fix: add bos token

README.md

@@ -1,226 +1,192 @@
----
-license: apache-2.0
-language:
-- zh
-- en
-pipeline_tag: text-generation
-library_name: transformers
----
-<div align="center">
-<img src="https://github.com/OpenBMB/MiniCPM/blob/main/assets/minicpm_logo.png?raw=true" width="500em" ></img> 
-</div>
-
-<p align="center">
-<a href="https://github.com/OpenBMB/MiniCPM/" target="_blank">GitHub Repo</a> |
-<a href="https://github.com/OpenBMB/MiniCPM/tree/main/report/MiniCPM_4_Technical_Report.pdf" target="_blank">Technical Report</a> 
-</p>
-<p align="center">
-👋 Join us on <a href="https://discord.gg/3cGQn9b3YM" target="_blank">Discord</a> and <a href="https://github.com/OpenBMB/MiniCPM/blob/main/assets/wechat.jpg" target="_blank">WeChat</a>
-</p>
-
-## What's New
-- [2025.06.06] **MiniCPM4** series are released! This model achieves ultimate efficiency improvements while maintaining optimal performance at the same scale! It can achieve over 5x generation acceleration on typical end-side chips! You can find technical report [here](https://github.com/OpenBMB/MiniCPM/tree/main/report/MiniCPM_4_Technical_Report.pdf).🔥🔥🔥
-
-## MiniCPM4 Series
-MiniCPM4 series are highly efficient large language models (LLMs) designed explicitly for end-side devices, which achieves this efficiency through systematic innovation in four key dimensions: model architecture, training data, training algorithms, and inference systems.
-- [MiniCPM4-8B](https://huggingface.co/openbmb/MiniCPM4-8B): The flagship of MiniCPM4, with 8B parameters, trained on 8T tokens.
-- [MiniCPM4-0.5B](https://huggingface.co/openbmb/MiniCPM4-0.5B): The small version of MiniCPM4, with 0.5B parameters, trained on 1T tokens. (**<-- you are here**)
-- [MiniCPM4-8B-Eagle-FRSpec](https://huggingface.co/openbmb/MiniCPM4-8B-Eagle-FRSpec): Eagle head for FRSpec, accelerating speculative inference for MiniCPM4-8B.
-- [MiniCPM4-8B-Eagle-FRSpec-QAT-cpmcu](https://huggingface.co/openbmb/MiniCPM4-8B-Eagle-FRSpec-QAT-cpmcu): Eagle head trained with QAT for FRSpec, efficiently integrate speculation and quantization to achieve ultra acceleration for MiniCPM4-8B.
-- [MiniCPM4-8B-Eagle-vLLM](https://huggingface.co/openbmb/MiniCPM4-8B-Eagle-vLLM): Eagle head in vLLM format, accelerating speculative inference for MiniCPM4-8B.
-- [MiniCPM4-8B-marlin-Eagle-vLLM](https://huggingface.co/openbmb/MiniCPM4-8B-marlin-Eagle-vLLM): Quantized Eagle head for vLLM format, accelerating speculative inference for MiniCPM4-8B.
-- [BitCPM4-0.5B](https://huggingface.co/openbmb/BitCPM4-0.5B): Extreme ternary quantization applied to MiniCPM4-0.5B compresses model parameters into ternary values, achieving a 90% reduction in bit width.
-- [BitCPM4-1B](https://huggingface.co/openbmb/BitCPM4-1B): Extreme ternary quantization applied to MiniCPM3-1B compresses model parameters into ternary values, achieving a 90% reduction in bit width.
-- [MiniCPM4-Survey](https://huggingface.co/openbmb/MiniCPM4-Survey): Based on MiniCPM4-8B, accepts users' quiries as input and autonomously generate trustworthy, long-form survey papers.
-- [MiniCPM4-MCP](https://huggingface.co/openbmb/MiniCPM4-MCP): Based on MiniCPM4-8B, accepts users' queries and available MCP tools as input and autonomously calls relevant MCP tools to satisfy users' requirements.
-
-## Introduction
-MiniCPM 4 is an extremely efficient edge-side large model that has undergone efficient optimization across four dimensions: model architecture, learning algorithms, training data, and inference systems, achieving ultimate efficiency improvements.
-
-- 🏗️ **Efficient Model Architecture:**
-  - InfLLM v2 -- Trainable Sparse Attention Mechanism: Adopts a trainable sparse attention mechanism architecture where each token only needs to compute relevance with less than 5% of tokens in 128K long text processing, significantly reducing computational overhead for long texts
-
-- 🧠 **Efficient Learning Algorithms:**
-  - Model Wind Tunnel 2.0 -- Efficient Predictable Scaling: Introduces scaling prediction methods for performance of downstream tasks, enabling more precise model training configuration search
-  - BitCPM -- Ultimate Ternary Quantization: Compresses model parameter bit-width to 3 values, achieving 90% extreme model bit-width reduction
-  - Efficient Training Engineering Optimization: Adopts FP8 low-precision computing technology combined with Multi-token Prediction training strategy
-
-- 📚 **High-Quality Training Data:**
-  - UltraClean -- High-quality Pre-training Data Filtering and Generation: Builds iterative data cleaning strategies based on efficient data verification, open-sourcing high-quality Chinese and English pre-training dataset [UltraFinweb](https://huggingface.co/datasets/openbmb/Ultra-FineWeb)
-  - UltraChat v2 -- High-quality Supervised Fine-tuning Data Generation: Constructs large-scale high-quality supervised fine-tuning datasets covering multiple dimensions including knowledge-intensive data, reasoning-intensive data, instruction-following data, long text understanding data, and tool calling data
-
-- ⚡ **Efficient Inference System:**
-  - CPM.cu -- Lightweight and Efficient CUDA Inference Framework: Integrates sparse attention, model quantization, and speculative sampling to achieve efficient prefilling and decoding
-  - ArkInfer -- Cross-platform Deployment System: Supports efficient deployment across multiple backend environments, providing flexible cross-platform adaptation capabilities
-
-## Usage
-### Inference with Transformers
-```python
-from transformers import AutoModelForCausalLM, AutoTokenizer
-import torch
-torch.manual_seed(0)
-
-path = 'openbmb/MiniCPM4-0.5B'
-device = "cuda"
-tokenizer = AutoTokenizer.from_pretrained(path)
-model = AutoModelForCausalLM.from_pretrained(path, torch_dtype=torch.bfloat16, device_map=device, trust_remote_code=True)
-
-# User can directly use the chat interface
-responds, history = model.chat(tokenizer, "Write an article about Artificial Intelligence.", temperature=0.7, top_p=0.7)
-print(responds)
-
-# User can also use the generate interface
-# messages = [
-#     {"role": "user", "content": "Write an article about Artificial Intelligence."},
-# ]
-# prompt_text = tokenizer.apply_chat_template(
-#     messages,
-#     tokenize=False,
-#     add_generation_prompt=True,
-# )
-# model_inputs = tokenizer([prompt_text], return_tensors="pt").to(device)
-
-# model_outputs = model.generate(
-#     **model_inputs,
-#     max_new_tokens=1024,
-#     top_p=0.7,
-#     temperature=0.7
-# )
-# output_token_ids = [
-#     model_outputs[i][len(model_inputs[i]):] for i in range(len(model_inputs['input_ids']))
-# ]
-
-# responses = tokenizer.batch_decode(output_token_ids, skip_special_tokens=True)[0]
-# print(responses)
-```
-
-### Inference with [SGLang](https://github.com/sgl-project/sglang)
-
-For now, you need to install our forked version of SGLang.
-```bash
-git clone -b openbmb https://github.com/OpenBMB/sglang.git
-cd sglang
-
-pip install --upgrade pip
-pip install -e "python[all]"
-```
-
-You can start the inference server by running the following command:
-```bash
-python -m sglang.launch_server --model openbmb/MiniCPM4-0.5B --trust-remote-code --port 30000 --chat-template chatml
-```
-
-Then you can use the chat interface by running the following command:
-```python
-import openai
-
-client = openai.Client(base_url=f"http://localhost:30000/v1", api_key="None")
-
-response = client.chat.completions.create(
-    model="openbmb/MiniCPM4-0.5B",
-    messages=[
-        {"role": "user", "content": "Write an article about Artificial Intelligence."},
-    ],
-    temperature=0.7,
-    max_tokens=1024,
-)
-
-print(response.choices[0].message.content)
-```
-
-### Inference with [vLLM](https://github.com/vllm-project/vllm)
-For now, you need to install the latest version of vLLM.
-```
-pip install -U vllm \
-    --pre \
-    --extra-index-url https://wheels.vllm.ai/nightly
-```
-
-Then you can inference MiniCPM4-0.5B with vLLM:
-```python
-from transformers import AutoTokenizer
-from vllm import LLM, SamplingParams
-
-model_name = "openbmb/MiniCPM4-0.5B"
-prompt = [{"role": "user", "content": "Please recommend 5 tourist attractions in Beijing. "}]
-
-tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
-input_text = tokenizer.apply_chat_template(prompt, tokenize=False, add_generation_prompt=True)
-
-llm = LLM(
-    model=model_name,
-    trust_remote_code=True,
-    max_num_batched_tokens=32768, 
-    dtype="bfloat16", 
-    gpu_memory_utilization=0.8, 
-)
-sampling_params = SamplingParams(top_p=0.7, temperature=0.7, max_tokens=1024, repetition_penalty=1.02)
-
-outputs = llm.generate(prompts=input_text, sampling_params=sampling_params)
-
-print(outputs[0].outputs[0].text)
-```
-
-Also, you can start the inference server by running the following command:
-> **Note**: In vLLM's chat API, `add_special_tokens` is `False` by default. This means important special tokens—such as the beginning-of-sequence (BOS) token—will not be added automatically. To ensure the input prompt is correctly formatted for the model, you should explicitly set `extra_body={"add_special_tokens": True}`.
-
-```bash
-vllm serve openbmb/MiniCPM4-0.5B 
-```
-
-Then you can use the chat interface by running the following code:
-
-```python
-import openai
-
-client = openai.Client(base_url="http://localhost:8000/v1", api_key="EMPTY")
-
-response = client.chat.completions.create(
-    model="openbmb/MiniCPM4-0.5B",
-    messages=[
-        {"role": "user", "content": "Write an article about Artificial Intelligence."},
-    ],
-    temperature=0.7,
-    max_tokens=1024,
-    extra_body=dict(add_special_tokens=True),  # Ensures special tokens are added for chat template
-    
-)
-
-print(response.choices[0].message.content)
-```
-
-
-## Evaluation Results
-On two typical end-side chips, Jetson AGX Orin and RTX 4090, MiniCPM4 demonstrates significantly faster processing speed compared to similar-size models in long text processing tasks. As text length increases, MiniCPM4's efficiency advantage becomes more pronounced. On the Jetson AGX Orin platform, compared to Qwen3-8B, MiniCPM4 achieves approximately 7x decoding speed improvement.
-
-![benchmark](https://github.com/OpenBMB/MiniCPM/blob/main/assets/minicpm4/efficiency.png?raw=true)
-
-#### Comprehensive Evaluation
-MiniCPM4 launches end-side versions with 8B and 0.5B parameter scales, both achieving best-in-class performance in their respective categories.
-
-![benchmark](https://github.com/OpenBMB/MiniCPM/blob/main/assets/minicpm4/benchmark.png?raw=true)
-
-#### Long Text Evaluation
-MiniCPM4 is pre-trained on 32K long texts and achieves length extension through YaRN technology. In the 128K long text needle-in-a-haystack task, MiniCPM4 demonstrates outstanding performance.
-
-![long-niah](https://github.com/OpenBMB/MiniCPM/blob/main/assets/minicpm4/128k-niah.png?raw=true)
-
-## Statement
-- As a language model, MiniCPM generates content by learning from a vast amount of text. 
-- However, it does not possess the ability to comprehend or express personal opinions or value judgments. 
-- Any content generated by MiniCPM does not represent the viewpoints or positions of the model developers. 
-- Therefore, when using content generated by MiniCPM, users should take full responsibility for evaluating and verifying it on their own.
-
-## LICENSE
-- This repository and MiniCPM models are released under the [Apache-2.0](https://github.com/OpenBMB/MiniCPM/blob/main/LICENSE) License. 
-
-## Citation
-- Please cite our [paper](https://github.com/OpenBMB/MiniCPM/tree/main/report/MiniCPM_4_Technical_Report.pdf) if you find our work valuable.
-
-```bibtex
-@article{minicpm4,
-  title={{MiniCPM4}: Ultra-Efficient LLMs on End Devices},
-  author={MiniCPM Team},
-  year={2025}
-}
-```
+---
+license: apache-2.0
+language:
+- zh
+- en
+pipeline_tag: text-generation
+library_name: transformers
+---
+<div align="center">
+<img src="https://github.com/OpenBMB/MiniCPM/blob/main/assets/minicpm_logo.png?raw=true" width="500em" ></img> 
+</div>
+
+<p align="center">
+<a href="https://github.com/OpenBMB/MiniCPM/" target="_blank">GitHub Repo</a> |
+<a href="https://github.com/OpenBMB/MiniCPM/tree/main/report/MiniCPM_4_Technical_Report.pdf" target="_blank">Technical Report</a> 
+</p>
+<p align="center">
+👋 Join us on <a href="https://discord.gg/3cGQn9b3YM" target="_blank">Discord</a> and <a href="https://github.com/OpenBMB/MiniCPM/blob/main/assets/wechat.jpg" target="_blank">WeChat</a>
+</p>
+
+## What's New
+- [2025.06.06] **MiniCPM4** series are released! This model achieves ultimate efficiency improvements while maintaining optimal performance at the same scale! It can achieve over 5x generation acceleration on typical end-side chips! You can find technical report [here](https://github.com/OpenBMB/MiniCPM/tree/main/report/MiniCPM_4_Technical_Report.pdf).🔥🔥🔥
+
+## MiniCPM4 Series
+MiniCPM4 series are highly efficient large language models (LLMs) designed explicitly for end-side devices, which achieves this efficiency through systematic innovation in four key dimensions: model architecture, training data, training algorithms, and inference systems.
+- [MiniCPM4-8B](https://huggingface.co/openbmb/MiniCPM4-8B): The flagship of MiniCPM4, with 8B parameters, trained on 8T tokens.
+- [MiniCPM4-0.5B](https://huggingface.co/openbmb/MiniCPM4-0.5B): The small version of MiniCPM4, with 0.5B parameters, trained on 1T tokens. (**<-- you are here**)
+- [MiniCPM4-8B-Eagle-FRSpec](https://huggingface.co/openbmb/MiniCPM4-8B-Eagle-FRSpec): Eagle head for FRSpec, accelerating speculative inference for MiniCPM4-8B.
+- [MiniCPM4-8B-Eagle-FRSpec-QAT-cpmcu](https://huggingface.co/openbmb/MiniCPM4-8B-Eagle-FRSpec-QAT-cpmcu): Eagle head trained with QAT for FRSpec, efficiently integrate speculation and quantization to achieve ultra acceleration for MiniCPM4-8B.
+- [MiniCPM4-8B-Eagle-vLLM](https://huggingface.co/openbmb/MiniCPM4-8B-Eagle-vLLM): Eagle head in vLLM format, accelerating speculative inference for MiniCPM4-8B.
+- [MiniCPM4-8B-marlin-Eagle-vLLM](https://huggingface.co/openbmb/MiniCPM4-8B-marlin-Eagle-vLLM): Quantized Eagle head for vLLM format, accelerating speculative inference for MiniCPM4-8B.
+- [BitCPM4-0.5B](https://huggingface.co/openbmb/BitCPM4-0.5B): Extreme ternary quantization applied to MiniCPM4-0.5B compresses model parameters into ternary values, achieving a 90% reduction in bit width.
+- [BitCPM4-1B](https://huggingface.co/openbmb/BitCPM4-1B): Extreme ternary quantization applied to MiniCPM3-1B compresses model parameters into ternary values, achieving a 90% reduction in bit width.
+- [MiniCPM4-Survey](https://huggingface.co/openbmb/MiniCPM4-Survey): Based on MiniCPM4-8B, accepts users' quiries as input and autonomously generate trustworthy, long-form survey papers.
+- [MiniCPM4-MCP](https://huggingface.co/openbmb/MiniCPM4-MCP): Based on MiniCPM4-8B, accepts users' queries and available MCP tools as input and autonomously calls relevant MCP tools to satisfy users' requirements.
+
+## Introduction
+MiniCPM 4 is an extremely efficient edge-side large model that has undergone efficient optimization across four dimensions: model architecture, learning algorithms, training data, and inference systems, achieving ultimate efficiency improvements.
+
+- 🏗️ **Efficient Model Architecture:**
+  - InfLLM v2 -- Trainable Sparse Attention Mechanism: Adopts a trainable sparse attention mechanism architecture where each token only needs to compute relevance with less than 5% of tokens in 128K long text processing, significantly reducing computational overhead for long texts
+
+- 🧠 **Efficient Learning Algorithms:**
+  - Model Wind Tunnel 2.0 -- Efficient Predictable Scaling: Introduces scaling prediction methods for performance of downstream tasks, enabling more precise model training configuration search
+  - BitCPM -- Ultimate Ternary Quantization: Compresses model parameter bit-width to 3 values, achieving 90% extreme model bit-width reduction
+  - Efficient Training Engineering Optimization: Adopts FP8 low-precision computing technology combined with Multi-token Prediction training strategy
+
+- 📚 **High-Quality Training Data:**
+  - UltraClean -- High-quality Pre-training Data Filtering and Generation: Builds iterative data cleaning strategies based on efficient data verification, open-sourcing high-quality Chinese and English pre-training dataset [UltraFinweb](https://huggingface.co/datasets/openbmb/Ultra-FineWeb)
+  - UltraChat v2 -- High-quality Supervised Fine-tuning Data Generation: Constructs large-scale high-quality supervised fine-tuning datasets covering multiple dimensions including knowledge-intensive data, reasoning-intensive data, instruction-following data, long text understanding data, and tool calling data
+
+- ⚡ **Efficient Inference System:**
+  - CPM.cu -- Lightweight and Efficient CUDA Inference Framework: Integrates sparse attention, model quantization, and speculative sampling to achieve efficient prefilling and decoding
+  - ArkInfer -- Cross-platform Deployment System: Supports efficient deployment across multiple backend environments, providing flexible cross-platform adaptation capabilities
+
+## Usage
+### Inference with Transformers
+```python
+from transformers import AutoModelForCausalLM, AutoTokenizer
+import torch
+torch.manual_seed(0)
+
+path = 'openbmb/MiniCPM4-0.5B'
+device = "cuda"
+tokenizer = AutoTokenizer.from_pretrained(path)
+model = AutoModelForCausalLM.from_pretrained(path, torch_dtype=torch.bfloat16, device_map=device, trust_remote_code=True)
+
+# User can directly use the chat interface
+responds, history = model.chat(tokenizer, "Write an article about Artificial Intelligence.", temperature=0.7, top_p=0.7)
+print(responds)
+
+# User can also use the generate interface
+# messages = [
+#     {"role": "user", "content": "Write an article about Artificial Intelligence."},
+# ]
+# model_inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(device)
+
+# model_outputs = model.generate(
+#     model_inputs,
+#     max_new_tokens=1024,
+#     top_p=0.7,
+#     temperature=0.7
+# )
+# output_token_ids = [
+#     model_outputs[i][len(model_inputs[i]):] for i in range(len(model_inputs))
+# ]
+
+# responses = tokenizer.batch_decode(output_token_ids, skip_special_tokens=True)[0]
+# print(responses)
+```
+
+### Inference with [SGLang](https://github.com/sgl-project/sglang)
+
+For now, you need to install our forked version of SGLang.
+```bash
+git clone -b openbmb https://github.com/OpenBMB/sglang.git
+cd sglang
+
+pip install --upgrade pip
+pip install -e "python[all]"
+```
+
+You can start the inference server by running the following command:
+```bash
+python -m sglang.launch_server --model openbmb/MiniCPM4-8B --trust-remote-code --port 30000 --chat-template chatml
+```
+
+Then you can use the chat interface by running the following command:
+```python
+import openai
+
+client = openai.Client(base_url=f"http://localhost:30000/v1", api_key="None")
+
+response = client.chat.completions.create(
+    model="openbmb/MiniCPM4-8B",
+    messages=[
+        {"role": "user", "content": "Write an article about Artificial Intelligence."},
+    ],
+    temperature=0.7,
+    max_tokens=1024,
+)
+
+print(response.choices[0].message.content)
+```
+
+### Inference with [vLLM](https://github.com/vllm-project/vllm)
+For now, you need to install the latest version of vLLM.
+```
+pip install -U vllm \
+    --pre \
+    --extra-index-url https://wheels.vllm.ai/nightly
+```
+
+Then you can inference MiniCPM4-8B with vLLM:
+```python
+from transformers import AutoTokenizer
+from vllm import LLM, SamplingParams
+
+model_name = "openbmb/MiniCPM4-8B"
+prompt = [{"role": "user", "content": "Please recommend 5 tourist attractions in Beijing. "}]
+
+tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
+input_text = tokenizer.apply_chat_template(prompt, tokenize=False, add_generation_prompt=True)
+
+llm = LLM(
+    model=model_name,
+    trust_remote_code=True,
+    max_num_batched_tokens=32768, 
+    dtype="bfloat16", 
+    gpu_memory_utilization=0.8, 
+)
+sampling_params = SamplingParams(top_p=0.7, temperature=0.7, max_tokens=1024, repetition_penalty=1.02)
+
+outputs = llm.generate(prompts=input_text, sampling_params=sampling_params)
+
+print(outputs[0].outputs[0].text)
+```
+
+## Evaluation Results
+On two typical end-side chips, Jetson AGX Orin and RTX 4090, MiniCPM4 demonstrates significantly faster processing speed compared to similar-size models in long text processing tasks. As text length increases, MiniCPM4's efficiency advantage becomes more pronounced. On the Jetson AGX Orin platform, compared to Qwen3-8B, MiniCPM4 achieves approximately 7x decoding speed improvement.
+
+![benchmark](https://github.com/OpenBMB/MiniCPM/blob/main/assets/minicpm4/efficiency.png?raw=true)
+
+#### Comprehensive Evaluation
+MiniCPM4 launches end-side versions with 8B and 0.5B parameter scales, both achieving best-in-class performance in their respective categories.
+
+![benchmark](https://github.com/OpenBMB/MiniCPM/blob/main/assets/minicpm4/benchmark.png?raw=true)
+
+#### Long Text Evaluation
+MiniCPM4 is pre-trained on 32K long texts and achieves length extension through YaRN technology. In the 128K long text needle-in-a-haystack task, MiniCPM4 demonstrates outstanding performance.
+
+![long-niah](https://github.com/OpenBMB/MiniCPM/blob/main/assets/minicpm4/128k-niah.png?raw=true)
+
+## Statement
+- As a language model, MiniCPM generates content by learning from a vast amount of text. 
+- However, it does not possess the ability to comprehend or express personal opinions or value judgments. 
+- Any content generated by MiniCPM does not represent the viewpoints or positions of the model developers. 
+- Therefore, when using content generated by MiniCPM, users should take full responsibility for evaluating and verifying it on their own.
+
+## LICENSE
+- This repository and MiniCPM models are released under the [Apache-2.0](https://github.com/OpenBMB/MiniCPM/blob/main/LICENSE) License. 
+
+## Citation
+- Please cite our [paper](https://github.com/OpenBMB/MiniCPM/tree/main/report/MiniCPM_4_Technical_Report.pdf) if you find our work valuable.
+
+```bibtex
+@article{minicpm4,
+  title={{MiniCPM4}: Ultra-Efficient LLMs on End Devices},
+  author={MiniCPM Team},
+  year={2025}
+}
+```

modeling_minicpm.py

@@ -24,7 +24,7 @@ import torch.utils.checkpoint
 from torch import nn
 from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
 from transformers.activations import ACT2FN
-from transformers.cache_utils import Cache, DynamicCache, CacheLayerMixin, DynamicLayer
+from transformers.cache_utils import Cache, DynamicCache
 from transformers.modeling_attn_mask_utils import (
     AttentionMaskConverter,
     _prepare_4d_attention_mask,
@@ -52,9 +52,493 @@ from .configuration_minicpm import MiniCPMConfig
 try:
     from flash_attn import flash_attn_func, flash_attn_varlen_func
     from flash_attn.bert_padding import index_first_axis, pad_input, unpad_input  # noqa
+    from infllm_v2 import (
+        infllmv2_attn_stage1,
+        infllmv2_attn_varlen_func,
+        infllmv2_attn_with_kvcache,
+        max_pooling_1d,
+    )
 except:
     pass
 
+from functools import lru_cache
+
+
+def compressed_attention(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    kernel_size: int,
+    kernel_stride: int,
+    block_size: int,
+    topk: int,
+    cu_seqlens_q: torch.Tensor,
+    cu_seqlens_k: torch.Tensor,
+    max_seqlen_q: int,
+    max_seqlen_k: int,
+    sm_scale: float = None,
+    init_blocks: int = 1,
+    local_blocks: int = 2,
+    parallel_topk_compute: Union[str, bool] = 'auto',
+    total_seq_lens=-1,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    """Attention between query and compressed key and value. Compute attention output and topk block idx used in topk_sparse_attention.
+
+    Args:
+        q (torch.Tensor): shape [total_q_len, num_q_heads, head_dim]
+        k (torch.Tensor): shape [total_kv_len, num_kv_heads, head_dim]
+        v (torch.Tensor): shape [total_kv_len, num_kv_heads, head_dim]
+        kernel_size (int): kernel size in compress_key_value
+        kernel_stride (int): stride of compress_key_value
+        block_size (int): key value block size for topk sparse attention.
+        topk (int): number of blocks for each query.
+        cu_seqlens_q (torch.Tensor): shape [batch_size + 1], similar to cu_seqlens_q in flash_attn_func_varlen.
+        cu_seqlens_k (torch.Tensor): shape [batch_size + 1], similar to cu_seqlens_k in flash_attn_func_varlen.
+        max_seqlen_q (int): max q len of the batch.
+        max_seqlen_k (int): max k len of the batch.
+        sm_scale (float, optional): softmax scale. Defaults to None, means 1/sqrt(head_dim).
+        init_blocks (int, optional): Number of init blocks for each query. Defaults to 1.
+        local_blocks (int, optional): Number of local blocks for each query. Defaults to 2.
+        parallel_topk_compute (str, optional): Only set it to False when the sequence length is too long. This can avoid a current bug.
+            We'll fix this issue later. Defaults to auto, it will be set to False when the sequence length is greater than 32k and True otherwise.
+
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: attention output and topk_idx used in topk_sparse_attention
+    """
+    with torch.no_grad():
+        cache_len = 0
+        batch_size = cu_seqlens_q.shape[0] - 1
+        if total_seq_lens == -1:
+            total_seq_lens = max_seqlen_q
+            q_idx = torch.cat(
+                [
+                    torch.arange(cu_seqlens_q[i + 1] - cu_seqlens_q[i], device=q.device) + total_seq_lens - (cu_seqlens_q[i + 1] - cu_seqlens_q[i])
+                    for i in range(batch_size)
+                ],
+                dim=0,
+            )
+            q_idx = q_idx // block_size
+
+        else:
+            cache_len = total_seq_lens - max_seqlen_q
+            assert batch_size == 1, 'batch_size must be 1 when total_seq_lens is set'
+            q_idx = torch.tensor([total_seq_lens - 1], device=q.device, dtype=torch.int32) // block_size
+
+        score = infllmv2_attn_stage1(
+            q.contiguous(),
+            k.contiguous(),
+            v.contiguous(),
+            cu_seqlens_q=cu_seqlens_q,
+            cu_seqlens_k=cu_seqlens_k,
+            max_seqlen_q=max_seqlen_q,
+            max_seqlen_k=max_seqlen_k,
+            causal=q_idx.shape[0] > 1)
+        score = score[:, :q_idx.shape[0], :]
+
+        # Replace transform_score with max_pooling_1d
+        block_score = max_pooling_1d(
+            score.contiguous(),
+            cache_len=cache_len,
+            local_blocks=local_blocks,
+            init_blocks=init_blocks,
+            block_size=block_size,
+            stride=kernel_stride,
+        )
+        # get topk
+        topk = min(topk, block_score.shape[-1])
+        topk_idx = block_score.topk(topk, dim=-1).indices.sort(-1).values
+        topk_idx[topk_idx >= q_idx[None, :, None]] = -1
+        topk_idx = topk_idx.to(torch.int32)
+
+    return topk_idx
+
+
+@lru_cache(maxsize=16)
+def calc_chunks_with_stride(cu_seqlen, chunk_size, kernel_stride):
+    """
+    Compute the chunks that require Sparse attention, with stride support.
+
+    Args:
+        cu_seqlen (torch.Tensor): Cumulative sequence lengths for each sample.
+        chunk_size (int): Chunk size used for Sparse attention.
+        kernel_stride (int): Stride size when sliding over the sequence.
+
+    Returns:
+        filtered_indices (torch.Tensor): Indices used to directly index into the key/value tensors.
+        cu_seqlens_compressed (torch.Tensor): Cumulative sequence lengths after compression.
+    """
+    # 1. Compute the length of each sequence
+    batch_sizes = cu_seqlen[1:] - cu_seqlen[:-1]
+
+    # 2. Compute the start positions of chunks for each sequence (with stride)
+    max_seq_len = torch.max(batch_sizes)
+    max_num_chunks_per_seq = (max_seq_len - chunk_size) // kernel_stride + 1
+    chunk_start_offsets = torch.arange(0, max_num_chunks_per_seq * kernel_stride, kernel_stride, device=cu_seqlen.device)
+    seq_starts = cu_seqlen[:-1]
+    chunk_start_in_seq = seq_starts[:, None] + chunk_start_offsets[None, :]  # [batch_size, max_num_chunks_per_seq]
+
+    # 3. Filter out chunks that exceed sequence length or are smaller than the full chunk size
+    chunk_end_in_seq = chunk_start_in_seq + chunk_size
+    valid_chunk_mask = (chunk_end_in_seq <= (seq_starts[:, None] + batch_sizes[:, None]))
+
+    # 4. Filter valid chunk start positions using the valid_chunk_mask
+    valid_chunk_starts = chunk_start_in_seq[valid_chunk_mask]  # [num_valid_chunks]
+    del chunk_start_in_seq
+    # 5. Generate filtered_indices
+    chunk_indices = torch.arange(
+        0, chunk_size, device=cu_seqlen.device
+    )[None, :]  # [1, chunk_size]
+    filtered_indices = valid_chunk_starts[:, None] + chunk_indices  # [num_valid_chunks, chunk_size]
+    filtered_indices = filtered_indices.view(-1)  # Flatten to 1D indices
+
+    # 6. Compute compressed cumulative sequence lengths
+    num_filtered_chunks_per_batch = valid_chunk_mask.sum(dim=1)  # Number of valid chunks per batch
+    cu_seqlens_compressed = torch.zeros(
+        len(cu_seqlen), dtype=torch.int32, device=cu_seqlen.device
+    )
+    cu_seqlens_compressed[1:] = num_filtered_chunks_per_batch.cumsum(dim=0)
+    del num_filtered_chunks_per_batch, chunk_start_offsets, seq_starts, chunk_end_in_seq, valid_chunk_mask, chunk_indices
+    return filtered_indices, cu_seqlens_compressed
+
+
+class CompressK(torch.nn.Module):
+    def __init__(self, head_num_k, head_dim, kernel_size, kernel_stride=16):
+        """
+        Module for compressing key (K) representations.
+
+        Args:
+            head_num_k (int): Number of key attention heads.
+            head_dim (int): Dimension of each attention head.
+            kernel_size (int): Size of each chunk used for compression.
+            kernel_stride (int, optional): Stride used when dividing input into chunks. Default is 16.
+        """
+        super().__init__()
+        self.kernel_size = kernel_size
+        self.head_num_k = head_num_k
+        self.head_dim = head_dim
+        self.kernel_stride = kernel_stride
+
+    def forward(self, k: torch.Tensor, cu_seqlens):
+        """
+        Forward pass for compressing the key (K) tensor.
+
+        Args:
+            k (torch.Tensor): Input key tensor of shape (total_seq_len, num_heads, head_dim).
+            cu_seqlens (torch.Tensor): Cumulative sequence lengths for each sample in the batch, typically used for handling variable-length sequences.
+
+        Returns:
+            compress_k (torch.Tensor): Compressed key tensor.
+            cu_seqlens_compressed (torch.Tensor): Updated cumulative sequence lengths after compression.
+
+        """
+        # Compute chunk-related metadata, with stride support
+        filtered_k_indices, cu_seqlens_compressed = calc_chunks_with_stride(
+            cu_seqlens, self.kernel_size, self.kernel_stride
+        )
+
+        # Extract filtered key vectors
+        filtered_k = k.index_select(0, filtered_k_indices.view(-1))
+
+        # split
+        filtered_k = filtered_k.view(filtered_k.shape[0] // self.kernel_size, self.kernel_size, self.head_num_k, self.head_dim)  # [l, block_size,h,d]
+
+        compressed_k = filtered_k.mean(dim=1)
+        return compressed_k, cu_seqlens_compressed
+
+
+class DynamicCacheQKV(DynamicCache):
+    """
+    A cache that grows dynamically as more tokens are generated. This is the default for generative models.
+
+    It stores the Key and Value states as a list of tensors, one for each layer. The expected shape for each tensor is
+    `[batch_size, num_heads, seq_len, head_dim]`.
+
+    Example:
+        ```python
+        >>> from transformers import AutoTokenizer, AutoModelForCausalLM, DynamicCache
+
+        >>> model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2-0.5B-Instruct")
+        >>> tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen2-0.5B-Instruct")
+
+        >>> inputs = tokenizer(text="My name is Qwen2", return_tensors="pt")
+
+        >>> # Prepare a cache class and pass it to model's forward
+        >>> past_key_values = DynamicCache()
+        >>> outputs = model(**inputs, past_key_values=past_key_values, use_cache=True)
+        >>> outputs.past_key_values # access cache filled with key/values from generation
+        DynamicCache()
+        ```
+    """
+    def __init__(self, num_hidden_layers: Optional[int] = None) -> None:
+        super().__init__()
+        if num_hidden_layers is None:
+            self.key_cache: List[torch.Tensor] = []
+            self.value_cache: List[torch.Tensor] = []
+            self.compress_k_cache: List[torch.Tensor] = []
+            self.no_compress_k_cache: List[torch.Tensor] = []
+            self.cached_compressed_cu_seqlens: List[torch.Tensor] = []
+            self.no_rope_key_cache: List[torch.Tensor] = []
+        else:
+            self.key_cache: List[torch.Tensor] = [[] for _ in range(num_hidden_layers)]
+            self.value_cache: List[torch.Tensor] = [[] for _ in range(num_hidden_layers)]
+            self.compress_k_cache: List[torch.Tensor] = [[] for _ in range(num_hidden_layers)]
+            self.no_compress_k_cache: List[torch.Tensor] = [[] for _ in range(num_hidden_layers)]
+            self.cached_compressed_cu_seqlens: List[torch.Tensor] = [[] for _ in range(num_hidden_layers)]
+            self.no_rope_key_cache: List[torch.Tensor] = [[] for _ in range(num_hidden_layers)]
+        self._seen_tokens = 0  # Used in `generate` to keep tally of how many tokens the cache has seen
+
+    def __getitem__(self, layer_idx: int) -> List[Tuple[torch.Tensor]]:
+        """
+        Support for backwards-compatible `past_key_value` indexing, e.g. `past_key_value[0][0].shape[2]` to get the
+        sequence length.
+        """
+        if layer_idx < len(self):
+            return (self.key_cache[layer_idx], self.value_cache[layer_idx])
+        else:
+            raise KeyError(f'Cache only has {len(self)} layers, attempted to access layer with index {layer_idx}')
+
+    def __iter__(self):
+        """
+        Support for backwards-compatible `past_key_value` iteration, e.g. `for x in past_key_value:` to iterate over
+        keys and values
+        """
+        for layer_idx in range(len(self)):
+            yield (self.key_cache[layer_idx], self.value_cache[layer_idx])
+
+    def __len__(self):
+        """
+        Support for backwards-compatible `past_key_value` length, e.g. `len(past_key_value)`. This value corresponds
+        to the number of layers in the model.
+        """
+        return len(self.key_cache)
+
+    def update(
+        self,
+        key_states: torch.Tensor,
+        value_states: torch.Tensor,
+        layer_idx: int,
+        cache_kwargs: Optional[Dict[str, Any]] = None
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Updates the cache with the new `key_states` and `value_states` for the layer `layer_idx`.
+
+        Parameters:
+            key_states (`torch.Tensor`):
+                The new key states to cache.
+            value_states (`torch.Tensor`):
+                The new value states to cache.
+            layer_idx (`int`):
+                The index of the layer to cache the states for.
+            cache_kwargs (`Dict[str, Any]`, `optional`):
+                Additional arguments for the cache subclass. No additional arguments are used in `DynamicCache`.
+
+        Return:
+            A tuple containing the updated key and value states.
+        """
+        # Update the number of seen tokens
+        if layer_idx == 0:
+            self._seen_tokens += key_states.shape[-2]
+
+        # Update the cache
+        if len(self.key_cache) <= layer_idx:
+            self.key_cache.append(key_states)
+            self.value_cache.append(value_states)
+
+        # content on layer cache can be a tensor and checking not tensor causes errors
+        # so we explicitly check for the empty list
+        elif self.key_cache[layer_idx] == []:
+            self.key_cache[layer_idx] = key_states
+            self.value_cache[layer_idx] = value_states
+
+        else:
+            self.key_cache[layer_idx] = torch.cat([self.key_cache[layer_idx], key_states], dim=-2)
+            self.value_cache[layer_idx] = torch.cat([self.value_cache[layer_idx], value_states], dim=-2)
+        return self.key_cache[layer_idx], self.value_cache[layer_idx]
+
+    def update_no_rope_key(
+            self,
+            key_states: torch.Tensor,
+            layer_idx: int,
+            cache_kwargs: Optional[Dict[str, Any]] = None):
+
+        # Update the cache
+        if len(self.no_rope_key_cache) <= layer_idx:
+            self.no_rope_key_cache.append(key_states)
+
+        # content on layer cache can be a tensor and checking not tensor causes errors
+        # so we explicitly check for the empty list
+        elif self.no_rope_key_cache[layer_idx] == []:
+            self.no_rope_key_cache[layer_idx] = key_states
+        else:
+            self.no_rope_key_cache[layer_idx] = torch.cat([self.no_rope_key_cache[layer_idx], key_states], dim=1)
+        return self.no_rope_key_cache[layer_idx]
+
+    def update_compress_k(
+        self,
+        key_states: torch.Tensor,
+        layer_idx: int,
+        cache_kwargs: Optional[Dict[str, Any]] = None
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Updates the cache with the new `key_states` and `value_states` for the layer `layer_idx`.
+
+        Parameters:
+            key_states (`torch.Tensor`):
+                The new key states to cache.
+            value_states (`torch.Tensor`):
+                The new value states to cache.
+            layer_idx (`int`):
+                The index of the layer to cache the states for.
+            cache_kwargs (`Dict[str, Any]`, `optional`):
+                Additional arguments for the cache subclass. No additional arguments are used in `DynamicCache`.
+
+        Return:
+            A tuple containing the updated key and value states.
+        """
+
+        # Update the cache
+        if len(self.compress_k_cache) <= layer_idx:
+            self.compress_k_cache.append(key_states)
+
+        # content on layer cache can be a tensor and checking not tensor causes errors
+        # so we explicitly check for the empty list
+        elif self.compress_k_cache[layer_idx] == []:
+            self.compress_k_cache[layer_idx] = key_states
+        else:
+            self.compress_k_cache[layer_idx] = torch.cat([self.compress_k_cache[layer_idx], key_states], dim=0)
+        return self.compress_k_cache[layer_idx]
+
+    def update_no_compress_k(
+        self,
+        key_states: torch.Tensor,
+        layer_idx: int,
+        kernel_size: int = 32,
+        kernel_stride: int = 16,
+        cache_kwargs: Optional[Dict[str, Any]] = None
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Updates the cache with the new `key_states` and `value_states` for the layer `layer_idx`.
+
+        Parameters:
+            key_states (`torch.Tensor`):
+                The new key states to cache.
+            value_states (`torch.Tensor`):
+                The new value states to cache.
+            layer_idx (`int`):
+                The index of the layer to cache the states for.
+            cache_kwargs (`Dict[str, Any]`, `optional`):
+                Additional arguments for the cache subclass. No additional arguments are used in `DynamicCache`.
+
+        Return:
+            A tuple containing the updated key and value states.
+        """
+        # Update the cache
+        if len(self.no_compress_k_cache) <= layer_idx:
+            self.no_compress_k_cache.append(key_states)
+
+        # content on layer cache can be a tensor and checking not tensor causes errors
+        # so we explicitly check for the empty list
+        elif self.no_compress_k_cache[layer_idx] == []:
+            self.no_compress_k_cache[layer_idx] = key_states
+        else:
+            self.no_compress_k_cache[layer_idx] = torch.cat([self.no_compress_k_cache[layer_idx], key_states], dim=0)
+
+        current_len = self.no_compress_k_cache[layer_idx].shape[0]
+
+        if current_len >= kernel_size:
+            k_chunk = self.no_compress_k_cache[layer_idx][:kernel_size]
+            self.no_compress_k_cache[layer_idx] = self.no_compress_k_cache[layer_idx][kernel_stride:]
+            return k_chunk
+        else:
+            return None
+
+    def get_seq_length(self, layer_idx: Optional[int] = 0) -> int:
+        """Returns the sequence length of the cached states. A layer index can be optionally passed."""
+        # TODO: deprecate this function in favor of `cache_position`
+        if len(self.key_cache) <= layer_idx or (len(self.key_cache) > layer_idx and self.key_cache[layer_idx] == []):
+            return 0
+        return self.key_cache[layer_idx].shape[-2]
+
+    def get_max_length(self) -> Optional[int]:
+        """Returns the maximum sequence length of the cached states. DynamicCache does not have a maximum length."""
+        return None
+
+    def to_legacy_cache(self) -> Tuple[Tuple[torch.Tensor], Tuple[torch.Tensor]]:
+        """Converts the `DynamicCache` instance into the its equivalent in the legacy cache format. Used for
+        backward compatibility."""
+        legacy_cache = ()
+        for layer_idx in range(len(self)):
+            legacy_cache += ((self.key_cache[layer_idx], self.value_cache[layer_idx]),)
+        return legacy_cache
+
+    # @classmethod
+    # def from_legacy_cache(
+    #     cls, past_key_values: Optional[Tuple[Tuple[torch.FloatTensor]]] = None, num_hidden_layers: int = None
+    # ) -> "DynamicCacheQKV":
+    #     """Converts a cache in the legacy cache format into an equivalent `DynamicCache`. Used for
+    #     backward compatibility."""
+    #     cache = cls(num_hidden_layers)
+    #     if past_key_values is not None:
+    #         for layer_idx in range(len(past_key_values)):
+    #             key_states, value_states, query_status = past_key_values[layer_idx]
+    #             cache.update(key_states, value_states, query_status,layer_idx)
+    #     return cache
+
+    def crop(self, max_length: int):
+        """Crop the past key values up to a new `max_length` in terms of tokens. `max_length` can also be
+        negative to remove `max_length` tokens. This is used in assisted decoding and contrastive search."""
+        # In case it is negative
+        if max_length < 0:
+            max_length = self.get_seq_length() - abs(max_length)
+
+        if self.get_seq_length() <= max_length:
+            return
+
+        self._seen_tokens = max_length
+        for idx in range(len(self.key_cache)):
+            if self.key_cache[idx] != []:
+                self.key_cache[idx] = self.key_cache[idx][..., :max_length, :]
+                self.value_cache[idx] = self.value_cache[idx][..., :max_length, :]
+
+    def batch_split(self, full_batch_size: int, split_size: int, num_hidden_layers: int) -> List['DynamicCacheQKV']:
+        """Split the current instance into a list of `DynamicCache` by the batch size. This will be used by
+        `_split_model_inputs()` in `generation.utils`"""
+        out = []
+        for i in range(0, full_batch_size, split_size):
+            current_split = DynamicCacheQKV(num_hidden_layers)
+            current_split._seen_tokens = self._seen_tokens
+            current_split.key_cache = [tensor[i: i + split_size] for tensor in self.key_cache]
+            current_split.value_cache = [tensor[i: i + split_size] for tensor in self.value_cache]
+            out.append(current_split)
+        return out
+
+    @classmethod
+    def from_batch_splits(cls, splits: List['DynamicCacheQKV'], num_hidden_layers: int) -> 'DynamicCacheQKV':
+        """This is the opposite of the above `batch_split()` method. This will be used by `stack_model_outputs` in
+        `generation.utils`"""
+        cache = cls(num_hidden_layers)
+        for idx in range(len(splits[0])):
+            key_cache = [current.key_cache[idx] for current in splits if current.key_cache[idx] != []]
+            value_cache = [current.key_cache[idx] for current in splits if current.key_cache[idx] != []]
+            query_cache = [current.key_cache[idx] for current in splits if current.key_cache[idx] != []]
+            if key_cache != []:
+                layer_keys = torch.cat(key_cache, dim=0)
+                layer_values = torch.cat(value_cache, dim=0)
+                layer_query = torch.cat(query_cache, dim=0)
+                cache.update(layer_keys, layer_values, idx, query_states=layer_query)
+        return cache
+
+    def batch_repeat_interleave(self, repeats: int):
+        """Repeat the cache `repeats` times in the batch dimension. Used in contrastive search."""
+        for layer_idx in range(len(self)):
+            self.key_cache[layer_idx] = self.key_cache[layer_idx].repeat_interleave(repeats, dim=0)
+            self.value_cache[layer_idx] = self.value_cache[layer_idx].repeat_interleave(repeats, dim=0)
+
+    def batch_select_indices(self, indices: torch.Tensor):
+        """Only keep the `indices` in the batch dimension of the cache. Used in contrastive search."""
+        for layer_idx in range(len(self)):
+            self.key_cache[layer_idx] = self.key_cache[layer_idx][indices, ...]
+            self.value_cache[layer_idx] = self.value_cache[layer_idx][indices, ...]
 
 
 # This makes `_prepare_4d_causal_attention_mask` a leaf function in the FX graph.
@@ -83,6 +567,22 @@ def _get_unpad_data(attention_mask):
     )
 
 
+def _expand_mask(mask: torch.Tensor, dtype: torch.dtype, tgt_len: Optional[int] = None):
+    warnings.warn(
+        'Calling `transformers.models.minicpm.modeling_minicpm._prepare_4d_attention_mask` is deprecated and will be removed in v4.37. Use `transformers.modeling_attn_mask_utils._prepare_4d_attention_mask'
+    )
+    return _prepare_4d_attention_mask(mask=mask, dtype=dtype, tgt_len=tgt_len)
+
+
+def _make_causal_mask(
+    input_ids_shape: torch.Size, dtype: torch.dtype, device: torch.device, past_key_values_length: int = 0
+):
+    warnings.warn(
+        'Calling `transformers.models.minicpm.modeling_minicpm._make_causal_mask` is deprecated and will be removed in v4.37. Use `transformers.models.minicpm.modeling_minicpm.AttentionMaskConverter._make_causal_mask'
+    )
+    return AttentionMaskConverter._make_causal_mask(
+        input_ids_shape=input_ids_shape, dtype=dtype, device=device, past_key_values_length=past_key_values_length
+    )
 
 
 # @torch.jit.script  # type: ignore
@@ -296,21 +796,6 @@ class MiniCPMMLP(nn.Module):
 
         return down_proj
 
-def _unpad_one_tensor(hidden_states, attention_mask):
-    # Unpad the hidden states using the indices
-    indices, cu_seqlens, max_seqlen_in_batch = _get_unpad_data(attention_mask)
-    batch_size, seq_len = hidden_states.shape[:2]
-    
-    # Get the remaining dimensions
-    remaining_dims = hidden_states.shape[2:]
-    
-    # Reshape to (batch_size * seq_len, *remaining_dims)
-    reshaped_states = hidden_states.reshape(batch_size * seq_len, *remaining_dims)
-    
-    # Apply unpadding using indices
-    unpadded_states = index_first_axis(reshaped_states, indices)
-    
-    return unpadded_states, indices, cu_seqlens, max_seqlen_in_batch
 
 def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
     """
@@ -442,7 +927,15 @@ class MiniCPMAttention(nn.Module):
         key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
         value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
 
-        kv_seq_len = position_ids.max().item() + 1
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            if self.layer_idx is None:
+                raise ValueError(
+                    f'The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} '
+                    'for auto-regressive decoding with k/v caching, please make sure to initialize the attention class '
+                    'with a layer index.'
+                )
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
         cos, sin = self.rotary_emb(value_states.to(torch.float32), seq_len=kv_seq_len)
 
         query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
@@ -544,7 +1037,9 @@ class MiniCPMFlashAttention2(MiniCPMAttention):
         key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
         value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
 
-        kv_seq_len = position_ids.max().item() + 1
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
         cos, sin = self.rotary_emb(value_states.to(torch.float32), seq_len=kv_seq_len)
         query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
 
@@ -692,6 +1187,504 @@ class MiniCPMFlashAttention2(MiniCPMAttention):
         )
 
 
+class MiniCPMInfLLMv2Attention(MiniCPMAttention):
+    """
+    MiniCPM flash attention module. This module inherits from `MiniCPMAttention` as the weights of the module stays
+    untouched. The only required change would be on the forward pass where it needs to correctly call the public API of
+    flash attention and deal with padding tokens in case the input contains any of them.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.config._attn_implementation == 'flash_attention_2', 'Only flash_attention_2 is supported for sparse attention'
+        # TODO: Should be removed once Flash Attention for RoCm is bumped to 2.1.
+        # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignment, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+        # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+        self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+        #  -------sparse-------
+        self.kernel_size = self.config.sparse_config.get('kernel_size', 32)
+        self.kernel_stride = self.config.sparse_config.get('kernel_stride', 16)
+        self.init_blocks = self.config.sparse_config.get('init_blocks', 1)
+        self.block_size = self.config.sparse_config.get('block_size', 64)
+        self.window_size = self.config.sparse_config.get('window_size', 2048)
+        self.dense_len = self.config.sparse_config.get('dense_len', 8192)
+
+        self.local_blocks = self.window_size // self.block_size  # local_blocks
+        self.topk = self.config.sparse_config.get('topk', 64)
+        self.use_nope = self.config.sparse_config.get('use_nope', False)
+        self.compress_k = CompressK(self.num_key_value_heads, self.head_dim, kernel_size=self.kernel_size, kernel_stride=self.kernel_stride)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.LongTensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_value: Optional[Cache] = None,
+        output_attentions: bool = False,
+        use_cache: bool = False,
+        **kwargs,
+    ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+        # MiniCPMFlashAttention2 attention does not support output_attentions
+        if 'padding_mask' in kwargs:
+            warnings.warn(
+                'Passing `padding_mask` is deprecated and will be removed in v4.37. Please make sure use `attention_mask` instead.`'
+            )
+
+            # overwrite attention_mask with padding_mask
+            attention_mask = kwargs.pop('padding_mask')
+
+        output_attentions = False
+
+        bsz, q_len, _ = hidden_states.size()
+        assert bsz == 1, 'Only batch_size=1 is supported at the moment.'
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        # !save no rope
+        if self.use_nope:
+            query_states_no_rope = query_states.view(bsz, q_len, self.num_heads, self.head_dim)
+            key_states_no_rope = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim)
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        # therefore we just need to keep the original shape
+        query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+        cos, sin = self.rotary_emb(value_states.to(torch.float32), seq_len=kv_seq_len)
+        query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
+
+        if past_key_value is not None:
+            cache_kwargs = {'sin': sin, 'cos': cos}  # Specific to RoPE models
+            key_states, value_states = past_key_value.update(key_states, value_states, self.layer_idx, cache_kwargs)
+
+        # TODO: These transpose are quite inefficient but Flash Attention requires the layout [batch_size, sequence_length, num_heads, head_dim]. We would need to refactor the KV cache
+        # to be able to avoid many of these transpose/reshape/view.
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+        value_states = value_states.transpose(1, 2)
+        if self.use_nope:
+            no_rope_param = {
+                'key_states_no_rope': key_states_no_rope,
+                'query_states_no_rope': query_states_no_rope,
+            }
+            if kv_seq_len <= self.dense_len:
+                past_key_value.update_no_rope_key(key_states_no_rope, self.layer_idx)
+        else:
+            no_rope_param = None
+
+        dropout_rate = self.attention_dropout if self.training else 0.0
+
+        # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+        # therefore the input hidden states gets silently casted in float32. Hence, we need
+        # cast them back in the correct dtype just to be sure everything works as expected.
+        # This might slowdown training & inference so it is recommended to not cast the LayerNorms
+        # in fp32. (MiniCPMRMSNorm handles it correctly)
+
+        input_dtype = query_states.dtype
+        if input_dtype == torch.float32:
+            # Handle the case where the model is quantized
+            if hasattr(self.config, '_pre_quantization_dtype'):
+                target_dtype = self.config._pre_quantization_dtype
+            else:
+                target_dtype = self.q_proj.weight.dtype
+
+            logger.warning_once(
+                f'The input hidden states seems to be silently casted in float32, this might be related to'
+                f' the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in'
+                f' {target_dtype}.'
+            )
+
+            query_states = query_states.to(target_dtype)
+            key_states = key_states.to(target_dtype)
+            value_states = value_states.to(target_dtype)
+        if kv_seq_len < self.dense_len:
+            attn_output = self._flash_attention_forward_dense(
+                query_states, key_states, value_states, attention_mask, q_len, dropout=dropout_rate)
+        elif past_key_value is None or q_len != 1:    # prefilling
+            attn_output = self._flash_attention_forward(
+                query_states, key_states, value_states, attention_mask, q_len, dropout=dropout_rate,
+                no_rope_param=no_rope_param,  # if past_key_value is not None else None,
+                past_key_value=past_key_value)
+        else:
+            attn_output = self._flash_attention_forward_with_kv_cache(
+                query_states, key_states, value_states, attention_mask, q_len, dropout=dropout_rate, no_rope_param=no_rope_param, past_key_value=past_key_value)
+
+        attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+        attn_output = self.o_proj(attn_output)
+
+        if not output_attentions:
+            attn_weights = None
+
+        return attn_output, attn_weights, past_key_value
+
+    def _flash_attention_forward(
+        self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None, no_rope_param=None, past_key_value=None
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`int`, *optional*):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+        """
+        if not self._flash_attn_uses_top_left_mask:
+            causal = self.is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in MiniCPMFlashAttention2 __init__.
+            causal = self.is_causal and query_length != 1
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+            if no_rope_param is not None:
+                # nope unpad
+                no_rope_param['query_states_no_rope'] = no_rope_param['query_states_no_rope'].squeeze(0)
+                no_rope_param['key_states_no_rope'] = no_rope_param['key_states_no_rope'].squeeze(0)
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+            attn_output_unpad = self.sparse_forward(
+                query_states,
+                key_states,
+                value_states,
+                cu_seqlens_q,
+                cu_seqlens_k,
+                max_seqlen_in_batch_q,
+                max_seqlen_in_batch_k,
+                no_rope_param=no_rope_param,
+                past_key_value=past_key_value,
+            )
+
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            raise ValueError('Need attention mask')
+
+        return attn_output
+
+    def _flash_attention_forward_with_kv_cache(
+        self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None, no_rope_param=None, past_key_value=None
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`int`, *optional*):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+        """
+        if not self._flash_attn_uses_top_left_mask:
+            causal = self.is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in MiniCPMFlashAttention2 __init__.
+            causal = self.is_causal and query_length != 1
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+
+            batch_size = query_states.shape[0]
+
+            # query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+            #     query_states, key_states, value_states, attention_mask, query_length=query_length
+            # )
+
+            assert batch_size == 1, 'Only batch_size=1 is supported at the moment.'
+            # prepare past kv ,new kv
+            new_q = query_states
+
+            new_k = key_states[:, -1:, :, :].contiguous()
+            new_v = value_states[:, -1:, :, :].contiguous()
+
+            past_k = key_states[:, :-1, :, :].contiguous()
+            past_v = value_states[:, :-1, :, :].contiguous()
+            if no_rope_param is not None:
+                # nope unpad
+                no_rope_param['query_states_no_rope'] = no_rope_param['query_states_no_rope'].squeeze(0)
+                no_rope_param['key_states_no_rope'] = no_rope_param['key_states_no_rope'].squeeze(0)
+
+            attn_output = self.sparse_forward_with_kv_cache(
+                past_k=past_k, past_v=past_v, new_k=new_k, new_v=new_v, new_q=new_q, batch_size=batch_size, no_rope_param=no_rope_param, past_key_value=past_key_value)
+
+            # attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            raise ValueError('need attention mask')
+
+        return attn_output
+
+    def sparse_forward(self,
+                       query_layer,
+                       key_layer,
+                       value_layer,
+                       cu_seqlens_q,
+                       cu_seqlens_k,
+                       max_seqlen_in_batch_q,
+                       max_seqlen_in_batch_k,
+                       no_rope_param=None,
+                       past_key_value=None):
+        stage1_k = key_layer if no_rope_param is None else no_rope_param['key_states_no_rope']
+        compressed_k, compressed_cu_seqlens = self.compress_k(stage1_k, cu_seqlens_k)
+        compressed_v = compressed_k.clone()
+        if past_key_value is not None:
+            # Compute the start indices of keys (k) that were not compressed, Only batch_size=1 is supported at the moment.
+            no_compress_k_start = compressed_k.shape[0] * self.kernel_stride
+            past_key_value.update_compress_k(
+                compressed_k, self.layer_idx
+            )
+            past_key_value.update_no_compress_k(
+                key_layer[no_compress_k_start:], self.layer_idx, no_compress_k_start)
+            past_key_value.cached_compressed_cu_seqlens.append(compressed_cu_seqlens)
+        compressed_seqlens = compressed_cu_seqlens[1:] - compressed_cu_seqlens[:-1]
+        topk_idx = compressed_attention(
+            query_layer if no_rope_param is None else no_rope_param['query_states_no_rope'],
+            compressed_k,
+            compressed_v,
+            self.kernel_size,
+            self.kernel_stride,
+            self.block_size,
+            self.topk,
+            cu_seqlens_q,
+            compressed_cu_seqlens,
+            max_seqlen_in_batch_q,
+            compressed_seqlens.max().item(),
+            None,
+            init_blocks=self.init_blocks,
+            local_blocks=self.local_blocks,
+        )
+
+        topk_attn_output = infllmv2_attn_varlen_func(
+            query_layer,
+            key_layer,
+            value_layer,
+            cu_seqlens_q,
+            cu_seqlens_k,
+            max_seqlen_in_batch_q,
+            max_seqlen_in_batch_k,
+            dropout_p=0.0,
+            deterministic=False,
+            softmax_scale=None,
+            causal=True,
+            return_attn_probs=False,
+            block_window_size=self.window_size // self.block_size,
+            topk_idx=topk_idx
+        )
+
+        return topk_attn_output
+
+    def sparse_forward_with_kv_cache(self, past_k=None, past_v=None, new_k=None, new_v=None, new_q=None, batch_size=None, no_rope_param=None, past_key_value=None):
+
+        # stage1_k = new_k.squeeze(0) if no_rope_param is None else no_rope_param['key_states_no_rope']
+        if past_k.shape[1] + new_k.shape[1] == self.dense_len and (past_key_value.compress_k_cache == [] or len(past_key_value.compress_k_cache) < self.layer_idx + 1 or past_key_value.compress_k_cache[self.layer_idx] == []):
+            if no_rope_param is not None:
+                stage1_k = past_key_value.no_rope_key_cache[self.layer_idx].squeeze(0).contiguous()  # just batch_size ==1
+            else:
+                stage1_k = torch.cat([past_k, new_k], dim=1).contiguous().squeeze(0).contiguous()  # just batch_size ==1
+            compressed_k, compressed_cu_seqlens = self.compress_k(stage1_k, torch.tensor([0, stage1_k.shape[0]], device=stage1_k.device, dtype=torch.int32))  # just batch_size ==1
+
+            # Compute the start indices of keys (k) that were not compressed, Only batch_size=1 is supported at the moment.
+            no_compress_k_start = compressed_k.shape[0] * self.kernel_stride
+            past_key_value.update_compress_k(
+                compressed_k, self.layer_idx
+            )
+            past_key_value.update_no_compress_k(
+                stage1_k[no_compress_k_start:], self.layer_idx, no_compress_k_start)
+            past_key_value.cached_compressed_cu_seqlens.append(compressed_cu_seqlens)
+
+        else:
+            stage1_k = new_k.squeeze(0) if no_rope_param is None else no_rope_param['key_states_no_rope']
+            no_compress_k = past_key_value.update_no_compress_k(
+                stage1_k, self.layer_idx, kernel_stride=self.kernel_stride, kernel_size=self.kernel_size)
+            if no_compress_k is not None:
+                compressed_k = no_compress_k.mean(dim=0, keepdim=True)  # [1, n_heads_k, head_dim]
+
+                compressed_k = past_key_value.update_compress_k(
+                    compressed_k, self.layer_idx)  # [seqlen, nheads_k, head_dim]
+
+                past_key_value.cached_compressed_cu_seqlens[self.layer_idx][-1] += 1    # !Increment the last entry in sequence lengths by 1; currently supports only batch_size = 1
+                compressed_cu_seqlens = past_key_value.cached_compressed_cu_seqlens[self.layer_idx]
+            else:
+                compressed_k = past_key_value.compress_k_cache[self.layer_idx]  # [seqlen, nheads_k, head_dim]
+                compressed_cu_seqlens = past_key_value.cached_compressed_cu_seqlens[self.layer_idx]
+
+        compressed_v = compressed_k.clone()
+
+        compressed_seqlens = compressed_cu_seqlens[1:] - compressed_cu_seqlens[:-1]
+        torch.cuda.synchronize()
+        # Manually verify that the lengths match
+        assert compressed_k.shape[0] == compressed_seqlens.sum().item(), 'The length of compressed_k does not match the sum of compressed_seqlens'
+        topk_idx = compressed_attention(
+            new_q.squeeze(0).contiguous() if no_rope_param is None else no_rope_param['query_states_no_rope'],
+            compressed_k,
+            compressed_v,
+            self.kernel_size,
+            self.kernel_stride,
+            self.block_size,
+            self.topk,
+            torch.tensor([0, 1], device=compressed_k.device, dtype=torch.int32),
+            compressed_cu_seqlens,
+            1,
+            compressed_seqlens.max().item(),
+            None,
+            init_blocks=self.init_blocks,
+            local_blocks=self.local_blocks,
+            total_seq_lens=past_k.shape[1] + 1,  # !Only batch_size=1 is supported at the moment.
+        )
+
+        repeat_times = 1
+        if repeat_times > 1:
+            new_q = new_q.repeat_interleave(repeat_times, dim=-2)
+        else:
+            new_q = new_q
+
+        cache_batch_idx = torch.arange(batch_size, device=new_q.device, dtype=torch.int32)
+
+        seqlen_k = past_k.shape[1] + new_k.shape[1]  # !Only batch_size=1 is supported at the moment.
+        seqlens_k = torch.full((batch_size,), seqlen_k - 1, dtype=torch.int32, device=new_q.device)
+
+        past_k = torch.cat([past_k, torch.zeros_like(new_k, dtype=new_k.dtype)], dim=1).contiguous()   # Append one zero vector to avoid potential out-of-bounds access
+        past_v = torch.cat([past_v, torch.zeros_like(new_v, dtype=new_v.dtype)], dim=1).contiguous()   # Append one zero vector to avoid potential out-of-bounds access
+        topk_attn_output = infllmv2_attn_with_kvcache(
+            q=new_q,
+            k_cache=past_k,
+            v_cache=past_v,
+            topk_idx=topk_idx,
+            block_window_size=self.window_size // self.block_size,
+            k=new_k,                   # [batch_size, 1, nheads_k, d]
+            v=new_v,                   # [batch_size, 1, nheads_k, d]
+            cache_seqlens=seqlens_k,   # current_seqlens_k-1
+            rotary_cos=None,           # No rotary embeddings
+            rotary_sin=None,           # No rotary embeddings
+            cache_batch_idx=cache_batch_idx,
+            causal=False,              # Renaming to match function signature
+        )
+        return topk_attn_output
+
+    def _flash_attention_forward_dense(
+        self, query_states, key_states, value_states, attention_mask, query_length, dropout=0.0, softmax_scale=None
+    ):
+        """
+        Calls the forward method of Flash Attention - if the input hidden states contain at least one padding token
+        first unpad the input, then computes the attention scores and pad the final attention scores.
+
+        Args:
+            query_states (`torch.Tensor`):
+                Input query states to be passed to Flash Attention API
+            key_states (`torch.Tensor`):
+                Input key states to be passed to Flash Attention API
+            value_states (`torch.Tensor`):
+                Input value states to be passed to Flash Attention API
+            attention_mask (`torch.Tensor`):
+                The padding mask - corresponds to a tensor of size `(batch_size, seq_len)` where 0 stands for the
+                position of padding tokens and 1 for the position of non-padding tokens.
+            dropout (`int`, *optional*):
+                Attention dropout
+            softmax_scale (`float`, *optional*):
+                The scaling of QK^T before applying softmax. Default to 1 / sqrt(head_dim)
+        """
+        if not self._flash_attn_uses_top_left_mask:
+            causal = self.is_causal
+        else:
+            # TODO: Remove the `query_length != 1` check once Flash Attention for RoCm is bumped to 2.1. For details, please see the comment in MiniCPMFlashAttention2 __init__.
+            causal = self.is_causal and query_length != 1
+        # Contains at least one padding token in the sequence
+        if attention_mask is not None:
+            batch_size = query_states.shape[0]
+            query_states, key_states, value_states, indices_q, cu_seq_lens, max_seq_lens = self._upad_input(
+                query_states, key_states, value_states, attention_mask, query_length
+            )
+
+            cu_seqlens_q, cu_seqlens_k = cu_seq_lens
+            max_seqlen_in_batch_q, max_seqlen_in_batch_k = max_seq_lens
+            attn_output_unpad = flash_attn_varlen_func(
+                query_states,
+                key_states,
+                value_states,
+                cu_seqlens_q=cu_seqlens_q,
+                cu_seqlens_k=cu_seqlens_k,
+                max_seqlen_q=max_seqlen_in_batch_q,
+                max_seqlen_k=max_seqlen_in_batch_k,
+                dropout_p=dropout,
+                softmax_scale=softmax_scale,
+                causal=causal,
+            )
+
+            attn_output = pad_input(attn_output_unpad, indices_q, batch_size, query_length)
+        else:
+            attn_output = flash_attn_func(
+                query_states, key_states, value_states, dropout, softmax_scale=softmax_scale, causal=causal
+            )
+
+        return attn_output
+
+    def _upad_input(self, query_layer, key_layer, value_layer, attention_mask, query_length):
+        indices_k, cu_seqlens_k, max_seqlen_in_batch_k = _get_unpad_data(attention_mask)
+        batch_size, kv_seq_len, num_key_value_heads, head_dim = key_layer.shape
+
+        key_layer = index_first_axis(
+            key_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        value_layer = index_first_axis(
+            value_layer.reshape(batch_size * kv_seq_len, num_key_value_heads, head_dim), indices_k
+        )
+        if query_length == kv_seq_len:
+            query_layer = index_first_axis(
+                query_layer.reshape(batch_size * kv_seq_len, self.num_heads, head_dim), indices_k
+            )
+            cu_seqlens_q = cu_seqlens_k
+            max_seqlen_in_batch_q = max_seqlen_in_batch_k
+            indices_q = indices_k
+        elif query_length == 1:
+            max_seqlen_in_batch_q = 1
+            cu_seqlens_q = torch.arange(
+                batch_size + 1, dtype=torch.int32, device=query_layer.device
+            )  # There is a memcpy here, that is very bad.
+            indices_q = cu_seqlens_q[:-1]
+            query_layer = query_layer.squeeze(1)
+        else:
+            # The -q_len: slice assumes left padding.
+            attention_mask = attention_mask[:, -query_length:]
+            query_layer, indices_q, cu_seqlens_q, max_seqlen_in_batch_q = unpad_input(query_layer, attention_mask)
+
+        return (
+            query_layer,
+            key_layer,
+            value_layer,
+            indices_q,
+            (cu_seqlens_q, cu_seqlens_k),
+            (max_seqlen_in_batch_q, max_seqlen_in_batch_k),
+        )
+
+
 class MiniCPMSdpaAttention(MiniCPMAttention):
     """
     MiniCPM attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
@@ -734,7 +1727,9 @@ class MiniCPMSdpaAttention(MiniCPMAttention):
         key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
         value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
 
-        kv_seq_len = position_ids.max().item() + 1
+        kv_seq_len = key_states.shape[-2]
+        if past_key_value is not None:
+            kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
         cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
 
         query_states, key_states = apply_rotary_pos_emb(query_states, key_states, cos, sin, position_ids)
@@ -788,7 +1783,10 @@ class MiniCPMDecoderLayer(nn.Module):
     def __init__(self, config: MiniCPMConfig, layer_idx: int):
         super().__init__()
         self.hidden_size = config.hidden_size
-        self.self_attn = MINICPM_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)
+        if config.sparse_config is not None and torch.cuda.is_available():
+            raise NotImplementedError("MiniCPM4-0.5B does not support sparse attention yet.")
+        else:
+            self.self_attn = MINICPM_ATTENTION_CLASSES[config._attn_implementation](config=config, layer_idx=layer_idx)
 
         self.mlp = MiniCPMMLP(config)
         self.input_layernorm = MiniCPMRMSNorm(config.hidden_size, eps=config.rms_norm_eps)
@@ -1054,10 +2052,11 @@ class MiniCPMModel(MiniCPMPreTrainedModel):
                 raise ValueError(
                     'You must use the new past_key_values format, such as the Cache class, instead of the old tuple format.'
                 )
+                past_key_values = DynamicCache.from_legacy_cache(past_key_values)
 
-            # Calculate the usable length of past key values
-            past_key_values_length = past_key_values.get_seq_length() if isinstance(past_key_values, Cache) else 0
-
+            past_key_values_length = past_key_values.get_usable_length(seq_length)
+            if self.config.sparse_config is not None and torch.cuda.is_available() and past_key_values_length == 0:
+                past_key_values = DynamicCacheQKV()
 
         if position_ids is None:
             device = input_ids.device if input_ids is not None else inputs_embeds.device
@@ -1283,16 +2282,12 @@ class MiniCPMForCausalLM(MiniCPMPreTrainedModel):
     ):
         if past_key_values is not None:
             if isinstance(past_key_values, Cache):
-                # Use the new Cache class methods
                 cache_length = past_key_values.get_seq_length()
-
-                
-                past_length = cache_length
-                max_cache_length = None
+                past_length = past_key_values.seen_tokens
+                max_cache_length = None     # past_key_values.get_max_length()
             else:
-                raise ValueError(
-                    'You must use the new past_key_values format, such as the Cache class, instead of the old tuple format.'
-                )
+                cache_length = past_length = past_key_values[0][0].shape[2]
+                max_cache_length = None
 
             # Keep only the unprocessed tokens:
             # 1 - If the length of the attention_mask exceeds the length of input_ids, then we are in a setting where