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
+10 -8
View File
@@ -52,7 +52,7 @@ class Lfm2MLP(nn.Module):
def __init__( def __init__(
self, self,
dim: int, dim: int,
ff_dim: int, intermediate_size: int,
multiple_of: int, multiple_of: int,
auto_adjust_ff_dim: bool, auto_adjust_ff_dim: bool,
ffn_dim_multiplier: float | None, ffn_dim_multiplier: float | None,
@@ -61,21 +61,23 @@ class Lfm2MLP(nn.Module):
): ):
super().__init__() super().__init__()
if auto_adjust_ff_dim: if auto_adjust_ff_dim:
ff_dim = int(2 * ff_dim / 3) intermediate_size = int(2 * intermediate_size / 3)
# custom dim factor multiplier # custom dim factor multiplier
if ffn_dim_multiplier is not None: if ffn_dim_multiplier is not None:
ff_dim = int(ffn_dim_multiplier * ff_dim) intermediate_size = int(ffn_dim_multiplier * intermediate_size)
ff_dim = multiple_of * ((ff_dim + multiple_of - 1) // multiple_of) intermediate_size = multiple_of * (
(intermediate_size + multiple_of - 1) // multiple_of
)
self.w13 = MergedColumnParallelLinear( self.w13 = MergedColumnParallelLinear(
input_size=dim, input_size=dim,
output_sizes=[ff_dim] * 2, output_sizes=[intermediate_size] * 2,
bias=False, bias=False,
quant_config=quant_config, quant_config=quant_config,
prefix=f"{prefix}.w13", prefix=f"{prefix}.w13",
) )
self.w2 = RowParallelLinear( self.w2 = RowParallelLinear(
input_size=ff_dim, input_size=intermediate_size,
output_size=dim, output_size=dim,
bias=False, bias=False,
quant_config=quant_config, quant_config=quant_config,
@@ -212,7 +214,7 @@ class Lfm2AttentionDecoderLayer(nn.Module):
self.feed_forward = Lfm2MLP( self.feed_forward = Lfm2MLP(
dim=config.block_dim, dim=config.block_dim,
ff_dim=config.block_ff_dim, intermediate_size=config.intermediate_size,
multiple_of=config.block_multiple_of, multiple_of=config.block_multiple_of,
auto_adjust_ff_dim=config.block_auto_adjust_ff_dim, auto_adjust_ff_dim=config.block_auto_adjust_ff_dim,
ffn_dim_multiplier=config.block_ffn_dim_multiplier, ffn_dim_multiplier=config.block_ffn_dim_multiplier,
@@ -262,7 +264,7 @@ class Lfm2ShortConvDecoderLayer(nn.Module):
self.feed_forward = Lfm2MLP( self.feed_forward = Lfm2MLP(
dim=config.block_dim, dim=config.block_dim,
ff_dim=config.block_ff_dim, intermediate_size=config.intermediate_size,
multiple_of=config.block_multiple_of, multiple_of=config.block_multiple_of,
auto_adjust_ff_dim=config.block_auto_adjust_ff_dim, auto_adjust_ff_dim=config.block_auto_adjust_ff_dim,
ffn_dim_multiplier=config.block_ffn_dim_multiplier, ffn_dim_multiplier=config.block_ffn_dim_multiplier,
+23 -21
View File
@@ -4,6 +4,7 @@
from collections.abc import Iterable from collections.abc import Iterable
from itertools import islice from itertools import islice
from typing import TYPE_CHECKING
import torch import torch
from torch import nn 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.backend import AttentionMetadata
from vllm.v1.attention.backends.mamba2_attn import Mamba2AttentionMetadata from vllm.v1.attention.backends.mamba2_attn import Mamba2AttentionMetadata
# Only used for type hinting. # Only used for type hinting.
class Plamo2Config(PretrainedConfig): # type: ignore if TYPE_CHECKING:
model_type: str = "plamo2"
hidden_size: int class Plamo2Config(PretrainedConfig): # type: ignore
num_hidden_layers: int model_type: str = "plamo2"
rms_norm_eps: float
# Attention hidden_size: int
num_attention_heads: int num_hidden_layers: int
hidden_size_per_head: int rms_norm_eps: float
num_key_value_heads: int # Attention
# Mamba num_attention_heads: int
mamba_d_state: int hidden_size_per_head: int
mamba_d_conv: int num_key_value_heads: int
mamba_num_heads: int # Mamba
mamba_step: int mamba_d_state: int
# MLP mamba_d_conv: int
intermediate_size: int mamba_num_heads: int
# Tokenizer mamba_step: int
vocab_size: 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 assert config.mamba_step > 1
if config.num_hidden_layers <= (config.mamba_step // 2): if config.num_hidden_layers <= (config.mamba_step // 2):
@@ -502,7 +504,7 @@ direct_register_custom_op(
class DenseMLP(nn.Module): class DenseMLP(nn.Module):
def __init__( def __init__(
self, self,
config: Plamo2Config, config: "Plamo2Config",
quant_config: QuantizationConfig | None = None, quant_config: QuantizationConfig | None = None,
prefix: str = "", prefix: str = "",
) -> None: ) -> None:
+23 -22
View File
@@ -4,7 +4,7 @@
from collections.abc import Iterable from collections.abc import Iterable
from itertools import islice from itertools import islice
from typing import Any from typing import TYPE_CHECKING, Any
import torch import torch
from torch import nn from torch import nn
@@ -46,28 +46,29 @@ from vllm.model_executor.models.utils import (
from vllm.model_executor.utils import set_weight_attrs from vllm.model_executor.utils import set_weight_attrs
from vllm.sequence import IntermediateTensors from vllm.sequence import IntermediateTensors
# Only used for type hinting. # Only used for type hinting.
class Plamo3Config(PretrainedConfig): # type: ignore if TYPE_CHECKING:
model_type: str = "plamo3"
hidden_size: int class Plamo3Config(PretrainedConfig): # type: ignore
num_hidden_layers: int model_type: str = "plamo3"
rms_norm_eps: float
# Attention hidden_size: int
num_attention_heads: int num_hidden_layers: int
head_dim: int rms_norm_eps: float
num_key_value_heads: int # Attention
# vllm rename `sliding_window` attr to `interleaved_sliding_window` num_attention_heads: int
# if `sliding_window` is list head_dim: int
interleaved_sliding_window: list[int | None] num_key_value_heads: int
sliding_window_pattern: int # vllm rename `sliding_window` attr to `interleaved_sliding_window`
rope_parameters: dict[str, Any] # if `sliding_window` is list
rope_local_theta: int interleaved_sliding_window: list[int | None]
# MLP sliding_window_pattern: int
intermediate_size: int rope_parameters: dict[str, Any]
# Tokenizer rope_local_theta: int
vocab_size: int # MLP
intermediate_size: int
# Tokenizer
vocab_size: int
def rms_norm_weight_loader(offset: float) -> LoaderFunction: def rms_norm_weight_loader(offset: float) -> LoaderFunction:
@@ -80,7 +81,7 @@ def rms_norm_weight_loader(offset: float) -> LoaderFunction:
class DenseMLP(nn.Module): class DenseMLP(nn.Module):
def __init__( def __init__(
self, self,
config: Plamo3Config, config: "Plamo3Config",
quant_config: QuantizationConfig | None = None, quant_config: QuantizationConfig | None = None,
prefix: str = "", prefix: str = "",
) -> None: ) -> None: