File size: 10,668 Bytes
8e9f35a | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 | """Public Sol-Attn interface."""
from __future__ import annotations
import functools
import torch
BLOCK_SIZE = 64
_CUTE_BACKENDS = {
(9, 0): "cute_sm90",
(10, 0): "cute_sm100",
(12, 0): "cute_sm120",
}
_compiled = {}
def _validate_inputs(
q,
k,
v,
thresh_type,
sink_tokens=0,
sink_start=None,
):
if q.ndim != 4 or q.shape != k.shape or q.shape != v.shape:
raise ValueError("q, k, and v must share shape [B, T, H, 128]")
if q.shape[1] == 0 or q.shape[3] != 128:
raise ValueError("Sol-Attn requires T > 0 and head dimension 128")
if any(x.dtype != torch.bfloat16 for x in (q, k, v)):
raise TypeError("q, k, and v must use torch.bfloat16")
if q.device.type != "cuda" or k.device != q.device or v.device != q.device:
raise ValueError("q, k, and v must be on the same CUDA device")
if not (q.is_contiguous() and k.is_contiguous() and v.is_contiguous()):
raise ValueError("q, k, and v must be contiguous BTHD tensors")
if thresh_type not in ("diag", "exact"):
raise ValueError("thresh_type must be 'diag' or 'exact'")
if not isinstance(sink_tokens, int):
raise TypeError("sink_tokens must be an integer")
if not 0 <= sink_tokens <= q.shape[1]:
raise ValueError("sink_tokens must be in [0, T]")
if sink_start is not None:
if not isinstance(sink_start, int):
raise TypeError("sink_start must be an integer or None")
if not 0 <= sink_start <= q.shape[1]:
raise ValueError("sink_start must be in [0, T]")
if sink_start + sink_tokens > q.shape[1]:
raise ValueError("sink_start + sink_tokens must be <= T")
return tuple(torch.cuda.get_device_capability(q.device))
@functools.lru_cache(maxsize=1)
def _cute_runtime_available() -> bool:
"""Whether the optional CuTe DSL runtime can be imported."""
try:
import cuda.bindings.driver # noqa: F401
import cutlass.cute # noqa: F401
except ImportError:
return False
return True
def _backend_for_arch(
arch: tuple[int, int],
*,
cute_available: bool | None = None,
) -> str:
"""Select CuTe when specialized and available, otherwise Triton."""
if arch[0] < 8:
raise RuntimeError(
"Sol-Attn requires an NVIDIA GPU with compute capability >= 8.0; "
f"got SM{arch[0]}{arch[1]}"
)
cute_backend = _CUTE_BACKENDS.get(arch)
if cute_backend is not None:
available = (
_cute_runtime_available()
if cute_available is None
else cute_available
)
if available:
return cute_backend
return "triton"
def _validate_cute(arch, tokens, kv_splits):
if arch != (9, 0) and kv_splits != 1:
raise ValueError("kv_splits=2/4 is currently available on SM90 only")
route_groups = ((tokens + 63) // 64 + 63) // 64
if kv_splits > route_groups:
raise ValueError("each KV split must contain at least one N64 route group")
def _stream(device):
import cuda.bindings.driver as cuda
return cuda.CUstream(torch.cuda.current_stream(device).cuda_stream)
def _to_cute_tensors(tensors):
from .common import to_cute_tensor
return [to_cute_tensor(x) for x in tensors]
def _sink_block_range(tokens, sink_start, sink_tokens):
blocks = (tokens + BLOCK_SIZE - 1) // BLOCK_SIZE
if not sink_tokens:
return blocks, blocks
start = tokens - sink_tokens if sink_start is None else sink_start
return (
start // BLOCK_SIZE,
(start + sink_tokens + BLOCK_SIZE - 1) // BLOCK_SIZE,
)
def _compile_sm90(
key,
tensors,
scale,
tokens,
kv_splits,
sink_range,
stream,
):
import cutlass.cute as cute
from .sm90 import make_kernel
operator = make_kernel(tokens, kv_splits)
args = _to_cute_tensors(tensors)
compiled = cute.compile(
operator,
*args,
scale,
sink_range,
stream=stream,
options="--enable-tvm-ffi",
)
_compiled[key] = compiled
return compiled, args
def _compile_sm100(
key,
tensors,
scale,
sink_start_block,
sink_end_block,
stream,
):
import cutlass.cute as cute
from .sm100 import forward
args = _to_cute_tensors(tensors)
compiled = cute.compile(
forward,
*args,
scale,
sink_start_block,
sink_end_block,
stream=stream,
options="--enable-tvm-ffi",
)
_compiled[key] = compiled
return compiled, args
def _compile_sm120(
key,
tensors,
scale,
sink_start_block,
sink_end_block,
stream,
):
import cutlass.cute as cute
from .sm120 import make_kernel
operator = make_kernel()
args = _to_cute_tensors(tensors)
compiled = cute.compile(
operator,
*args,
scale,
sink_start_block,
sink_end_block,
stream=stream,
options="--enable-tvm-ffi",
)
_compiled[key] = compiled
return compiled, args
def _sol_attn_cute(
q,
k,
v,
*,
arch,
scale,
tau,
thresh_type,
kv_splits,
sink_tokens,
sink_start,
):
from .preprocess import prepare
batch, tokens, heads, _ = q.shape
with torch.cuda.device(q.device):
kc, vc, threshold = prepare(
q,
k,
v,
scale=scale,
tau=tau,
thresh_type=thresh_type,
)
output = torch.empty_like(v)
lse = torch.empty(
(batch, tokens, heads),
device=q.device,
dtype=torch.float32,
)
stream = _stream(q.device)
key = (q.device.index, arch, batch, tokens, heads, kv_splits)
if arch == (9, 0):
if sink_tokens:
sink_start_block, sink_end_block = _sink_block_range(
tokens,
sink_start,
sink_tokens,
)
sink_range = sink_start_block | (sink_end_block << 16)
else:
sink_range = 0
tensors = [q, k, v, output, kc, vc, threshold, lse]
if kv_splits > 1:
tensors.extend(
[
torch.empty(
(batch, tokens, kv_splits * heads, 128),
device=q.device,
dtype=torch.bfloat16,
),
torch.empty(
(batch, tokens, kv_splits * heads),
device=q.device,
dtype=torch.float32,
),
]
)
compiled = _compiled.get(key)
if compiled is None:
compiled, args = _compile_sm90(
key,
tensors,
scale,
tokens,
kv_splits,
sink_range,
stream,
)
else:
args = _to_cute_tensors(tensors)
compiled(
*args,
scale,
sink_range,
stream=stream,
)
elif arch == (10, 0):
sink_start_block, sink_end_block = _sink_block_range(
tokens,
sink_start,
sink_tokens,
)
tensors = [q, k, v, output, kc, vc, threshold, lse]
compiled = _compiled.get(key)
if compiled is None:
compiled, args = _compile_sm100(
key,
tensors,
scale,
sink_start_block,
sink_end_block,
stream,
)
else:
args = _to_cute_tensors(tensors)
compiled(
*args,
scale,
sink_start_block,
sink_end_block,
stream=stream,
)
else:
sink_start_block, sink_end_block = _sink_block_range(
tokens,
sink_start,
sink_tokens,
)
tensors = [q, k, v, output, kc, vc, threshold, lse]
compiled = _compiled.get(key)
if compiled is None:
compiled, args = _compile_sm120(
key,
tensors,
scale,
sink_start_block,
sink_end_block,
stream,
)
else:
args = _to_cute_tensors(tensors)
compiled(
*args,
scale,
sink_start_block,
sink_end_block,
stream=stream,
)
return output
def sol_attn(
q: torch.Tensor,
k: torch.Tensor,
v: torch.Tensor,
*,
scale: float | None = None,
tau: float = 1.0,
thresh_type: str = "diag",
kv_splits: int = 1,
sink_tokens: int = 0,
sink_start: int | None = None,
) -> torch.Tensor:
"""Compute noncausal Sol-Attn for contiguous BF16 BTHD tensors.
``sink_start`` and ``sink_tokens`` keep every KV block overlapping the
corresponding contiguous token range exact for all queries. Omitting
``sink_start`` places the range at the token suffix.
"""
arch = _validate_inputs(
q,
k,
v,
thresh_type,
sink_tokens,
sink_start,
)
if kv_splits not in (1, 2, 4):
raise ValueError("kv_splits must be 1, 2, or 4")
backend = _backend_for_arch(arch)
scale = q.shape[-1] ** -0.5 if scale is None else float(scale)
tau = float(tau)
if backend == "triton":
if kv_splits != 1:
raise ValueError("kv_splits=2/4 is currently available on SM90 only")
from .triton_ref import sol_attn as triton_sol_attn
return triton_sol_attn(
q,
k,
v,
scale=scale,
tau=tau,
thresh_type=thresh_type,
sink_tokens=sink_tokens,
sink_start=sink_start,
)
_validate_cute(arch, q.shape[1], kv_splits)
return _sol_attn_cute(
q,
k,
v,
arch=arch,
scale=scale,
tau=tau,
thresh_type=thresh_type,
kv_splits=kv_splits,
sink_tokens=sink_tokens,
sink_start=sink_start,
)
__all__ = ["sol_attn"]
|