biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

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

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
Harry Mellor
2025-11-19 18:06:36 +01:00
committed by GitHub
parent d44e9df7d4
commit a8b70304d6
104 changed files with 542 additions and 910 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
vllm/model_executor/models/config.py
View File

@@ -8,6 +8,7 @@ import vllm.envs as envs
from vllm.logger import init_logger
from vllm.model_executor.models import ModelRegistry
from vllm.platforms import current_platform
from vllm.transformers_utils.config import set_default_rope_theta
from vllm.utils.math_utils import cdiv, round_up
from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
from vllm.v1.kv_cache_interface import FullAttentionSpec, MambaSpec, MLAAttentionSpec
@@ -46,8 +47,7 @@ class GteNewModelConfig(VerifyAndUpdateConfig):
"head_size": head_dim,
"rotary_dim": getattr(config, "rotary_emb_dim", head_dim),
"max_position": config.max_position_embeddings,
"base": config.rope_theta,
"rope_scaling": getattr(config, "rope_scaling", None),
"rope_parameters": config.rope_parameters,
}
@@ -78,12 +78,13 @@ class JinaRobertaModelConfig(VerifyAndUpdateConfig):
if not model_config.enforce_eager:
max_position = round_up(max_position, 8)
set_default_rope_theta(config, default_theta=config.rotary_emb_base)
config.rotary_kwargs = {
"head_size": head_dim,
"rotary_dim": getattr(config, "rotary_emb_dim", head_dim),
"max_position": max_position,
"base": getattr(config, "rope_theta", config.rotary_emb_base),
"rope_scaling": getattr(config, "rope_scaling", None),
"rope_parameters": config.rope_parameters,
}
@@ -117,18 +118,20 @@ class NomicBertModelConfig(VerifyAndUpdateConfig):
head_dim = config.hidden_size // config.num_attention_heads
rotary_emb_dim = int(head_dim * config.rotary_emb_fraction)
max_trained_positions = getattr(config, "max_trained_positions", 2048)
set_default_rope_theta(config, default_theta=config.rotary_emb_base)
config.rotary_kwargs = {
"head_size": head_dim,
"rotary_dim": rotary_emb_dim,
"max_position": max_trained_positions,
"base": getattr(config, "rope_theta", config.rotary_emb_base),
"rope_scaling": getattr(config, "rope_scaling", None),
"rope_parameters": config.rope_parameters,
}
# we ignore config.rotary_scaling_factor so that for datasets shorter
# than max_trained_positions 2048, the results are consistent
# with SentenceTransformer.
# The context extension uses vllm style rope_theta and rope_scaling.
# The context extension uses vllm style rope_theta and rope_parameters.
# See #17785 #18755
if (
not vllm_config.model_config.hf_overrides
@@ -172,7 +175,7 @@ class NomicBertModelConfig(VerifyAndUpdateConfig):
if hasattr(hf_text_config, "max_model_len"):
delattr(hf_text_config, "max_model_len")
hf_text_config.max_position_embeddings = max_trained_positions
hf_text_config.rope_scaling = config.rotary_kwargs["rope_scaling"]
hf_text_config.rope_parameters = config.rotary_kwargs["rope_parameters"]
# The priority of sentence_bert_config.json is higher
# than max_position_embeddings
@@ -246,8 +249,7 @@ class SnowflakeGteNewModelConfig(VerifyAndUpdateConfig):
"head_size": head_dim,
"rotary_dim": getattr(config, "rotary_emb_dim", head_dim),
"max_position": config.max_position_embeddings,
"base": config.rope_theta,
"rope_scaling": getattr(config, "rope_scaling", None),
"rope_parameters": config.rope_parameters,
}