Refactor: Extract DeepSeek common utilities into shared module (#16969)

python/sglang/srt/models/deepseek_v2.py

@@ -128,15 +128,19 @@ from sglang.srt.models.deepseek_common.deepseek_weight_loader import (
     DeepseekV2WeightLoaderMixin,
 )
 from sglang.srt.models.deepseek_common.utils import (
+    FORWARD_ABSORB_CORE_ATTENTION_BACKENDS,
     _device_sm,
+    _get_llama_4_scaling,
     _is_cpu,
     _is_cpu_amx_available,
+    _is_cublas_ge_129,
     _is_cuda,
     _is_gfx95_supported,
     _is_hip,
     _is_npu,
     _use_aiter,
     _use_aiter_gfx95,
+    yarn_get_mscale,
 )
 from sglang.srt.server_args import get_global_server_args
 from sglang.srt.speculative.spec_info import SpeculativeAlgorithm
@@ -146,7 +150,6 @@ from sglang.srt.utils import (
     add_prefix,
     get_bool_env_var,
     is_non_idle_and_non_empty,
-    is_nvidia_cublas_cu12_version_ge_12_9,
     log_info_on_rank0,
     make_layers,
     use_intel_amx_backend,
@@ -193,8 +196,6 @@ elif _is_npu:
 else:
     pass
 
-_is_cublas_ge_129 = is_nvidia_cublas_cu12_version_ge_12_9()
-
 logger = logging.getLogger(__name__)
 
 
@@ -1049,24 +1050,6 @@ class DeepseekV2MoE(nn.Module):
         state.hidden_states_mlp_output = final_hidden_states
 
 
-def yarn_get_mscale(scale: float = 1, mscale: float = 1) -> float:
-    import math
-
-    if scale <= 1:
-        return 1.0
-    return 0.1 * mscale * math.log(scale) + 1.0
-
-
-def _get_llama_4_scaling(
-    original_max_position_embeddings: int, scaling_beta: float, positions: torch.Tensor
-) -> torch.Tensor:
-    scaling = 1 + scaling_beta * torch.log(
-        1 + torch.floor(positions / original_max_position_embeddings)
-    )
-    # Broadcast over num_heads and head_dim
-    return scaling[..., None, None]
-
-
 class DeepseekV2AttentionMLA(nn.Module, DeepseekMHAForwardMixin):
 
     def __init__(