File size: 7,684 Bytes
ba798d3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 | # Adapted from LongCat-Video.
# Original copyright (c) 2025 Meituan.
# Licensed under the MIT License.
#
# Modifications for Light Interaction:
# - adapted the sparse-attention kernel to HY-WorldPlay
# - integrated it with AR denoising and KV-cache recomputation paths
#
# Modifications copyright (c) 2026 Jiacheng Lu and contributors.
import torch
import triton
import triton.language as tl
import os
# Autotune configs for the aligned sparse-attention kernel.
# In this variant, BLOCK_N must match the logical sparse block size.
configs_fwd_bsa_align = [
triton.Config({}, num_stages=3, num_warps=8),
# triton.Config({}, num_stages=3, num_warps=4),
# triton.Config({}, num_stages=4, num_warps=4),
# triton.Config({}, num_stages=4, num_warps=8),
]
fwd_bsa_reevaluate_keys = ['Q_LEN', 'K_LEN', 'BLOCK_M', 'SPARSITY'] if os.environ.get('TRITON_REEVALUATE_KEY', '0') == '1' else []
@triton.autotune(list(configs_fwd_bsa_align), key=fwd_bsa_reevaluate_keys)
@triton.jit
def _attn_fwd_bsa_align(
Q, K, V, sm_scale, M, Out,
block_indices, block_indices_lens,
stride_qz, stride_qh, stride_qm, stride_qk,
stride_kz, stride_kh, stride_kn, stride_kk,
stride_vz, stride_vh, stride_vn, stride_vk,
stride_oz, stride_oh, stride_om, stride_ok,
stride_bz, stride_bh, stride_bm, stride_bs,
stride_lz, stride_lh, stride_lm,
H,
Q_LEN, K_LEN,
HEAD_DIM: tl.constexpr,
BLOCK_M: tl.constexpr,
BLOCK_N: tl.constexpr,
SPARSITY: tl.constexpr,
):
# Grid ID
start_m = tl.program_id(0)
off_hz = tl.program_id(1)
off_z = off_hz // H
off_h = off_hz % H
# Offsets
q_offset = off_z.to(tl.int64) * stride_qz + off_h.to(tl.int64) * stride_qh
k_offset = off_z.to(tl.int64) * stride_kz + off_h.to(tl.int64) * stride_kh
v_offset = off_z.to(tl.int64) * stride_vz + off_h.to(tl.int64) * stride_vh
o_offset = off_z.to(tl.int64) * stride_oz + off_h.to(tl.int64) * stride_oh
b_offset = off_z.to(tl.int64) * stride_bz + off_h.to(tl.int64) * stride_bh
l_offset = off_z.to(tl.int64) * stride_lz + off_h.to(tl.int64) * stride_lh
# Block Pointers
# Q: [BLOCK_M, HEAD_DIM]
Q_block_ptr = tl.make_block_ptr(
base=Q + q_offset,
shape=(Q_LEN, HEAD_DIM),
strides=(stride_qm, stride_qk),
offsets=(start_m * BLOCK_M, 0),
block_shape=(BLOCK_M, HEAD_DIM),
order=(1, 0),
)
V_block_ptr = tl.make_block_ptr(
base=V + v_offset,
shape=(K_LEN, HEAD_DIM),
strides=(stride_vn, stride_vk),
offsets=(0, 0),
block_shape=(BLOCK_N, HEAD_DIM),
order=(1, 0),
)
KT_block_ptr = tl.make_block_ptr(
base=K + k_offset,
shape=(HEAD_DIM, K_LEN),
strides=(stride_kk, stride_kn),
offsets=(0, 0),
block_shape=(HEAD_DIM, BLOCK_N),
order=(0, 1),
)
O_block_ptr = tl.make_block_ptr(
base=Out + o_offset,
shape=(Q_LEN, HEAD_DIM),
strides=(stride_om, stride_ok),
offsets=(start_m * BLOCK_M, 0),
block_shape=(BLOCK_M, HEAD_DIM),
order=(1, 0),
)
# Indices pointers
block_indices += b_offset + start_m * stride_bm
block_indices_lens += l_offset + start_m * stride_lm
# Accumulators
m_i = tl.zeros([BLOCK_M], dtype=tl.float32) - float("inf")
l_i = tl.zeros([BLOCK_M], dtype=tl.float32) + 1.0
acc = tl.zeros([BLOCK_M, HEAD_DIM], dtype=tl.float32)
# Pre-computation
qk_scale = sm_scale * 1.44269504
# Load Q (SRAM persistent)
q = tl.load(Q_block_ptr)
# Load Loop Length
S = tl.load(block_indices_lens)
# Aligned sparse-block loop. Each sparse index maps to one BLOCK_N-sized
# physical K/V block.
for i in range(S):
block_id = tl.load(block_indices + i * stride_bs).to(tl.int32)
lo = block_id * BLOCK_N
KT_curr = tl.advance(KT_block_ptr, (0, lo))
V_curr = tl.advance(V_block_ptr, (lo, 0))
kT = tl.load(KT_curr)
qkT = tl.dot(q, kT) # [BLOCK_M, BLOCK_N]
m_ij = tl.maximum(m_i, tl.max(qkT, 1) * qk_scale)
alpha = tl.math.exp2(m_i - m_ij)
acc = acc * alpha[:, None]
qkT = qkT * qk_scale - m_ij[:, None]
p = tl.math.exp2(qkT)
l_ij = tl.sum(p, 1)
l_i = l_i * alpha + l_ij
m_i = m_ij
v = tl.load(V_curr)
acc = tl.dot(p.to(v.dtype), v, acc)
# ============================================================
# Epilogue
# ============================================================
m_i += tl.math.log2(l_i)
acc = acc / l_i[:, None]
# Store M
offs_m = start_m * BLOCK_M + tl.arange(0, BLOCK_M)
m_ptrs = M + off_hz * Q_LEN + offs_m
tl.store(m_ptrs, m_i)
# Store O
tl.store(O_block_ptr, acc.to(Out.type.element_ty))
def longcat_flash_attn(q, k, v, block_indices, block_counts, sm_scale=None, logical_block_size=128):
"""Triton-based block-sparse flash attention.
Args:
q: Query ``(B, H, Q_len, D)`` (contiguous).
k: Key ``(B, H, K_len, D)`` (contiguous).
v: Value ``(B, H, K_len, D)`` (contiguous).
block_indices: Sparse block indices ``(B, H, Nq_blk, K_blk)`` int32.
block_counts: Number of valid blocks per query ``(B, H, Nq_blk)`` int32.
sm_scale: Scaling factor (default ``1/sqrt(D)``).
logical_block_size: Sparse block size used during index generation
(must evenly divide *K_len*).
Returns:
Attention output ``(B, H, Q_len, D)``.
Raises:
ValueError: If *K_len* is not divisible by *logical_block_size*
or if any block index is out of bounds.
"""
q = q.contiguous()
k = k.contiguous()
v = v.contiguous()
block_indices = block_indices.contiguous()
block_counts = block_counts.contiguous()
B, H, Q_len, HEAD_DIM = q.shape
K_len = k.shape[2]
if sm_scale is None:
sm_scale = 1.0 / (HEAD_DIM ** 0.5)
BLOCK_N = logical_block_size
if K_len % BLOCK_N != 0:
raise ValueError(f"K Length ({K_len}) must be divisible by logical_block_size ({BLOCK_N})")
if block_indices.numel() and block_indices.max() >= (K_len // BLOCK_N):
raise ValueError("Sparse indices out of bounds.")
# 3. Buffer
o = torch.empty_like(q)
m_buffer = torch.empty((B, H, Q_len), device=q.device, dtype=torch.float32)
# 4. Grid
BLOCK_M = 128
grid = lambda META: (
triton.cdiv(Q_len, BLOCK_M),
B * H
)
_attn_fwd_bsa_align[grid](
Q=q, K=k, V=v, sm_scale=sm_scale, M=m_buffer, Out=o,
block_indices=block_indices,
block_indices_lens=block_counts,
stride_qz=q.stride(0), stride_qh=q.stride(1), stride_qm=q.stride(2), stride_qk=q.stride(3),
stride_kz=k.stride(0), stride_kh=k.stride(1), stride_kn=k.stride(2), stride_kk=k.stride(3),
stride_vz=v.stride(0), stride_vh=v.stride(1), stride_vn=v.stride(2), stride_vk=v.stride(3),
stride_oz=o.stride(0), stride_oh=o.stride(1), stride_om=o.stride(2), stride_ok=o.stride(3),
stride_bz=block_indices.stride(0), stride_bh=block_indices.stride(1),
stride_bm=block_indices.stride(2), stride_bs=block_indices.stride(3),
stride_lz=block_counts.stride(0), stride_lh=block_counts.stride(1),
stride_lm=block_counts.stride(2),
H=H,
Q_LEN=Q_len,
K_LEN=K_len,
HEAD_DIM=HEAD_DIM,
BLOCK_M=BLOCK_M,
BLOCK_N=BLOCK_N,
SPARSITY=0
)
return o
|