_init_tensor(wsp_ggml_gallocr_t galloc, struct wsp_ggml_tensor * tensor, struct tensor_alloc * tensor_alloc) {
+    int buffer_id = tensor_alloc->buffer_id;
+    assert(tensor->data || tensor->view_src || wsp_ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], tensor) <= tensor_alloc->size_max);
+
+    if (tensor->view_src != NULL) {
+        if (tensor->buffer == NULL) {
+            assert(tensor_alloc->offset == SIZE_MAX);
+            if (tensor->view_src->buffer == NULL) {
+                // this tensor was allocated without ggml-backend
+                return;
+            }
+            wsp_ggml_backend_view_init(tensor);
+        }
+    } else {
+        if (tensor->data == NULL) {
+            assert(tensor_alloc->offset != SIZE_MAX);
+            assert(wsp_ggml_backend_buffer_get_alloc_size(galloc->buffers[buffer_id], tensor) <= tensor_alloc->size_max);
+            void * base = wsp_ggml_backend_buffer_get_base(galloc->buffers[buffer_id]);
+            void * addr = (char *)base + tensor_alloc->offset;
+            wsp_ggml_backend_tensor_alloc(galloc->buffers[buffer_id], tensor, addr);
+        } else {
+            if (tensor->buffer == NULL) {
+                // this tensor was allocated without ggml-backend
+                return;
+            }
+        }
+    }
+}
+
+static bool wsp_ggml_gallocr_node_needs_realloc(wsp_ggml_gallocr_t galloc, struct wsp_ggml_tensor * node, struct tensor_alloc * talloc) {
+    size_t node_size = 0;
+    if (!node->data && !node->view_src) {
+        WSP_GGML_ASSERT(talloc->buffer_id >= 0); // prevent segfault when misusing the API
+        node_size = wsp_ggml_backend_buft_get_alloc_size(galloc->bufts[talloc->buffer_id], node);
+    }
+    return talloc->size_max >= node_size;
+}
+
+static bool wsp_ggml_gallocr_needs_realloc(wsp_ggml_gallocr_t galloc, struct wsp_ggml_cgraph * graph) {
+    if (galloc->n_nodes != graph->n_nodes) {
+#ifndef NDEBUG
+        WSP_GGML_LOG_DEBUG("%s: graph has different number of nodes\n", __func__);
+#endif
+        return true;
+    }
+
+    if (galloc->n_leafs != graph->n_leafs) {
+#ifndef NDEBUG
+        WSP_GGML_LOG_DEBUG("%s: graph has different number of leafs\n", __func__);
+#endif
+        return true;
+    }
+
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct wsp_ggml_tensor * node = graph->nodes[i];
+        struct node_alloc * node_alloc = &galloc->node_allocs[i];
+
+        if (!wsp_ggml_gallocr_node_needs_realloc(galloc, node, &node_alloc->dst)) {
+#ifndef NDEBUG
+            WSP_GGML_LOG_DEBUG("%s: node %s is not valid\n", __func__, node->name);
+#endif
+            return true;
+        }
+
+        for (int j = 0; j < WSP_GGML_MAX_SRC; j++) {
+            struct wsp_ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                continue;
+            }
+            if (!wsp_ggml_gallocr_node_needs_realloc(galloc, src, &node_alloc->src[j])) {
+#ifndef NDEBUG
+                WSP_GGML_LOG_DEBUG("%s: src %d (%s) of node %s is not valid\n", __func__, j, src->name, node->name);
+#endif
+                return true;
+            }
+        }
+    }
+
+    return false;
+}
+
+bool wsp_ggml_gallocr_alloc_graph(wsp_ggml_gallocr_t galloc, struct wsp_ggml_cgraph * graph) {
+    if (wsp_ggml_gallocr_needs_realloc(galloc, graph)) {
+        if (galloc->n_buffers == 1) {
+#ifndef NDEBUG
+            WSP_GGML_LOG_DEBUG("%s: reallocating buffers automatically\n", __func__);
+#endif
+            if (!wsp_ggml_gallocr_reserve(galloc, graph)) {
+                return false;
+            }
+        } else {
+#ifndef NDEBUG
+            WSP_GGML_LOG_DEBUG("%s: cannot reallocate multi buffer graph automatically, call reserve\n", __func__);
+#endif
+            return false;
+        }
+    }
+
+    // reset buffers
+    for (int i = 0; i < galloc->n_buffers; i++) {
+        if (galloc->buffers[i] != NULL) {
+            wsp_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 wsp_ggml_tensor * leaf = graph->leafs[i];
+        struct leaf_alloc * leaf_alloc = &galloc->leaf_allocs[i];
+        wsp_ggml_gallocr_init_tensor(galloc, leaf, &leaf_alloc->leaf);
+    }
+    // nodes
+    for (int i = 0; i < graph->n_nodes; i++) {
+        struct wsp_ggml_tensor * node = graph->nodes[i];
+        struct node_alloc * node_alloc = &galloc->node_allocs[i];
+        for (int j = 0; j < WSP_GGML_MAX_SRC; j++) {
+            struct wsp_ggml_tensor * src = node->src[j];
+            if (src == NULL) {
+                continue;
+            }
+            wsp_ggml_gallocr_init_tensor(galloc, src, &node_alloc->src[j]);
+        }
+        wsp_ggml_gallocr_init_tensor(galloc, node, &node_alloc->dst);
+    }
+
+    return true;
+}
+
+size_t wsp_ggml_gallocr_get_buffer_size(wsp_ggml_gallocr_t galloc, int buffer_id) {
+    WSP_GGML_ASSERT(buffer_id >= 0 && buffer_id < galloc->n_buffers);
+
+    if (galloc->buffers[buffer_id] == NULL) {
+        return 0;
+    }
+
+    for (int i = 0; i < buffer_id; i++) {
+        if (galloc->buffers[i] == galloc->buffers[buffer_id]) {
+            // this buffer is the same as a previous one due to the same buffer type being used multiple times
+            // only return the buffer size the first time it appears to avoid double counting
+            return 0;
+        }
+    }
+
+    return wsp_ggml_backend_buffer_get_size(galloc->buffers[buffer_id]);
+}
+
+// utils
+
+static bool alloc_tensor_range(struct wsp_ggml_context * ctx,
+        struct wsp_ggml_tensor * first, struct wsp_ggml_tensor * last,
+        wsp_ggml_backend_buffer_type_t buft, size_t size,
+        wsp_ggml_backend_buffer_t ** buffers, size_t * n_buffers) {
+    wsp_ggml_backend_buffer_t buffer = wsp_ggml_backend_buft_alloc_buffer(buft, size);
+    if (buffer == NULL) {
+#ifndef NDEBUG
+        WSP_GGML_LOG_DEBUG("%s: failed to allocate %s buffer of size %zu\n", __func__, wsp_ggml_backend_buft_name(buft), size);
+#endif
+        for (size_t i = 0; i < *n_buffers; i++) {
+            wsp_ggml_backend_buffer_free((*buffers)[i]);
+        }
+        free(*buffers);
+        return false;
+    }
+
+    struct wsp_ggml_tallocr tallocr = wsp_ggml_tallocr_new(buffer);
+
+    for (struct wsp_ggml_tensor * t = first; t != last; t = wsp_ggml_get_next_tensor(ctx, t)) {
+        if (t->data == NULL) {
+            if (t->view_src == NULL) {
+                wsp_ggml_tallocr_alloc(&tallocr, t);
+            } else if (t->buffer == NULL) {
+                wsp_ggml_backend_view_init(t);
+            }
+        } else {
+            if (t->view_src != NULL && t->buffer == NULL) {
+                // view of a pre-allocated tensor
+                wsp_ggml_backend_view_init(t);
+            }
+        }
+    }
+
+    *buffers = realloc(*buffers, sizeof(wsp_ggml_backend_buffer_t) * (*n_buffers + 1));
+    (*buffers)[(*n_buffers)++] = buffer;
+
+    return true;
+}
+
+wsp_ggml_backend_buffer_t wsp_ggml_backend_alloc_ctx_tensors_from_buft(struct wsp_ggml_context * ctx, wsp_ggml_backend_buffer_type_t buft) {
+    WSP_GGML_ASSERT(wsp_ggml_get_no_alloc(ctx) == true);
+
+    size_t alignment = wsp_ggml_backend_buft_get_alignment(buft);
+    size_t max_size = wsp_ggml_backend_buft_get_max_size(buft);
+
+    wsp_ggml_backend_buffer_t * buffers = NULL;
+    size_t n_buffers = 0;
+
+    size_t cur_buf_size = 0;
+    struct wsp_ggml_tensor * first = wsp_ggml_get_first_tensor(ctx);
+    for (struct wsp_ggml_tensor * t = first; t != NULL; t = wsp_ggml_get_next_tensor(ctx, t)) {
+        size_t this_size = 0;
+        if (t->data == NULL && t->view_src == NULL) {
+            this_size = WSP_GGML_PAD(wsp_ggml_backend_buft_get_alloc_size(buft, t), alignment);
+        }
+
+        if (this_size > max_size) {
+            WSP_GGML_LOG_ERROR("%s: tensor %s is too large to fit in a %s buffer (tensor size: %zu, max buffer size: %zu)\n",
+                    __func__, t->name,
+                    wsp_ggml_backend_buft_name(buft),
+                    this_size, max_size);
+            for (size_t i = 0; i < n_buffers; i++) {
+                wsp_ggml_backend_buffer_free(buffers[i]);
+            }
+            free(buffers);
+            return NULL;
+        }
+
+        if ((cur_buf_size + this_size) > max_size) {
+            // allocate tensors in the current buffer
+            if (!alloc_tensor_range(ctx, first, t, buft, cur_buf_size, &buffers, &n_buffers)) {
+                return NULL;
+            }
+            first = t;
+            cur_buf_size = this_size;
+        } else {
+            cur_buf_size += this_size;
+        }
+    }
+
+    // allocate remaining tensors
+    if (cur_buf_size > 0) {
+        if (!alloc_tensor_range(ctx, first, NULL, buft, cur_buf_size, &buffers, &n_buffers)) {
+            return NULL;
+        }
+    }
+
+    if (n_buffers == 0) {
+#ifndef NDEBUG
+        WSP_GGML_LOG_DEBUG("%s: all tensors in the context are already allocated\n", __func__);
+#endif
+        return NULL;
+    }
+
+    wsp_ggml_backend_buffer_t buffer;
+    if (n_buffers == 1) {
+        buffer = buffers[0];
+    } else {
+        buffer = wsp_ggml_backend_multi_buffer_alloc_buffer(buffers, n_buffers);
+    }
+    free(buffers);
+    return buffer;
 }
 
-size_t wsp_ggml_allocr_alloc_graph(struct wsp_ggml_allocr * alloc, struct wsp_ggml_cgraph * graph) {
-    return wsp_ggml_allocr_alloc_graph_tensors_n(alloc, &graph, 1, NULL, NULL);
+wsp_ggml_backend_buffer_t wsp_ggml_backend_alloc_ctx_tensors(struct wsp_ggml_context * ctx, wsp_ggml_backend_t backend) {
+    return wsp_ggml_backend_alloc_ctx_tensors_from_buft(ctx, wsp_ggml_backend_get_default_buffer_type(backend));
 }