File size: 12,078 Bytes
23d26f4 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 |
# Copyright (c) 2024, Tri Dao, Albert Gu.
"""We want triton==2.1.0 or triton==2.2.0 or triton==2.3.0 for this
"""
import math
import torch
import torch.nn.functional as F
import triton
import triton.language as tl
from einops import rearrange, repeat
from .softplus import softplus
@triton.heuristics({"HAS_DT_BIAS": lambda args: args["dt_bias_ptr"] is not None})
@triton.heuristics({"HAS_D": lambda args: args["D_ptr"] is not None})
@triton.heuristics({"HAS_Z": lambda args: args["z_ptr"] is not None})
@triton.heuristics(
{
"HAS_STATE_BATCH_INDICES": lambda args: args["state_batch_indices_ptr"]
is not None
}
)
@triton.heuristics(
{"BLOCK_SIZE_DSTATE": lambda args: triton.next_power_of_2(args["dstate"])}
)
@triton.jit
def _selective_scan_update_kernel(
# Pointers to matrices
state_ptr,
x_ptr,
dt_ptr,
dt_bias_ptr,
A_ptr,
B_ptr,
C_ptr,
D_ptr,
z_ptr,
out_ptr,
state_batch_indices_ptr,
# Matrix dimensions
batch,
nheads,
dim,
dstate,
nheads_ngroups_ratio,
# Strides
stride_state_batch,
stride_state_head,
stride_state_dim,
stride_state_dstate,
stride_x_batch,
stride_x_head,
stride_x_dim,
stride_dt_batch,
stride_dt_head,
stride_dt_dim,
stride_dt_bias_head,
stride_dt_bias_dim,
stride_A_head,
stride_A_dim,
stride_A_dstate,
stride_B_batch,
stride_B_group,
stride_B_dstate,
stride_C_batch,
stride_C_group,
stride_C_dstate,
stride_D_head,
stride_D_dim,
stride_z_batch,
stride_z_head,
stride_z_dim,
stride_out_batch,
stride_out_head,
stride_out_dim,
# Meta-parameters
DT_SOFTPLUS: tl.constexpr,
TIE_HDIM: tl.constexpr,
BLOCK_SIZE_M: tl.constexpr,
HAS_DT_BIAS: tl.constexpr,
HAS_D: tl.constexpr,
HAS_Z: tl.constexpr,
HAS_STATE_BATCH_INDICES: tl.constexpr,
BLOCK_SIZE_DSTATE: tl.constexpr,
):
pid_m = tl.program_id(axis=0)
pid_b = tl.program_id(axis=1)
pid_h = tl.program_id(axis=2)
if HAS_STATE_BATCH_INDICES:
state_batch_indices_ptr += pid_b
state_batch_idx = tl.load(state_batch_indices_ptr)
state_ptr += state_batch_idx * stride_state_batch + pid_h * stride_state_head
else:
state_ptr += pid_b * stride_state_batch + pid_h * stride_state_head
x_ptr += pid_b * stride_x_batch + pid_h * stride_x_head
dt_ptr += pid_b * stride_dt_batch + pid_h * stride_dt_head
if HAS_DT_BIAS:
dt_bias_ptr += pid_h * stride_dt_bias_head
A_ptr += pid_h * stride_A_head
B_ptr += pid_b * stride_B_batch + (pid_h // nheads_ngroups_ratio) * stride_B_group
C_ptr += pid_b * stride_C_batch + (pid_h // nheads_ngroups_ratio) * stride_C_group
if HAS_Z:
z_ptr += pid_b * stride_z_batch + pid_h * stride_z_head
out_ptr += pid_b * stride_out_batch + pid_h * stride_out_head
offs_m = pid_m * BLOCK_SIZE_M + tl.arange(0, BLOCK_SIZE_M)
offs_n = tl.arange(0, BLOCK_SIZE_DSTATE)
state_ptrs = state_ptr + (
offs_m[:, None] * stride_state_dim + offs_n[None, :] * stride_state_dstate
)
x_ptrs = x_ptr + offs_m * stride_x_dim
dt_ptrs = dt_ptr + offs_m * stride_dt_dim
if HAS_DT_BIAS:
dt_bias_ptrs = dt_bias_ptr + offs_m * stride_dt_bias_dim
if HAS_D:
D_ptr += pid_h * stride_D_head
A_ptrs = A_ptr + (
offs_m[:, None] * stride_A_dim + offs_n[None, :] * stride_A_dstate
)
B_ptrs = B_ptr + offs_n * stride_B_dstate
C_ptrs = C_ptr + offs_n * stride_C_dstate
if HAS_D:
D_ptrs = D_ptr + offs_m * stride_D_dim
if HAS_Z:
z_ptrs = z_ptr + offs_m * stride_z_dim
out_ptrs = out_ptr + offs_m * stride_out_dim
state = tl.load(
state_ptrs, mask=(offs_m[:, None] < dim) & (offs_n[None, :] < dstate), other=0.0
)
x = tl.load(x_ptrs, mask=offs_m < dim, other=0.0).to(tl.float32)
if not TIE_HDIM:
dt = tl.load(dt_ptrs, mask=offs_m < dim, other=0.0).to(tl.float32)
if HAS_DT_BIAS:
dt += tl.load(dt_bias_ptrs, mask=offs_m < dim, other=0.0).to(tl.float32)
if DT_SOFTPLUS:
dt = tl.where(dt <= 20.0, softplus(dt), dt)
A = tl.load(
A_ptrs, mask=(offs_m[:, None] < dim) & (offs_n[None, :] < dstate), other=0.0
).to(tl.float32)
dA = tl.exp(A * dt[:, None])
else:
dt = tl.load(dt_ptr).to(tl.float32)
if HAS_DT_BIAS:
dt += tl.load(dt_bias_ptr).to(tl.float32)
if DT_SOFTPLUS:
dt = tl.where(dt <= 20.0, softplus(dt), dt)
A = tl.load(A_ptr).to(tl.float32)
dA = tl.exp(A * dt) # scalar, not a matrix
B = tl.load(B_ptrs, mask=offs_n < dstate, other=0.0).to(tl.float32)
C = tl.load(C_ptrs, mask=offs_n < dstate, other=0.0).to(tl.float32)
if HAS_D:
D = tl.load(D_ptrs, mask=offs_m < dim, other=0.0).to(tl.float32)
if HAS_Z:
z = tl.load(z_ptrs, mask=offs_m < dim, other=0.0).to(tl.float32)
if not TIE_HDIM:
dB = B[None, :] * dt[:, None]
else:
dB = B * dt # vector of size (dstate,)
state = state * dA + dB * x[:, None]
tl.store(
state_ptrs, state, mask=(offs_m[:, None] < dim) & (offs_n[None, :] < dstate)
)
out = tl.sum(state * C[None, :], axis=1)
if HAS_D:
out += x * D
if HAS_Z:
out *= z * tl.sigmoid(z)
tl.store(out_ptrs, out, mask=offs_m < dim)
def selective_state_update(
state,
x,
dt,
A,
B,
C,
D=None,
z=None,
dt_bias=None,
dt_softplus=False,
state_batch_indices=None,
):
"""
Argument:
state: (batch, dim, dstate) or (batch, nheads, dim, dstate)
x: (batch, dim) or (batch, nheads, dim)
dt: (batch, dim) or (batch, nheads, dim)
A: (dim, dstate) or (nheads, dim, dstate)
B: (batch, dstate) or (batch, ngroups, dstate)
C: (batch, dstate) or (batch, ngroups, dstate)
D: (dim,) or (nheads, dim)
z: (batch, dim) or (batch, nheads, dim)
dt_bias: (dim,) or (nheads, dim)
Return:
out: (batch, dim) or (batch, nheads, dim)
"""
has_heads = state.dim() > 3
if state.dim() == 3:
state = state.unsqueeze(1)
if x.dim() == 2:
x = x.unsqueeze(1)
if dt.dim() == 2:
dt = dt.unsqueeze(1)
if A.dim() == 2:
A = A.unsqueeze(0)
if B.dim() == 2:
B = B.unsqueeze(1)
if C.dim() == 2:
C = C.unsqueeze(1)
if D is not None and D.dim() == 1:
D = D.unsqueeze(0)
if z is not None and z.dim() == 2:
z = z.unsqueeze(1)
if dt_bias is not None and dt_bias.dim() == 1:
dt_bias = dt_bias.unsqueeze(0)
_, nheads, dim, dstate = state.shape
batch = x.shape[0]
if x.shape != (batch, nheads, dim):
print(f"{state.shape} {x.shape} {batch} {nheads} {dim}")
assert x.shape == (batch, nheads, dim)
assert dt.shape == x.shape
assert A.shape == (nheads, dim, dstate)
ngroups = B.shape[1]
assert nheads % ngroups == 0, "nheads must be divisible by ngroups"
assert B.shape == (batch, ngroups, dstate)
assert C.shape == B.shape
if D is not None:
assert D.shape == (nheads, dim)
if z is not None:
assert z.shape == x.shape
if dt_bias is not None:
assert dt_bias.shape == (nheads, dim)
if state_batch_indices is not None:
assert state_batch_indices.shape == (batch,)
out = torch.empty_like(x)
grid = lambda META: (triton.cdiv(dim, META["BLOCK_SIZE_M"]), batch, nheads)
z_strides = (z.stride(0), z.stride(1), z.stride(2)) if z is not None else (0, 0, 0)
# We don't want autotune since it will overwrite the state
# We instead tune by hand.
BLOCK_SIZE_M, num_warps = (
(32, 4)
if dstate <= 16
else (
(16, 4)
if dstate <= 32
else ((8, 4) if dstate <= 64 else ((4, 4) if dstate <= 128 else ((4, 8))))
)
)
tie_hdim = (
A.stride(-1) == 0
and A.stride(-2) == 0
and dt.stride(-1) == 0
and dt_bias.stride(-1) == 0
)
with torch.cuda.device(x.device.index):
_selective_scan_update_kernel[grid](
state,
x,
dt,
dt_bias,
A,
B,
C,
D,
z,
out,
state_batch_indices,
batch,
nheads,
dim,
dstate,
nheads // ngroups,
state.stride(0),
state.stride(1),
state.stride(2),
state.stride(3),
x.stride(0),
x.stride(1),
x.stride(2),
dt.stride(0),
dt.stride(1),
dt.stride(2),
*(dt_bias.stride(0), dt_bias.stride(1)) if dt_bias is not None else 0,
A.stride(0),
A.stride(1),
A.stride(2),
B.stride(0),
B.stride(1),
B.stride(2),
C.stride(0),
C.stride(1),
C.stride(2),
*(D.stride(0), D.stride(1)) if D is not None else 0,
z_strides[0],
z_strides[1],
z_strides[2],
out.stride(0),
out.stride(1),
out.stride(2),
dt_softplus,
tie_hdim,
BLOCK_SIZE_M,
num_warps=num_warps,
)
if not has_heads:
out = out.squeeze(1)
return out
def selective_state_update_ref(
state, x, dt, A, B, C, D=None, z=None, dt_bias=None, dt_softplus=False
):
"""
Argument:
state: (batch, dim, dstate) or (batch, nheads, dim, dstate)
x: (batch, dim) or (batch, nheads, dim)
dt: (batch, dim) or (batch, nheads, dim)
A: (dim, dstate) or (nheads, dim, dstate)
B: (batch, dstate) or (batch, ngroups, dstate)
C: (batch, dstate) or (batch, ngroups, dstate)
D: (dim,) or (nheads, dim)
z: (batch, dim) or (batch, nheads, dim)
dt_bias: (dim,) or (nheads, dim)
Return:
out: (batch, dim) or (batch, nheads, dim)
"""
has_heads = state.dim() > 3
if state.dim() == 3:
state = state.unsqueeze(1)
if x.dim() == 2:
x = x.unsqueeze(1)
if dt.dim() == 2:
dt = dt.unsqueeze(1)
if A.dim() == 2:
A = A.unsqueeze(0)
if B.dim() == 2:
B = B.unsqueeze(1)
if C.dim() == 2:
C = C.unsqueeze(1)
if D is not None and D.dim() == 1:
D = D.unsqueeze(0)
if z is not None and z.dim() == 2:
z = z.unsqueeze(1)
if dt_bias is not None and dt_bias.dim() == 1:
dt_bias = dt_bias.unsqueeze(0)
batch, nheads, dim, dstate = state.shape
assert x.shape == (batch, nheads, dim)
assert dt.shape == x.shape
assert A.shape == (nheads, dim, dstate)
ngroups = B.shape[1]
assert nheads % ngroups == 0, "nheads must be divisible by ngroups"
assert B.shape == (batch, ngroups, dstate)
assert C.shape == B.shape
if D is not None:
assert D.shape == (nheads, dim)
if z is not None:
assert z.shape == x.shape
if dt_bias is not None:
assert dt_bias.shape == (nheads, dim)
dt = dt + dt_bias
dt = F.softplus(dt) if dt_softplus else dt
dA = torch.exp(
rearrange(dt, "b h d -> b h d 1") * A
) # (batch, nheads, dim, dstate)
B = repeat(B, "b g n -> b (g h) n", h=nheads // ngroups) # (batch, nheads, dstate)
C = repeat(C, "b g n -> b (g h) n", h=nheads // ngroups) # (batch, nheads, dstate)
dB = rearrange(dt, "b h d -> b h d 1") * rearrange(
B, "b h n -> b h 1 n"
) # (batch, nheads, dim, dstate)
state.copy_(
state * dA + dB * rearrange(x, "b h d -> b h d 1")
) # (batch, dim, dstate
out = torch.einsum("bhdn,bhn->bhd", state.to(C.dtype), C)
if D is not None:
out += (x * D).to(out.dtype)
out = (out if z is None else out * F.silu(z)).to(x.dtype)
if not has_heads:
out = out.squeeze(1)
return out
|