A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 12, 07:15 UTC
Fix RoPE failures in Transformers nightly (#29700)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Harry Mellor committed
Nov 30, 2025 at 14:29 UTC
cd719de5cb2d0a7c8e1c93fedd206ca8eaa8c517
2 files changed
+2
-19
vllm/transformers_utils/config.py
+2
-17
@@ -300,25 +300,10 @@ def set_default_rope_theta(config: PretrainedConfig, default_theta: float) -> No
300
301
def patch_rope_parameters(config: PretrainedConfig) -> None:
302
"""Provide backwards compatibility for RoPE."""
303
- # Patch rope_parameters differently based on Transformers version
304
- if Version(version("transformers")) >= Version("5.0.0.dev0"):
305
- from transformers.modeling_rope_utils import (
306
- rope_config_validation,
307
- standardize_rope_params,
308
- )
309
-
310
- # When Transformers v5 is installed, legacy rope_theta may be present
311
- # when using custom code models written for Transformers v4
312
- if (rope_theta := getattr(config, "rope_theta", None)) is not None:
313
- standardize_rope_params(config, rope_theta=rope_theta)
314
- rope_config_validation(config)
315
- # Delete rope_theta to avoid confusion in downstream code
316
- del config.rope_theta
317
- else:
318
- # When Transformers v4 is installed, legacy rope_scaling may be present
303
+ if Version(version("transformers")) < Version("5.0.0.dev0"):
304
+ # Transformers v4 installed, legacy config fields may be present
305
if (rope_scaling := getattr(config, "rope_scaling", None)) is not None:
306
config.rope_parameters = rope_scaling
321
- # When Transformers v4 is installed, legacy rope_theta may be present
307
if (rope_theta := getattr(config, "rope_theta", None)) is not None:
308
if not hasattr(config, "rope_parameters"):
309
config.rope_parameters = {"rope_type": "default"}
vllm/transformers_utils/configs/qwen3_next.py
-2
@@ -17,7 +17,6 @@
17
"""Qwen3-Next model configuration"""
18
19
from transformers.configuration_utils import PretrainedConfig, layer_type_validation
20
-from transformers.modeling_rope_utils import rope_config_validation
20
from transformers.utils import logging
21
22
logger = logging.get_logger(__name__)
@@ -245,7 +244,6 @@ class Qwen3NextConfig(PretrainedConfig):
244
self.attention_bias = attention_bias
245
self.attention_dropout = attention_dropout
246
self.head_dim = head_dim
248
- rope_config_validation(self)
247
248
self.layer_types = layer_types
249
if self.layer_types is None: