Fix Plamo 2/3 & LFM2 for Transformers v5 (#38090)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2026-03-25 12:29:49 +00:00
committed by GitHub
parent 1e673a43ce
commit d6bb2a9d9a
3 changed files with 56 additions and 51 deletions
+23 -21
View File
@@ -4,6 +4,7 @@
from collections.abc import Iterable
from itertools import islice
from typing import TYPE_CHECKING
import torch
from torch import nn
@@ -71,30 +72,31 @@ from vllm.utils.torch_utils import direct_register_custom_op
from vllm.v1.attention.backend import AttentionMetadata
from vllm.v1.attention.backends.mamba2_attn import Mamba2AttentionMetadata
# Only used for type hinting.
class Plamo2Config(PretrainedConfig): # type: ignore
model_type: str = "plamo2"
if TYPE_CHECKING:
hidden_size: int
num_hidden_layers: int
rms_norm_eps: float
# Attention
num_attention_heads: int
hidden_size_per_head: int
num_key_value_heads: int
# Mamba
mamba_d_state: int
mamba_d_conv: int
mamba_num_heads: int
mamba_step: int
# MLP
intermediate_size: int
# Tokenizer
vocab_size: int
class Plamo2Config(PretrainedConfig): # type: ignore
model_type: str = "plamo2"
hidden_size: int
num_hidden_layers: int
rms_norm_eps: float
# Attention
num_attention_heads: int
hidden_size_per_head: int
num_key_value_heads: int
# Mamba
mamba_d_state: int
mamba_d_conv: int
mamba_num_heads: int
mamba_step: int
# MLP
intermediate_size: int
# Tokenizer
vocab_size: int
def is_mamba(config: Plamo2Config, i: int) -> bool:
def is_mamba(config: "Plamo2Config", i: int) -> bool:
assert config.mamba_step > 1
if config.num_hidden_layers <= (config.mamba_step // 2):
@@ -502,7 +504,7 @@ direct_register_custom_op(
class DenseMLP(nn.Module):
def __init__(
self,
config: Plamo2Config,
config: "Plamo2Config",
quant_config: QuantizationConfig | None = None,
prefix: str = "",
) -> None: