Update rope_scaling to rope_parameters in preparation for Transformers v5 (#28542)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-11-19 09:06:36 -08:00
committed by GitHub
parent d44e9df7d4
commit a8b70304d6
104 changed files with 542 additions and 910 deletions
+3 -16
View File
@@ -23,7 +23,6 @@
from collections.abc import Iterable
from itertools import islice
from typing import Any
import torch
from torch import nn
@@ -52,6 +51,7 @@ from vllm.model_executor.model_loader.weight_utils import (
maybe_remap_kv_scale_name,
)
from vllm.sequence import IntermediateTensors
from vllm.transformers_utils.config import set_default_rope_theta
from .interfaces import SupportsLoRA, SupportsPP
from .utils import (
@@ -110,8 +110,6 @@ class Exaone4Attention(nn.Module):
hidden_size: int,
num_heads: int,
num_kv_heads: int,
rope_theta: float = 1000000,
rope_scaling: dict[str, Any] | None = None,
max_position_embeddings: int = 8192,
quant_config: QuantizationConfig | None = None,
bias: bool = False,
@@ -141,7 +139,6 @@ class Exaone4Attention(nn.Module):
self.q_size = self.num_heads * self.head_dim
self.kv_size = self.num_kv_heads * self.head_dim
self.scaling = self.head_dim**-0.5
self.rope_theta = rope_theta
self.max_position_embeddings = max_position_embeddings
self.qkv_proj = QKVParallelLinear(
@@ -176,12 +173,12 @@ class Exaone4Attention(nn.Module):
# apply rotary embeddings to every layer in full attention models
self.apply_rope_all_layers = "sliding_attention" not in config.layer_types
set_default_rope_theta(config, default_theta=1000000)
self.rotary_emb = get_rope(
self.head_dim,
rotary_dim=self.head_dim,
max_position=max_position_embeddings,
base=rope_theta,
rope_scaling=rope_scaling,
rope_parameters=config.rope_parameters,
is_neox_style=is_neox_style,
)
self.attn = Attention(
@@ -227,14 +224,6 @@ class Exaone4DecoderLayer(nn.Module):
) -> None:
super().__init__()
self.hidden_size = config.hidden_size
rope_theta = getattr(config, "rope_theta", 1000000)
rope_scaling = getattr(config, "rope_scaling", None)
if rope_scaling is not None and getattr(
config, "original_max_position_embeddings", None
):
rope_scaling["original_max_position_embeddings"] = (
config.original_max_position_embeddings
)
max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
# Support abacusai/Smaug-72B-v0.1 with attention_bias
# Support internlm/internlm-7b with bias
@@ -249,8 +238,6 @@ class Exaone4DecoderLayer(nn.Module):
num_kv_heads=getattr(
config, "num_key_value_heads", config.num_attention_heads
),
rope_theta=rope_theta,
rope_scaling=rope_scaling,
max_position_embeddings=max_position_embeddings,
quant_config=quant_config,
bias=attention_bias,