Fix integer overflows in attention & cache ops (#1514)

This commit is contained in:
Woosuk Kwon
2023-10-31 15:19:30 -07:00
committed by GitHub
parent 9cabcb7645
commit 0ce8647dc5
5 changed files with 53 additions and 47 deletions
+36 -36
View File
@@ -55,26 +55,26 @@ template<typename scalar_t>
__global__ void copy_blocks_kernel(
int64_t* key_cache_ptrs,
int64_t* value_cache_ptrs,
const int* __restrict__ block_mapping,
const int64_t* __restrict__ block_mapping,
const int numel_per_block) {
const int layer_idx = blockIdx.x;
const int pair_idx = blockIdx.y;
scalar_t* key_cache = reinterpret_cast<scalar_t*>(key_cache_ptrs[layer_idx]);
scalar_t* value_cache = reinterpret_cast<scalar_t*>(value_cache_ptrs[layer_idx]);
int src_block_number = block_mapping[2 * pair_idx];
int dst_block_number = block_mapping[2 * pair_idx + 1];
int64_t src_block_number = block_mapping[2 * pair_idx];
int64_t dst_block_number = block_mapping[2 * pair_idx + 1];
const int src_block_offset = src_block_number * numel_per_block;
const int dst_block_offset = dst_block_number * numel_per_block;
const int64_t src_block_offset = src_block_number * numel_per_block;
const int64_t dst_block_offset = dst_block_number * numel_per_block;
for (int i = threadIdx.x; i < numel_per_block; i += blockDim.x) {
int src_offset = src_block_offset + i;
int dst_offset = dst_block_offset + i;
int64_t src_offset = src_block_offset + i;
int64_t dst_offset = dst_block_offset + i;
key_cache[dst_offset] = key_cache[src_offset];
}
for (int i = threadIdx.x; i < numel_per_block; i += blockDim.x) {
int src_offset = src_block_offset + i;
int dst_offset = dst_block_offset + i;
int64_t src_offset = src_block_offset + i;
int64_t dst_offset = dst_block_offset + i;
value_cache[dst_offset] = value_cache[src_offset];
}
}
@@ -102,15 +102,15 @@ void copy_blocks(
value_cache_ptrs[layer_idx] = reinterpret_cast<int64_t>(value_caches[layer_idx].data_ptr());
}
// Create block mapping array.
std::vector<int> block_mapping_vec;
std::vector<int64_t> block_mapping_vec;
for (const auto& pair : block_mapping) {
int src_block_number = pair.first;
for (int dst_block_number : pair.second) {
int64_t src_block_number = pair.first;
for (int64_t dst_block_number : pair.second) {
block_mapping_vec.push_back(src_block_number);
block_mapping_vec.push_back(dst_block_number);
}
}
int* block_mapping_array = block_mapping_vec.data();
int64_t* block_mapping_array = block_mapping_vec.data();
int num_pairs = block_mapping_vec.size() / 2;
// Move the data structures to the GPU.
@@ -120,7 +120,7 @@ void copy_blocks(
torch::Tensor value_cache_ptrs_tensor = torch::from_blob(
value_cache_ptrs, {num_layers}, torch::kInt64).to(cache_device);
torch::Tensor block_mapping_tensor = torch::from_blob(
block_mapping_array, {2 * num_pairs}, torch::kInt).to(cache_device);
block_mapping_array, {2 * num_pairs}, torch::kInt64).to(cache_device);
// Launch the kernel.
const int numel_per_block = key_caches[0][0].numel();
@@ -132,7 +132,7 @@ void copy_blocks(
vllm::copy_blocks_kernel<scalar_t><<<grid, block, 0, stream>>>(
key_cache_ptrs_tensor.data_ptr<int64_t>(),
value_cache_ptrs_tensor.data_ptr<int64_t>(),
block_mapping_tensor.data_ptr<int>(),
block_mapping_tensor.data_ptr<int64_t>(),
numel_per_block);
}));
}
@@ -141,46 +141,46 @@ namespace vllm {
template<typename scalar_t>
__global__ void reshape_and_cache_kernel(
const scalar_t* __restrict__ key, // [num_tokens, num_heads, head_size]
const scalar_t* __restrict__ value, // [num_tokens, num_heads, head_size]
scalar_t* __restrict__ key_cache, // [num_blocks, num_heads, head_size/x, block_size, x]
scalar_t* __restrict__ value_cache, // [num_blocks, num_heads, head_size, block_size]
const int* __restrict__ slot_mapping, // [num_tokens]
const scalar_t* __restrict__ key, // [num_tokens, num_heads, head_size]
const scalar_t* __restrict__ value, // [num_tokens, num_heads, head_size]
scalar_t* __restrict__ key_cache, // [num_blocks, num_heads, head_size/x, block_size, x]
scalar_t* __restrict__ value_cache, // [num_blocks, num_heads, head_size, block_size]
const int64_t* __restrict__ slot_mapping, // [num_tokens]
const int key_stride,
const int value_stride,
const int num_heads,
const int head_size,
const int block_size,
const int x) {
const int token_idx = blockIdx.x;
const int slot_idx = slot_mapping[token_idx];
const int64_t token_idx = blockIdx.x;
const int64_t slot_idx = slot_mapping[token_idx];
if (slot_idx < 0) {
// Padding token that should be ignored.
return;
}
const int block_idx = slot_idx / block_size;
const int block_offset = slot_idx % block_size;
const int64_t block_idx = slot_idx / block_size;
const int64_t block_offset = slot_idx % block_size;
const int n = num_heads * head_size;
for (int i = threadIdx.x; i < n; i += blockDim.x) {
const int src_key_idx = token_idx * key_stride + i;
const int src_value_idx = token_idx * value_stride + i;
const int64_t src_key_idx = token_idx * key_stride + i;
const int64_t src_value_idx = token_idx * value_stride + i;
const int head_idx = i / head_size;
const int head_offset = i % head_size;
const int x_idx = head_offset / x;
const int x_offset = head_offset % x;
const int tgt_key_idx = block_idx * num_heads * (head_size / x) * block_size * x
+ head_idx * (head_size / x) * block_size * x
+ x_idx * block_size * x
+ block_offset * x
+ x_offset;
const int tgt_value_idx = block_idx * num_heads * head_size * block_size
+ head_idx * head_size * block_size
+ head_offset * block_size
+ block_offset;
const int64_t tgt_key_idx = block_idx * num_heads * (head_size / x) * block_size * x
+ head_idx * (head_size / x) * block_size * x
+ x_idx * block_size * x
+ block_offset * x
+ x_offset;
const int64_t tgt_value_idx = block_idx * num_heads * head_size * block_size
+ head_idx * head_size * block_size
+ head_offset * block_size
+ block_offset;
key_cache[tgt_key_idx] = key[src_key_idx];
value_cache[tgt_value_idx] = value[src_value_idx];
}
@@ -216,7 +216,7 @@ void reshape_and_cache(
value.data_ptr<scalar_t>(),
key_cache.data_ptr<scalar_t>(),
value_cache.data_ptr<scalar_t>(),
slot_mapping.data_ptr<int>(),
slot_mapping.data_ptr<int64_t>(),
key_stride,
value_stride,
num_heads,