ViperEkura
d5067af064
- NSDMI null/-1 defaults for AttentionParams/FP8Params pointer+flag members: partially packed structs can no longer hold garbage non-null pointers that gate optional paths (root cause class of the paged test bug); still aggregates, still trivially copyable - move per-lane ldmatrix wrappers (ldsm_x2/x4) from fp8/gemm.cuh to common/mma.cuh as ldmatrix_x2_lane/x4_lane, next to the single-address variants - DEVICE_FORCEINLINE macro in common/mma.cuh (matches layout_policies.cuh, internal linkage) - frag_addr now delegates to tile_at: the swizzle math has one source - operand layouts as CUTLASS-style RowMajor/ColMajor tags threaded from launch_fp8_gemm through the kernel to load_operand_tile; B's operand view via transpose_layout_t; call sites read <Fmt, false, RowMajor, ColMajor> instead of <Fmt, false, false, true>
96 lines
3.3 KiB
C++
96 lines
3.3 KiB
C++
#pragma once
|
|
|
|
// Pure POD header
|
|
|
|
namespace astrai {
|
|
namespace attention {
|
|
|
|
// Tensor layout for Q/K/V tensors passed to attention kernels.
|
|
// Internally, kernels always operate on BHLD [batch, n_heads, seq_len, head_dim].
|
|
// When the caller passes BLHD, dims 1 and 2 are transposed at entry.
|
|
enum TensorLayout : int {
|
|
BHLD = 0, // [batch, n_heads, seq_len, head_dim]
|
|
BLHD = 1, // [batch, seq_len, n_heads, head_dim]
|
|
};
|
|
|
|
// Split-KV workspace cap: max decode splits per (batch, q_head).
|
|
constexpr int MAX_SPLITS = 32;
|
|
|
|
|
|
// Unified attention params covering BOTH addressing modes:
|
|
// - Contiguous K/V: dense [batch, kv_head, kv_len, head_dim] tensors (k/v).
|
|
// - Paged (SGLang-style): flat pool [size, kv_head, head_dim] + req_to_token.
|
|
// Each kernel selects the addressing via a KVSource policy (see
|
|
// layout_policies.cuh); a given call only touches the fields of one mode, so
|
|
// this is a POD shared by both paths rather than two parallel structs that
|
|
// drift out of sync.
|
|
//
|
|
// Pointer/flag members carry default member initializers: the pointers gate
|
|
// optional paths via null checks (new_k_ptr, mask, o_part, ...), so a stack
|
|
// `AttentionParams<T> p;` left partially packed must never see garbage
|
|
// non-null pointers or a garbage use_mask/causal_offset — that class of bug
|
|
// reads through wild addresses. NSDMI keeps the struct an aggregate (C++17)
|
|
// and trivially copyable, so `= {}`, memcpy-style packing and by-value kernel
|
|
// params all behave exactly as before.
|
|
template<typename T, typename AT = float>
|
|
struct AttentionParams {
|
|
// Shape
|
|
int batch;
|
|
int q_head;
|
|
int kv_head;
|
|
int head_dim;
|
|
int q_len; // Per-request in contiguous mode; total_q in paged mode.
|
|
int kv_len; // Contiguous mode; paged mode uses kv_indptr.
|
|
|
|
// Attention behavior
|
|
float scale;
|
|
// -1 = non-causal; >=0 = absolute position of first Q token
|
|
int causal_offset = -1;
|
|
int use_mask = 0;
|
|
|
|
// pointers
|
|
const T* __restrict__ q_ptr = nullptr;
|
|
const T* __restrict__ k_ptr = nullptr;
|
|
const T* __restrict__ v_ptr = nullptr;
|
|
const T* __restrict__ new_k_ptr = nullptr;
|
|
const T* __restrict__ new_v_ptr = nullptr;
|
|
T* __restrict__ o_ptr = nullptr;
|
|
const bool* __restrict__ mask = nullptr;
|
|
|
|
// strides
|
|
int q_b_stride;
|
|
int q_h_stride;
|
|
int q_l_stride;
|
|
int q_d_stride;
|
|
|
|
int kv_b_stride;
|
|
int kv_h_stride;
|
|
int kv_l_stride;
|
|
int kv_d_stride;
|
|
|
|
int new_kv_b_stride;
|
|
int new_kv_h_stride;
|
|
|
|
int mask_b_stride;
|
|
int mask_h_stride;
|
|
int mask_l_stride;
|
|
|
|
// Paged K/V addressing
|
|
const int* __restrict__ req_to_token = nullptr; // [num_reqs, max_context_len]
|
|
const int* __restrict__ req_pool_indices = nullptr; // [batch]
|
|
const int* __restrict__ kv_indptr = nullptr; // [batch + 1]
|
|
const int* __restrict__ qo_indptr = nullptr; // [batch + 1] or nullptr for decode
|
|
const int* __restrict__ q_tile_to_batch = nullptr; // [num_q_tiles], prefill only
|
|
const int* __restrict__ q_tile_to_index = nullptr; // [num_q_tiles], prefill only
|
|
int num_q_tiles;
|
|
int max_context_len; // req_to_token stride (dim 1)
|
|
|
|
// Decode split-KV workspace
|
|
int num_splits;
|
|
AT* __restrict__ o_part = nullptr;
|
|
AT* __restrict__ ml_part = nullptr;
|
|
};
|
|
|
|
} // namespace attention
|
|
} // namespace astrai
|