| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from dataclasses import dataclass |
| 5 | from enum import Enum |
| 6 | |
| 7 | import torch |
| 8 | import torch.nn as nn |
| 9 | |
| 10 | from vllm import envs |
| 11 | from vllm.model_executor.layers.fused_moe.fused_moe import try_get_optimal_moe_config |
| 12 | from vllm.platforms import current_platform |
| 13 | from vllm.utils.math_utils import next_power_of_2 |
| 14 | |
| 15 | _lora_aux_cuda_stream: torch.cuda.Stream | None = None |
| 16 | |
| 17 | |
| 18 | def _get_lora_aux_cuda_stream() -> torch.cuda.Stream | None: |
| 19 | if not envs.VLLM_LORA_ENABLE_DUAL_STREAM: |
| 20 | return None |
| 21 | global _lora_aux_cuda_stream |
| 22 | if _lora_aux_cuda_stream is None and current_platform.is_cuda_alike(): |
| 23 | _lora_aux_cuda_stream = torch.cuda.Stream() |
| 24 | return _lora_aux_cuda_stream |
| 25 | |
| 26 | |
| 27 | class LoRAMappingType(Enum): |
| 28 | LANGUAGE = 1 |
| 29 | TOWER = 2 |
| 30 | CONNECTOR = 3 |
| 31 | |
| 32 | |
| 33 | @dataclass |
| 34 | class LoRAMapping: |
| 35 | index_mapping: tuple[int, ...] |
| 36 | prompt_mapping: tuple[int, ...] |
| 37 | is_prefill: bool = False |
| 38 | type: LoRAMappingType = LoRAMappingType.LANGUAGE |
| 39 | |
| 40 | def __post_init__(self): |
| 41 | self.index_mapping = tuple(self.index_mapping) |
| 42 | self.prompt_mapping = tuple(self.prompt_mapping) |
| 43 | |
| 44 | |
| 45 | def _get_lora_device(base_layer: nn.Module) -> torch.device: |
| 46 | # code borrowed from https://github.com/fmmoret/vllm/blob/fm-support-lora-on-quantized-models/vllm/lora/layers.py#L34 |
| 47 | """Returns the device for where to place the LoRA tensors.""" |
| 48 | if hasattr(base_layer, "routed_experts"): |
| 49 | base_layer = base_layer.routed_experts |
| 50 | |
| 51 | # unquantizedLinear |
| 52 | if hasattr(base_layer, "weight"): |
| 53 | return base_layer.weight.device |
| 54 | # Compressed Tensor |
| 55 | elif hasattr(base_layer, "weight_packed"): |
| 56 | return base_layer.weight_packed.device |
| 57 | # GPTQ/AWQ |
| 58 | elif hasattr(base_layer, "qweight"): |
| 59 | return base_layer.qweight.device |
| 60 | # INC WNA16 (AutoRound) |
| 61 | elif hasattr(base_layer, "ark_linear"): |
| 62 | return base_layer.ark_linear.qweight.device |
| 63 | # MoE layer |
| 64 | elif hasattr(base_layer, "w2_weight"): |
| 65 | return base_layer.w2_weight.device |
| 66 | # MoE Compressed Tensor |
| 67 | elif hasattr(base_layer, "w2_weight_packed"): |
| 68 | return base_layer.w2_weight_packed.device |
| 69 | # MoE GPTQ/AWQ/GGUF |
| 70 | elif hasattr(base_layer, "w2_qweight"): |
| 71 | return base_layer.w2_qweight.device |
| 72 | else: |
| 73 | raise ValueError(f"Unsupported base layer: {base_layer}") |
| 74 | |
| 75 | |
| 76 | def _not_fully_sharded_can_replace(can_replace): |
| 77 | """Decorator which adds the condition of not using fully sharded loras |
| 78 | intended to wrap can_replace_layer() |
| 79 | """ |
| 80 | |
| 81 | def dec(*args, **kwargs): |
| 82 | decorate = kwargs.pop("decorate") if "decorate" in kwargs else True |
| 83 | condition = not kwargs["lora_config"].fully_sharded_loras if decorate else True |
| 84 | return can_replace(*args, **kwargs) and condition |
| 85 | |
| 86 | return dec |
| 87 | |
| 88 | |
| 89 | def _fully_sharded_can_replace(can_replace): |
| 90 | """Decorator which adds the condition of fully sharded loras |
| 91 | intended to wrap can_replace_layer() |
| 92 | """ |
| 93 | |
| 94 | def dec(*args, **kwargs): |
| 95 | return ( |
| 96 | can_replace(*args, **kwargs) and kwargs["lora_config"].fully_sharded_loras |
| 97 | ) |
| 98 | |
| 99 | return dec |
| 100 | |
| 101 | |
| 102 | def try_get_optimal_moe_lora_config( |
| 103 | op_type: str, |
| 104 | w1_shape: tuple[int, ...], |
| 105 | w2_shape: tuple[int, ...], |
| 106 | rank: int, |
| 107 | top_k: int, |
| 108 | dtype: str | None, |
| 109 | M: int, |
| 110 | ) -> dict[str, int | None]: |
| 111 | # LoRA shrink/expand operates on bf16/fp16 adapters regardless of the |
| 112 | # base MoE weight's block-wise quantization, so block_shape is omitted |
| 113 | # from the config lookup — the non-quantized branch in get_default_config |
| 114 | # ignores it anyway. |
| 115 | raw_config = try_get_optimal_moe_config(w1_shape, w2_shape, top_k, dtype, M) |
| 116 | config: dict[str, int | None] = dict(raw_config) |
| 117 | if op_type in [ |
| 118 | "fused_moe_lora_w13_shrink", |
| 119 | "fused_moe_lora_w2_shrink", |
| 120 | ]: |
| 121 | block_size_n = config.get("BLOCK_SIZE_N") |
| 122 | config["BLOCK_SIZE_N"] = min( |
| 123 | block_size_n if block_size_n is not None else 64, |
| 124 | next_power_of_2(rank), |
| 125 | ) |
| 126 | elif op_type in [ |
| 127 | "fused_moe_lora_w13_expand", |
| 128 | "fused_moe_lora_w2_expand", |
| 129 | ]: |
| 130 | block_size_k = config.get("BLOCK_SIZE_K") |
| 131 | config["BLOCK_SIZE_K"] = max( |
| 132 | 16, |
| 133 | min( |
| 134 | block_size_k if block_size_k is not None else 32, |
| 135 | next_power_of_2(rank), |
| 136 | ), |
| 137 | ) |
| 138 | return config |