Upload folder using huggingface_hub

README.md

@@ -12,33 +12,30 @@ base_model:
 <div align="center">
 
 # Baichuan-M2-32B
-
 [![License](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
 [![Hugging Face](https://img.shields.io/badge/🤗%20Hugging%20Face-Model-yellow)](https://huggingface.co/baichuan-inc/Baichuan-M2-32B)
 
 </div>
 
-## 🌟 模型简介
-
-Baichuan-M2-32B 是百川智能推出的医疗增强推理模型，这是百川开源发布的第二个医疗增强模型，专为真实世界的医疗推理任务设计。该模型基于 Qwen2.5-32B 基座，通过创新的大型验证器系统（Large Verifier System）从真实世界的医疗问题出发，进行医疗领域后训练对齐，在保持模型通用能力的同时，实现了医疗效果的突破性提升。
+## 🌟 Model Overview
 
-**模型特点：**
+Baichuan-M2-32B is BaiChuan AI's medical-enhanced reasoning model, the second medical model released by BaiChuan. Designed for real-world medical reasoning tasks, this model builds upon Qwen2.5-32B with an innovative Large Verifier System. Through domain-specific fine-tuning on real-world medical questions, it achieves breakthrough medical performance while maintaining strong general capabilities.
 
-Baichuan-M2 采用了三个核心技术创新：首先通过**大型验证器系统**，结合医疗场景特点设计了全面的医疗验证体系，包含患者模拟器和多维度验证机制；其次通过**医疗领域适应性增强**的中期训练（Mid-Training），在保持通用能力的同时实现轻量高效的医疗领域适应；最后采用**多阶段强化学习**策略，将复杂的 RL 任务分解为层次化的训练阶段，逐步提升模型的医学常识、推理和患者交互能力。
+**Model Features:**
 
-**核心亮点：**
-- 🏆 **全球最强医疗开源模型**：在 HealthBench 评测集上超越所有开源模型及众多前沿闭源模型，是最接近 GPT-5 医疗能力的开源大模型
-- 🧠 **医生思维对齐**：基于真实病例数据和患者模拟器训练，具备临床诊断思维和鲁棒的医患交互能力
-- ⚡ **高效部署与推理**：支持 4bit 量化在 RTX4090 单卡部署，MTP 版本单用户场景下 token 吞吐提升 58.5%
+Baichuan-M2 incorporates three core technical innovations: First, through the Large Verifier System, it combines medical scenario characteristics to design a comprehensive medical verification framework, including patient simulators and multi-dimensional verification mechanisms; second, through medical domain adaptation enhancement via Mid-Training, it achieves lightweight and efficient medical domain adaptation while preserving general capabilities; finally, it employs a multi-stage reinforcement learning strategy, decomposing complex RL tasks into hierarchical training stages to progressively enhance the model's medical knowledge, reasoning, and patient interaction capabilities.
 
+**Core Highlights:**
+- 🏆 **World's Leading Open-Source Medical Model**: Outperforms all open-source models and many proprietary models on HealthBench, achieving medical capabilities closest to GPT-5
+- 🧠 **Doctor-Thinking Alignment**: Trained on real clinical cases and patient simulators, with clinical diagnostic thinking and robust patient interaction capabilities
+- ⚡ **Efficient Deployment**: Supports 4-bit quantization for single-RTX4090 deployment, with 58.5% higher token throughput in MTP version for single-user scenarios
 
+## 📊 Performance Metrics
 
-## 📊 性能表现
+### HealthBench Scores
 
-### HealthBench指标
-
-| 模型名称 | HealthBench | HealthBench-Hard | HealthBench-Consensus |
-|----------|-------------|------------------|-----------------------|
+| Model Name | HealthBench | HealthBench-Hard | HealthBench-Consensus |
+|------------|-------------|------------------|-----------------------|
 | Baichuan-M2 | 60.1 | 34.7 | 91.5 |
 | gpt-oss-120b | 57.6 | 30 | 90 |
 | Qwen3-235B-A22B-Thinking-2507 | 55.2 | 25.9 | 90.6 |
@@ -47,80 +44,123 @@ Baichuan-M2 采用了三个核心技术创新：首先通过**大型验证器系
 | Kimi-K2 | 43 | 10.7 | 90.9 |
 | gpt-oss-20b | 42.5 | 10.8 | 82.6 |
 
-### 通用指标
+### General Performance
 
-| 评测集 | Baichuan-M2-32B | Qwen3-32B |
-|--------|-----------------|-----------|
+| Benchmark | Baichuan-M2-32B | Qwen3-32B |
+|-----------|-----------------|-----------|
 | AIME24 | 83.4 | 81.4 |
 | AIME25 | 72.9 | 72.9 |
 | Arena-Hard-v2.0 | 45.8 | 44.5 |
 | CFBench | 77.6 | 75.7 |
 | WritingBench | 8.56 | 7.90 |
 
-*备注：AIME 的 max_tokens 设为 64k，其他评测集设为 32k，温度统一为 0.6。*
-
+*Note: AIME uses max_tokens=64k, others use 32k; temperature=0.6 for all tests.*
 
-## 🛠️ 技术特色
+## 🔧 Technical Features
 
-### 大型验证器系统
-- **患者模拟器**：基于真实病例构建的虚拟患者系统
-- **多维度验证**：医学准确性、回答完整性、追问感知等 8 个维度
-- **动态评分**：实时生成评分标准，适应复杂临床环境
+### Large Verifier System
+- **Patient Simulator**: Virtual patient system based on real clinical cases
+- **Multi-Dimensional Verification**: 8 dimensions including medical accuracy, response completeness, and follow-up awareness
+- **Dynamic Scoring**: Real-time generation of adaptive evaluation criteria for complex clinical scenarios
+### Medical Domain Adaptation
+- **Mid-Training**: Medical knowledge injection while preserving general capabilities
+- **Reinforcement Learning**: Multi-stage RL strategy optimization
+- **General-Specialized Balance**: Carefully balanced medical, general, and mathematical composite training data
 
-### 医疗领域适应
-- **Mid-Training**：医疗知识注入的同时保持通用能力
-- **强化学习**：多阶段 RL 策略优化
-- **通专兼顾**：2:2:1 配比的医疗、通用、数学数据
+## ⚙️ Quick Start
 
-## 🔧 快速开始
-
-### 安装使用
-
-```bash
-# 安装依赖
-pip install transformers torch vllm sglang
-
-# Transformers 使用
+```python
+# 1. load model
 from transformers import AutoTokenizer, AutoModelForCausalLM
 model = AutoModelForCausalLM.from_pretrained("baichuan-inc/Baichuan-M2-32B", trust_remote_code=True)
+tokenizer = AutoTokenizer.from_pretrained("baichuan-inc/Baichuan-M2-32B")
+# 2. Input prompt text
+prompt = "Got a big swelling after a bug bite. Need help reducing it."
+# 3. Encode the input text for the model
+messages = [
+    {"role": "user", "content": prompt}
+]
+text = tokenizer.apply_chat_template(
+    messages,
+    tokenize=False,
+    add_generation_prompt=True,
+    thinking_mode='on' # on/off/auto 
+)
+model_inputs = tokenizer([text], return_tensors="pt").to(model.device)
+# 4. Generate text
+generated_ids = model.generate(
+    **model_inputs,
+    max_new_tokens=4096
+)
+output_ids = [
+    output_ids[len(input_ids):] for input_ids, output_ids in zip(model_inputs.input_ids, generated_ids)
+][0].tolist()
+# 5. parsing thinking content
+try:
+    # rindex finding 151668 (</think>)
+    index = len(output_ids) - output_ids[::-1].index(151668)
+except ValueError:
+    index = 0
+
+thinking_content = tokenizer.decode(output_ids[:index], skip_special_tokens=True).strip("\n")
+content = tokenizer.decode(output_ids[index:], skip_special_tokens=True).strip("\n")
+
+print("thinking content:", thinking_content)
+print("content:", content)
 
-# vLLM 使用（推荐）
-from vllm import LLM
-llm = LLM(model="baichuan-inc/Baichuan-M2-32B", trust_remote_code=True)
-
-# SGLang 使用
-python -m sglang.launch_server --model-path baichuan-inc/Baichuan-M2-32B
 ```
 
-## ⚠️ 使用须知
+For deployment, you can use `sglang>=0.4.6.post1` or `vllm>=0.9.0` or to create an OpenAI-compatible API endpoint:
+- SGLang:
+    ```shell
+    python -m sglang.launch_server --model-path baichuan-inc/Baichuan-M2-32B --reasoning-parser qwen3
+    ```
+- vLLM:
+    ```shell
+    vllm serve baichuan-inc/Baichuan-M2-32B  --reasoning-parser qwen3
+    ```
 
-1. **医疗免责声明**：本模型仅供研究和参考，不能替代专业医疗诊断和治疗建议
-2. **适用场景**：医学教育、健康咨询、临床辅助决策等
-3. **安全使用**：建议在专业医疗人员指导下使用
+## MTP inference with SGLang
 
-## 📄 许可证
-
-本项目采用 [Apache License 2.0](LICENSE) 开源协议，欢迎研究和商业使用。
-
-## 🤝 致谢
-
-- 基础模型：Qwen2.5-32B
-- 训练框架：VERL
-- 推理引擎：vLLM、SGLang
-- 量化方法：AutoRound、GPTQ、QuaRot、QQQ
+1. Replace the qwen2.py file in the sglang installation directory with draft/qwen2.py.
+2. Launch sglang:
+```
+python3 -m sglang.launch_server \
+--model Baichuan-M2-32B \
+--speculative-algorithm EAGLE3 \
+--speculative-draft-model-path Baichuan-M2-32B/draft \
+--speculative-num-steps 6 \
+--speculative-eagle-topk 10 \
+--speculative-num-draft-tokens 32 \
+--mem-fraction 0.9 \
+--cuda-graph-max-bs 2 \
+--reasoning-parser qwen3 \
+--dtype bfloat16
+```
 
-感谢开源社区的贡献，我们将持续回馈社区，推动医疗 AI 技术发展。
+## ⚠️ Usage Notices
+1. **Medical Disclaimer**: For research and reference only; cannot replace professional medical diagnosis or treatment
+2. **Intended Use Cases**: Medical education, health consultation, clinical decision support
+3. **Safe Use**: Recommended under guidance of medical professionals
 
-## 📞 联系我们
+## 📄 License
+Licensed under the [Apache License 2.0](LICENSE). Research and commercial use permitted.
 
-- 更多资源：[百川智能官网](https://www.baichuan-ai.com)
+## 🤝 Acknowledgements
+- Base Model: Qwen2.5-32B
+- Training Framework: verl
+- Inference Engines: vLLM, SGLang
+- Quantization: AutoRound, GPTQ
+Thank you to the open-source community. We commit to continuous contribution and advancement of healthcare AI.
 
-- 技术交流：[GitHub](https://github.com/baichuan-inc)
+## 📞 Contact Us
+- Resources: [BaiChuan AI Website](https://www.baichuan-ai.com)
+- Technical Support: [GitHub](https://github.com/baichuan-inc)
 
 ---
-
 <div align="center">
 
-**让AI助力医疗，让健康触手可及**
+**Empowering Healthcare with AI, Making Health Accessible to All**
+
+</div>
 
-</div>

draft/qwen2.py

@@ -0,0 +1,641 @@
+# Copyright 2023-2024 SGLang Team
+# 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.
+# ==============================================================================
+
+# Adapted from llama2.py
+# Modify details for the adaptation of Qwen2 model.
+"""Inference-only Qwen2 model compatible with HuggingFace weights."""
+import logging
+from typing import Any, Dict, Iterable, Optional, Tuple, Union, List
+
+import torch
+from torch import nn
+
+from sglang.srt.distributed import (
+    get_pp_group,
+    get_tensor_model_parallel_rank,
+    get_tensor_model_parallel_world_size,
+)
+from sglang.srt.layers.activation import SiluAndMul
+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 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.managers.schedule_batch import global_server_args_dict
+from sglang.srt.model_executor.forward_batch_info import ForwardBatch, PPProxyTensors
+from sglang.srt.model_loader.weight_utils import (
+    default_weight_loader,
+    kv_cache_scales_loader,
+)
+from sglang.srt.utils import add_prefix, make_layers
+
+Qwen2Config = None
+
+
+logger = logging.getLogger(__name__)
+
+
+class Qwen2MLP(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):
+        gate_up, _ = self.gate_up_proj(x)
+        x = self.act_fn(gate_up)
+        x, _ = self.down_proj(x)
+        return x
+
+
+class Qwen2Attention(nn.Module):
+    def __init__(
+        self,
+        hidden_size: int,
+        num_heads: int,
+        num_kv_heads: int,
+        head_dim: Optional[int] = None,
+        layer_id: int = 0,
+        rope_theta: float = 1000000,
+        rope_scaling: Optional[Dict[str, Any]] = None,
+        max_position_embeddings: int = 32768,
+        quant_config: Optional[QuantizationConfig] = None,
+        dual_chunk_attention_config: Optional[dict[str, Any]] = None,
+        prefix: str = "",
+    ) -> None:
+        super().__init__()
+        self.hidden_size = hidden_size
+        tp_size = get_tensor_model_parallel_world_size()
+        self.total_num_heads = num_heads
+        assert self.total_num_heads % tp_size == 0
+        self.num_heads = self.total_num_heads // tp_size
+        self.total_num_kv_heads = num_kv_heads
+        if self.total_num_kv_heads >= tp_size:
+            # Number of KV heads is greater than TP size, so we partition
+            # the KV heads across multiple tensor parallel GPUs.
+            assert self.total_num_kv_heads % tp_size == 0
+        else:
+            # Number of KV heads is less than TP size, so we replicate
+            # the KV heads across multiple tensor parallel GPUs.
+            assert tp_size % self.total_num_kv_heads == 0
+        self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
+        if head_dim is not None:
+            self.head_dim = head_dim
+        else:
+            self.head_dim = hidden_size // self.total_num_heads
+        self.q_size = self.num_heads * self.head_dim
+        self.kv_size = self.num_kv_heads * self.head_dim
+        self.scaling = self.head_dim**-0.5
+        self.rope_theta = rope_theta
+        self.max_position_embeddings = max_position_embeddings
+
+        self.qkv_proj = QKVParallelLinear(
+            hidden_size,
+            self.head_dim,
+            self.total_num_heads,
+            self.total_num_kv_heads,
+            bias=True,
+            quant_config=quant_config,
+            prefix=add_prefix("qkv_proj", prefix),
+        )
+        self.o_proj = RowParallelLinear(
+            self.total_num_heads * self.head_dim,
+            hidden_size,
+            bias=False,
+            quant_config=quant_config,
+            prefix=add_prefix("o_proj", prefix),
+        )
+
+        self.rotary_emb = get_rope(
+            self.head_dim,
+            rotary_dim=self.head_dim,
+            max_position=max_position_embeddings,
+            base=rope_theta,
+            rope_scaling=rope_scaling,
+            dual_chunk_attention_config=dual_chunk_attention_config,
+        )
+        self.attn = 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),
+        )
+
+    def forward(
+        self,
+        positions: torch.Tensor,
+        hidden_states: torch.Tensor,
+        forward_batch: ForwardBatch,
+    ) -> 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(q, k, v, forward_batch)
+        output, _ = self.o_proj(attn_output)
+        return output
+
+
+class Qwen2DecoderLayer(nn.Module):
+    def __init__(
+        self,
+        config: Qwen2Config,
+        layer_id: int = 0,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+        alt_stream: Optional[torch.cuda.Stream] = None,
+    ) -> None:
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        rope_theta = getattr(config, "rope_theta", 1000000)
+        rope_scaling = getattr(config, "rope_scaling", None)
+        max_position_embeddings = getattr(config, "max_position_embeddings", 32768)
+        head_dim = getattr(config, "head_dim", None)
+        dual_chunk_attention_config = getattr(
+            config, "dual_chunk_attention_config", None
+        )
+        self.self_attn = Qwen2Attention(
+            hidden_size=self.hidden_size,
+            num_heads=config.num_attention_heads,
+            num_kv_heads=config.num_key_value_heads,
+            head_dim=head_dim,
+            layer_id=layer_id,
+            rope_theta=rope_theta,
+            rope_scaling=rope_scaling,
+            max_position_embeddings=max_position_embeddings,
+            quant_config=quant_config,
+            dual_chunk_attention_config=dual_chunk_attention_config,
+            prefix=add_prefix("self_attn", prefix),
+        )
+        self.mlp = Qwen2MLP(
+            hidden_size=self.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,
+        residual: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        # Self Attention
+        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,
+        )
+
+        # Fully Connected
+        hidden_states, residual = self.post_attention_layernorm(hidden_states, residual)
+        hidden_states = self.mlp(hidden_states)
+        return hidden_states, residual
+
+
+class Qwen2Model(nn.Module):
+    def __init__(
+        self,
+        config: Qwen2Config,
+        quant_config: Optional[QuantizationConfig] = None,
+        prefix: str = "",
+        decoder_layer_type: type[nn.Module] = Qwen2DecoderLayer,
+        alt_stream: Optional[torch.cuda.Stream] = None,
+    ) -> None:
+        super().__init__()
+        self.config = config
+        self.padding_idx = config.pad_token_id
+        self.vocab_size = config.vocab_size
+        self.pp_group = get_pp_group()
+
+        if self.pp_group.is_first_rank:
+            self.embed_tokens = VocabParallelEmbedding(
+                config.vocab_size,
+                config.hidden_size,
+                quant_config=quant_config,
+                enable_tp=not global_server_args_dict["enable_dp_attention"],
+                prefix=add_prefix("embed_tokens", prefix),
+            )
+        else:
+            self.embed_tokens = PPMissingLayer()
+
+        # Use the provided decoder layer type or default to Qwen2DecoderLayer
+        decoder_layer_type = decoder_layer_type or Qwen2DecoderLayer
+        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 support
+        self.layers_to_capture = []
+
+    def get_input_embedding(self, input_ids: torch.Tensor) -> torch.Tensor:
+        if hasattr(self.config, "scale_emb"):
+            return self.get_input_embeddings()(input_ids) * self.config.scale_emb
+        else:
+            return self.get_input_embeddings()(input_ids)
+
+    def get_input_embeddings(self) -> nn.Embedding:
+        return self.embed_tokens
+
+    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 = []
+        for i in range(self.start_layer, self.end_layer):
+            if i in self.layers_to_capture:
+                aux_hidden_states.append(
+                    hidden_states + residual if residual is not None else hidden_states
+                )
+            layer = self.layers[i]
+            hidden_states, residual = layer(
+                positions,
+                hidden_states,
+                forward_batch,
+                residual,
+            )
+        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
+
+    # If this function is called, it should always initialize KV cache scale
+    # factors (or else raise an exception). Thus, handled exceptions should
+    # make sure to leave KV cache scale factors in a known good (dummy) state
+    def load_kv_cache_scales(self, quantization_param_path: str) -> None:
+        tp_size = get_tensor_model_parallel_world_size()
+        tp_rank = get_tensor_model_parallel_rank()
+        for layer_idx, scaling_factor in kv_cache_scales_loader(
+            quantization_param_path,
+            tp_rank,
+            tp_size,
+            self.config.num_hidden_layers,
+            self.config.__class__.model_type,
+        ):
+            if not isinstance(self.layers[layer_idx], nn.Identity):
+                layer_self_attn = self.layers[layer_idx].self_attn
+            if hasattr(layer_self_attn.attn, "k_scale"):
+                layer_self_attn.attn.k_scale = scaling_factor
+                layer_self_attn.attn.v_scale = scaling_factor
+            else:
+                raise RuntimeError(
+                    "Self attention has no KV cache scaling " "factor attribute!"
+                )
+
+
+class Qwen2ForCausalLM(nn.Module):
+    # BitandBytes specific attributes
+    default_bitsandbytes_target_modules = [
+        ".gate_proj.",
+        ".down_proj.",
+        ".up_proj.",
+        ".q_proj.",
+        ".k_proj.",
+        ".v_proj.",
+        ".o_proj.",
+    ]
+    bitsandbytes_stacked_params_mapping = {
+        # shard_name, weight_name, index
+        "q_proj": ("qkv_proj", 0),
+        "k_proj": ("qkv_proj", 1),
+        "v_proj": ("qkv_proj", 2),
+        "gate_proj": ("gate_up_proj", 0),
+        "up_proj": ("gate_up_proj", 1),
+    }
+
+    def __init__(
+        self,
+        config: Qwen2Config,
+        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 = Qwen2Model(
+            config, quant_config=quant_config, prefix=add_prefix("model", prefix)
+        )
+        self.capture_aux_hidden_states = False
+
+        # 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)
+
+    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
+
+    @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,
+    ):
+        start, end = split_interval
+        # embed
+        if start == 0:
+            if input_embeds is None:
+                forward_batch.hidden_states = self.model.embed_tokens(input_ids)
+            else:
+                forward_batch.hidden_states = input_embeds
+        # decoder layer
+        for i in range(start, end):
+            layer = self.model.layers[i]
+            forward_batch.hidden_states, forward_batch.residual = layer(
+                positions,
+                forward_batch.hidden_states,
+                forward_batch,
+                forward_batch.residual,
+            )
+
+        if end == self.model.config.num_hidden_layers:
+            # norm
+            hidden_states, _ = self.model.norm(
+                forward_batch.hidden_states, forward_batch.residual
+            )
+            forward_batch.hidden_states = hidden_states
+            # logits process
+            result = self.logits_processor(
+                input_ids, forward_batch.hidden_states, self.lm_head, forward_batch
+            )
+        else:
+            result = None
+
+        return result
+
+    @property
+    def start_layer(self):
+        return self.model.start_layer
+
+    @property
+    def end_layer(self):
+        return self.model.end_layer
+
+    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", "gate_proj", 0),
+            ("gate_up_proj", "up_proj", 1),
+        ]
+
+        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 "rotary_emb.cos_cached" in name or "rotary_emb.sin_cached" in name:
+                # Models trained using ColossalAI may include these tensors in
+                # the checkpoint. Skip them.
+                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
+            if name.startswith("model.vision_tower") and name not in params_dict:
+                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 load_kv_cache_scales(self, quantization_param_path: str) -> None:
+        self.model.load_kv_cache_scales(quantization_param_path)
+    
+    def set_eagle3_layers_to_capture(self, layer_ids: Optional[List[int]] = None):
+        if not self.pp_group.is_last_rank:
+            return
+
+        self.capture_aux_hidden_states = True
+        if layer_ids is None:
+            num_layers = self.config.num_hidden_layers
+            self.model.layers_to_capture = [
+                2,
+                num_layers // 2,
+                num_layers - 3,
+            ]  # Specific layers for EAGLE3 support
+        else:
+            self.model.layers_to_capture = [val + 1 for val in layer_ids]
+
+
+EntryClass = Qwen2ForCausalLM