File size: 12,124 Bytes
2976eec | 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 | # SPDX-License-Identifier: MIT
# Copyright (C) 2024-2026, Advanced Micro Devices, Inc. All rights reserved.
from typing import Optional
import torch
import triton
from ...utils._triton import arch_info as arch_info
from ...utils.logger import AiterTritonLogger
from ...utils.common_utils import serialize_dict, deserialize_str
from ..._triton_kernels.gemm.basic.gemm_a16wfp4 import (
_gemm_a16wfp4_kernel,
_gemm_a16wfp4_preshuffle_kernel,
_get_config,
)
from ..._triton_kernels.common.splitk_reduce import (
_gemm_splitk_reduce_kernel,
)
from ...gemm.basic.gemm_afp4wfp4 import (
get_splitk,
)
from ..._aiter_compat.torch_guard import torch_compile_guard
_LOGGER = AiterTritonLogger()
def gemm_a16wfp4_fake_tensor(
x: torch.Tensor,
w: torch.Tensor,
w_scales: torch.Tensor,
atomic_add: Optional[bool] = False,
dtype: Optional[torch.dtype] = torch.bfloat16,
y: Optional[torch.Tensor] = None,
config: Optional[str] = None,
) -> torch.Tensor:
if y is None:
M, _ = x.shape
N, _ = w.shape
return torch.zeros((M, N), dtype=dtype, device=x.device)
return y
@torch_compile_guard(gen_fake=gemm_a16wfp4_fake_tensor)
def gemm_a16wfp4_(
x: torch.Tensor,
w: torch.Tensor,
w_scales: torch.Tensor,
atomic_add: Optional[bool] = False,
dtype: Optional[torch.dtype] = torch.bfloat16,
y: Optional[torch.Tensor] = None,
config: Optional[str] = None,
) -> torch.Tensor:
"""
Computes matrix multiplication Y = X @ W^T with BF16 activations and FP4 weights.
Key parameters:
x (torch.Tensor): BF16/FP16 input matrix X with shape (M, K).
Quantized to MXFP4 on-the-fly during GEMM.
w (torch.Tensor): FP4 E2M1 weight matrix W with shape (N, K//2).
w_scales (torch.Tensor): E8M0 per-group scale for w with shape (N, K//32).
One scale per 32 elements in K dimension.
atomic_add (Optional[bool]): use atomic_add for reduction
dtype (Optional[torch.dtype]): Output datatype (BF16 or FP16).
y (Optional[torch.Tensor]): Pre-allocated output tensor with shape (M, N).
config (Optional[str]): Kernel tuning parameters (BLOCK_SIZE_M, BLOCK_SIZE_N,
BLOCK_SIZE_K, GROUP_SIZE_M, NUM_KSPLIT, SPLITK_BLOCK_SIZE).
Returns:
y (torch.Tensor): Output with shape (M, N).
"""
_LOGGER.info(
f"GEMM_A16WFP4: x={tuple(x.shape)} w={tuple(w.shape)} w_scale={tuple(w_scales.shape)} "
)
assert arch_info.is_fp4_avail(), "MXFP4 is not available on your device"
M, K = x.shape
N, K = w.shape
# inner kernel expects (K, N)
w = w.T
if config is None:
config, _ = _get_config(M, N, K)
else:
config = deserialize_str(config)
if y is None:
if atomic_add:
y = torch.zeros((M, N), dtype=dtype, device=x.device)
else:
y = torch.empty((M, N), dtype=dtype, device=x.device)
if config["NUM_KSPLIT"] > 1 and not atomic_add:
SPLITK_BLOCK_SIZE, BLOCK_SIZE_K, NUM_KSPLIT = get_splitk(
K, config["BLOCK_SIZE_K"], config["NUM_KSPLIT"]
)
config["SPLITK_BLOCK_SIZE"] = SPLITK_BLOCK_SIZE
config["BLOCK_SIZE_K"] = BLOCK_SIZE_K
config["NUM_KSPLIT"] = NUM_KSPLIT
if config["BLOCK_SIZE_K"] >= 2 * K:
config["BLOCK_SIZE_K"] = triton.next_power_of_2(2 * K)
config["SPLITK_BLOCK_SIZE"] = 2 * K
config["NUM_KSPLIT"] = 1
config["BLOCK_SIZE_K"] = max(config["BLOCK_SIZE_K"], 64)
if config["NUM_KSPLIT"] > 1 and not atomic_add:
y_pp = torch.empty(
(config["NUM_KSPLIT"], M, N), dtype=torch.float32, device=y.device
)
else:
config["SPLITK_BLOCK_SIZE"] = 2 * K
y_pp = None
grid = lambda META: ( # noqa: E731
(
META["NUM_KSPLIT"]
* triton.cdiv(M, META["BLOCK_SIZE_M"])
* triton.cdiv(N, META["BLOCK_SIZE_N"])
),
)
_gemm_a16wfp4_kernel[grid](
x,
w,
y if y_pp is None else y_pp,
w_scales,
M,
N,
K,
x.stride(0),
x.stride(1),
w.stride(0),
w.stride(1),
0 if y_pp is None else y_pp.stride(0),
y.stride(0) if y_pp is None else y_pp.stride(1),
y.stride(1) if y_pp is None else y_pp.stride(2),
w_scales.stride(0),
w_scales.stride(1),
ATOMIC_ADD=atomic_add,
**config,
)
if config["NUM_KSPLIT"] > 1 and not atomic_add:
REDUCE_BLOCK_SIZE_M = 16
REDUCE_BLOCK_SIZE_N = 64
# TODO: Need to debug - REDUCE_BLOCK_SIZE_N=128 with fp32 partials fails
# NOTE: REDUCE_BLOCK_SIZE_N=16 gives best perf with fp32 partials and
# REDUCE_BLOCK_SIZE_N=128 gives best perf with bf16 partials
ACTUAL_KSPLIT = triton.cdiv(K, (config["SPLITK_BLOCK_SIZE"] // 2))
grid_reduce = (
triton.cdiv(M, REDUCE_BLOCK_SIZE_M),
triton.cdiv(N, REDUCE_BLOCK_SIZE_N),
)
_gemm_splitk_reduce_kernel[grid_reduce](
y_pp,
y,
None,
M,
N,
y_pp.stride(0),
y_pp.stride(1),
y_pp.stride(2),
y.stride(0),
y.stride(1),
REDUCE_BLOCK_SIZE_M,
REDUCE_BLOCK_SIZE_N,
ACTUAL_KSPLIT,
triton.next_power_of_2(config["NUM_KSPLIT"]),
ADD_BIAS=False,
activation="",
use_activation=False,
KERNEL_NAME="_gemm_afp4wfp4_reduce_kernel",
)
return y
def gemm_a16wfp4(
x: torch.Tensor,
w: torch.Tensor,
w_scales: torch.Tensor,
atomic_add: Optional[bool] = False,
dtype: Optional[torch.dtype] = torch.bfloat16,
y: Optional[torch.Tensor] = None,
config: Optional[dict] = None,
) -> torch.Tensor:
config_hashable = serialize_dict(config) if config else None
return gemm_a16wfp4_(x, w, w_scales, atomic_add, dtype, y, config_hashable)
def gemm_a16wfp4_preshuffle_fake_tensor(
x: torch.Tensor,
w: torch.Tensor,
w_scales: torch.Tensor,
dtype: Optional[torch.dtype] = torch.bfloat16,
y: Optional[torch.Tensor] = None,
config: Optional[str] = None,
skip_reduce: Optional[bool] = False,
) -> torch.Tensor:
M, K = x.shape
N, _ = w.shape
config = deserialize_str(config)
num_ksplit = config["NUM_KSPLIT"]
block_size_k = config["BLOCK_SIZE_K"]
if num_ksplit > 1:
_, block_size_k, num_ksplit = get_splitk(K, block_size_k, num_ksplit)
if block_size_k >= 2 * K:
num_ksplit = 1
if num_ksplit > 1 and skip_reduce:
y_pp = torch.empty((num_ksplit, M, N), dtype=torch.float32, device=x.device)
return y_pp
return torch.empty((M, N), dtype=dtype, device=x.device)
@torch_compile_guard(gen_fake=gemm_a16wfp4_preshuffle_fake_tensor)
def gemm_a16wfp4_preshuffle_(
x: torch.Tensor,
w: torch.Tensor,
w_scales: torch.Tensor,
prequant: Optional[bool] = True,
dtype: Optional[torch.dtype] = torch.bfloat16,
y: Optional[torch.Tensor] = None,
config: Optional[str] = None,
skip_reduce: Optional[bool] = False,
) -> torch.Tensor:
"""
Computes matrix multiplication Y = X @ W^T with BF16 activations and FP4 weights.
Key parameters:
x (torch.Tensor): BF16/FP16 input matrix X with shape (M, K).
Quantized to MXFP4 on-the-fly during GEMM.
w (torch.Tensor): FP4 E2M1 weight matrix W with shape (N, K//2).
w_scales (torch.Tensor): E8M0 per-group scale for w with shape (M//32, K).
One scale per 32 elements in K dimension.
dtype (Optional[torch.dtype]): Output datatype (BF16 or FP16).
y (Optional[torch.Tensor]): Pre-allocated output tensor with shape (M, N).
config (Optional[str]): Kernel tuning parameters (BLOCK_SIZE_M, BLOCK_SIZE_N,
BLOCK_SIZE_K, GROUP_SIZE_M, NUM_KSPLIT, SPLITK_BLOCK_SIZE).
skip_reduce (Optional[bool]): skip reduction, y becomes (SPK, M, N) where SPK is determined by config
Returns:
y (torch.Tensor): Output with shape (M, N).
"""
_LOGGER.info(
f"GEMM_A16WFP4_PRESHUFFLE: x={tuple(x.shape)} w={tuple(w.shape)} w_scale={tuple(w_scales.shape)} "
)
assert arch_info.is_fp4_avail(), "MXFP4 is not available on your device"
assert prequant, "prequant == False is not supported yet"
M, K = x.shape
N, K = w.shape
N = N * 16
K = K // 16
if config is None:
config, _ = _get_config(M, N, K, True)
else:
config = deserialize_str(config)
if config["NUM_KSPLIT"] > 1:
SPLITK_BLOCK_SIZE, BLOCK_SIZE_K, NUM_KSPLIT = get_splitk(
K, config["BLOCK_SIZE_K"], config["NUM_KSPLIT"]
)
config["SPLITK_BLOCK_SIZE"] = SPLITK_BLOCK_SIZE
config["BLOCK_SIZE_K"] = BLOCK_SIZE_K
config["NUM_KSPLIT"] = NUM_KSPLIT
if config["BLOCK_SIZE_K"] >= 2 * K:
config["BLOCK_SIZE_K"] = triton.next_power_of_2(2 * K)
config["SPLITK_BLOCK_SIZE"] = 2 * K
config["NUM_KSPLIT"] = 1
config["BLOCK_SIZE_N"] = max(config["BLOCK_SIZE_N"], 32)
return_y_pp = config["NUM_KSPLIT"] > 1 and skip_reduce
if config["NUM_KSPLIT"] > 1:
y_pp = torch.empty(
(config["NUM_KSPLIT"], M, N), dtype=torch.float32, device=x.device
)
else:
config["SPLITK_BLOCK_SIZE"] = 2 * K
y_pp = None
if y is None and not return_y_pp:
y = torch.empty((M, N), dtype=dtype, device=x.device)
grid = lambda META: ( # noqa: E731
(
META["NUM_KSPLIT"]
* triton.cdiv(M, META["BLOCK_SIZE_M"])
* triton.cdiv(N, META["BLOCK_SIZE_N"])
),
)
_gemm_a16wfp4_preshuffle_kernel[grid](
x,
w,
y if y_pp is None else y_pp,
w_scales,
M,
N,
K,
x.stride(0),
x.stride(1),
w.stride(0),
w.stride(1),
0 if y_pp is None else y_pp.stride(0),
y.stride(0) if y_pp is None else y_pp.stride(1),
y.stride(1) if y_pp is None else y_pp.stride(2),
w_scales.stride(0),
w_scales.stride(1),
PREQUANT=prequant,
**config,
)
if return_y_pp:
return y_pp
elif config["NUM_KSPLIT"] > 1:
REDUCE_BLOCK_SIZE_M = 16
REDUCE_BLOCK_SIZE_N = 64
# TODO: Need to debug - REDUCE_BLOCK_SIZE_N=128 with fp32 partials fails
# NOTE: REDUCE_BLOCK_SIZE_N=16 gives best perf with fp32 partials and
# REDUCE_BLOCK_SIZE_N=128 gives best perf with bf16 partials
ACTUAL_KSPLIT = triton.cdiv(K, (config["SPLITK_BLOCK_SIZE"] // 2))
grid_reduce = (
triton.cdiv(M, REDUCE_BLOCK_SIZE_M),
triton.cdiv(N, REDUCE_BLOCK_SIZE_N),
)
_gemm_splitk_reduce_kernel[grid_reduce](
y_pp,
y,
None,
M,
N,
y_pp.stride(0),
y_pp.stride(1),
y_pp.stride(2),
y.stride(0),
y.stride(1),
REDUCE_BLOCK_SIZE_M,
REDUCE_BLOCK_SIZE_N,
ACTUAL_KSPLIT,
triton.next_power_of_2(config["NUM_KSPLIT"]),
ADD_BIAS=False,
activation="",
use_activation=False,
KERNEL_NAME="_gemm_afp4wfp4_reduce_kernel",
)
return y
def gemm_a16wfp4_preshuffle(
x: torch.Tensor,
w: torch.Tensor,
w_scales: torch.Tensor,
prequant: Optional[bool] = True,
dtype: Optional[torch.dtype] = torch.bfloat16,
y: Optional[torch.Tensor] = None,
config: Optional[dict] = None,
skip_reduce: Optional[bool] = False,
) -> torch.Tensor:
if config is None:
config_hashable = None
M, _ = x.shape
N, K = w.shape
N = N * 16
K = K // 16
config, _ = _get_config(M, N, K, True)
config_hashable = serialize_dict(config)
return gemm_a16wfp4_preshuffle_(
x, w, w_scales, prequant, dtype, y, config_hashable, skip_reduce
)
|