| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | import argparse |
| 5 | import copy |
| 6 | import json |
| 7 | import pickle |
| 8 | import time |
| 9 | from collections.abc import Callable |
| 10 | from dataclasses import dataclass |
| 11 | from enum import Enum, auto |
| 12 | from itertools import product |
| 13 | from pathlib import Path |
| 14 | from typing import Any |
| 15 | |
| 16 | import torch |
| 17 | import torch.utils.benchmark as TBenchmark |
| 18 | from torch.utils.benchmark import Measurement as TMeasurement |
| 19 | from utils import ArgPool, Bench, CudaGraphBenchParams |
| 20 | from weight_shapes import WEIGHT_SHAPES |
| 21 | |
| 22 | from vllm.lora.ops.triton_ops.utils import get_lora_op_configs |
| 23 | from vllm.triton_utils import HAS_TRITON, triton |
| 24 | |
| 25 | if HAS_TRITON: |
| 26 | from vllm.lora.ops.triton_ops import ( ## added fused_moe_lora |
| 27 | LoRAKernelMeta, |
| 28 | fused_moe_lora_expand, |
| 29 | fused_moe_lora_shrink, |
| 30 | lora_expand, |
| 31 | lora_shrink, |
| 32 | ) |
| 33 | from vllm.lora.ops.triton_ops.fused_moe_lora_op import ( |
| 34 | _LORA_PTR_DICT, ## added _LORA_PTR_DICT for fused_moe_lora |
| 35 | ) |
| 36 | from vllm.lora.ops.triton_ops.utils import _LORA_A_PTR_DICT, _LORA_B_PTR_DICT |
| 37 | from vllm import _custom_ops as ops |
| 38 | from vllm.utils.argparse_utils import FlexibleArgumentParser |
| 39 | from vllm.utils.math_utils import round_up |
| 40 | |
| 41 | DEFAULT_MODELS = list(WEIGHT_SHAPES.keys()) |
| 42 | DEFAULT_TP_SIZES = [1] |
| 43 | DEFAULT_BATCH_SIZES = [ |
| 44 | 1, |
| 45 | 16, |
| 46 | 32, |
| 47 | 64, |
| 48 | 128, |
| 49 | 192, |
| 50 | 256, |
| 51 | 320, |
| 52 | 384, |
| 53 | 448, |
| 54 | 512, |
| 55 | 640, |
| 56 | 768, |
| 57 | 896, |
| 58 | 1024, |
| 59 | 2048, |
| 60 | 3072, |
| 61 | 4096, |
| 62 | 5120, |
| 63 | 6144, |
| 64 | 7168, |
| 65 | 8192, |
| 66 | ] |
| 67 | DEFAULT_HIDDEN_SIZES = [1024, 2048, 4096, 8192, 16384] |
| 68 | DEFAULT_LORA_RANKS = [16] |
| 69 | DEFAULT_NUM_LORAS = [1, 2, 3, 4] |
| 70 | DEFAULT_SORT_BY_LORA_IDS = [False, True] |
| 71 | DEFAULT_SEQ_LENGTHS = [1] |
| 72 | DEFAULT_EXPAND_FN_ADD_INPUTS = [True, False] |
| 73 | DEFAULT_TOP_K_NUMS = [1] # Added for MoE LoRA top_k |
| 74 | DEFAULT_NUM_EXPERTS = [8] # Added for MoE LoRA num_experts |
| 75 | |
| 76 | |
| 77 | # Utilities |
| 78 | def dtype_to_str(dtype: torch.dtype): |
| 79 | if dtype == torch.float16: |
| 80 | return "f16" |
| 81 | if dtype == torch.bfloat16: |
| 82 | return "bf16" |
| 83 | if dtype == torch.float32: |
| 84 | return "f32" |
| 85 | raise ValueError(f"Unsupported dtype {dtype}") |
| 86 | |
| 87 | |
| 88 | def make_rand_lora_weight_tensor( |
| 89 | k: int, n: int, num_loras: int, dtype: torch.dtype, device: str = "cuda" |
| 90 | ) -> torch.Tensor: |
| 91 | # LoRA weights column major |
| 92 | return torch.rand((num_loras, n, k), dtype=dtype).to(device) |
| 93 | |
| 94 | |
| 95 | def make_rand_tensors( |
| 96 | a_shape: tuple[int, ...], |
| 97 | b_shape: tuple[int, ...], |
| 98 | c_shape: tuple[int, ...], |
| 99 | a_dtype: torch.dtype, |
| 100 | b_dtype: torch.dtype, |
| 101 | c_dtype: torch.dtype, |
| 102 | num_slices: int, |
| 103 | device: str = "cuda", |
| 104 | ) -> tuple[torch.Tensor, list[torch.Tensor], torch.Tensor]: |
| 105 | """Make LoRA input/output matrices.""" |
| 106 | A = torch.rand(a_shape, dtype=a_dtype).to(device) |
| 107 | |
| 108 | # LoRA weights column major |
| 109 | Bs = [torch.rand(b_shape, dtype=b_dtype).to(device) for _ in range(num_slices)] |
| 110 | |
| 111 | C = torch.zeros(c_shape, dtype=c_dtype).to(device) |
| 112 | return A, Bs, C |
| 113 | |
| 114 | |
| 115 | def make_prompt_lora_mapping( |
| 116 | num_prompts: int, num_active_loras: int, sort_by_lora_id: bool, device: str |
| 117 | ) -> torch.Tensor: |
| 118 | """All prompts are mapped to a LoRA ID in range [0, num_active_loras). |
| 119 | where 0 refers to first lora, 1 refers to second lora and so on. |
| 120 | """ |
| 121 | assert num_active_loras > 0 |
| 122 | |
| 123 | if not sort_by_lora_id: |
| 124 | return torch.randint(0, num_active_loras, (num_prompts,), dtype=torch.long) |
| 125 | |
| 126 | # Divide LoRAs equally and in order. |
| 127 | part_size = num_prompts // num_active_loras |
| 128 | part_size = max(part_size, 1) |
| 129 | |
| 130 | lora_id = 0 |
| 131 | prompt_lora_mapping = [] |
| 132 | while len(prompt_lora_mapping) < num_prompts: |
| 133 | prompt_lora_mapping.extend([lora_id] * part_size) |
| 134 | lora_id = lora_id + 1 if lora_id + 1 < num_active_loras else lora_id |
| 135 | return torch.tensor( |
| 136 | prompt_lora_mapping[:num_prompts], dtype=torch.long, device=device |
| 137 | ) |
| 138 | |
| 139 | |
| 140 | def make_token_lora_mapping( |
| 141 | num_tokens: int, |
| 142 | num_prompts: int, |
| 143 | prompt_lora_mapping: torch.Tensor, |
| 144 | seq_len_tensor: torch.Tensor, |
| 145 | device: str, |
| 146 | ): |
| 147 | """Make token_lora_mapping from prompt_lora_mapping and seq_lens_tensor.""" |
| 148 | assert prompt_lora_mapping.shape[0] == num_prompts |
| 149 | |
| 150 | # token to lora index mapping |
| 151 | token_lora_mapping = [0] * num_tokens |
| 152 | current_offset = 0 |
| 153 | for b_id in range(num_prompts): |
| 154 | lora_index = prompt_lora_mapping[b_id].item() |
| 155 | s = current_offset |
| 156 | e = s + seq_len_tensor[b_id].item() |
| 157 | token_lora_mapping[s:e] = [lora_index] * (e - s) |
| 158 | current_offset += seq_len_tensor[b_id].item() |
| 159 | |
| 160 | return torch.tensor(token_lora_mapping, dtype=torch.long, device=device) |
| 161 | |
| 162 | |
| 163 | def ref_group_gemm( |
| 164 | ref_out: torch.Tensor, |
| 165 | input: torch.Tensor, |
| 166 | lora_weights: list[torch.Tensor], |
| 167 | seq_lens_cpu: torch.Tensor, |
| 168 | prompt_lora_mapping_cpu: torch.Tensor, |
| 169 | scaling: float, |
| 170 | add_inputs: bool | None, |
| 171 | ): |
| 172 | """Torch group gemm reference implementation to test correctness of |
| 173 | benchmarking operations. |
| 174 | """ |
| 175 | batches = seq_lens_cpu.size(0) |
| 176 | out_list = [] |
| 177 | current_offset = 0 |
| 178 | for lora_index, b_length in zip(range(batches), seq_lens_cpu): |
| 179 | x = input[current_offset : b_length + current_offset, :] |
| 180 | current_offset += b_length |
| 181 | w = lora_weights[prompt_lora_mapping_cpu[lora_index]] |
| 182 | result = torch.nn.functional.linear(x, w) |
| 183 | result *= scaling |
| 184 | out_list.append(result) |
| 185 | |
| 186 | cat_result = torch.cat(out_list, dim=0) |
| 187 | |
| 188 | if add_inputs: |
| 189 | ref_out += cat_result |
| 190 | else: |
| 191 | ref_out.copy_(cat_result) |
| 192 | |
| 193 | |
| 194 | class OpType(Enum): |
| 195 | """LoRA Ops to benchmark and its properties.""" |
| 196 | |
| 197 | LORA_SHRINK = auto() |
| 198 | LORA_EXPAND = auto() |
| 199 | ## Adding support for fused moe lora |
| 200 | FUSED_MOE_LORA_GATE_UP_SHRINK = auto() ## Gate/Up projection variant with shrink |
| 201 | FUSED_MOE_LORA_GATE_UP_EXPAND = auto() ## Gate/Up projection variant with expand |
| 202 | FUSED_MOE_LORA_DOWN_SHRINK = auto() ## Down projection variant with shrink |
| 203 | FUSED_MOE_LORA_DOWN_EXPAND = auto() ## Down projection variant with expand |
| 204 | |
| 205 | @staticmethod |
| 206 | def from_str(s: str) -> "OpType": |
| 207 | if s.lower() == "lora_shrink": |
| 208 | return OpType.LORA_SHRINK |
| 209 | if s.lower() == "lora_expand": |
| 210 | return OpType.LORA_EXPAND |
| 211 | # Adding support for fused moe lora, both in gate_up and down |
| 212 | if s.lower() == "fused_moe_lora_gate_up_shrink": ## Gate/Up variant with shrink |
| 213 | return OpType.FUSED_MOE_LORA_GATE_UP_SHRINK |
| 214 | if s.lower() == "fused_moe_lora_gate_up_expand": ## Gate/Up variant with expand |
| 215 | return OpType.FUSED_MOE_LORA_GATE_UP_EXPAND |
| 216 | if s.lower() == "fused_moe_lora_down_shrink": ## Down variant with shrink |
| 217 | return OpType.FUSED_MOE_LORA_DOWN_SHRINK |
| 218 | if s.lower() == "fused_moe_lora_down_expand": ## Down variant with expand |
| 219 | return OpType.FUSED_MOE_LORA_DOWN_EXPAND |
| 220 | raise ValueError(f"Unrecognized str {s} to convert to OpType") |
| 221 | |
| 222 | def is_shrink_fn(self) -> bool: |
| 223 | return self in [OpType.LORA_SHRINK] |
| 224 | |
| 225 | def is_expand_fn(self) -> bool: |
| 226 | return self in [OpType.LORA_EXPAND] |
| 227 | |
| 228 | def is_fused_moe_lora_fn(self) -> bool: ## adding for fused MoE LoRA |
| 229 | return self in [ |
| 230 | OpType.FUSED_MOE_LORA_GATE_UP_SHRINK, |
| 231 | OpType.FUSED_MOE_LORA_DOWN_SHRINK, |
| 232 | OpType.FUSED_MOE_LORA_GATE_UP_EXPAND, |
| 233 | OpType.FUSED_MOE_LORA_DOWN_EXPAND, |
| 234 | ] |
| 235 | |
| 236 | def is_fused_moe_lora_gate_up_fn( |
| 237 | self, |
| 238 | ) -> bool: ## adding for fused MoE LoRA Gate/Up |
| 239 | return self in [ |
| 240 | OpType.FUSED_MOE_LORA_GATE_UP_SHRINK, |
| 241 | OpType.FUSED_MOE_LORA_GATE_UP_EXPAND, |
| 242 | ] |
| 243 | |
| 244 | def is_fused_moe_lora_down_fn(self) -> bool: ## adding for fused MoE LoRA Down |
| 245 | return self in [ |
| 246 | OpType.FUSED_MOE_LORA_DOWN_SHRINK, |
| 247 | OpType.FUSED_MOE_LORA_DOWN_EXPAND, |
| 248 | ] |
| 249 | |
| 250 | def is_fused_moe_lora_shrink_fn(self) -> bool: |
| 251 | return self in [ |
| 252 | OpType.FUSED_MOE_LORA_GATE_UP_SHRINK, |
| 253 | OpType.FUSED_MOE_LORA_DOWN_SHRINK, |
| 254 | ] |
| 255 | |
| 256 | def is_fused_moe_lora_expand_fn(self) -> bool: |
| 257 | return self in [ |
| 258 | OpType.FUSED_MOE_LORA_GATE_UP_EXPAND, |
| 259 | OpType.FUSED_MOE_LORA_DOWN_EXPAND, |
| 260 | ] |
| 261 | |
| 262 | def num_slices(self) -> list[int]: |
| 263 | if self.is_fused_moe_lora_gate_up_fn(): |
| 264 | return [2] |
| 265 | elif self.is_fused_moe_lora_down_fn(): |
| 266 | return [1] |
| 267 | return [1, 2, 3] |
| 268 | |
| 269 | def mkn( |
| 270 | self, batch_size: int, seq_length: int, hidden_size: int, lora_rank: int |
| 271 | ) -> tuple[int, int, int]: |
| 272 | num_tokens = batch_size * seq_length |
| 273 | if self.is_shrink_fn() or self.is_fused_moe_lora_fn(): |
| 274 | m = num_tokens |
| 275 | k = hidden_size |
| 276 | n = lora_rank |
| 277 | elif self.is_expand_fn(): |
| 278 | m = num_tokens |
| 279 | k = lora_rank |
| 280 | n = hidden_size |
| 281 | return m, k, n |
| 282 | |
| 283 | def matmul_dtypes( |
| 284 | self, op_dtype: torch.dtype |
| 285 | ) -> tuple[torch.dtype, torch.dtype, torch.dtype]: |
| 286 | """Return a type, b type and c type for A x B = C.""" |
| 287 | if self.is_shrink_fn(): |
| 288 | return op_dtype, op_dtype, torch.float32 |
| 289 | elif self.is_expand_fn(): |
| 290 | return torch.float32, op_dtype, op_dtype |
| 291 | else: |
| 292 | assert self.is_fused_moe_lora_fn() |
| 293 | return op_dtype, op_dtype, op_dtype |
| 294 | |
| 295 | def matmul_shapes_fused_moe_lora( |
| 296 | self, |
| 297 | m: int, |
| 298 | n: int, |
| 299 | k: int, |
| 300 | num_loras: int, |
| 301 | num_slices: int, |
| 302 | top_k_num: int, |
| 303 | num_experts: int, |
| 304 | ) -> tuple[tuple[int], tuple[int], tuple[int], tuple[int]]: |
| 305 | if self.is_fused_moe_lora_shrink_fn(): |
| 306 | input_shape = ( |
| 307 | (m * top_k_num, n) |
| 308 | if self in [OpType.FUSED_MOE_LORA_DOWN_SHRINK] |
| 309 | else (m, n) |
| 310 | ) |
| 311 | output_shape = (num_slices, m, top_k_num, k) |
| 312 | weight_shape = (num_loras, num_experts, k, n) |
| 313 | else: |
| 314 | assert self.is_fused_moe_lora_expand_fn() |
| 315 | input_shape = (num_slices, m, top_k_num, k) |
| 316 | output_shape = (m, top_k_num, n * num_slices) |
| 317 | weight_shape = (num_loras, num_experts, n, k) |
| 318 | return (input_shape, weight_shape, output_shape) |
| 319 | |
| 320 | def matmul_shapes( |
| 321 | self, |
| 322 | batch_size: int, |
| 323 | seq_length: int, |
| 324 | hidden_size: int, |
| 325 | lora_rank: int, |
| 326 | num_loras: int, |
| 327 | num_slices: int, |
| 328 | top_k_num: int | None = None, |
| 329 | num_experts: int | None = None, |
| 330 | ) -> tuple[tuple[int, ...], tuple[int, ...], tuple[int, ...]]: |
| 331 | """Given num_slices, return the shapes of the A, B, and C matrices |
| 332 | in A x B = C, for the op_type |
| 333 | """ |
| 334 | m, k, n = self.mkn(batch_size, seq_length, hidden_size, lora_rank) |
| 335 | |
| 336 | b_shape = (num_loras, n, k) # col-major |
| 337 | if self in [OpType.LORA_SHRINK]: |
| 338 | # LoRA shrink kernels support num_slices inherently in the kernel. |
| 339 | return ((m, k), b_shape, (num_slices, m, n)) |
| 340 | if self in [OpType.LORA_EXPAND]: |
| 341 | # LoRA expand kernels support num_slices inherently in the kernel |
| 342 | return ((num_slices, m, k), b_shape, (m, n * num_slices)) |
| 343 | if self.is_fused_moe_lora_fn(): |
| 344 | return self.matmul_shapes_fused_moe_lora( |
| 345 | m, |
| 346 | k, |
| 347 | n, |
| 348 | num_loras, |
| 349 | num_slices, |
| 350 | top_k_num, |
| 351 | num_experts, |
| 352 | ) |
| 353 | raise ValueError(f"Unrecognized op_type {self}") |
| 354 | |
| 355 | def bench_fn(self) -> Callable: |
| 356 | if self == OpType.LORA_SHRINK: |
| 357 | return lora_shrink |
| 358 | if self == OpType.LORA_EXPAND: |
| 359 | return lora_expand |
| 360 | if self in [ |
| 361 | OpType.FUSED_MOE_LORA_GATE_UP_SHRINK, |
| 362 | OpType.FUSED_MOE_LORA_DOWN_SHRINK, |
| 363 | ]: |
| 364 | return fused_moe_lora_shrink |
| 365 | if self in [ |
| 366 | OpType.FUSED_MOE_LORA_GATE_UP_EXPAND, |
| 367 | OpType.FUSED_MOE_LORA_DOWN_EXPAND, |
| 368 | ]: |
| 369 | return fused_moe_lora_expand |
| 370 | |
| 371 | raise ValueError(f"Unrecognized optype {self}") |
| 372 | |
| 373 | def run_ref_group_gemm( |
| 374 | self, |
| 375 | output: torch.Tensor, |
| 376 | input: torch.Tensor, |
| 377 | lora_weights: list[torch.Tensor], |
| 378 | **kwargs, |
| 379 | ) -> Callable: |
| 380 | """Each benchmark operation expects the input, lora_weights and outputs |
| 381 | in a slightly different format. Refer to self.matmul_shapes(). |
| 382 | run_ref_group_gemm accounts for those differences in executing a |
| 383 | reference group gemm for correctness testing. |
| 384 | """ |
| 385 | w_dtype = lora_weights[0].dtype |
| 386 | num_slices = len(lora_weights) |
| 387 | if self in [OpType.LORA_SHRINK]: |
| 388 | for slice_idx in range(num_slices): |
| 389 | ref_group_gemm( |
| 390 | ref_out=output[slice_idx, :], |
| 391 | input=input, |
| 392 | lora_weights=lora_weights[slice_idx], |
| 393 | **kwargs, |
| 394 | ) |
| 395 | elif self in [OpType.LORA_EXPAND]: |
| 396 | hidden_size = lora_weights[0].shape[1] |
| 397 | for slice_idx in range(num_slices): |
| 398 | slice_offset = slice_idx * hidden_size |
| 399 | ref_group_gemm( |
| 400 | ref_out=output[:, slice_offset : slice_offset + hidden_size], |
| 401 | input=input[slice_idx].clone().to(dtype=w_dtype), |
| 402 | lora_weights=lora_weights[slice_idx], |
| 403 | **kwargs, |
| 404 | ) |
| 405 | else: |
| 406 | raise ValueError(f"Unrecognized optype {self}") |
| 407 | |
| 408 | |
| 409 | @dataclass |
| 410 | class BenchmarkContext: |
| 411 | """LoRA benchmark context.""" |
| 412 | |
| 413 | batch_size: int |
| 414 | hidden_size: int |
| 415 | num_loras: int |
| 416 | num_active_loras: int |
| 417 | lora_rank: int |
| 418 | sort_by_lora_id: bool |
| 419 | dtype: torch.dtype |
| 420 | seq_length: int | None = None |
| 421 | num_experts: int | None = None # num_experts for MoE based ops |
| 422 | top_k_num: int | None = None # top_k for MoE based ops |
| 423 | num_slices: int | None = None # num_slices for slice based ops |
| 424 | |
| 425 | def with_seq_length(self, seq_length: int) -> "BenchmarkContext": |
| 426 | ctx = copy.copy(self) |
| 427 | ctx.seq_length = seq_length |
| 428 | return ctx |
| 429 | |
| 430 | def with_num_slices(self, num_slices: int) -> "BenchmarkContext": |
| 431 | ctx = copy.copy(self) |
| 432 | ctx.num_slices = num_slices |
| 433 | return ctx |
| 434 | |
| 435 | def bench_label(self) -> str: |
| 436 | return f"lora-{self.dtype}" |
| 437 | |
| 438 | def bench_sublabel(self, op_type: OpType) -> str: |
| 439 | m, k, n = op_type.mkn( |
| 440 | self.batch_size, self.seq_length, self.hidden_size, self.lora_rank |
| 441 | ) |
| 442 | desc = { |
| 443 | "bs": self.batch_size, |
| 444 | "sl": self.seq_length, |
| 445 | "m": m, |
| 446 | "k": k, |
| 447 | "n": n, |
| 448 | "num_loras": self.num_loras, |
| 449 | "sort_by_lora": self.sort_by_lora_id, |
| 450 | "num_slices": self.num_slices, |
| 451 | } |
| 452 | return json.dumps(desc) |
| 453 | |
| 454 | |
| 455 | @dataclass |
| 456 | class BenchmarkTensors: |
| 457 | """Input/Output tensors used for benchmarks.""" |
| 458 | |
| 459 | # matmul tensors |
| 460 | input: torch.Tensor |
| 461 | lora_weights_lst: list[torch.Tensor] |
| 462 | output: torch.Tensor |
| 463 | # LoRA kernel metadata |
| 464 | lora_kernel_meta: LoRAKernelMeta |
| 465 | # Metadata tensors used in testing correctness |
| 466 | seq_lens: torch.Tensor |
| 467 | prompt_lora_mapping: torch.Tensor |
| 468 | |
| 469 | def io_types(self) -> str: |
| 470 | return ( |
| 471 | f"{dtype_to_str(self.input.dtype)}x" |
| 472 | f"{dtype_to_str(self.lora_weights_lst[0].dtype)}=>" |
| 473 | f"{dtype_to_str(self.output.dtype)}" |
| 474 | ) |
| 475 | |
| 476 | def get_num_tokens(self, size: int, top_k_num: int, op_type: OpType): |
| 477 | return ( |
| 478 | size * top_k_num if op_type in [OpType.FUSED_MOE_LORA_DOWN_SHRINK] else size |
| 479 | ) |
| 480 | |
| 481 | @staticmethod |
| 482 | def make( |
| 483 | ctx: BenchmarkContext, op_type: OpType, device: str = "cuda" |
| 484 | ) -> "BenchmarkTensors": |
| 485 | # Make input / output matmul tensors. |
| 486 | a_shape, b_shape, c_shape = op_type.matmul_shapes( |
| 487 | ctx.batch_size, |
| 488 | ctx.seq_length, |
| 489 | ctx.hidden_size, |
| 490 | ctx.lora_rank, |
| 491 | ctx.num_loras, |
| 492 | ctx.num_slices, |
| 493 | ctx.top_k_num, |
| 494 | ctx.num_experts, |
| 495 | ) |
| 496 | a_type, b_type, c_type = op_type.matmul_dtypes(ctx.dtype) |
| 497 | input_tensor, lora_weights, output_tensor = make_rand_tensors( |
| 498 | a_shape, b_shape, c_shape, a_type, b_type, c_type, num_slices=ctx.num_slices |
| 499 | ) |
| 500 | |
| 501 | # Make metadata tensors. |
| 502 | # Keep the metadata tensors in the CPU for further processing if needed. |
| 503 | # The tensors get moved to the GPU before benchmarking. |
| 504 | assert ctx.num_active_loras <= ctx.num_loras |
| 505 | total_tokens = ctx.batch_size * ctx.seq_length |
| 506 | |
| 507 | # Make metadata tensors involved in correctness testing. |
| 508 | # Prepare seq lens tensor |
| 509 | seq_len_tensor = torch.randint( |
| 510 | ctx.seq_length, ctx.seq_length + 1, (ctx.batch_size,) |
| 511 | ) |
| 512 | assert total_tokens == seq_len_tensor.sum() |
| 513 | # Prepare prompt lora indices tensor |
| 514 | prompt_lora_indices_tensor = make_prompt_lora_mapping( |
| 515 | ctx.batch_size, ctx.num_active_loras, ctx.sort_by_lora_id, "cpu" |
| 516 | ) |
| 517 | |
| 518 | # Make LoRAKernelMeta |
| 519 | token_lora_indices_tensor = make_token_lora_mapping( |
| 520 | total_tokens, |
| 521 | ctx.batch_size, |
| 522 | prompt_lora_indices_tensor, |
| 523 | seq_len_tensor, |
| 524 | "cpu", |
| 525 | ) |
| 526 | lora_kernel_meta = LoRAKernelMeta.make( |
| 527 | max_loras=ctx.num_loras, |
| 528 | max_num_tokens=token_lora_indices_tensor.size(0), |
| 529 | device="cpu", |
| 530 | ) |
| 531 | lora_kernel_meta.prepare_tensors(token_lora_mapping=token_lora_indices_tensor) |
| 532 | |
| 533 | return BenchmarkTensors( |
| 534 | input_tensor, |
| 535 | lora_weights, |
| 536 | output_tensor, |
| 537 | lora_kernel_meta, |
| 538 | seq_len_tensor, |
| 539 | prompt_lora_indices_tensor, |
| 540 | ) |
| 541 | |
| 542 | def sanity_check(self, ctx: BenchmarkContext, op_type: OpType) -> None: |
| 543 | """Fails asserts when non-conformality is detected.""" |
| 544 | num_tokens = ( |
| 545 | self.input.shape[1] |
| 546 | if op_type.is_fused_moe_lora_expand_fn() |
| 547 | else self.input.shape[-2] |
| 548 | ) |
| 549 | # check metadata tensors |
| 550 | ## In down shrink case, each token is repeated top_k_num times |
| 551 | assert num_tokens == self.get_num_tokens( |
| 552 | torch.sum(self.seq_lens), ctx.top_k_num, op_type |
| 553 | ), f"Expected {num_tokens} tokens, but got {torch.sum(self.seq_lens)}" |
| 554 | num_seqs = self.seq_lens.shape[0] |
| 555 | # assert self.seq_start_loc.shape[0] == num_seqs |
| 556 | ## In down shrink case, each prompt corresponds to top_k_num sequences |
| 557 | assert self.prompt_lora_mapping.shape[0] == num_seqs |
| 558 | assert self.get_num_tokens( |
| 559 | self.lora_kernel_meta.token_lora_mapping.shape[0], ctx.top_k_num, op_type |
| 560 | ) |
| 561 | |
| 562 | def to_device(self, device: str): |
| 563 | """Transfer tensors to device if the tensors aren't already on the device.""" |
| 564 | |
| 565 | def to_device(tensor: torch.Tensor): |
| 566 | if tensor.device != device: |
| 567 | tensor = tensor.to(device=device) |
| 568 | return tensor |
| 569 | |
| 570 | self.input = to_device(self.input) |
| 571 | self.output = to_device(self.output) |
| 572 | self.seq_lens = to_device(self.seq_lens) |
| 573 | self.prompt_lora_mapping = to_device(self.prompt_lora_mapping) |
| 574 | for i in range(len(self.lora_weights_lst)): |
| 575 | self.lora_weights_lst[i] = to_device(self.lora_weights_lst[i]) |
| 576 | |
| 577 | # LoRA meta |
| 578 | for field_name in LoRAKernelMeta.__dataclass_fields__: |
| 579 | field = getattr(self.lora_kernel_meta, field_name) |
| 580 | assert isinstance(field, torch.Tensor) |
| 581 | setattr( |
| 582 | self.lora_kernel_meta, |
| 583 | field_name, |
| 584 | to_device(field) if field_name != "no_lora_flag_cpu" else field, |
| 585 | ) |
| 586 | |
| 587 | def metadata(self, ctx: BenchmarkContext, op_type: OpType) -> tuple[int, int, int]: |
| 588 | """Return num_seqs, num_tokens and max_seq_len.""" |
| 589 | num_seqs = self.seq_lens.shape[0] |
| 590 | num_tokens = self.get_num_tokens( |
| 591 | self.lora_kernel_meta.token_lora_mapping.shape[0], ctx.top_k_num, op_type |
| 592 | ) |
| 593 | max_seq_len = torch.max(self.seq_lens).item() |
| 594 | num_slices = len(self.lora_weights_lst) |
| 595 | return num_seqs, num_tokens, max_seq_len, num_slices |
| 596 | |
| 597 | def fused_moe_lora_data_prepare( |
| 598 | self, |
| 599 | block_size: int, |
| 600 | token_lora_mapping: torch.Tensor, |
| 601 | ctx: BenchmarkContext, |
| 602 | ): |
| 603 | def moe_lora_align_block_size( |
| 604 | topk_ids: torch.Tensor, |
| 605 | token_lora_mapping: torch.Tensor, |
| 606 | block_size: int, |
| 607 | num_experts: int, |
| 608 | max_loras: int, |
| 609 | expert_map: torch.Tensor | None = None, |
| 610 | pad_sorted_ids: bool = False, |
| 611 | ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]: |
| 612 | """Aligns tokens and experts into block-sized chunks for LoRA-based |
| 613 | mixture-of-experts (MoE) execution. |
| 614 | """ |
| 615 | max_num_tokens_padded = topk_ids.numel() + num_experts * (block_size - 1) |
| 616 | if pad_sorted_ids: |
| 617 | max_num_tokens_padded = round_up(max_num_tokens_padded, block_size) |
| 618 | sorted_ids = torch.empty( |
| 619 | (max_loras * max_num_tokens_padded,), |
| 620 | dtype=torch.int32, |
| 621 | device=topk_ids.device, |
| 622 | ) |
| 623 | max_num_m_blocks = triton.cdiv(max_num_tokens_padded, block_size) |
| 624 | # Expert ids must be set default to -1 to prevent a blank block |
| 625 | expert_ids = torch.empty( |
| 626 | (max_loras * max_num_m_blocks,), |
| 627 | dtype=torch.int32, |
| 628 | device=topk_ids.device, |
| 629 | ) |
| 630 | num_tokens_post_pad = torch.empty( |
| 631 | (max_loras), dtype=torch.int32, device=topk_ids.device |
| 632 | ) |
| 633 | |
| 634 | ops.moe_lora_align_block_size( |
| 635 | topk_ids, |
| 636 | token_lora_mapping, |
| 637 | num_experts, |
| 638 | block_size, |
| 639 | max_loras, |
| 640 | max_num_tokens_padded, |
| 641 | max_num_m_blocks, |
| 642 | sorted_ids, |
| 643 | expert_ids, |
| 644 | num_tokens_post_pad, |
| 645 | ) |
| 646 | if expert_map is not None: |
| 647 | expert_ids = expert_map[expert_ids] |
| 648 | |
| 649 | return sorted_ids, expert_ids, num_tokens_post_pad |
| 650 | |
| 651 | num_tokens = ctx.batch_size |
| 652 | curr_topk_ids = torch.randint( |
| 653 | 0, |
| 654 | ctx.num_experts, |
| 655 | (num_tokens, ctx.top_k_num), |
| 656 | device="cuda", |
| 657 | dtype=torch.int32, |
| 658 | ) |
| 659 | topk_weights = torch.randint( |
| 660 | 0, |
| 661 | ctx.num_experts, |
| 662 | (num_tokens, ctx.top_k_num), |
| 663 | device="cuda", |
| 664 | dtype=torch.int32, |
| 665 | ) |
| 666 | |
| 667 | (sorted_token_ids_lora, expert_ids_lora, num_tokens_post_padded_lora) = ( |
| 668 | moe_lora_align_block_size( |
| 669 | topk_ids=curr_topk_ids, |
| 670 | token_lora_mapping=token_lora_mapping, |
| 671 | block_size=block_size, |
| 672 | num_experts=ctx.num_experts, |
| 673 | max_loras=ctx.num_loras, |
| 674 | ) |
| 675 | ) |
| 676 | |
| 677 | sorted_token_ids = sorted_token_ids_lora.view(ctx.num_loras, -1) |
| 678 | expert_ids = expert_ids_lora.view(ctx.num_loras, -1) |
| 679 | num_tokens_post_padded = num_tokens_post_padded_lora |
| 680 | return (topk_weights, sorted_token_ids, expert_ids, num_tokens_post_padded) |
| 681 | |
| 682 | def as_lora_shrink_kwargs( |
| 683 | self, ctx: BenchmarkContext, op_type: OpType |
| 684 | ) -> dict[str, Any]: |
| 685 | self.sanity_check(ctx, op_type) |
| 686 | self.to_device(self.input.device) |
| 687 | |
| 688 | _, num_tokens, _, num_slices = self.metadata(ctx, op_type) |
| 689 | |
| 690 | # Sanity check matrix shapes. |
| 691 | i_shape, lw_shape, o_shape = ( |
| 692 | self.input.shape, |
| 693 | self.lora_weights_lst[0].shape, |
| 694 | self.output.shape, |
| 695 | ) |
| 696 | # Expected input shape [num_tokens, hidden_size] |
| 697 | assert len(i_shape) == 2 |
| 698 | assert i_shape[0] == num_tokens |
| 699 | hidden_size = i_shape[1] |
| 700 | # Expected lora weight shape [num_loras, lora_rank, hidden_size] |
| 701 | assert len(lw_shape) == 3 |
| 702 | assert lw_shape[2] == hidden_size |
| 703 | lora_rank = lw_shape[1] |
| 704 | # Expected output shape [num_slices, num_tokens, lora_rank] |
| 705 | assert len(o_shape) == 3 |
| 706 | assert o_shape == (num_slices, num_tokens, lora_rank) |
| 707 | |
| 708 | return { |
| 709 | "inputs": self.input, |
| 710 | "lora_a_weights": self.lora_weights_lst, |
| 711 | "output_tensor": self.output, |
| 712 | "token_lora_mapping": self.lora_kernel_meta.token_lora_mapping, |
| 713 | "token_indices_sorted_by_lora_ids": ( |
| 714 | self.lora_kernel_meta.token_indices_sorted_by_lora_ids |
| 715 | ), |
| 716 | "num_tokens_per_lora": self.lora_kernel_meta.num_tokens_per_lora, |
| 717 | "lora_token_start_loc": self.lora_kernel_meta.lora_token_start_loc, |
| 718 | "lora_ids": self.lora_kernel_meta.active_lora_ids, |
| 719 | "scaling": 1.0, |
| 720 | "no_lora_flag_cpu": self.lora_kernel_meta.no_lora_flag_cpu, |
| 721 | } |
| 722 | |
| 723 | def as_lora_expand_kwargs( |
| 724 | self, ctx: BenchmarkContext, op_type: OpType, add_inputs: bool |
| 725 | ) -> dict[str, Any]: |
| 726 | self.sanity_check(ctx, op_type) |
| 727 | self.to_device(self.input.device) |
| 728 | |
| 729 | _, num_tokens, _, num_slices = self.metadata(ctx, op_type) |
| 730 | |
| 731 | # Sanity check matrix shapes. |
| 732 | i_shape, lw_shape, o_shape = ( |
| 733 | self.input.shape, |
| 734 | self.lora_weights_lst[0].shape, |
| 735 | self.output.shape, |
| 736 | ) |
| 737 | # Expected input shape : [num_slices, num_tokens, lora_rank] |
| 738 | assert len(i_shape) == 3 |
| 739 | assert i_shape[0] == num_slices |
| 740 | assert i_shape[1] == num_tokens |
| 741 | lora_rank = i_shape[2] |
| 742 | # Expected lora weight shape : [num_lora, hidden_size, lora_rank] |
| 743 | assert len(lw_shape) == 3 |
| 744 | assert lw_shape[2] == lora_rank |
| 745 | hidden_size = lw_shape[1] |
| 746 | # Expected output shape : [num_tokens, hidden_size * num_slices] |
| 747 | assert len(o_shape) == 2 |
| 748 | assert o_shape == (num_tokens, hidden_size * num_slices) |
| 749 | |
| 750 | return { |
| 751 | "inputs": self.input, |
| 752 | "lora_b_weights": self.lora_weights_lst, |
| 753 | "output_tensor": self.output, |
| 754 | "token_lora_mapping": self.lora_kernel_meta.token_lora_mapping, |
| 755 | "token_indices_sorted_by_lora_ids": ( |
| 756 | self.lora_kernel_meta.token_indices_sorted_by_lora_ids |
| 757 | ), |
| 758 | "num_tokens_per_lora": self.lora_kernel_meta.num_tokens_per_lora, |
| 759 | "lora_token_start_loc": self.lora_kernel_meta.lora_token_start_loc, |
| 760 | "lora_ids": self.lora_kernel_meta.active_lora_ids, |
| 761 | "offset_start": 0, |
| 762 | "add_inputs": add_inputs, |
| 763 | "no_lora_flag_cpu": self.lora_kernel_meta.no_lora_flag_cpu, |
| 764 | } |
| 765 | |
| 766 | def as_fused_moe_lora_shrink_kwargs( |
| 767 | self, ctx: BenchmarkContext, op_type: OpType |
| 768 | ) -> dict[str, Any]: |
| 769 | self.sanity_check(ctx, op_type) |
| 770 | self.to_device(self.input.device) |
| 771 | |
| 772 | _, num_tokens, _, num_slices = self.metadata(ctx, op_type) |
| 773 | |
| 774 | # Sanity check matrix shapes. |
| 775 | i_shape, lw_shape, o_shape = ( |
| 776 | self.input.shape, |
| 777 | self.lora_weights_lst[0].shape, |
| 778 | self.output.shape, |
| 779 | ) |
| 780 | # Expected input shape : [num_tokens, hidden_size] for gate_up |
| 781 | # Expected input shape : [top_k_num * num_tokens, hidden_size] for down |
| 782 | assert len(i_shape) == 2 |
| 783 | assert i_shape[0] == num_tokens |
| 784 | hidden_size = i_shape[1] |
| 785 | # Expected lora weight shape [max_lora, num_experts, lora_rank, hidden_size] |
| 786 | assert len(lw_shape) == 4 |
| 787 | assert lw_shape[-1] == hidden_size |
| 788 | lora_rank = lw_shape[-2] |
| 789 | # Expected output shape : [num_slices, num_tokens, top_k_num, lora_rank] |
| 790 | assert len(o_shape) == 4 |
| 791 | assert ( |
| 792 | o_shape |
| 793 | == (num_slices, num_tokens // ctx.top_k_num, ctx.top_k_num, lora_rank) |
| 794 | if op_type in [OpType.FUSED_MOE_LORA_DOWN_SHRINK] |
| 795 | else o_shape == (num_slices, num_tokens, ctx.top_k_num, lora_rank) |
| 796 | ) |
| 797 | kernel_config = get_lora_op_configs( |
| 798 | op_type.name.lower(), |
| 799 | max_loras=lw_shape[0], |
| 800 | batch=num_tokens, |
| 801 | hidden_size=hidden_size, |
| 802 | rank=lora_rank, |
| 803 | num_slices=num_slices, |
| 804 | add_inputs=False, |
| 805 | ) |
| 806 | |
| 807 | (topk_weights, sorted_token_ids, expert_ids, num_tokens_post_padded) = ( |
| 808 | self.fused_moe_lora_data_prepare( |
| 809 | block_size=kernel_config["BLOCK_SIZE_M"], |
| 810 | token_lora_mapping=self.lora_kernel_meta.token_lora_mapping, |
| 811 | ctx=ctx, |
| 812 | ) |
| 813 | ) |
| 814 | |
| 815 | return { |
| 816 | "qcurr_hidden_states": self.input, |
| 817 | "lora_a_stacked": self.lora_weights_lst, |
| 818 | "a_intermediate_cache1": self.output, |
| 819 | "topk_weights": topk_weights, |
| 820 | "sorted_token_ids": sorted_token_ids, |
| 821 | "expert_ids": expert_ids, |
| 822 | "num_tokens_post_padded": num_tokens_post_padded, |
| 823 | "token_lora_mapping": self.lora_kernel_meta.token_lora_mapping, |
| 824 | "top_k_num": ctx.top_k_num, |
| 825 | "device": self.input.device, |
| 826 | "N": lora_rank, |
| 827 | "M": topk_weights.shape[0], |
| 828 | "EM": sorted_token_ids.shape[1], |
| 829 | "K": self.input.shape[1], |
| 830 | "num_tokens": num_tokens, |
| 831 | "num_experts": ctx.num_experts, |
| 832 | "num_slices": num_slices, |
| 833 | "shrink_block_size_m": kernel_config["BLOCK_SIZE_M"], |
| 834 | "shrink_block_size_n": kernel_config["BLOCK_SIZE_N"], |
| 835 | "shrink_block_size_k": kernel_config["BLOCK_SIZE_K"], |
| 836 | "shrink_group_size_m": kernel_config["GROUP_SIZE_M"], |
| 837 | "shrink_num_warps": kernel_config["NUM_WARPS"], |
| 838 | "shrink_num_stages": kernel_config["NUM_STAGES"], |
| 839 | "shrink_split_k": kernel_config.get("SPLIT_K", 1), |
| 840 | "mul_routed_weight": op_type.is_fused_moe_lora_down_fn(), |
| 841 | } |
| 842 | |
| 843 | def as_fused_moe_lora_expand_kwargs( |
| 844 | self, ctx: BenchmarkContext, op_type: OpType |
| 845 | ) -> dict[str, Any]: |
| 846 | self.sanity_check(ctx, op_type) |
| 847 | self.to_device(self.input.device) |
| 848 | |
| 849 | _, num_tokens, _, num_slices = self.metadata(ctx, op_type) |
| 850 | |
| 851 | # Sanity check matrix shapes. |
| 852 | i_shape, lw_shape, o_shape = ( |
| 853 | self.input.shape, |
| 854 | self.lora_weights_lst[0].shape, |
| 855 | self.output.shape, |
| 856 | ) |
| 857 | |
| 858 | # Expected input shape : [num_slices, num_tokens, top_k_num, lora_rank] |
| 859 | assert len(i_shape) == 4 |
| 860 | assert i_shape[0] == num_slices |
| 861 | assert i_shape[1] == num_tokens |
| 862 | lora_rank = i_shape[-1] |
| 863 | # Expected lora weight shape : [num_loras, num_experts, hidden_size, lora_rank] |
| 864 | assert len(lw_shape) == 4 |
| 865 | assert lw_shape[-1] == lora_rank |
| 866 | hidden_size = lw_shape[-2] |
| 867 | # Expected output shape : [num_tokens, top_k_num, hidden_size * num_slices] |
| 868 | assert len(o_shape) == 3 |
| 869 | assert o_shape == (num_tokens, ctx.top_k_num, hidden_size * num_slices) |
| 870 | |
| 871 | kernel_config = get_lora_op_configs( |
| 872 | op_type.name.lower(), |
| 873 | max_loras=lw_shape[0], |
| 874 | batch=num_tokens, |
| 875 | hidden_size=hidden_size, |
| 876 | rank=lora_rank, |
| 877 | num_slices=num_slices, |
| 878 | add_inputs=False, |
| 879 | ) |
| 880 | |
| 881 | (topk_weights, sorted_token_ids, expert_ids, num_tokens_post_padded) = ( |
| 882 | self.fused_moe_lora_data_prepare( |
| 883 | block_size=kernel_config["BLOCK_SIZE_M"], |
| 884 | token_lora_mapping=self.lora_kernel_meta.token_lora_mapping, |
| 885 | ctx=ctx, |
| 886 | ) |
| 887 | ) |
| 888 | |
| 889 | return { |
| 890 | "a_intermediate_cache1": self.input, |
| 891 | "lora_b_stacked": self.lora_weights_lst, |
| 892 | "output": self.output, |
| 893 | "topk_weights": topk_weights, |
| 894 | "sorted_token_ids": sorted_token_ids, |
| 895 | "expert_ids": expert_ids, |
| 896 | "num_tokens_post_padded": num_tokens_post_padded, |
| 897 | "token_lora_mapping": self.lora_kernel_meta.token_lora_mapping, |
| 898 | "top_k_num": ctx.top_k_num, |
| 899 | "device": self.input.device, |
| 900 | "N": lora_rank, |
| 901 | "M": topk_weights.shape[0], |
| 902 | "EM": sorted_token_ids.shape[1], |
| 903 | "K": self.input.shape[1], |
| 904 | "num_tokens": num_tokens, |
| 905 | "num_experts": ctx.num_experts, |
| 906 | "num_slices": num_slices, |
| 907 | "max_lora_rank": lora_rank, |
| 908 | "w1_output_dim_size": lw_shape[2], |
| 909 | "expand_block_size_m": kernel_config["BLOCK_SIZE_M"], |
| 910 | "expand_block_size_n": kernel_config["BLOCK_SIZE_N"], |
| 911 | "expand_block_size_k": kernel_config["BLOCK_SIZE_K"], |
| 912 | "expand_group_size_m": kernel_config["GROUP_SIZE_M"], |
| 913 | "expand_num_warps": kernel_config["NUM_WARPS"], |
| 914 | "expand_num_stages": kernel_config["NUM_STAGES"], |
| 915 | "expand_split_k": kernel_config.get("SPLIT_K", 1), |
| 916 | "mul_routed_weight": op_type.is_fused_moe_lora_down_fn(), |
| 917 | } |
| 918 | |
| 919 | def bench_fn_kwargs( |
| 920 | self, ctx: BenchmarkContext, op_type: OpType, add_inputs: bool | None = None |
| 921 | ) -> dict[str, Any]: |
| 922 | if op_type.is_shrink_fn() or op_type.is_fused_moe_lora_fn(): |
| 923 | assert add_inputs is None |
| 924 | else: |
| 925 | assert add_inputs is not None |
| 926 | |
| 927 | if op_type == OpType.LORA_SHRINK: |
| 928 | return self.as_lora_shrink_kwargs(ctx, op_type) |
| 929 | if op_type == OpType.LORA_EXPAND: |
| 930 | return self.as_lora_expand_kwargs(ctx, op_type, add_inputs) |
| 931 | if op_type.is_fused_moe_lora_shrink_fn(): |
| 932 | return self.as_fused_moe_lora_shrink_kwargs(ctx, op_type) |
| 933 | if op_type.is_fused_moe_lora_expand_fn(): |
| 934 | return self.as_fused_moe_lora_expand_kwargs(ctx, op_type) |
| 935 | raise ValueError(f"Unrecognized optype {self}") |
| 936 | |
| 937 | def test_correctness( |
| 938 | self, op_type: OpType, expand_fn_add_inputs: bool | None |
| 939 | ) -> bool: |
| 940 | """Test correctness of op_type implementation against a grouped gemm |
| 941 | reference implementation. |
| 942 | """ |
| 943 | seq_lens_cpu = self.seq_lens.to(device="cpu") |
| 944 | prompt_lora_mapping_cpu = self.prompt_lora_mapping.to(device="cpu") |
| 945 | ref_output = self.output.clone() |
| 946 | |
| 947 | self.output.zero_() |
| 948 | op_type.bench_fn()(**self.bench_fn_kwargs(op_type, expand_fn_add_inputs)) |
| 949 | |
| 950 | op_type.run_ref_group_gemm( |
| 951 | ref_output, |
| 952 | self.input, |
| 953 | self.lora_weights_lst, |
| 954 | seq_lens_cpu=seq_lens_cpu, |
| 955 | prompt_lora_mapping_cpu=prompt_lora_mapping_cpu, |
| 956 | scaling=1.0, |
| 957 | add_inputs=expand_fn_add_inputs, |
| 958 | ) |
| 959 | |
| 960 | rtol, atol = { |
| 961 | torch.float16: (6e-2, 6e-2), |
| 962 | torch.bfloat16: (6e-2, 6e-2), |
| 963 | torch.float32: (1e-2, 1e-2), |
| 964 | }[self.output.dtype] |
| 965 | |
| 966 | return torch.allclose(ref_output, self.output, rtol=rtol, atol=atol) |
| 967 | |
| 968 | |
| 969 | def bench_optype( |
| 970 | ctx: BenchmarkContext, |
| 971 | arg_pool_size: int, |
| 972 | op_type: OpType, |
| 973 | cuda_graph_nops: int | None = None, |
| 974 | expand_fn_add_inputs: bool | None = None, |
| 975 | test_correctness: bool = False, |
| 976 | ) -> TMeasurement: |
| 977 | assert arg_pool_size >= 1 |
| 978 | if op_type.is_shrink_fn() or op_type.is_fused_moe_lora_fn(): |
| 979 | assert expand_fn_add_inputs is None |
| 980 | else: |
| 981 | assert expand_fn_add_inputs is not None |
| 982 | |
| 983 | # BenchmarkContext -> BenchmarkTensors |
| 984 | bench_tensors: list[BenchmarkTensors] = [ |
| 985 | BenchmarkTensors.make(ctx, op_type) for _ in range(arg_pool_size) |
| 986 | ] |
| 987 | for bt in bench_tensors: |
| 988 | bt.sanity_check(ctx, op_type) |
| 989 | |
| 990 | # Test correctness of our implementation. |
| 991 | if test_correctness: |
| 992 | assert op_type in [OpType.LORA_SHRINK, OpType.LORA_EXPAND], ( |
| 993 | f"Correctness testing is not supported for {op_type.name}." |
| 994 | ) |
| 995 | assert all( |
| 996 | [ |
| 997 | bt.test_correctness(ctx, op_type, expand_fn_add_inputs) |
| 998 | for bt in bench_tensors |
| 999 | ] |
| 1000 | ) |
| 1001 | |
| 1002 | # BenchmarkTensors -> dict (kwargs) |
| 1003 | kwargs_list = [ |
| 1004 | bt.bench_fn_kwargs(ctx, op_type, add_inputs=expand_fn_add_inputs) |
| 1005 | for bt in bench_tensors |
| 1006 | ] |
| 1007 | |
| 1008 | # Clear LoRA optimization hash-maps. |
| 1009 | _LORA_A_PTR_DICT.clear() |
| 1010 | _LORA_B_PTR_DICT.clear() |
| 1011 | _LORA_PTR_DICT.clear() |
| 1012 | # Run bench function so that _LORA_A_PTR_DICT and _LORA_B_PTR_DICT are set up |
| 1013 | for kwargs in kwargs_list: |
| 1014 | op_type.bench_fn()(**kwargs) |
| 1015 | torch.accelerator.synchronize() |
| 1016 | |
| 1017 | # Merge into a single kwargs and qualify arguments as ArgPool |
| 1018 | kwargs = {k: ArgPool([]) for k in kwargs_list[0]} |
| 1019 | for _kwargs in kwargs_list: |
| 1020 | for k, v in _kwargs.items(): |
| 1021 | kwargs[k].values.append(v) |
| 1022 | |
| 1023 | describe_args = ( |
| 1024 | f"add_inputs={expand_fn_add_inputs}" if expand_fn_add_inputs is not None else "" |
| 1025 | ) |
| 1026 | description = f"{op_type.name}({describe_args}) ({bench_tensors[0].io_types()})" |
| 1027 | |
| 1028 | cuda_graph_params = None |
| 1029 | if cuda_graph_nops: |
| 1030 | cuda_graph_params = CudaGraphBenchParams(cuda_graph_nops) |
| 1031 | timer = None |
| 1032 | with Bench( |
| 1033 | cuda_graph_params, |
| 1034 | ctx.bench_label(), |
| 1035 | ctx.bench_sublabel(op_type), |
| 1036 | description, |
| 1037 | op_type.bench_fn(), |
| 1038 | **kwargs, |
| 1039 | ) as bench: |
| 1040 | timer = bench.run() |
| 1041 | return timer |
| 1042 | |
| 1043 | |
| 1044 | def bench_torch_mm( |
| 1045 | ctx: BenchmarkContext, |
| 1046 | arg_pool_size: int, |
| 1047 | op_type: OpType, |
| 1048 | cuda_graph_nops: int | None = None, |
| 1049 | ) -> TMeasurement: |
| 1050 | """Benchmark basic torch.mm as a roofline. |
| 1051 | |
| 1052 | When all the input tokens have the same LoRA ID, the LoRA kernels are just |
| 1053 | a matmul. This torch.mm benchmark serves as a roofline for that case. |
| 1054 | |
| 1055 | input op_type is used in determining the m, k, n dimensions for the matmul. |
| 1056 | """ |
| 1057 | batch_size, hidden_size, lora_rank, seq_length, dtype = ( |
| 1058 | ctx.batch_size, |
| 1059 | ctx.hidden_size, |
| 1060 | ctx.lora_rank, |
| 1061 | ctx.seq_length, |
| 1062 | ctx.dtype, |
| 1063 | ) |
| 1064 | |
| 1065 | m, k, n = op_type.mkn(batch_size, seq_length, hidden_size, lora_rank) |
| 1066 | # For a fairer comparison. |
| 1067 | n = n * ctx.num_slices |
| 1068 | |
| 1069 | # Get matmul input and output tensors for A x B = C |
| 1070 | As, Bs, Cs = [], [], [] |
| 1071 | for _ in range(arg_pool_size): |
| 1072 | As.append(torch.rand((m, k), dtype=dtype).to("cuda")) |
| 1073 | Bs.append(torch.rand((n, k), dtype=dtype).to("cuda").t()) |
| 1074 | Cs.append(torch.rand((m, n), dtype=dtype).to("cuda")) |
| 1075 | |
| 1076 | # Make torch.mm kwargs |
| 1077 | mm_kwargs = {"input": ArgPool(As), "mat2": ArgPool(Bs), "out": ArgPool(Cs)} |
| 1078 | |
| 1079 | description = ( |
| 1080 | f"single-lora roofline using torch.mm ({dtype_to_str(dtype)}" |
| 1081 | f"x{dtype_to_str(dtype)}" |
| 1082 | f"=>{dtype_to_str(dtype)})" |
| 1083 | ) |
| 1084 | cuda_graph_params = None |
| 1085 | if cuda_graph_nops: |
| 1086 | cuda_graph_params = CudaGraphBenchParams(cuda_graph_nops) |
| 1087 | with Bench( |
| 1088 | cuda_graph_params, |
| 1089 | ctx.bench_label(), |
| 1090 | ctx.bench_sublabel(op_type), |
| 1091 | description, |
| 1092 | torch.mm, |
| 1093 | **mm_kwargs, |
| 1094 | ) as bench: |
| 1095 | return bench.run() |
| 1096 | |
| 1097 | |
| 1098 | # runner |
| 1099 | def use_cuda_graph_recommendation() -> str: |
| 1100 | return """ |
| 1101 | Triton kernels have a significant launch overhead with |
| 1102 | launched directly via python. This overhead is more noticeable |
| 1103 | for small the problem sizes. For these cases, it is recommended |
| 1104 | to use the script with `--cuda-graph-nops N` to benchmark N |
| 1105 | consecutive invocations of the benchmarking operations from |
| 1106 | inside a CUDA Graph. Note that the returned measurement is for N |
| 1107 | invocations of the operation. |
| 1108 | """ |
| 1109 | |
| 1110 | |
| 1111 | def print_timers(timers: list[TMeasurement], args: argparse.Namespace | None = None): |
| 1112 | compare = TBenchmark.Compare(timers) |
| 1113 | compare.print() |
| 1114 | |
| 1115 | if args and args.cuda_graph_nops: |
| 1116 | print( |
| 1117 | f"Note : The timings reported above is for {args.cuda_graph_nops} " |
| 1118 | "consecutive invocations of the benchmarking functions. " |
| 1119 | f"Please divide by {args.cuda_graph_nops} for single invocation " |
| 1120 | "timings." |
| 1121 | ) |
| 1122 | |
| 1123 | print( |
| 1124 | "Note on Comparison with torch.mm : The torch.mm numbers are " |
| 1125 | "benchmark numbers of a simple matmul emulating the single lora " |
| 1126 | "case. It is provided as a roofline for comparing our LoRA Kernel " |
| 1127 | "implementations. It is expected that the LoRA kernels will be " |
| 1128 | "slower than torch.mm in cases where num_loras is big. But for " |
| 1129 | "small num_loras the goal should be to match the torch.mm numbers." |
| 1130 | ) |
| 1131 | |
| 1132 | |
| 1133 | def run(args: argparse.Namespace, bench_ctxs: list[BenchmarkContext]): |
| 1134 | if args.cuda_graph_nops is not None: |
| 1135 | assert args.cuda_graph_nops > 0 |
| 1136 | print(f"Benchmarking {args.cuda_graph_nops} invocations inside a CUDA Graph") |
| 1137 | else: |
| 1138 | print(f"CUDA Graphs not enabled.\n{use_cuda_graph_recommendation()}") |
| 1139 | |
| 1140 | timers = [] |
| 1141 | for bench_ctx in bench_ctxs: |
| 1142 | for seq_len in args.seq_lengths: |
| 1143 | bench_ops: list[OpType] = args.op_types |
| 1144 | seq_len_timers = [] |
| 1145 | for bench_op in bench_ops: |
| 1146 | for num_slices in bench_op.num_slices(): |
| 1147 | _ctx = bench_ctx.with_seq_length(seq_len).with_num_slices( |
| 1148 | num_slices |
| 1149 | ) |
| 1150 | # Benchmark torch.mm as a roofline |
| 1151 | seq_len_timers.append( |
| 1152 | bench_torch_mm( |
| 1153 | _ctx, args.arg_pool_size, bench_op, args.cuda_graph_nops |
| 1154 | ) |
| 1155 | ) |
| 1156 | |
| 1157 | # Benchmark bench_op |
| 1158 | expand_fn_add_inputs = ( |
| 1159 | [None] |
| 1160 | if bench_op.is_shrink_fn() or bench_op.is_fused_moe_lora_fn() |
| 1161 | else args.expand_fn_add_inputs |
| 1162 | ) |
| 1163 | for add_input_arg in expand_fn_add_inputs: |
| 1164 | seq_len_timers.append( |
| 1165 | bench_optype( |
| 1166 | _ctx, |
| 1167 | args.arg_pool_size, |
| 1168 | bench_op, |
| 1169 | args.cuda_graph_nops, |
| 1170 | add_input_arg, |
| 1171 | args.test_correctness, |
| 1172 | ) |
| 1173 | ) |
| 1174 | |
| 1175 | print_timers(seq_len_timers) |
| 1176 | timers.extend(seq_len_timers) |
| 1177 | |
| 1178 | # Result stdout dump |
| 1179 | print("== All Results ====") |
| 1180 | print_timers(timers, args) |
| 1181 | |
| 1182 | if args.output_directory: |
| 1183 | # Result file dump |
| 1184 | od = Path(args.output_directory) |
| 1185 | if not od.exists(): |
| 1186 | od.mkdir() |
| 1187 | |
| 1188 | timestamp = int(time.time()) |
| 1189 | pkl_file = od / f"lora_bench-{timestamp}.pkl" |
| 1190 | print(f"Writing benchmarks to {pkl_file}") |
| 1191 | with open(pkl_file, "wb") as f: |
| 1192 | pickle.dump(timers, f) |
| 1193 | |
| 1194 | |
| 1195 | def as_benchmark_contexts( |
| 1196 | hidden_sizes: list[int], lora_ranks: list[int], args: argparse.Namespace |
| 1197 | ) -> list[BenchmarkContext]: |
| 1198 | ctxs: list[BenchmarkContext] = [] |
| 1199 | for ( |
| 1200 | batch_size, |
| 1201 | hidden_size, |
| 1202 | lora_rank, |
| 1203 | num_loras, |
| 1204 | sort_by_lora_id, |
| 1205 | top_k_num, |
| 1206 | num_experts, |
| 1207 | ) in product( # noqa |
| 1208 | args.batch_sizes, |
| 1209 | list(hidden_sizes), |
| 1210 | lora_ranks, |
| 1211 | args.num_loras, |
| 1212 | args.sort_by_lora_id, |
| 1213 | args.top_k_nums, |
| 1214 | args.num_experts, |
| 1215 | ): |
| 1216 | ctxs.append( |
| 1217 | BenchmarkContext( |
| 1218 | batch_size=batch_size, |
| 1219 | hidden_size=hidden_size, |
| 1220 | lora_rank=lora_rank, |
| 1221 | num_loras=num_loras, |
| 1222 | num_active_loras=args.num_active_loras |
| 1223 | if args.num_active_loras |
| 1224 | else num_loras, |
| 1225 | # To be filled based on the OpType to benchmark |
| 1226 | seq_length=None, |
| 1227 | sort_by_lora_id=sort_by_lora_id, |
| 1228 | dtype=args.dtype, |
| 1229 | top_k_num=top_k_num, |
| 1230 | num_experts=num_experts, |
| 1231 | # To be filled based on the OpType to benchmark |
| 1232 | num_slices=None, |
| 1233 | ) |
| 1234 | ) |
| 1235 | |
| 1236 | return ctxs |
| 1237 | |
| 1238 | |
| 1239 | def run_list_bench(args: argparse.Namespace): |
| 1240 | print(args) |
| 1241 | |
| 1242 | print( |
| 1243 | "List bench :\n" |
| 1244 | f" Hidden Sizes {args.hidden_sizes}" |
| 1245 | f" LoRA Ranks {args.lora_ranks}" |
| 1246 | ) |
| 1247 | |
| 1248 | # Get all benchmarking contexts |
| 1249 | bench_contexts: list[BenchmarkContext] = as_benchmark_contexts( |
| 1250 | hidden_sizes=args.hidden_sizes, lora_ranks=args.lora_ranks, args=args |
| 1251 | ) |
| 1252 | |
| 1253 | run(args, bench_contexts) |
| 1254 | |
| 1255 | |
| 1256 | def run_range_bench(args: argparse.Namespace): |
| 1257 | print(args) |
| 1258 | |
| 1259 | hidden_sizes = list( |
| 1260 | range( |
| 1261 | args.hidden_sizes_start, |
| 1262 | args.hidden_sizes_end + 1, |
| 1263 | args.hidden_sizes_increment, |
| 1264 | ) |
| 1265 | ) |
| 1266 | lora_ranks = list( |
| 1267 | range(args.lora_ranks_start, args.lora_ranks_end + 1, args.lora_ranks_increment) |
| 1268 | ) |
| 1269 | |
| 1270 | print(f"Range bench :\n Hidden Sizes {hidden_sizes} LoRA Ranks {lora_ranks}") |
| 1271 | |
| 1272 | # Get all benchmarking contexts |
| 1273 | bench_contexts: list[BenchmarkContext] = as_benchmark_contexts( |
| 1274 | hidden_sizes=hidden_sizes, lora_ranks=lora_ranks, args=args |
| 1275 | ) |
| 1276 | |
| 1277 | run(args, bench_contexts) |
| 1278 | |
| 1279 | |
| 1280 | def run_model_bench(args: argparse.Namespace): |
| 1281 | print(args) |
| 1282 | |
| 1283 | def hidden_sizes_from_model(model: str, tp_size: int) -> set[int]: |
| 1284 | hidden_sizes = set() |
| 1285 | for KN, tp_split_dim in WEIGHT_SHAPES[model]: |
| 1286 | KN[tp_split_dim] = KN[tp_split_dim] // tp_size |
| 1287 | hidden_sizes.add(KN[1]) |
| 1288 | return hidden_sizes |
| 1289 | |
| 1290 | # Get all hidden sizes |
| 1291 | hidden_sizes: set[int] = set() |
| 1292 | for model_name, tp_size in product(args.models, args.tp_sizes): |
| 1293 | hidden_sizes = hidden_sizes.union(hidden_sizes_from_model(model_name, tp_size)) |
| 1294 | |
| 1295 | print(f"Model bench :\n Hidden Sizes {hidden_sizes} LoRA Ranks {args.lora_ranks}") |
| 1296 | |
| 1297 | # Get all benchmarking contexts |
| 1298 | bench_contexts: list[BenchmarkContext] = as_benchmark_contexts( |
| 1299 | hidden_sizes=hidden_sizes, lora_ranks=args.lora_ranks, args=args |
| 1300 | ) |
| 1301 | |
| 1302 | run(args, bench_contexts) |
| 1303 | |
| 1304 | |
| 1305 | if __name__ == "__main__": |
| 1306 | |
| 1307 | def to_torch_dtype(dt): |
| 1308 | if dt == "torch.float16": |
| 1309 | return torch.float16 |
| 1310 | if dt == "torch.bfloat16": |
| 1311 | return torch.bfloat16 |
| 1312 | raise ValueError("unsupported dtype") |
| 1313 | |
| 1314 | def get_bool(s: str) -> bool: |
| 1315 | return s.lower() in ["true", "1"] |
| 1316 | |
| 1317 | def add_common_command_args(p: argparse.ArgumentParser): |
| 1318 | p.add_argument( |
| 1319 | "--dtype", |
| 1320 | type=to_torch_dtype, |
| 1321 | required=True, |
| 1322 | help="Available options are ['torch.float16', 'torch.bfloat16']", |
| 1323 | ) |
| 1324 | |
| 1325 | p.add_argument( |
| 1326 | "--arg-pool-size", |
| 1327 | type=int, |
| 1328 | default=32, |
| 1329 | help="Run profiles with a pool of input/output/meta tensors instead" |
| 1330 | "of simply reusing the same tensors for all runs. A bigger arg-pool" |
| 1331 | "mitigates hardware caching effects during benchmarking.", |
| 1332 | ) |
| 1333 | |
| 1334 | p.add_argument( |
| 1335 | "--cuda-graph-nops", |
| 1336 | type=int, |
| 1337 | help=( |
| 1338 | "when set profiling is done using cudagraph, " |
| 1339 | "with the given number of operations in a graph." |
| 1340 | "Note that the measurement returned is the time " |
| 1341 | "taken for N consecutive executions of the benchmarking " |
| 1342 | "functions, where N is the value of this argument." |
| 1343 | ), |
| 1344 | ) |
| 1345 | p.add_argument("--num-loras", nargs="+", type=int, default=DEFAULT_NUM_LORAS) |
| 1346 | p.add_argument( |
| 1347 | "--num-active-loras", |
| 1348 | type=int, |
| 1349 | default=None, |
| 1350 | help="Active LoRAs. When None, all LoRAs are active", |
| 1351 | ) |
| 1352 | p.add_argument( |
| 1353 | "--sort-by-lora-id", |
| 1354 | nargs="+", |
| 1355 | type=get_bool, |
| 1356 | default=DEFAULT_SORT_BY_LORA_IDS, |
| 1357 | ) |
| 1358 | p.add_argument( |
| 1359 | "--op-types", nargs="+", type=OpType.from_str, default=list(OpType) |
| 1360 | ) |
| 1361 | p.add_argument( |
| 1362 | "--seq-lengths", nargs="+", type=int, default=DEFAULT_SEQ_LENGTHS |
| 1363 | ) |
| 1364 | p.add_argument( |
| 1365 | "--batch-sizes", nargs="+", type=int, default=DEFAULT_BATCH_SIZES |
| 1366 | ) |
| 1367 | p.add_argument( |
| 1368 | "--expand-fn-add-inputs", |
| 1369 | nargs="+", |
| 1370 | type=get_bool, |
| 1371 | default=DEFAULT_EXPAND_FN_ADD_INPUTS, |
| 1372 | ) |
| 1373 | p.add_argument( |
| 1374 | "-o", |
| 1375 | "--output-directory", |
| 1376 | type=str, |
| 1377 | help=( |
| 1378 | "Output directory to store a the list of benchmarking" |
| 1379 | "TMeasurement objects as a pickle file" |
| 1380 | ), |
| 1381 | ) |
| 1382 | |
| 1383 | p.add_argument( |
| 1384 | "--test-correctness", |
| 1385 | action="store_true", |
| 1386 | help=( |
| 1387 | "When enabled, the benchmarking functions are tested" |
| 1388 | "for correctness before the actual benchmarking" |
| 1389 | ), |
| 1390 | ) |
| 1391 | |
| 1392 | p.add_argument( |
| 1393 | "--top-k-nums", |
| 1394 | nargs="+", |
| 1395 | type=int, |
| 1396 | default=DEFAULT_TOP_K_NUMS, |
| 1397 | help="Top-K values for MoE LoRA operations", |
| 1398 | ) |
| 1399 | |
| 1400 | p.add_argument( |
| 1401 | "--num-experts", |
| 1402 | nargs="+", |
| 1403 | type=int, |
| 1404 | default=DEFAULT_NUM_EXPERTS, |
| 1405 | help="Number of experts for MoE LoRA operations", |
| 1406 | ) |
| 1407 | |
| 1408 | parser = FlexibleArgumentParser( |
| 1409 | description=f""" |
| 1410 | Benchmark LoRA kernels: |
| 1411 | {use_cuda_graph_recommendation()} |
| 1412 | |
| 1413 | list_bench example: |
| 1414 | python3 benchmarks/kernels/benchmark_lora.py list_bench --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --hidden-sizes 2048 --lora-ranks 16 --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 |
| 1415 | |
| 1416 | model_bench example: |
| 1417 | python3 benchmarks/kernels/benchmark_lora.py model_bench --models meta-llama/Llama-3-8b --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --lora-ranks 16 --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 |
| 1418 | |
| 1419 | range_bench example: |
| 1420 | python3 benchmarks/kernels/benchmark_lora.py range_bench --arg-pool-size 32 --batch-sizes 1 16 32 --dtype torch.float16 --num-loras 1 4 --op-types lora_shrink lora_expand --seq-lengths 1 16 --sort-by-lora-id 1 --cuda-graph-nops 32 --hidden-sizes-start 1024 --hidden-sizes-end 4096 --hidden-sizes-increment 1024 --lora-ranks-start 8 --lora-ranks-end 24 --lora-ranks-increment 8 |
| 1421 | """, # noqa: E501 |
| 1422 | formatter_class=argparse.RawTextHelpFormatter, |
| 1423 | ) |
| 1424 | |
| 1425 | subparsers = parser.add_subparsers(dest="cmd", required=True) |
| 1426 | |
| 1427 | list_parser = subparsers.add_parser("list_bench") |
| 1428 | list_parser.add_argument( |
| 1429 | "--hidden-sizes", nargs="+", type=int, default=DEFAULT_HIDDEN_SIZES |
| 1430 | ) |
| 1431 | list_parser.add_argument( |
| 1432 | "--lora-ranks", nargs="+", type=int, default=DEFAULT_LORA_RANKS |
| 1433 | ) |
| 1434 | add_common_command_args(list_parser) |
| 1435 | list_parser.set_defaults(func=run_list_bench) |
| 1436 | |
| 1437 | range_parser = subparsers.add_parser("range_bench") |
| 1438 | range_parser.add_argument("--hidden-sizes-start", type=int, required=True) |
| 1439 | range_parser.add_argument("--hidden-sizes-end", type=int, required=True) |
| 1440 | range_parser.add_argument("--hidden-sizes-increment", type=int, required=True) |
| 1441 | range_parser.add_argument("--lora-ranks-start", type=int, required=True) |
| 1442 | range_parser.add_argument("--lora-ranks-end", type=int, required=True) |
| 1443 | range_parser.add_argument("--lora-ranks-increment", type=int, required=True) |
| 1444 | add_common_command_args(range_parser) |
| 1445 | range_parser.set_defaults(func=run_range_bench) |
| 1446 | |
| 1447 | model_parser = subparsers.add_parser("model_bench") |
| 1448 | model_parser.add_argument( |
| 1449 | "--models", |
| 1450 | nargs="+", |
| 1451 | type=str, |
| 1452 | default=DEFAULT_MODELS, |
| 1453 | choices=WEIGHT_SHAPES.keys(), |
| 1454 | ) |
| 1455 | model_parser.add_argument( |
| 1456 | "--tp-sizes", nargs="+", type=int, default=DEFAULT_TP_SIZES |
| 1457 | ) |
| 1458 | model_parser.add_argument( |
| 1459 | "--lora-ranks", nargs="+", type=int, default=DEFAULT_LORA_RANKS |
| 1460 | ) |
| 1461 | add_common_command_args(model_parser) |
| 1462 | model_parser.set_defaults(func=run_model_bench) |
| 1463 | |
| 1464 | args = parser.parse_args() |
| 1465 | args.func(args) |