File size: 10,539 Bytes
df13683 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 | # Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
import torch
import triton
import triton.language as tl
from fla.ops.utils import prepare_chunk_indices
from fla.utils import IS_AMD, autotune_cache_kwargs, get_multiprocessor_count, input_guard
NUM_WARPS_AUTOTUNE = [2, 4, 8, 16] if IS_AMD else [2, 4, 8, 16, 32]
def k_update_ref(k: torch.Tensor, a: torch.Tensor, ka: torch.Tensor) -> torch.Tensor:
return k.addcmul(k * (a - 1), ka)
@triton.heuristics({'IS_VARLEN': lambda args: args['cu_seqlens'] is not None})
@triton.autotune(
configs=[
triton.Config({}, num_warps=w, num_stages=s)
for w in NUM_WARPS_AUTOTUNE
for s in [1, 2, 3]
],
key=['BD'],
**autotune_cache_kwargs,
)
@triton.jit
def k_update_fwd_kernel_short(
k, a, ka, out,
cu_seqlens,
T, D,
BD: tl.constexpr,
IS_VARLEN: tl.constexpr,
):
i_b, i_t = tl.program_id(0), tl.program_id(1)
if IS_VARLEN:
bos = tl.load(cu_seqlens + i_b).to(tl.int32)
eos = tl.load(cu_seqlens + i_b + 1).to(tl.int32)
g_t = bos + i_t
if g_t >= eos:
return
offset = g_t * D
else:
g_t = i_t
offset = i_b * T * D + g_t * D
o_d = tl.arange(0, BD)
m_d = o_d < D
off = offset + o_d
b_k = tl.load(k + off, mask=m_d, other=0.).to(tl.float32)
b_a = tl.load(a + off, mask=m_d, other=0.).to(tl.float32)
b_ka = tl.load(ka + o_d, mask=m_d, eviction_policy='evict_last').to(tl.float32)
out_val = b_k * (1 + (b_a - 1) * b_ka)
tl.store(out + off, out_val.to(out.dtype.element_ty), mask=m_d)
@triton.heuristics({'IS_VARLEN': lambda args: args['cu_seqlens'] is not None})
@triton.autotune(
configs=[
triton.Config({}, num_warps=w, num_stages=s)
for w in NUM_WARPS_AUTOTUNE
for s in [1, 2, 3]
],
key=['BD', 'BT'],
**autotune_cache_kwargs,
)
@triton.jit
def k_update_fwd_kernel_long(
k, a, ka, out,
cu_seqlens, chunk_indices,
T, D,
BD: tl.constexpr, BT: tl.constexpr,
IS_VARLEN: tl.constexpr,
):
i_d, i_t_blk, i_b = tl.program_id(0), tl.program_id(1), tl.program_id(2)
if IS_VARLEN:
i_n, i_t_blk = tl.load(chunk_indices + i_t_blk * 2).to(tl.int32), \
tl.load(chunk_indices + i_t_blk * 2 + 1).to(tl.int32)
bos = tl.load(cu_seqlens + i_n).to(tl.int32)
eos = tl.load(cu_seqlens + i_n + 1).to(tl.int32)
t_start = i_t_blk * BT
t_end = tl.minimum(t_start + BT, eos - bos)
else:
bos = i_b * T
eos = (i_b + 1) * T
t_start = i_t_blk * BT
t_end = tl.minimum(t_start + BT, T)
o_d = i_d * BD + tl.arange(0, BD)
m_d = o_d < D
for t in range(t_start, t_end):
global_t = bos + t
off = global_t * D + o_d
b_k = tl.load(k + off, mask=m_d, other=0.).to(tl.float32)
b_a = tl.load(a + off, mask=m_d, other=0.).to(tl.float32)
b_ka = tl.load(ka + o_d, mask=m_d, eviction_policy='evict_last').to(tl.float32)
out_val = b_k * (1 + (b_a - 1) * b_ka)
tl.store(out + off, out_val.to(out.dtype.element_ty), mask=m_d)
@triton.heuristics({'IS_VARLEN': lambda args: args['cu_seqlens'] is not None})
@triton.autotune(
configs=[
triton.Config({'BT': BT}, num_warps=w, num_stages=s)
for w in NUM_WARPS_AUTOTUNE
for s in [1, 2, 3]
for BT in [2, 4, 8]
],
key=['BD'],
**autotune_cache_kwargs,
)
@triton.jit
def k_update_bwd_kernel_short(
grad_out, k, a, ka,
dk, da, dka,
cu_seqlens,
T, D,
BT: tl.constexpr,
BD: tl.constexpr,
IS_VARLEN: tl.constexpr,
):
i_b, i_t_base = tl.program_id(0), tl.program_id(1) * BT
if IS_VARLEN:
bos = tl.load(cu_seqlens + i_b).to(tl.int32)
eos = tl.load(cu_seqlens + i_b + 1).to(tl.int32)
seq_len = eos - bos
else:
bos = i_b * T
eos = (i_b + 1) * T
seq_len = T
t_vec = i_t_base + tl.arange(0, BT)
mask_t = t_vec < seq_len
global_t_vec = bos + t_vec
o_d = tl.arange(0, BD)[None, :]
m_d = o_d < D
off = global_t_vec[:, None] * D + o_d
b_go = tl.load(grad_out + off, mask=mask_t[:, None] & m_d, other=0.).to(tl.float32)
b_k = tl.load(k + off, mask=mask_t[:, None] & m_d, other=0.).to(tl.float32)
b_a = tl.load(a + off, mask=mask_t[:, None] & m_d, other=0.).to(tl.float32)
b_ka = tl.load(ka + o_d, mask=m_d, eviction_policy='evict_last').to(tl.float32) # [1, BD]
dk_vec = b_go * (1 + (b_a - 1) * b_ka)
da_vec = b_go * b_k * b_ka
dka_vec = b_go * b_k * (b_a - 1)
tl.store(dk + off, dk_vec.to(dk.dtype.element_ty), mask=mask_t[:, None] & m_d)
tl.store(da + off, da_vec.to(da.dtype.element_ty), mask=mask_t[:, None] & m_d)
tl.store(dka + off, dka_vec.to(dka.dtype.element_ty), mask=mask_t[:, None] & m_d)
@triton.heuristics({'IS_VARLEN': lambda args: args['cu_seqlens'] is not None})
@triton.autotune(
configs=[
triton.Config({}, num_warps=w, num_stages=s)
for w in NUM_WARPS_AUTOTUNE
for s in [1, 2, 3]
],
key=['BD', 'BT'],
**autotune_cache_kwargs,
)
@triton.jit
def k_update_bwd_kernel_long(
grad_out, k, a, ka,
dk, da, dka,
cu_seqlens, chunk_indices,
T, D,
BD: tl.constexpr, BT: tl.constexpr,
IS_VARLEN: tl.constexpr,
):
i_d, i_t_blk, i_b = tl.program_id(0), tl.program_id(1), tl.program_id(2)
if IS_VARLEN:
i_n, i_t_blk = tl.load(chunk_indices + i_t_blk * 2).to(tl.int32), \
tl.load(chunk_indices + i_t_blk * 2 + 1).to(tl.int32)
bos = tl.load(cu_seqlens + i_n).to(tl.int32)
eos = tl.load(cu_seqlens + i_n + 1).to(tl.int32)
t_start = i_t_blk * BT
t_end = tl.minimum(t_start + BT, eos - bos)
else:
bos = i_b * T
eos = (i_b + 1) * T
t_start = i_t_blk * BT
t_end = tl.minimum(t_start + BT, T)
o_d = i_d * BD + tl.arange(0, BD)
m_d = o_d < D
for t in range(t_start, t_end):
global_t = bos + t
off = global_t * D + o_d
b_go = tl.load(grad_out + off, mask=m_d, other=0.).to(tl.float32)
b_k = tl.load(k + off, mask=m_d, other=0.).to(tl.float32)
b_a = tl.load(a + off, mask=m_d, other=0.).to(tl.float32)
b_ka = tl.load(ka + o_d, mask=m_d, eviction_policy='evict_last').to(tl.float32)
tl.store(dk + off, (b_go * (1 + (b_a - 1) * b_ka)).to(dk.dtype.element_ty), mask=m_d)
tl.store(da + off, (b_go * b_k * b_ka).to(da.dtype.element_ty), mask=m_d)
tl.store(dka + off, (b_go * b_k * (b_a - 1)).to(dka.dtype.element_ty), mask=m_d)
def k_update_fwd(
k: torch.Tensor,
a: torch.Tensor,
ka: torch.Tensor,
cu_seqlens: torch.Tensor | None = None,
cu_seqlens_cpu: torch.LongTensor | None = None,
) -> torch.Tensor:
B, T, D = k.shape
out = torch.empty_like(k)
use_short = T <= 512
if use_short:
if cu_seqlens is not None:
N = len(cu_seqlens) - 1
else:
N = B
BD = triton.next_power_of_2(D)
grid = (N, T)
k_update_fwd_kernel_short[grid](
k, a, ka, out,
cu_seqlens,
T, D,
BD=BD,
)
else:
BT = min(64, triton.next_power_of_2(
triton.cdiv(max(16, B * T), get_multiprocessor_count(k.device.index)),
))
if cu_seqlens is not None:
chunk_idx = prepare_chunk_indices(cu_seqlens, BT, cu_seqlens_cpu=cu_seqlens_cpu)
NT = len(chunk_idx)
N = len(cu_seqlens) - 1
else:
chunk_idx = None
NT = triton.cdiv(T, BT)
N = B
BD = triton.next_power_of_2(D)
def grid(meta):
return (triton.cdiv(D, meta['BD']), NT, N)
k_update_fwd_kernel_long[grid](
k, a, ka, out,
cu_seqlens, chunk_idx,
T, D,
BD=BD, BT=BT,
)
return out, use_short, N, T
def k_update_bwd(
grad_out: torch.Tensor,
k: torch.Tensor,
a: torch.Tensor,
ka: torch.Tensor,
cu_seqlens: torch.Tensor | None,
use_short: bool,
N: int,
T: int,
cu_seqlens_cpu: torch.LongTensor | None = None,
):
B, _, D = grad_out.shape
dk = torch.empty_like(k)
da = torch.empty_like(a)
dka_tmp = torch.empty_like(k, dtype=torch.float32)
if use_short:
BD = triton.next_power_of_2(D)
def grid(meta): return (N, triton.cdiv(T, meta['BT']))
k_update_bwd_kernel_short[grid](
grad_out, k, a, ka,
dk, da, dka_tmp,
cu_seqlens,
T, D,
BD=BD,
)
else:
BT = min(64, triton.next_power_of_2(
triton.cdiv(max(16, B * T), get_multiprocessor_count(grad_out.device.index)),
))
if cu_seqlens is not None:
chunk_idx = prepare_chunk_indices(cu_seqlens, BT, cu_seqlens_cpu=cu_seqlens_cpu)
NT = len(chunk_idx)
else:
chunk_idx = None
NT = triton.cdiv(T, BT)
BD = triton.next_power_of_2(D)
def grid(meta):
return (triton.cdiv(D, meta['BD']), NT, N)
k_update_bwd_kernel_long[grid](
grad_out, k, a, ka,
dk, da, dka_tmp,
cu_seqlens, chunk_idx,
T, D,
BD=BD, BT=BT,
)
if dka_tmp.dim() == 3:
dka = dka_tmp.sum(dim=(0, 1), keepdim=True).type_as(ka)
else:
dka = dka_tmp.sum(dim=(0, 1)).type_as(ka)
return dk, da, dka
class KUpdateFunction(torch.autograd.Function):
@staticmethod
@input_guard
def forward(ctx, k, a, ka, cu_seqlens=None, cu_seqlens_cpu=None):
out, use_short, N, T = k_update_fwd(k, a, ka, cu_seqlens, cu_seqlens_cpu=cu_seqlens_cpu)
ctx.save_for_backward(k, a, ka)
ctx.use_short = use_short
ctx.N = N
ctx.T = T
ctx.cu_seqlens = cu_seqlens
ctx.cu_seqlens_cpu = cu_seqlens_cpu
return out
@staticmethod
@input_guard
def backward(ctx, grad_output):
k, a, ka = ctx.saved_tensors
dk, da, dka = k_update_bwd(
grad_output, k, a, ka,
ctx.cu_seqlens,
ctx.use_short,
ctx.N,
ctx.T,
cu_seqlens_cpu=ctx.cu_seqlens_cpu,
)
return dk, da, dka, None, None
def fused_k_rwkv7(k, a, ka, cu_seqlens=None, cu_seqlens_cpu=None):
if k.shape[1] == 1:
return k_update_ref(k, a, ka)
return KUpdateFunction.apply(k, a, ka, cu_seqlens, cu_seqlens_cpu)
|