From 77aee20259d1cdd6c337ed55d9b77ed78ebe70a8 Mon Sep 17 00:00:00 2001 From: zql <37731799+zqlcode@users.noreply.github.com> Date: Sun, 6 Sep 2026 03:57:27 +0800 Subject: [PATCH] [Model] Add support for Nanbeige4.2 (#32151) Co-authored-by: root Co-authored-by: Xinyuan Tong --- python/sglang/srt/configs/__init__.py | 2 + python/sglang/srt/configs/model_config.py | 3 + python/sglang/srt/configs/nanbeige.py | 217 ++++++ .../srt/function_call/function_call_parser.py | 1 + .../model_runner_components/layer_setup.py | 32 +- .../spec_aux_hidden_state.py | 3 + python/sglang/srt/models/nanbeige.py | 626 ++++++++++++++++++ python/sglang/srt/parser/reasoning_parser.py | 1 + .../srt/utils/hf_transformers/common.py | 2 + 9 files changed, 881 insertions(+), 6 deletions(-) create mode 100644 python/sglang/srt/configs/nanbeige.py create mode 100644 python/sglang/srt/models/nanbeige.py diff --git a/python/sglang/srt/configs/__init__.py b/python/sglang/srt/configs/__init__.py index 1c0ccf924..f0d3b94d0 100644 --- a/python/sglang/srt/configs/__init__.py +++ b/python/sglang/srt/configs/__init__.py @@ -52,6 +52,7 @@ from sglang.srt.configs.muse_glimmer import ( MuseGlimmerAssistantConfig, MuseGlimmerConfig, ) +from sglang.srt.configs.nanbeige import NanbeigeConfig from sglang.srt.configs.nano_nemotron_vl import ( NemotronH_Nano_Omni_Reasoning_V3_Config, NemotronH_Nano_VL_V2_Config, @@ -130,6 +131,7 @@ __all__ = [ "NemotronHPuzzleConfig", "NemotronH_Nano_VL_V2_Config", "NemotronH_Nano_Omni_Reasoning_V3_Config", + "NanbeigeConfig", "JetNemotronConfig", "JetVLMConfig", "MiniCPMHybridConfig", diff --git a/python/sglang/srt/configs/model_config.py b/python/sglang/srt/configs/model_config.py index abaffcf95..18cf06d5a 100644 --- a/python/sglang/srt/configs/model_config.py +++ b/python/sglang/srt/configs/model_config.py @@ -1148,6 +1148,9 @@ class ModelConfig: if "IQuestLoopCoderForCausalLM" in self.hf_config.architectures: loop_num = getattr(self.hf_text_config, "loop_num", 1) self.num_attention_layers = int(self.num_hidden_layers * int(loop_num)) + if "NanbeigeForCausalLM" in self.hf_config.architectures: + num_loops = getattr(self.hf_text_config, "num_loops", 1) + self.num_attention_layers = int(self.num_hidden_layers * int(num_loops)) if "WhisperForConditionalGeneration" in self.hf_config.architectures: # Whisper has unique layer ID scheme: # - Encoder self-attention: 0 to encoder_layers-1 (no KV cache) diff --git a/python/sglang/srt/configs/nanbeige.py b/python/sglang/srt/configs/nanbeige.py new file mode 100644 index 000000000..0efd4ec39 --- /dev/null +++ b/python/sglang/srt/configs/nanbeige.py @@ -0,0 +1,217 @@ +# coding=utf-8 +# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved. +# +# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX +# and OPT implementations in this library. It has been modified from its +# original forms to accommodate minor architectural differences compared +# to GPT-NeoX and OPT used by the Meta AI team that trained the model. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""Nanbeige model configuration""" + +from transformers.configuration_utils import PretrainedConfig +from transformers.utils import logging + +logger = logging.get_logger(__name__) + + +class NanbeigeConfig(PretrainedConfig): + r""" + This is the configuration class to store the configuration of a [`NanbeigeModel`]. It is used to instantiate an Nanbeige + model according to the specified arguments, defining the model architecture. Instantiating a configuration with the + defaults will yield a similar configuration to that of the LLaMA-7B. + + Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the + documentation from [`PretrainedConfig`] for more information. + + + Args: + vocab_size (`int`, *optional*, defaults to 32000): + Vocabulary size of the Nanbeige model. Defines the number of different tokens that can be represented by the + `inputs_ids` passed when calling [`NanbeigeModel`] + hidden_size (`int`, *optional*, defaults to 4096): + Dimension of the hidden representations. + intermediate_size (`int`, *optional*, defaults to 11008): + Dimension of the MLP representations. + num_hidden_layers (`int`, *optional*, defaults to 32): + Number of hidden layers in the Transformer decoder. + num_attention_heads (`int`, *optional*, defaults to 32): + Number of attention heads for each attention layer in the Transformer decoder. + num_key_value_heads (`int`, *optional*): + This is the number of key_value heads that should be used to implement Grouped Query Attention. If + `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if + `num_key_value_heads=1 the model will use Multi Query Attention (MQA) otherwise GQA is used. When + converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed + by meanpooling all the original heads within that group. For more details checkout [this + paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to + `num_attention_heads`. + hidden_act (`str` or `function`, *optional*, defaults to `"silu"`): + The non-linear activation function (function or string) in the decoder. + max_position_embeddings (`int`, *optional*, defaults to 2048): + The maximum sequence length that this model might ever be used with. Llama 1 supports up to 2048 tokens, + Llama 2 up to 4096, CodeLlama up to 16384. + initializer_range (`float`, *optional*, defaults to 0.02): + The standard deviation of the truncated_normal_initializer for initializing all weight matrices. + rms_norm_eps (`float`, *optional*, defaults to 1e-06): + The epsilon used by the rms normalization layers. + use_cache (`bool`, *optional*, defaults to `True`): + Whether or not the model should return the last key/values attentions (not used by all models). Only + relevant if `config.is_decoder=True`. + pad_token_id (`int`, *optional*): + Padding token id. + bos_token_id (`int`, *optional*, defaults to 1): + Beginning of stream token id. + eos_token_id (`int`, *optional*, defaults to 2): + End of stream token id. + pretraining_tp (`int`, *optional*, defaults to 1): + Experimental feature. Tensor parallelism rank used during pretraining. Please refer to [this + document](https://huggingface.co/docs/transformers/main/perf_train_gpu_many#tensor-parallelism) to understand more about it. This value is + necessary to ensure exact reproducibility of the pretraining results. Please refer to [this + issue](https://github.com/pytorch/pytorch/issues/76232). + tie_word_embeddings (`bool`, *optional*, defaults to `False`): + Whether to tie weight embeddings + rope_theta (`float`, *optional*, defaults to 10000.0): + The base period of the RoPE embeddings. + rope_scaling (`Dict`, *optional*): + Dictionary containing the scaling configuration for the RoPE embeddings. Currently supports two scaling + strategies: linear and dynamic. Their scaling factor must be a float greater than 1. The expected format is + `{"type": strategy name, "factor": scaling factor}`. When using this flag, don't update + `max_position_embeddings` to the expected new maximum. See the following thread for more information on how + these scaling strategies behave: + https://www.reddit.com/r/LocalLLaMA/comments/14mrgpr/dynamically_scaled_rope_further_increases/. This is an + experimental feature, subject to breaking API changes in future versions. + attention_bias (`bool`, defaults to `False`, *optional*, defaults to `False`): + Whether to use a bias in the query, key, value and output projection layers during self-attention. + attention_dropout (`float`, *optional*, defaults to 0.0): + The dropout ratio for the attention probabilities. + num_loops (`int`, *optional*, defaults to 1): + The number of loops for the loop model. + loop_loss_weights (`List[float]`, *optional*, defaults to `[]`): + The weights for the loop loss. + skip_loop_final_norm (`bool`, *optional*, defaults to `False`): + Whether to skip final norm after each loop (except the last one). + + ```python + >>> from configuration_nanbeige import NanbeigeConfig + >>> from modeling_nanbeige import NanbeigeModel + + >>> # Initializing a Nanbeige nanbeige-7b style configuration + >>> configuration = NanbeigeConfig() + + >>> # Initializing a model from the nanbeige-7b style configuration + >>> model = NanbeigeModel(configuration) + + >>> # Accessing the model configuration + >>> configuration = model.config + ```""" + + model_type = "nanbeige" + keys_to_ignore_at_inference = ["past_key_values"] + + def __init__( + self, + vocab_size=166144, + hidden_size=3072, + intermediate_size=10752, + num_hidden_layers=22, + num_attention_heads=48, + num_key_value_heads=None, + head_dim=None, + hidden_act="silu", + max_position_embeddings=4096, + initializer_range=0.02, + rms_norm_eps=1e-5, + use_cache=True, + pad_token_id=None, + bos_token_id=1, + eos_token_id=2, + pretraining_tp=1, + tie_word_embeddings=False, + rope_theta=50000.0, + rope_scaling=None, + attention_bias=False, + attention_dropout=0.0, + num_loops=2, + loop_loss_weights=None, + skip_loop_final_norm=False, + **kwargs, + ): + self.vocab_size = vocab_size + self.max_position_embeddings = max_position_embeddings + self.hidden_size = hidden_size + self.intermediate_size = intermediate_size + self.num_hidden_layers = num_hidden_layers + self.num_attention_heads = num_attention_heads + + # Add head_dim logic + if head_dim is not None: + self.head_dim = head_dim + else: + self.head_dim = hidden_size // num_attention_heads + + # for backward compatibility + if num_key_value_heads is None: + num_key_value_heads = num_attention_heads + + self.num_key_value_heads = num_key_value_heads + self.hidden_act = hidden_act + self.initializer_range = initializer_range + self.rms_norm_eps = rms_norm_eps + self.pretraining_tp = pretraining_tp + self.use_cache = use_cache + self.rope_theta = rope_theta + self.rope_scaling = rope_scaling + self._rope_scaling_validation() + self.attention_bias = attention_bias + self.attention_dropout = attention_dropout + + self.num_loops = num_loops + self.loop_loss_weights = ( + loop_loss_weights if loop_loss_weights is not None else [] + ) + self.skip_loop_final_norm = skip_loop_final_norm + + super().__init__( + pad_token_id=pad_token_id, + bos_token_id=bos_token_id, + eos_token_id=eos_token_id, + tie_word_embeddings=tie_word_embeddings, + **kwargs, + ) + + def _rope_scaling_validation(self): + """ + Validate the `rope_scaling` configuration. + """ + if self.rope_scaling is None: + return + + if not isinstance(self.rope_scaling, dict) or len(self.rope_scaling) != 2: + raise ValueError( + "`rope_scaling` must be a dictionary with two fields, `type` and `factor`, " + f"got {self.rope_scaling}" + ) + rope_scaling_type = self.rope_scaling.get("type", None) + rope_scaling_factor = self.rope_scaling.get("factor", None) + if rope_scaling_type is None or rope_scaling_type not in ["linear", "dynamic"]: + raise ValueError( + f"`rope_scaling`'s type field must be one of ['linear', 'dynamic'], got {rope_scaling_type}" + ) + if ( + rope_scaling_factor is None + or not isinstance(rope_scaling_factor, float) + or rope_scaling_factor <= 1.0 + ): + raise ValueError( + f"`rope_scaling`'s factor field must be a float > 1, got {rope_scaling_factor}" + ) diff --git a/python/sglang/srt/function_call/function_call_parser.py b/python/sglang/srt/function_call/function_call_parser.py index 12dce448b..5eeb12098 100644 --- a/python/sglang/srt/function_call/function_call_parser.py +++ b/python/sglang/srt/function_call/function_call_parser.py @@ -97,6 +97,7 @@ class FunctionCallParser: "step3p5": Qwen3CoderDetector, "minimax-m2": MinimaxM2Detector, "minimax-m3": MinimaxM3Detector, + "nanbeige": Qwen3CoderDetector, "trinity": TrinityDetector, "interns1": InternlmDetector, "hermes": HermesDetector, diff --git a/python/sglang/srt/model_executor/model_runner_components/layer_setup.py b/python/sglang/srt/model_executor/model_runner_components/layer_setup.py index 2ec90a90f..d421be458 100644 --- a/python/sglang/srt/model_executor/model_runner_components/layer_setup.py +++ b/python/sglang/srt/model_executor/model_runner_components/layer_setup.py @@ -18,12 +18,23 @@ class AttentionAndMoeLayers(NamedTuple): mha_companion_layers: list[Any] +def _get_loop_num(hf_config: Any) -> int: + # Nanbeige uses num_loops; IQuestLoopCoder uses loop_num. + return int(getattr(hf_config, "loop_num", getattr(hf_config, "num_loops", 1)) or 1) + + def compute_attention_and_moe_layers(layer_model: Any) -> AttentionAndMoeLayers: attention_layers: list[Any] = [] moe_layers: list[Any] = [] moe_fusions: list[Any] = [] dsa_indexers: list[Any] = [] mha_companion_layers: list[Any] = [] + + # Loop models (Nanbeige / IQuestLoopCoder) store one RadixAttention per loop + # in a ModuleList. Prefill CUDA graph indexes by layer_id, so expand and + # reorder to a dense [0..N) list. + has_loop_attn = False + layers = layer_model.layers if isinstance(layers, nn.ModuleDict): layers = layers.values() @@ -62,11 +73,16 @@ def compute_attention_and_moe_layers(layer_model: Any) -> AttentionAndMoeLayers: # Mamba layer with split op support - store the layer itself attn_layer = layer - # Keep these lists aligned with global layer ids. Pipeline-parallel - # models retain placeholders outside the local stage, while real - # attention modules use their global layer_id during graph replay. - attention_layers.append(attn_layer) - mha_companion_layers.append(mha_companion_layer) + if isinstance(attn_layer, nn.ModuleList): + attention_layers.extend(attn_layer) + mha_companion_layers.extend([mha_companion_layer] * len(attn_layer)) + has_loop_attn = True + else: + # Keep these lists aligned with global layer ids. Pipeline-parallel + # models retain placeholders outside the local stage, while real + # attention modules use their global layer_id during graph replay. + attention_layers.append(attn_layer) + mha_companion_layers.append(mha_companion_layer) moe_block = None moe_fusion = None @@ -93,6 +109,10 @@ def compute_attention_and_moe_layers(layer_model: Any) -> AttentionAndMoeLayers: dsa_indexer = layer.self_attn.indexer dsa_indexers.append(dsa_indexer) + # Reorder so attention_layers[i] matches RadixAttention.layer_id. + if has_loop_attn: + attention_layers.sort(key=lambda x: x.layer_id) + return AttentionAndMoeLayers( attention_layers, moe_layers, @@ -132,7 +152,7 @@ def resolve_layer_indices( num_effective_layers = pp_range.end_layer - pp_range.start_layer # For LoopCoder models, each loop has its own layer_id, so we need to multiply by loop_num - loop_num = getattr(model_config.hf_config, "loop_num", 1) + loop_num = _get_loop_num(model_config.hf_config) if loop_num > 1: num_effective_layers = num_effective_layers * loop_num diff --git a/python/sglang/srt/model_executor/model_runner_components/spec_aux_hidden_state.py b/python/sglang/srt/model_executor/model_runner_components/spec_aux_hidden_state.py index eb1028a41..76a4a0386 100644 --- a/python/sglang/srt/model_executor/model_runner_components/spec_aux_hidden_state.py +++ b/python/sglang/srt/model_executor/model_runner_components/spec_aux_hidden_state.py @@ -158,6 +158,9 @@ def _resolve_dflash_aux_hidden_state( f"in config. Got target={target_num_layers}." ) target_num_layers = int(target_num_layers) + # Loop models: target layer ids span num_hidden_layers * num_loops. + num_loops = getattr(model_config.hf_text_config, "num_loops", 1) + target_num_layers = target_num_layers * int(num_loops) if ( trained_target_layers is not None diff --git a/python/sglang/srt/models/nanbeige.py b/python/sglang/srt/models/nanbeige.py new file mode 100644 index 000000000..be9d6d5de --- /dev/null +++ b/python/sglang/srt/models/nanbeige.py @@ -0,0 +1,626 @@ +import logging +from typing import Iterable, List, Optional, Tuple, Union + +import torch +from torch import nn + +from sglang.srt.configs import NanbeigeConfig +from sglang.srt.distributed import get_pp_group +from sglang.srt.layers.activation import SiluAndMul +from sglang.srt.layers.dp_attention import is_dp_attention_enabled +from sglang.srt.layers.layernorm import RMSNorm +from sglang.srt.layers.linear import ( + MergedColumnParallelLinear, + QKVParallelLinear, + RowParallelLinear, +) +from sglang.srt.layers.logits_processor import LogitsProcessor +from sglang.srt.layers.pooler import EmbeddingPoolerOutput, Pooler, PoolingType +from sglang.srt.layers.quantization.base_config import QuantizationConfig +from sglang.srt.layers.radix_attention import RadixAttention +from sglang.srt.layers.rotary_embedding import get_rope +from sglang.srt.layers.utils import PPMissingLayer, get_layer_id +from sglang.srt.layers.vocab_parallel_embedding import ( + ParallelLMHead, + VocabParallelEmbedding, +) +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.runtime_context import get_parallel +from sglang.srt.utils import add_prefix, make_layers + +logger = logging.getLogger(__name__) + + +class NanbeigeRMSNorm(nn.Module): + def __init__(self, hidden_size, eps=1e-6): + """ + NanbeigeRMSNorm is equivalent to T5LayerNorm + """ + super().__init__() + self.weight = nn.Parameter(torch.ones(hidden_size)) + self.variance_epsilon = eps + + def forward(self, hidden_states): + input_dtype = hidden_states.dtype + hidden_states = hidden_states.to(torch.float32) + variance = hidden_states.pow(2).mean(-1, keepdim=True) + hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon) + return self.weight * hidden_states.to(input_dtype) + + +class NanbeigeMLP(nn.Module): + def __init__( + self, + hidden_size: int, + intermediate_size: int, + hidden_act: str, + quant_config: Optional[QuantizationConfig] = None, + prefix: str = "", + ) -> None: + super().__init__() + self.gate_up_proj = MergedColumnParallelLinear( + hidden_size, + [intermediate_size] * 2, + bias=False, + quant_config=quant_config, + prefix=add_prefix("gate_up_proj", prefix), + ) + self.down_proj = RowParallelLinear( + intermediate_size, + hidden_size, + bias=False, + quant_config=quant_config, + prefix=add_prefix("down_proj", prefix), + ) + if hidden_act != "silu": + raise ValueError( + f"Unsupported activation: {hidden_act}. Only silu is supported for now." + ) + self.act_fn = SiluAndMul() + + def forward(self, x, use_reduce_scatter: bool = False): + gate_up, _ = self.gate_up_proj(x) + x = self.act_fn(gate_up) + x, _ = self.down_proj( + x, + skip_all_reduce=use_reduce_scatter, + ) + return x + + +class NanbeigeAttention(nn.Module): + """Multi-headed attention from 'Attention Is All You Need' paper""" + + def __init__( + self, + config: NanbeigeConfig, + layer_id: Optional[int] = None, + quant_config: Optional[QuantizationConfig] = None, + prefix: str = "", + ): + super().__init__() + self.config = config + self.layer_id = layer_id + tp_size = get_parallel().tp_size + + self.attention_dropout = config.attention_dropout + self.hidden_size = config.hidden_size + + self.total_num_heads = config.num_attention_heads + assert self.total_num_heads % tp_size == 0, ( + "num_attention_heads must be divisible by tp_size." + ) + self.num_heads = self.total_num_heads // tp_size + self.head_dim = getattr( + config, "head_dim", self.hidden_size // self.total_num_heads + ) + + self.total_num_kv_heads = config.num_key_value_heads + assert self.total_num_kv_heads >= tp_size, ( + "num_key_value_heads must be greater than tp_size." + ) + assert self.total_num_kv_heads % tp_size == 0, ( + "num_key_value_heads must be divisible by tp_size." + ) + self.num_kv_heads = config.num_key_value_heads // tp_size + + self.q_size = self.num_heads * self.head_dim + self.kv_size = self.num_kv_heads * self.head_dim + + self.max_position_embeddings = config.max_position_embeddings + self.rope_theta = config.rope_theta + self.is_causal = True + self.scaling = self.head_dim**-0.5 + self.total_layers = config.num_hidden_layers + self.num_loops = config.num_loops + + self.qkv_proj = QKVParallelLinear( + self.hidden_size, + self.head_dim, + self.total_num_heads, + self.total_num_kv_heads, + bias=False, + quant_config=quant_config, + prefix=add_prefix("qkv_proj", prefix), + ) + + self.o_proj = RowParallelLinear( + self.total_num_heads * self.head_dim, + self.hidden_size, + bias=False, + quant_config=quant_config, + prefix=add_prefix("o_proj", prefix), + ) + + self.attn = nn.ModuleList() + base_layer_id = layer_id + for loop_idx in range(self.num_loops): + layer_id = base_layer_id + loop_idx * self.total_layers + self.attn.append( + RadixAttention( + self.num_heads, + self.head_dim, + self.scaling, + num_kv_heads=self.num_kv_heads, + layer_id=layer_id, + quant_config=quant_config, + prefix=add_prefix("attn", prefix), + ) + ) + + self.rotary_emb = get_rope( + self.head_dim, + rotary_dim=self.head_dim, + max_position=self.max_position_embeddings, + base=self.rope_theta, + rope_scaling=self.config.rope_scaling, + ) + + def forward( + self, + positions: torch.Tensor, + hidden_states: torch.Tensor, + forward_batch: ForwardBatch, + loop_idx: int, + ) -> torch.Tensor: + qkv, _ = self.qkv_proj(hidden_states) + q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) + q, k = self.rotary_emb(positions, q, k) + attn_output = self.attn[loop_idx](q, k, v, forward_batch) + output, _ = self.o_proj(attn_output) + return output + + +class NanbeigeDecoderLayer(nn.Module): + def __init__( + self, + config: NanbeigeConfig, + layer_id: int = 0, + quant_config: Optional[QuantizationConfig] = None, + prefix: str = "", + alt_stream: Optional[torch.cuda.Stream] = None, + ) -> None: + super().__init__() + + self.self_attn = NanbeigeAttention( + config=config, + layer_id=layer_id, + quant_config=quant_config, + prefix=add_prefix("self_attn", prefix), + ) + + self.mlp = NanbeigeMLP( + hidden_size=config.hidden_size, + intermediate_size=config.intermediate_size, + hidden_act=config.hidden_act, + quant_config=quant_config, + prefix=add_prefix("mlp", prefix), + ) + self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + self.post_attention_layernorm = RMSNorm( + config.hidden_size, eps=config.rms_norm_eps + ) + + def forward( + self, + positions: torch.Tensor, + hidden_states: torch.Tensor, + forward_batch: ForwardBatch, + loop_idx: int, + residual: Optional[torch.Tensor], + ) -> Tuple[torch.Tensor, torch.Tensor]: + + if residual is None: + residual = hidden_states + hidden_states = self.input_layernorm(hidden_states) + else: + hidden_states, residual = self.input_layernorm(hidden_states, residual) + hidden_states = self.self_attn( + positions=positions, + hidden_states=hidden_states, + forward_batch=forward_batch, + loop_idx=loop_idx, + ) + + # Fully Connected + hidden_states, residual = self.post_attention_layernorm(hidden_states, residual) + hidden_states = self.mlp(hidden_states) + + return hidden_states, residual + + +class NanbeigeModel(nn.Module): + def __init__( + self, + config: NanbeigeConfig, + quant_config: Optional[QuantizationConfig] = None, + prefix: str = "", + decoder_layer_type: type[nn.Module] = NanbeigeDecoderLayer, + alt_stream: Optional[torch.cuda.Stream] = None, + ) -> None: + super().__init__() + self.config = config + self.vocab_size = config.vocab_size + self.pp_group = get_pp_group() + pp_size = self.pp_group.world_size + assert pp_size == 1, ( + "The NanbeigeModel only supports a pipeline parallelism (PP) value of 1." + ) + + if self.pp_group.is_first_rank: + self.embed_tokens = VocabParallelEmbedding( + config.vocab_size, + config.hidden_size, + quant_config=quant_config, + use_attn_tp_group=is_dp_attention_enabled(), + prefix=add_prefix("embed_tokens", prefix), + ) + else: + self.embed_tokens = PPMissingLayer() + + # Use the provided decoder layer type or default to NanbeigeDecoderLayer + decoder_layer_type = decoder_layer_type or NanbeigeDecoderLayer + self.layers, self.start_layer, self.end_layer = make_layers( + config.num_hidden_layers, + lambda idx, prefix: decoder_layer_type( + layer_id=idx, + config=config, + quant_config=quant_config, + prefix=prefix, + alt_stream=alt_stream, + ), + pp_rank=self.pp_group.rank_in_group, + pp_size=self.pp_group.world_size, + prefix=add_prefix("layers", prefix), + ) + if self.pp_group.is_last_rank: + self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + else: + self.norm = PPMissingLayer(return_tuple=True) + + # For EAGLE3 / DFLASH: capture *before* unrolled layer id (ids already +1'd). + self.layers_to_capture = [] + + def get_input_embedding(self, input_ids): + return self.embed_tokens(input_ids) + + def forward( + self, + input_ids: torch.Tensor, + positions: torch.Tensor, + forward_batch: ForwardBatch, + input_embeds: torch.Tensor = None, + pp_proxy_tensors: Optional[PPProxyTensors] = None, + ) -> Union[torch.Tensor, PPProxyTensors]: + if self.pp_group.is_first_rank: + if input_embeds is None: + hidden_states = self.embed_tokens(input_ids) + else: + hidden_states = input_embeds + residual = None + else: + assert pp_proxy_tensors is not None + hidden_states = pp_proxy_tensors["hidden_states"] + residual = pp_proxy_tensors["residual"] + + aux_hidden_states = [] + # Draft target_layer_ids are in unrolled depth space: + # logical_id = loop_idx * num_hidden_layers + physical_i + # e.g. num_hidden_layers=22, num_loops=2 → ids in [0, 44). + # DFLASH style: setter stores k+1; we capture before that unrolled step. + num_physical_layers = self.config.num_hidden_layers + for loop_idx in range(self.config.num_loops): + for i in range(self.start_layer, self.end_layer): + logical_id = loop_idx * num_physical_layers + i + if logical_id in self.layers_to_capture: + aux_hidden_states.append( + hidden_states + residual + if residual is not None + else hidden_states.clone() + ) + layer = self.layers[i] + hidden_states, residual = layer( + positions, + hidden_states, + forward_batch, + loop_idx, + residual, + ) + + # Match the reference HF semantics for Nanbeige "loop models": + # - At the end of each full loop (except the last), convert the + # (hidden_states, residual) representation into real hidden_states + # by applying the missing residual addition. + # - If skip_loop_final_norm=False, HF applies RMSNorm after each loop + # (including intermediate loops) before entering the next loop. + if loop_idx != self.config.num_loops - 1: + if residual is not None: + hidden_states = hidden_states + residual + residual = None + if not self.config.skip_loop_final_norm: + hidden_states = self.norm(hidden_states) + + if not self.pp_group.is_last_rank: + return PPProxyTensors( + { + "hidden_states": hidden_states, + "residual": residual, + } + ) + else: + if hidden_states.shape[0] != 0: + if residual is None: + hidden_states = self.norm(hidden_states) + else: + hidden_states, _ = self.norm(hidden_states, residual) + + if len(aux_hidden_states) == 0: + return hidden_states + + return hidden_states, aux_hidden_states + + +class NanbeigeForCausalLM(nn.Module): + def __init__( + self, + config: NanbeigeConfig, + quant_config: Optional[QuantizationConfig] = None, + prefix: str = "", + ) -> None: + super().__init__() + self.pp_group = get_pp_group() + self.config = config + self.quant_config = quant_config + self.model = NanbeigeModel( + config, quant_config=quant_config, prefix=add_prefix("model", prefix) + ) + + # handle the lm head on different pp ranks + if self.pp_group.is_last_rank: + if self.pp_group.world_size == 1 and config.tie_word_embeddings: + self.lm_head = self.model.embed_tokens + else: + self.lm_head = ParallelLMHead( + config.vocab_size, + config.hidden_size, + quant_config=quant_config, + prefix=add_prefix("lm_head", prefix), + ) + else: + # ranks other than the last rank will have a placeholder layer + self.lm_head = PPMissingLayer() + + # perform weight tying for PP + if self.pp_group.world_size > 1 and config.tie_word_embeddings: + if self.pp_group.is_first_rank: + self.pp_group.send( + self.model.embed_tokens.weight, dst=self.pp_group.last_rank + ) + else: + emb_token_weight = self.pp_group.recv( + size=(config.vocab_size, config.hidden_size), + dtype=next(self.model.parameters()).dtype, + src=self.pp_group.first_rank, + ) + self.lm_head.weight.copy_(emb_token_weight) + + self.logits_processor = LogitsProcessor(config) + self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=True) + # For EAGLE3 support + self.capture_aux_hidden_states = False + + def get_input_embedding(self, input_ids: torch.Tensor) -> torch.Tensor: + return self.model.get_input_embedding(input_ids) + + def get_input_embeddings(self) -> nn.Embedding: + return self.model.embed_tokens + + @torch.no_grad() + def forward( + self, + input_ids: torch.Tensor, + positions: torch.Tensor, + forward_batch: ForwardBatch, + input_embeds: torch.Tensor = None, + get_embedding: bool = False, + pp_proxy_tensors: Optional[PPProxyTensors] = None, + ) -> torch.Tensor: + hidden_states = self.model( + input_ids, + positions, + forward_batch, + input_embeds, + pp_proxy_tensors=pp_proxy_tensors, + ) + aux_hidden_states = None + if self.capture_aux_hidden_states: + hidden_states, aux_hidden_states = hidden_states + + if self.pp_group.is_last_rank: + if not get_embedding: + return self.logits_processor( + input_ids, + hidden_states, + self.lm_head, + forward_batch, + aux_hidden_states, + ) + else: + return self.pooler(hidden_states, forward_batch) + else: + return hidden_states + + @property + def start_layer(self): + return self.model.start_layer + + @property + def end_layer(self): + return self.model.end_layer + + @torch.no_grad() + def forward_split_prefill( + self, + input_ids: torch.Tensor, + positions: torch.Tensor, + forward_batch: ForwardBatch, + split_interval: Tuple[int, int], # [start, end) 0-based + input_embeds: torch.Tensor = None, + ): + assert False, "NanbeigeModel does not support split_prefill." + + def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]): + stacked_params_mapping = [ + # (param_name, shard_name, shard_id) + (".qkv_proj", ".q_proj", "q"), + (".qkv_proj", ".k_proj", "k"), + (".qkv_proj", ".v_proj", "v"), + (".gate_up_proj", ".up_proj", 1), + (".gate_up_proj", ".gate_proj", 0), + ] + + params_dict = dict(self.named_parameters()) + for name, loaded_weight in weights: + layer_id = get_layer_id(name) + if ( + layer_id is not None + and hasattr(self.model, "start_layer") + and ( + layer_id < self.model.start_layer + or layer_id >= self.model.end_layer + ) + ): + continue + + if "rotary_emb.inv_freq" in name or "projector" in name: + continue + if self.config.tie_word_embeddings and "lm_head.weight" in name: + if self.pp_group.world_size > 1 and self.pp_group.is_last_rank: + # Handle pp weight tying here + # find the embed_tokens.weight in the weights + embed_token_weights = next( + filter(lambda x: x[0] == "model.embed_tokens.weight", weights) + )[1] + loaded_weight = embed_token_weights + else: + continue + + for param_name, weight_name, shard_id in stacked_params_mapping: + if weight_name not in name: + continue + name = name.replace(weight_name, param_name) + # Skip loading extra bias for GPTQ models. + if name.endswith(".bias") and name not in params_dict: + continue + if name not in params_dict: + continue + param = params_dict[name] + weight_loader = param.weight_loader + weight_loader(param, loaded_weight, shard_id) + break + else: + # Skip loading extra bias for GPTQ models. + if name.endswith(".bias") and name not in params_dict: + continue + + if name in params_dict.keys(): + param = params_dict[name] + weight_loader = getattr( + param, "weight_loader", default_weight_loader + ) + weight_loader(param, loaded_weight) + else: + logger.warning(f"Parameter {name} not found in params_dict") + + def get_embed_and_head(self): + return self.model.embed_tokens.weight, self.lm_head.weight + + def set_embed_and_head(self, embed, head): + del self.model.embed_tokens.weight + del self.lm_head.weight + self.model.embed_tokens.weight = embed + self.lm_head.weight = head + torch.cuda.empty_cache() + torch.cuda.synchronize() + + def set_dflash_layers_to_capture(self, layer_ids: List[int]): + if not self.pp_group.is_last_rank: + return + + if layer_ids is None: + raise ValueError( + "DFLASH requires explicit layer_ids for aux hidden capture." + ) + + self.capture_aux_hidden_states = True + # Unrolled ids: capture before step (k+1) == after HF-style layer k. + self.model.layers_to_capture = [val + 1 for val in layer_ids] + + +class NanbeigeForSequenceClassification(nn.Module): + def __init__( + self, + config: NanbeigeConfig, + quant_config: Optional[QuantizationConfig] = None, + prefix: str = "", + ) -> None: + super().__init__() + self.config = config + self.quant_config = quant_config + self.model = NanbeigeModel( + config, quant_config=quant_config, prefix=add_prefix("model", prefix) + ) + self.score = nn.Linear(config.hidden_size, config.num_labels, bias=False) + self.pooler = Pooler(pooling_type=PoolingType.LAST, normalize=False) + + self.eos_token_id = config.eos_token_id + + @torch.no_grad() + def forward( + self, + input_ids: torch.Tensor, + positions: torch.Tensor, + forward_batch: ForwardBatch, + input_embeds: torch.Tensor = None, + get_embedding: bool = True, + ) -> EmbeddingPoolerOutput: + assert get_embedding, ( + "NanbeigeForSequenceClassification is only used for embedding" + ) + + hidden_states = self.model(input_ids, positions, forward_batch, input_embeds) + logits = self.score(hidden_states) + pooled_logits = self.pooler(logits, forward_batch).embeddings + + return EmbeddingPoolerOutput(pooled_logits) + + def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]): + # Filter out lm_head weights of NanbeigeForCausalLM + filtered_weights = [ + (name, w) for name, w in weights if not name.startswith("lm_head") + ] + return NanbeigeForCausalLM.load_weights(self, filtered_weights) + + +EntryClass = [NanbeigeForCausalLM, NanbeigeForSequenceClassification] diff --git a/python/sglang/srt/parser/reasoning_parser.py b/python/sglang/srt/parser/reasoning_parser.py index 1691781e9..a460e2ab0 100644 --- a/python/sglang/srt/parser/reasoning_parser.py +++ b/python/sglang/srt/parser/reasoning_parser.py @@ -2042,6 +2042,7 @@ class ReasoningParser: "minimax": Qwen3Detector, "minimax-append-think": MiniMaxAppendThinkDetector, "minimax-m3": MiniMaxM3Detector, + "nanbeige": Qwen3Detector, "step3": DeepSeekR1Detector, "step3p5": DeepSeekR1Detector, "mistral": MistralDetector, diff --git a/python/sglang/srt/utils/hf_transformers/common.py b/python/sglang/srt/utils/hf_transformers/common.py index 552c6a799..13e4f897c 100644 --- a/python/sglang/srt/utils/hf_transformers/common.py +++ b/python/sglang/srt/utils/hf_transformers/common.py @@ -63,6 +63,7 @@ from sglang.srt.configs import ( MultiModalityConfig, MuseGlimmerAssistantConfig, MuseGlimmerConfig, + NanbeigeConfig, NemotronH_Nano_Omni_Reasoning_V3_Config, NemotronH_Nano_VL_V2_Config, NemotronHConfig, @@ -131,6 +132,7 @@ _CONFIG_REGISTRY: Dict[str, Type[PretrainedConfig]] = { NemotronH_Nano_Omni_Reasoning_V3_Config, NemotronHConfig, NemotronHPuzzleConfig, + NanbeigeConfig, DeepseekVLV2Config, Qwen3_5Config, Qwen3_5MoeConfig,