fix: use get_rope_config() to support models without rope_parameters (#21135)

python/sglang/srt/models/baichuan.py

@@ -48,6 +48,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.utils import add_prefix, is_npu
+from sglang.srt.utils.hf_transformers_utils import get_rope_config
 
 _is_npu = is_npu()
 
@@ -229,7 +230,7 @@ class BaiChuanDecoderLayer(nn.Module):
     ):
         super().__init__()
         self.hidden_size = config.hidden_size
-        rope_theta = config.rope_parameters["rope_theta"]
+        rope_theta, _ = get_rope_config(config)
         max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
         self.self_attn = BaiChuanAttention(
             hidden_size=self.hidden_size,

python/sglang/srt/models/deepseek.py

@@ -49,6 +49,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.utils import add_prefix, cpu_has_amx_support, is_cpu
+from sglang.srt.utils.hf_transformers_utils import get_rope_config
 
 _is_cpu_amx_available = cpu_has_amx_support()
 _is_cpu = is_cpu()
@@ -310,8 +311,7 @@ class DeepseekDecoderLayer(nn.Module):
     ) -> None:
         super().__init__()
         self.hidden_size = config.hidden_size
-        rope_theta = config.rope_parameters["rope_theta"]
+        rope_theta, rope_scaling = get_rope_config(config)
-        rope_scaling = config.rope_parameters
         max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
         self.self_attn = DeepseekAttention(
             hidden_size=self.hidden_size,

python/sglang/srt/models/ernie4.py

@@ -43,6 +43,7 @@ from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.deepseek_v2 import DeepseekV2MLP as Ernie4MLP
 from sglang.srt.models.llama import LlamaAttention as Ernie4Attention
 from sglang.srt.utils import add_prefix, make_layers
+from sglang.srt.utils.hf_transformers_utils import get_rope_config
 
 
 class MoEGate(nn.Module):
@@ -155,8 +156,7 @@ class Ernie4DecoderLayer(nn.Module):
         is_mtp: bool = False,
     ):
         super().__init__()
-        rope_theta = config.rope_parameters["rope_theta"]
+        rope_theta, rope_scaling = get_rope_config(config)
-        rope_scaling = config.rope_parameters
         rope_is_neox_style = getattr(config, "rope_is_neox_style", False)
         # Self attention.
         self.self_attn = Ernie4Attention(

python/sglang/srt/models/exaone.py

@@ -40,6 +40,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.utils import add_prefix
+from sglang.srt.utils.hf_transformers_utils import get_rope_config
 
 
 class ExaoneGatedMLP(nn.Module):
@@ -182,8 +183,7 @@ class ExaoneDecoderLayer(nn.Module):
     ) -> None:
         super().__init__()
         self.hidden_size = config.hidden_size
-        rope_theta = config.rope_parameters["rope_theta"]
+        rope_theta, rope_scaling = get_rope_config(config)
-        rope_scaling = config.rope_parameters
         if rope_scaling is not None and getattr(
             config, "original_max_position_embeddings", None
         ):

python/sglang/srt/models/glm4.py

@@ -52,6 +52,7 @@ from sglang.srt.model_loader.weight_utils import (
     kv_cache_scales_loader,
 )
 from sglang.srt.utils import add_prefix, make_layers
+from sglang.srt.utils.hf_transformers_utils import get_rope_config
 
 Glm4Config = None
 
@@ -217,9 +218,10 @@ class Glm4DecoderLayer(nn.Module):
     ) -> None:
         super().__init__()
         self.hidden_size = config.hidden_size
-        rope_theta = config.rope_parameters["rope_theta"]
+        rope_theta, rope_scaling = get_rope_config(config)
-        rope_scaling = config.rope_parameters
+        partial_rotary_factor = (rope_scaling or {}).get("partial_rotary_factor")
-        partial_rotary_factor = config.rope_parameters.get("partial_rotary_factor", 0.5)
+        if partial_rotary_factor is None:
+            partial_rotary_factor = getattr(config, "partial_rotary_factor", 0.5)
         bias = getattr(config, "attention_bias", True)
         max_position_embeddings = getattr(config, "max_position_embeddings", 32768)
         head_dim = getattr(config, "head_dim", None)

python/sglang/srt/models/glm4_moe.py

@@ -94,6 +94,7 @@ from sglang.srt.utils import (
     log_info_on_rank0,
     make_layers,
 )
+from sglang.srt.utils.hf_transformers_utils import get_rope_config
 
 _is_hip = is_hip()
 _is_cuda = is_cuda()
@@ -684,11 +685,10 @@ class Glm4MoeDecoderLayer(nn.Module):
         nn.Module.__init__(self)
         self.hidden_size = config.hidden_size
         self.config = config
-        rope_theta = config.rope_parameters["rope_theta"]
+        rope_theta, rope_scaling = get_rope_config(config)
-        rope_scaling = config.rope_parameters
+        partial_rotary_factor = (rope_scaling or {}).get("partial_rotary_factor")
-        partial_rotary_factor = getattr(
+        if partial_rotary_factor is None:
-            getattr(config, "rope_parameters", None), "partial_rotary_factor", None
+            partial_rotary_factor = getattr(config, "partial_rotary_factor", 0.5)
-        ) or getattr(config, "partial_rotary_factor", 0.5)
         max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
         head_dim = getattr(
             config, "head_dim", config.hidden_size // config.num_attention_heads

python/sglang/srt/models/grok.py

@@ -61,6 +61,7 @@ from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.loader import DefaultModelLoader
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.utils import add_prefix, is_npu
+from sglang.srt.utils.hf_transformers_utils import get_rope_config
 
 _is_npu = is_npu()
 
@@ -477,11 +478,7 @@ class Grok1DecoderLayer(nn.Module):
         self.layer_id = layer_id
         self.alt_stream = alt_stream or torch.cuda.Stream()
 
-        rope_params = getattr(config, "rope_parameters", None)
+        rope_theta, _ = get_rope_config(config)
-        if rope_params and "rope_theta" in rope_params:
-            rope_theta = rope_params["rope_theta"]
-        else:
-            rope_theta = getattr(config, "rope_theta", 10000)
         self.self_attn = Grok1Attention(
             config=config,
             hidden_size=self.hidden_size,

python/sglang/srt/models/hunyuan.py

@@ -53,6 +53,7 @@ from sglang.srt.model_loader.weight_utils import (
     maybe_remap_kv_scale_name,
 )
 from sglang.srt.utils import is_hip
+from sglang.srt.utils.hf_transformers_utils import get_rope_config
 
 expert_distribution_recorder = ExpertDistributionRecorder()
 
@@ -402,8 +403,7 @@ class HunYuanDecoderLayer(nn.Module):
             if isinstance(config.intermediate_size, int)
             else config.intermediate_size[layer_id]
         )
-        rope_theta = config.rope_parameters["rope_theta"]
+        rope_theta, rope_scaling = get_rope_config(config)
-        rope_scaling = config.rope_parameters
         if rope_scaling is not None and getattr(
             config, "original_max_position_embeddings", None
         ):

python/sglang/srt/models/iquest_loopcoder.py

@@ -39,6 +39,7 @@ from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.models.llama import LlamaMLP as LoopCoderMLP
 from sglang.srt.utils import add_prefix, make_layers
+from sglang.srt.utils.hf_transformers_utils import get_rope_config
 
 logger = logging.getLogger(__name__)
 
@@ -166,8 +167,7 @@ class LoopCoderAttention(nn.Module):
             prefix=add_prefix("o_proj", prefix),
         )
 
-        rope_theta = config.rope_parameters["rope_theta"]
+        rope_theta, rope_scaling = get_rope_config(config)
-        rope_scaling = config.rope_parameters
         max_position_embeddings = getattr(
             config, "max_position_embeddings", max_position
         )

python/sglang/srt/models/llada2.py

@@ -84,6 +84,7 @@ from sglang.srt.utils import (
     is_npu,
     make_layers,
 )
+from sglang.srt.utils.hf_transformers_utils import get_rope_config
 
 LoraConfig = None
 logger = logging.getLogger(__name__)
@@ -486,12 +487,13 @@ class LLaDA2MoeAttention(nn.Module):
             self.rotary_dim = config.rotary_dim
         else:
             self.rotary_dim = self.head_dim
+        rope_theta, rope_scaling = get_rope_config(config)
         self.rotary_emb = get_rope(
             self.head_dim,
             rotary_dim=self.rotary_dim,
             max_position=config.max_position_embeddings,
-            base=config.rope_parameters["rope_theta"],
+            base=rope_theta,
-            rope_scaling=config.rope_parameters,
+            rope_scaling=rope_scaling,
         )
 
         self.attn = RadixAttention(

python/sglang/srt/models/minicpm.py

@@ -38,6 +38,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.utils import add_prefix
+from sglang.srt.utils.hf_transformers_utils import get_rope_config
 
 
 class MiniCPMMLP(nn.Module):
@@ -176,8 +177,7 @@ class MiniCPMDecoderLayer(nn.Module):
         super().__init__()
         self.config = config
         self.hidden_size = config.hidden_size
-        rope_theta = config.rope_parameters["rope_theta"]
+        rope_theta, rope_scaling = get_rope_config(config)
-        rope_scaling = config.rope_parameters
         max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
         self.self_attn = MiniCPMAttention(
             hidden_size=self.hidden_size,

python/sglang/srt/models/minicpm3.py

@@ -40,6 +40,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.utils import add_prefix, is_cuda
+from sglang.srt.utils.hf_transformers_utils import get_rope_config
 
 if is_cuda():
     from sgl_kernel import bmm_fp8 as _raw_bmm_fp8
@@ -305,8 +306,7 @@ class MiniCPM3DecoderLayer(nn.Module):
         super().__init__()
         self.config = config
         self.hidden_size = config.hidden_size
-        rope_theta = config.rope_parameters["rope_theta"]
+        rope_theta, rope_scaling = get_rope_config(config)
-        rope_scaling = config.rope_parameters
         max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
         self.self_attn = MiniCPM3AttentionMLA(
             config=config,

python/sglang/srt/models/orion.py

@@ -35,6 +35,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.utils import add_prefix, make_layers
+from sglang.srt.utils.hf_transformers_utils import get_rope_config
 
 
 class OrionMLP(nn.Module):
@@ -165,8 +166,7 @@ class OrionDecoderLayer(nn.Module):
     ) -> None:
         super().__init__()
         self.hidden_size = config.hidden_size
-        rope_theta = config.rope_parameters["rope_theta"]
+        rope_theta, rope_scaling = get_rope_config(config)
-        rope_scaling = config.rope_parameters
         max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
         self.self_attn = OrionAttention(
             hidden_size=self.hidden_size,

python/sglang/srt/models/qwen.py

@@ -40,6 +40,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.utils import add_prefix
+from sglang.srt.utils.hf_transformers_utils import get_rope_config
 
 
 class QWenMLP(nn.Module):
@@ -162,8 +163,7 @@ class QWenBlock(nn.Module):
         super().__init__()
         self.ln_1 = RMSNorm(config.hidden_size, eps=config.layer_norm_epsilon)
 
-        rope_theta = config.rope_parameters["rope_theta"]
+        rope_theta, rope_scaling = get_rope_config(config)
-        rope_scaling = config.rope_parameters
         self.attn = QWenAttention(
             config.hidden_size,
             config.num_attention_heads,

python/sglang/srt/models/solar.py

@@ -55,6 +55,7 @@ from sglang.srt.model_loader.weight_utils import (
     kv_cache_scales_loader,
 )
 from sglang.srt.utils import add_prefix, make_layers
+from sglang.srt.utils.hf_transformers_utils import get_rope_config
 
 
 class SolarMLP(nn.Module):
@@ -194,8 +195,7 @@ class SolarDecoderLayer(nn.Module):
     ) -> None:
         super().__init__()
         self.hidden_size = config.hidden_size
-        rope_theta = config.rope_parameters["rope_theta"]
+        rope_theta, rope_scaling = get_rope_config(config)
-        rope_scaling = config.rope_parameters
 
         if rope_scaling is not None and getattr(
             config, "original_max_position_embeddings", None

python/sglang/srt/models/step3_vl.py

@@ -61,6 +61,7 @@ from sglang.srt.managers.schedule_batch import (
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.utils import add_prefix, log_info_on_rank0, make_layers
+from sglang.srt.utils.hf_transformers_utils import get_rope_config
 
 logger = logging.getLogger(__name__)
 
@@ -290,8 +291,7 @@ class Step3TextDecoderLayer(nn.Module):
     ) -> None:
         super().__init__()
         self.hidden_size = config.hidden_size
-        rope_theta = config.rope_parameters["rope_theta"]
+        rope_theta, rope_scaling = get_rope_config(config)
-        rope_scaling = config.rope_parameters
         max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
         head_dim = getattr(
             config, "head_dim", config.hidden_size // config.num_attention_heads

python/sglang/srt/models/xverse.py

@@ -41,6 +41,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
 from sglang.srt.model_executor.model_runner import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.utils import add_prefix
+from sglang.srt.utils.hf_transformers_utils import get_rope_config
 
 
 class XverseMLP(nn.Module):
@@ -181,8 +182,7 @@ class XverseDecoderLayer(nn.Module):
     ) -> None:
         super().__init__()
         self.hidden_size = config.hidden_size
-        rope_theta = config.rope_parameters["rope_theta"]
+        rope_theta, rope_scaling = get_rope_config(config)
-        rope_scaling = config.rope_parameters
         if rope_scaling is not None and getattr(
             config, "original_max_position_embeddings", None
         ):

python/sglang/srt/models/xverse_moe.py

@@ -49,6 +49,7 @@ from sglang.srt.layers.vocab_parallel_embedding import (
 from sglang.srt.model_executor.forward_batch_info import ForwardBatch
 from sglang.srt.model_loader.weight_utils import default_weight_loader
 from sglang.srt.utils import add_prefix, is_npu
+from sglang.srt.utils.hf_transformers_utils import get_rope_config
 
 
 class XverseMLP(nn.Module):
@@ -291,8 +292,7 @@ class XverseDecoderLayer(nn.Module):
     ) -> None:
         super().__init__()
         self.hidden_size = config.hidden_size
-        rope_theta = config.rope_parameters["rope_theta"]
+        rope_theta, rope_scaling = get_rope_config(config)
-        rope_scaling = config.rope_parameters
         max_position_embeddings = getattr(config, "max_position_embeddings", 8192)
         num_key_value_heads = getattr(
             config, "num_key_value_heads", config.num_attention_heads