Skip to content

vllm.lora.request

Classes:

LoRARequest

Bases: Struct

Request for a LoRA adapter.

lora_int_id must be globally unique for a given adapter. This is currently not enforced in vLLM.

If True, forces reloading the adapter even if one

with the same lora_int_id already exists in the cache. This replaces the existing adapter in-place. If False (default), only loads if the adapter is not already loaded.

Methods:

  • __eq__ –

    Overrides the equality method to compare LoRARequest

  • __hash__ –

    Overrides the hash method to hash LoRARequest instances

Attributes:

Source code in vllm/lora/request.py 
 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
class LoRARequest(
    msgspec.Struct,
    omit_defaults=True,  # type: ignore[call-arg]
    array_like=True,
):  # type: ignore[call-arg]
    """Request for a LoRA adapter.

    lora_int_id must be globally unique for a given adapter.
    This is currently not enforced in vLLM.

    load_inplace: If True, forces reloading the adapter even if one
        with the same lora_int_id already exists in the cache. This replaces
        the existing adapter in-place. If False (default), only loads if the
        adapter is not already loaded.
    """

    lora_name: str
    lora_int_id: int
    lora_path: str = ""
    base_model_name: str | None = msgspec.field(default=None)
    tensorizer_config_dict: dict | None = None
    load_inplace: bool = False
    is_3d_lora_weight: bool = False
    """Whether this adapter's MoE weights are stored in the 3D fused
    `gate_up_proj` / `down_proj` layout (one fused tensor per layer) or the
    2D per-expert split layout (separate `gate_proj` / `up_proj` / `down_proj`
    tensors per expert). Only consulted when the engine is started with
    `enable_mixed_moe_lora_format=True`; otherwise it is ignored and the
    on-disk format is inferred from the base model."""

    def __post_init__(self):
        if self.lora_int_id < 1:
            raise ValueError(f"id must be > 0, got {self.lora_int_id}")

        # Ensure lora_path is not empty
        assert self.lora_path, "lora_path cannot be empty"

    @property
    def adapter_id(self):
        return self.lora_int_id

    @property
    def name(self):
        return self.lora_name

    @property
    def path(self):
        return self.lora_path

    def __eq__(self, value: object) -> bool:
        """Overrides the equality method to compare LoRARequest
        instances based on lora_name. This allows for identification
        and comparison lora adapter across engines.
        """
        return isinstance(value, self.__class__) and self.lora_name == value.lora_name

    def __hash__(self) -> int:
        """Overrides the hash method to hash LoRARequest instances
        based on lora_name. This ensures that LoRARequest instances
        can be used in hash-based collections such as sets and dictionaries,
        identified by their names across engines.
        """
        return hash(self.lora_name)

is_3d_lora_weight = False class-attribute instance-attribute

Whether this adapter's MoE weights are stored in the 3D fused gate_up_proj / down_proj layout (one fused tensor per layer) or the 2D per-expert split layout (separate gate_proj / up_proj / down_proj tensors per expert). Only consulted when the engine is started with enable_mixed_moe_lora_format=True; otherwise it is ignored and the on-disk format is inferred from the base model.

__eq__(value)

Overrides the equality method to compare LoRARequest instances based on lora_name. This allows for identification and comparison lora adapter across engines.

Source code in vllm/lora/request.py 
57
58
59
60
61
62
def __eq__(self, value: object) -> bool:
    """Overrides the equality method to compare LoRARequest
    instances based on lora_name. This allows for identification
    and comparison lora adapter across engines.
    """
    return isinstance(value, self.__class__) and self.lora_name == value.lora_name

__hash__()

Overrides the hash method to hash LoRARequest instances based on lora_name. This ensures that LoRARequest instances can be used in hash-based collections such as sets and dictionaries, identified by their names across engines.

Source code in vllm/lora/request.py 
64
65
66
67
68
69
70
def __hash__(self) -> int:
    """Overrides the hash method to hash LoRARequest instances
    based on lora_name. This ensures that LoRARequest instances
    can be used in hash-based collections such as sets and dictionaries,
    identified by their names across engines.
    """
    return hash(self.lora_name)