Commit 99b940b2 authored by PanZezhong's avatar PanZezhong
Browse files

issue/847 paged attention prefill一段式接口

parent e13ad8f9
......@@ -2,6 +2,7 @@
#include "../tensor.hpp"
#include <optional>
#include <type_traits>
namespace infinicore {
......@@ -24,6 +25,15 @@ inline void hash_combine(size_t &seed, Tensor tensor) {
}
}
// Specialization for optional
template <typename T>
inline void hash_combine(size_t &seed, const std::optional<T> &opt) {
hash_combine(seed, opt.has_value());
if (opt) {
hash_combine(seed, *opt);
}
}
// Specialization for std::string
inline void hash_combine(size_t &seed, const std::string &str) {
hash_combine(seed, std::hash<std::string>{}(str));
......
......@@ -6,6 +6,7 @@
#include "ops/matmul.hpp"
#include "ops/ones.hpp"
#include "ops/paged_attention.hpp"
#include "ops/paged_attention_prefill.hpp"
#include "ops/paged_caching.hpp"
#include "ops/random_sample.hpp"
#include "ops/rearrange.hpp"
......
#pragma once
#include "../device.hpp"
#include "common/op.hpp"
#include <optional>
namespace infinicore::op {
class PagedAttentionPrefill {
public:
using schema = void (*)(Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, Tensor, std::optional<Tensor>, float);
static void execute(Tensor out, Tensor q, Tensor k_cache, Tensor v_cache, Tensor block_tables, Tensor cache_lens, Tensor seq_lens, Tensor seq_offsets, std::optional<Tensor> alibi_slopes, float);
static common::OpDispatcher<schema> &dispatcher();
};
Tensor paged_attention_prefill(Tensor q, Tensor k_cache, Tensor v_cache, Tensor block_tables, Tensor cache_lens, Tensor seq_lens, Tensor seq_offsets, std::optional<Tensor> alibi_slopes, float scale);
void paged_attention_prefill_(Tensor out, Tensor q, Tensor k_cache, Tensor v_cache, Tensor block_tables, Tensor cache_lens, Tensor seq_lens, Tensor seq_offsets, std::optional<Tensor> alibi_slopes, float scale);
} // namespace infinicore::op
......@@ -45,6 +45,7 @@ from infinicore.ops.matmul import matmul
from infinicore.ops.mul import mul
from infinicore.ops.narrow import narrow
from infinicore.ops.paged_attention import paged_attention
from infinicore.ops.paged_attention_prefill import paged_attention_prefill
from infinicore.ops.paged_caching import paged_caching
from infinicore.ops.rearrange import rearrange
from infinicore.ops.squeeze import squeeze
......@@ -119,6 +120,7 @@ __all__ = [
"from_torch",
"paged_caching",
"paged_attention",
"paged_attention_prefill",
"ones",
"strided_empty",
"strided_from_blob",
......
from infinicore.lib import _infinicore
from infinicore.tensor import Tensor
def paged_attention_prefill(
q: Tensor,
k_cache: Tensor,
v_cache: Tensor,
block_tables: Tensor,
cache_lens: Tensor,
seq_lens: Tensor,
seq_offsets: Tensor,
alibi_slopes: Tensor | None = None,
scale: float = 1.0,
*,
out: Tensor | None = None,
):
if out is None:
return Tensor(
_infinicore.paged_attention_prefill(
q._underlying,
k_cache._underlying,
v_cache._underlying,
block_tables._underlying,
cache_lens._underlying,
seq_lens._underlying,
seq_offsets._underlying,
alibi_slopes._underlying if alibi_slopes is not None else None,
scale,
)
)
_infinicore.paged_attention_prefill_(
out._underlying,
q._underlying,
k_cache._underlying,
v_cache._underlying,
block_tables._underlying,
cache_lens._underlying,
seq_lens._underlying,
seq_offsets._underlying,
alibi_slopes._underlying if alibi_slopes is not None else None,
scale,
)
return out
......@@ -16,7 +16,7 @@ thread_local common::OpCache<size_t, infiniopPagedAttentionDescriptor_t> caches(
});
void calculate(Tensor out, Tensor q, Tensor k_cache, Tensor v_cache, Tensor block_tables, Tensor cache_lens, std::optional<Tensor> alibi_slopes, float scale) {
size_t seed = hash_combine(out, q, k_cache, v_cache, block_tables, cache_lens);
size_t seed = hash_combine(out, q, k_cache, v_cache, block_tables, cache_lens, alibi_slopes, scale);
auto device = context::getDevice();
auto &cache = caches.getCache(device);
......
#include "infinicore/ops/paged_attention_prefill.hpp"
#include "../../utils.hpp"
namespace infinicore::op {
common::OpDispatcher<PagedAttentionPrefill::schema> &PagedAttentionPrefill::dispatcher() {
static common::OpDispatcher<PagedAttentionPrefill::schema> dispatcher_;
return dispatcher_;
};
void PagedAttentionPrefill::execute(Tensor out, Tensor q, Tensor k_cache, Tensor v_cache, Tensor block_tables, Tensor cache_lens, Tensor seq_lens, Tensor seq_offsets, std::optional<Tensor> alibi_slopes, float scale) {
INFINICORE_ASSERT_TENSORS_SAME_DEVICE(out, q, k_cache, v_cache, block_tables, cache_lens);
infinicore::context::setDevice(out->device());
dispatcher().lookup(out->device().getType())(out, q, k_cache, v_cache, block_tables, cache_lens, seq_lens, seq_offsets, alibi_slopes, scale);
}
Tensor paged_attention_prefill(Tensor q, Tensor k_cache, Tensor v_cache, Tensor block_tables, Tensor cache_lens, Tensor seq_lens, Tensor seq_offsets, std::optional<Tensor> alibi_slopes, float scale) {
auto out = Tensor::empty(q->shape(), q->dtype(), q->device());
paged_attention_prefill_(out, q, k_cache, v_cache, block_tables, cache_lens, seq_lens, seq_offsets, alibi_slopes, scale);
return out;
}
void paged_attention_prefill_(Tensor out, Tensor q, Tensor k_cache, Tensor v_cache, Tensor block_tables, Tensor cache_lens, Tensor seq_lens, Tensor seq_offsets, std::optional<Tensor> alibi_slopes, float scale) {
PagedAttentionPrefill::execute(out, q, k_cache, v_cache, block_tables, cache_lens, seq_lens, seq_offsets, alibi_slopes, scale);
}
} // namespace infinicore::op
#include "../../utils.hpp"
#include "infinicore/common/hash.hpp"
#include "infinicore/ops/common/cache.hpp"
#include "infinicore/ops/paged_attention_prefill.hpp"
#include <infiniop.h>
namespace infinicore::op::paged_attention_prefill_impl::infiniop {
thread_local common::OpCache<size_t, infiniopPagedAttentionPrefillDescriptor_t> caches(
100, // capacity
[](infiniopPagedAttentionPrefillDescriptor_t &desc) {
if (desc != nullptr) {
INFINICORE_CHECK_ERROR(infiniopDestroyPagedAttentionPrefillDescriptor(desc));
desc = nullptr;
}
});
void calculate(Tensor out, Tensor q, Tensor k_cache, Tensor v_cache, Tensor block_tables, Tensor cache_lens, Tensor seq_lens, Tensor seq_offsets, std::optional<Tensor> alibi_slopes, float scale) {
size_t seed = hash_combine(out, q, k_cache, v_cache, block_tables, cache_lens, seq_lens, seq_offsets, alibi_slopes, scale);
auto device = context::getDevice();
auto &cache = caches.getCache(device);
auto desc_opt = cache.get(seed);
infiniopPagedAttentionPrefillDescriptor_t desc = nullptr;
if (!desc_opt) {
INFINICORE_CHECK_ERROR(infiniopCreatePagedAttentionPrefillDescriptor(
context::getInfiniopHandle(device), &desc,
out->desc(), q->desc(), k_cache->desc(), v_cache->desc(), block_tables->desc(),
cache_lens->desc(), seq_lens->desc(), seq_offsets->desc(),
alibi_slopes.has_value() ? alibi_slopes.value()->desc() : nullptr,
scale));
cache.put(seed, desc);
} else {
desc = *desc_opt;
}
size_t workspace_size = 0;
INFINICORE_CHECK_ERROR(infiniopGetPagedAttentionPrefillWorkspaceSize(desc, &workspace_size));
std::shared_ptr<Memory> workspace = context::allocateMemory(workspace_size);
INFINICORE_CHECK_ERROR(infiniopPagedAttentionPrefill(
desc, workspace->data(), workspace_size,
out->data(), q->data(), k_cache->data(), v_cache->data(), block_tables->data(), cache_lens->data(), seq_lens->data(), seq_offsets->data(),
alibi_slopes.has_value() ? alibi_slopes.value()->data() : nullptr,
context::getStream()));
}
static bool registered = []() {
PagedAttentionPrefill::dispatcher().registerAll(&calculate, false);
return true;
}();
} // namespace infinicore::op::paged_attention_prefill_impl::infiniop
......@@ -4,10 +4,10 @@
namespace op::paged_attention_prefill::cuda {
// 辅助函数:二分查找确定当前 global_token_idx 属于哪个 sequence
__device__ __forceinline__ int find_seq_id(int token_idx, const int64_t *offset, int num_seqs) {
int low = 0, high = num_seqs - 1;
__device__ __forceinline__ size_t find_seq_id(size_t token_idx, const int64_t *offset, size_t num_seqs) {
size_t low = 0, high = num_seqs - 1;
while (low <= high) {
int mid = (low + high) >> 1;
size_t mid = (low + high) >> 1;
if (token_idx >= offset[mid] && token_idx < offset[mid + 1]) {
return mid;
} else if (token_idx < offset[mid]) {
......@@ -32,22 +32,22 @@ __global__ void pagedAttentionPrefillKernel(
const size_t num_seqs) {
// --- 使用 2D Grid 坐标 ---
const int global_token_idx = blockIdx.x; // 展平后的全局 token 索引
const int head_idx = blockIdx.y; // Head 索引
const int dim_idx = threadIdx.x; // Head 内部维度
const size_t global_token_idx = blockIdx.x; // 展平后的全局 token 索引
const size_t head_idx = blockIdx.y; // Head 索引
const size_t dim_idx = threadIdx.x; // Head 内部维度
if (dim_idx >= head_size) {
return;
}
// --- 通过二分查找 offset 找到所属的 seq_idx ---
int seq_idx = find_seq_id(global_token_idx, offset_, num_seqs);
size_t seq_idx = find_seq_id(global_token_idx, offset_, num_seqs);
// --- 获取该 Sequence 本次 Prefill 的长度
const int64_t cur_new_len = seq_lens_[seq_idx];
// --- 该 token 在当前序列中的相对位置
int q_token_idx = global_token_idx - offset_[seq_idx];
size_t q_token_idx = global_token_idx - offset_[seq_idx];
const Tdata *q_ptr_base = q_ + global_token_idx * num_heads * head_size + head_idx * head_size;
Tdata *out_ptr = out_ + global_token_idx * num_heads * head_size + head_idx * head_size;
......@@ -65,14 +65,14 @@ __global__ void pagedAttentionPrefillKernel(
// Pass 1: 计算 Score 并找最大值
Tcompute max_score = -FLT_MAX;
for (int t = 0; t <= causal_limit; ++t) {
for (size_t t = 0; t <= causal_limit; ++t) {
const int64_t b_idx = t / block_size;
const int64_t t_off = t % block_size;
const int64_t physical_block_id = block_table[b_idx];
const Tdata *k_vec = k_cache_ + physical_block_id * kv_block_stride + kv_head_idx * kv_head_stride + t_off * head_size;
Tcompute score = 0.0f;
for (int d = 0; d < head_size; ++d) {
for (size_t d = 0; d < head_size; ++d) {
score += static_cast<Tcompute>(q_ptr_base[d]) * static_cast<Tcompute>(k_vec[d]);
}
score *= static_cast<Tcompute>(scale);
......@@ -86,14 +86,14 @@ __global__ void pagedAttentionPrefillKernel(
// Pass 2: 计算 Sum of Exp
Tcompute sum_exp = 0.0f;
for (int t = 0; t <= causal_limit; ++t) {
for (size_t t = 0; t <= causal_limit; ++t) {
const int64_t b_idx = t / block_size;
const int64_t t_off = t % block_size;
const int64_t physical_block_id = block_table[b_idx];
const Tdata *k_vec = k_cache_ + physical_block_id * kv_block_stride + kv_head_idx * kv_head_stride + t_off * head_size;
Tcompute score = 0.0f;
for (int d = 0; d < head_size; ++d) {
for (size_t d = 0; d < head_size; ++d) {
score += static_cast<Tcompute>(q_ptr_base[d]) * static_cast<Tcompute>(k_vec[d]);
}
score *= static_cast<Tcompute>(scale);
......@@ -106,14 +106,14 @@ __global__ void pagedAttentionPrefillKernel(
// Pass 3: 加权求和得到输出
Tcompute acc = 0.0f;
Tcompute inv_sum = 1.0f / (sum_exp + 1e-6f);
for (int t = 0; t <= causal_limit; ++t) {
for (size_t t = 0; t <= causal_limit; ++t) {
const int64_t b_idx = t / block_size;
const int64_t t_off = t % block_size;
const int64_t physical_block_id = block_table[b_idx];
const Tdata *k_vec = k_cache_ + physical_block_id * kv_block_stride + kv_head_idx * kv_head_stride + t_off * head_size;
Tcompute score = 0.0f;
for (int d = 0; d < head_size; ++d) {
for (size_t d = 0; d < head_size; ++d) {
score += static_cast<Tcompute>(q_ptr_base[d]) * static_cast<Tcompute>(k_vec[d]);
}
score *= static_cast<Tcompute>(scale);
......
......@@ -55,7 +55,7 @@ target("infiniop-nvidia")
end
end
add_cuflags("-Xcompiler=-Wno-error=deprecated-declarations")
add_cuflags("-Xcompiler=-Wno-error=deprecated-declarations", "-Xcompiler=-Wno-error=unused-function")
local arch_opt = get_config("cuda_arch")
if arch_opt and type(arch_opt) == "string" then
......
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment