Unsloth的 MCP 服务器 - 一个使 LLM 微调速度提高 2 倍且内存减少 80% 的库。
Unsloth 是一个可以显著提高微调大型语言模型效率的库：
速度：与标准方法相比，微调速度提高 2 倍
内存：VRAM 使用量减少 80%，允许在消费级 GPU 上对更大的模型进行微调
上下文长度：上下文长度最多增加 13 倍（例如，80GB GPU 上的 Llama 3.3 有 89K 个令牌）
准确性：模型质量或性能没有损失
Unsloth 通过用 OpenAI 的 Triton 语言编写的自定义 CUDA 内核、优化的反向传播和动态 4 位量化实现这些改进。
Related MCP server: TOON MCP Server
针对 Llama、Mistral、Phi、Gemma 等模型进行优化微调
4 位量化，实现高效训练
扩展上下文长度支持
用于模型加载、微调和推理的简单 API
导出为各种格式（GGUF、Hugging Face 等）
安装 Unsloth： pip install unsloth
安装并构建服务器：
cd unsloth-server
npm install
npm run build添加到 MCP 设置：
{
"mcpServers": {
"unsloth-server": {
"command": "node",
"args": ["/path/to/unsloth-server/build/index.js"],
"env": {
"HUGGINGFACE_TOKEN": "your_token_here" // Optional
},
"disabled": false,
"autoApprove": []
}
}
}验证 Unsloth 是否已正确安装在您的系统上。
参数：无
例子：
const result = await use_mcp_tool({
server_name: "unsloth-server",
tool_name: "check_installation",
arguments: {}
});获取 Unsloth 支持的所有模型的列表，包括 Llama、Mistral、Phi 和 Gemma 变体。
参数：无
例子：
const result = await use_mcp_tool({
server_name: "unsloth-server",
tool_name: "list_supported_models",
arguments: {}
});加载具有 Unsloth 优化的预训练模型，以实现更快的推理和微调。
参数：
model_name （必需）：要加载的模型的名称（例如，“unsloth/Llama-3.2-1B”）
max_seq_length （可选）：模型的最大序列长度（默认值：2048）
load_in_4bit （可选）：是否以 4 位量化加载模型（默认值：true）
use_gradient_checkpointing （可选）：是否使用梯度检查点来节省内存（默认值：true）
例子：
const result = await use_mcp_tool({
server_name: "unsloth-server",
tool_name: "load_model",
arguments: {
model_name: "unsloth/Llama-3.2-1B",
max_seq_length: 4096,
load_in_4bit: true
}
});使用 LoRA/QLoRA 技术通过 Unsloth 优化对模型进行微调。
参数：
model_name （必需）：需要微调的模型的名称
dataset_name （必需）：用于微调的数据集的名称
output_dir （必需）：保存微调模型的目录
max_seq_length （可选）：训练的最大序列长度（默认值：2048）
lora_rank （可选）：LoRA 微调的排名（默认值：16）
lora_alpha （可选）：LoRA 微调的 Alpha（默认值：16）
batch_size （可选）：训练的批次大小（默认值：2）
gradient_accumulation_steps （可选）：梯度累积步数（默认值：4）
learning_rate （可选）：训练的学习率（默认值：2e-4）
max_steps （可选）：最大训练步数（默认值：100）
dataset_text_field （可选）：数据集中包含文本的字段（默认值：“text”）
load_in_4bit （可选）：是否使用4位量化（默认值：true）
例子：
const result = await use_mcp_tool({
server_name: "unsloth-server",
tool_name: "finetune_model",
arguments: {
model_name: "unsloth/Llama-3.2-1B",
dataset_name: "tatsu-lab/alpaca",
output_dir: "./fine-tuned-model",
max_steps: 100,
batch_size: 2,
learning_rate: 2e-4
}
});使用经过微调的 Unsloth 模型生成文本。
参数：
model_path （必需）：微调模型的路径
prompt （必需）：提示文本生成
max_new_tokens （可选）：要生成的最大令牌数（默认值：256）
temperature （可选）：文本生成的温度（默认值：0.7）
top_p （可选）：文本生成的 Top-p（默认值：0.9）
例子：
const result = await use_mcp_tool({
server_name: "unsloth-server",
tool_name: "generate_text",
arguments: {
model_path: "./fine-tuned-model",
prompt: "Write a short story about a robot learning to paint:",
max_new_tokens: 512,
temperature: 0.8
}
});将经过微调的 Unsloth 模型导出为各种格式以供部署。
参数：
model_path （必需）：微调模型的路径
export_format （必需）：导出格式（gguf、ollama、vllm、huggingface）
output_path （必需）：保存导出模型的路径
quantization_bits （可选）：量化位（用于 GGUF 导出）（默认值：4）
例子：
const result = await use_mcp_tool({
server_name: "unsloth-server",
tool_name: "export_model",
arguments: {
model_path: "./fine-tuned-model",
export_format: "gguf",
output_path: "./exported-model.gguf",
quantization_bits: 4
}
});您可以通过正确格式化自定义数据集并将其托管在 Hugging Face 上或提供本地路径来使用它们：
const result = await use_mcp_tool({
server_name: "unsloth-server",
tool_name: "finetune_model",
arguments: {
model_name: "unsloth/Llama-3.2-1B",
dataset_name: "json",
data_files: {"train": "path/to/your/data.json"},
output_dir: "./fine-tuned-model"
}
});对于有限硬件上的大型模型：
减少批次大小并增加梯度积累步骤
使用 4 位量化
启用梯度检查点
如果可能的话减少序列长度
CUDA 内存不足：减少批量大小，使用 4 位量化，或尝试更小的模型
导入错误：确保安装了正确版本的 torch、transformers 和 unsloth
未找到模型：请检查您是否正在使用受支持的模型名称或是否有权访问私有模型
Python：3.10、3.11 或 3.12（不是 3.13）
CUDA：建议使用 11.8 或 12.1+
PyTorch：推荐 2.0+
模型 | 显存 | 不懒惰的速度 | 显存减少 | 上下文长度 |
骆驼 3.3 (70B) | 80GB | 速度快 2 倍 | >75% | 13倍长 |
骆驼 3.1 (8B) | 80GB | 速度快 2 倍 | >70% | 12倍长 |
米斯特拉尔 v0.3 (7B) | 80GB | 速度快 2.2 倍 | 减少 75% | - |
Python 3.10-3.12
支持 CUDA 的 NVIDIA GPU（推荐）
Node.js 和 npm
Apache-2.0
Check if Unsloth is properly installed
| Name | Required | Description | Default |
|---|---|---|---|
No parameters | |||
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries full burden for behavioral disclosure. It states what the tool does but doesn't describe what 'properly installed' means, what checks are performed, what output format to expect, or whether this has side effects. For a diagnostic tool with zero annotation coverage, this leaves significant behavioral questions unanswered.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is a single, efficient sentence that directly states the tool's purpose with zero wasted words. It's appropriately sized for a simple diagnostic tool and front-loads the essential information.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a zero-parameter diagnostic tool with no output schema, the description states the basic purpose adequately. However, it doesn't explain what constitutes 'properly installed' or what format the result will take, leaving the agent uncertain about how to interpret the tool's output. Given the simplicity of the tool, this is minimally viable but has clear gaps.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The tool has zero parameters with 100% schema description coverage, so the baseline is 4. The description appropriately doesn't discuss parameters since none exist, and the schema already fully documents this.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the tool's purpose: checking if Unsloth is properly installed. It uses a specific verb ('check') and identifies the target resource ('Unsloth installation'). However, it doesn't differentiate from siblings like 'list_supported_models' which might also involve installation status checks.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description provides no guidance on when to use this tool versus alternatives. It doesn't mention prerequisites, timing considerations, or relationships to sibling tools like 'list_supported_models' or 'load_model' that might be used before or after installation checks.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Export a fine-tuned Unsloth model to various formats
| Name | Required | Description | Default |
|---|---|---|---|
| model_path | Yes | Path to the fine-tuned model | |
| export_format | Yes | Format to export to (gguf, ollama, vllm, huggingface) | |
| output_path | Yes | Path to save the exported model | |
| quantization_bits | No | Bits for quantization (for GGUF export) |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
No annotations are provided, so the description carries the full burden of behavioral disclosure. It states the tool exports models but doesn't describe what the export entails (e.g., file creation, format conversion, potential data loss, permissions required, or rate limits). For a tool with 4 parameters and no annotations, this leaves significant gaps in understanding its behavior and side effects.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is a single, efficient sentence that front-loads the core purpose ('Export a fine-tuned Unsloth model') and adds necessary detail ('to various formats'). There is no wasted verbiage, and it's appropriately sized for the tool's complexity.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given 4 parameters with full schema coverage but no annotations and no output schema, the description is minimally adequate. It covers the basic purpose but lacks behavioral context, usage guidelines, and output details. For a tool that likely involves file operations and format conversions, more completeness would be beneficial, but it meets a bare minimum.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 100%, so the schema already documents all parameters thoroughly. The description adds minimal value beyond the schema by mentioning 'various formats,' which aligns with the 'export_format' enum but doesn't provide additional syntax or usage details. Baseline 3 is appropriate as the schema does the heavy lifting.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the action ('Export') and resource ('a fine-tuned Unsloth model'), specifying the target formats ('various formats'). It distinguishes from siblings like 'finetune_model' or 'load_model' by focusing on export rather than creation or loading. However, it doesn't explicitly differentiate from all siblings (e.g., 'generate_text' is clearly different, but the distinction is implicit rather than explicit).
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description provides no guidance on when to use this tool versus alternatives. It doesn't mention prerequisites (e.g., needing a fine-tuned model first), exclusions (e.g., not for raw models), or comparisons to sibling tools like 'list_supported_models' for checking export options. Usage is implied by the action but lacks explicit context.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Fine-tune a model with Unsloth optimizations
| Name | Required | Description | Default |
|---|---|---|---|
| model_name | Yes | Name of the model to fine-tune | |
| dataset_name | Yes | Name of the dataset to use for fine-tuning | |
| output_dir | Yes | Directory to save the fine-tuned model | |
| max_seq_length | No | Maximum sequence length for training | |
| lora_rank | No | Rank for LoRA fine-tuning | |
| lora_alpha | No | Alpha for LoRA fine-tuning | |
| batch_size | No | Batch size for training | |
| gradient_accumulation_steps | No | Number of gradient accumulation steps | |
| learning_rate | No | Learning rate for training | |
| max_steps | No | Maximum number of training steps | |
| dataset_text_field | No | Field in the dataset containing the text | |
| load_in_4bit | No | Whether to use 4-bit quantization |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
No annotations are provided, so the description carries the full burden of behavioral disclosure. It mentions 'Unsloth optimizations' but doesn't explain what this entails (e.g., performance improvements, memory efficiency). Critical behavioral traits like computational cost, time requirements, potential data destruction, or output format are omitted. The description adds minimal context beyond the basic action.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is a single, efficient sentence with zero waste. It is front-loaded with the core action ('fine-tune a model') and includes a relevant detail ('with Unsloth optimizations') that adds value without verbosity. Every word earns its place.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the complexity of a fine-tuning operation (12 parameters, no annotations, no output schema), the description is incomplete. It lacks information on behavioral aspects (e.g., runtime, resource usage), output (what is returned), and usage guidelines. The high parameter count and absence of structured support fields make this inadequate for such a complex tool.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 100%, so the schema already documents all 12 parameters thoroughly (e.g., 'model_name', 'dataset_name', 'lora_rank'). The description adds no additional meaning about parameters, such as typical values, constraints, or relationships between them. Baseline score of 3 is appropriate as the schema does the heavy lifting.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the action ('fine-tune') and resource ('a model') with a specific optimization method ('with Unsloth optimizations'). It distinguishes from siblings like 'export_model' or 'generate_text' by focusing on training rather than inference or export. However, it doesn't explicitly differentiate from 'load_model' in terms of when to use each.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description provides no guidance on when to use this tool versus alternatives. It doesn't mention prerequisites (e.g., needing a pre-trained model and dataset), when not to use it (e.g., for inference tasks), or refer to sibling tools like 'list_supported_models' for model selection. Usage is implied but not explicitly stated.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Generate text using a fine-tuned Unsloth model
| Name | Required | Description | Default |
|---|---|---|---|
| model_path | Yes | Path to the fine-tuned model | |
| prompt | Yes | Prompt for text generation | |
| max_new_tokens | No | Maximum number of tokens to generate | |
| temperature | No | Temperature for text generation | |
| top_p | No | Top-p for text generation |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description carries the full burden but only states the basic function without disclosing behavioral traits like computational cost, rate limits, error handling, or output format. It mentions 'fine-tuned Unsloth model' but doesn't explain implications for performance or compatibility.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is a single, efficient sentence that directly states the tool's purpose without unnecessary words. It is appropriately sized and front-loaded, with every part contributing essential information.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the complexity of a text generation tool with 5 parameters, no annotations, and no output schema, the description is incomplete. It fails to address key aspects like what the output looks like, error conditions, or how it integrates with sibling tools (e.g., requiring 'load_model' first).
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 100%, so the schema fully documents all 5 parameters. The description adds no additional meaning beyond what's in the schema, such as typical values for 'temperature' or 'top_p', or how 'model_path' relates to other tools. Baseline 3 is appropriate as the schema handles parameter documentation.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the action ('Generate text') and specifies the resource ('using a fine-tuned Unsloth model'), which distinguishes it from siblings like 'finetune_model' or 'export_model'. However, it doesn't explicitly differentiate from 'load_model' in terms of when to use each for text generation.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
No guidance is provided on when to use this tool versus alternatives like 'load_model' (which might be a prerequisite) or other text generation methods. The description lacks context about prerequisites, such as needing a loaded model, or exclusions.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
List all models supported by Unsloth
| Name | Required | Description | Default |
|---|---|---|---|
No parameters | |||
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
No annotations are provided, so the description carries the full burden. It states what the tool does but lacks behavioral details such as whether it requires authentication, has rate limits, returns paginated results, or what format the output takes. This is a significant gap for a tool with no annotation coverage.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is a single, efficient sentence that directly states the tool's purpose without any unnecessary words. It is appropriately sized and front-loaded, making it easy to understand at a glance.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the tool's simplicity (0 parameters, no output schema, no annotations), the description is minimally adequate. It explains what the tool does but lacks context on output format, behavioral traits, or usage guidelines. This makes it complete enough for basic understanding but with clear gaps in practical application.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The tool has 0 parameters, and schema description coverage is 100%, so there are no parameters to document. The description doesn't need to add parameter semantics, and it appropriately doesn't mention any. A baseline of 4 is applied as per the rules for zero parameters.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the verb ('List') and resource ('all models supported by Unsloth'), providing a specific purpose. However, it doesn't differentiate from sibling tools like 'check_installation' or 'load_model' which might also involve model information, so it doesn't reach the highest score.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
No guidance is provided on when to use this tool versus alternatives. Sibling tools like 'check_installation' or 'load_model' might overlap in context, but the description doesn't mention any prerequisites, exclusions, or specific scenarios for usage.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Load a pretrained model with Unsloth optimizations
| Name | Required | Description | Default |
|---|---|---|---|
| model_name | Yes | Name of the model to load (e.g., "unsloth/Llama-3.2-1B") | |
| max_seq_length | No | Maximum sequence length for the model | |
| load_in_4bit | No | Whether to load the model in 4-bit quantization | |
| use_gradient_checkpointing | No | Whether to use gradient checkpointing to save memory |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries the full burden of behavioral disclosure. It mentions 'Unsloth optimizations' but doesn't explain what this entails (e.g., memory efficiency, speed improvements) or potential side effects (e.g., memory usage, time to load, compatibility issues). It lacks details on error handling, performance implications, or what 'load' means operationally (e.g., into memory for subsequent use).
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
The description is a single, efficient sentence with zero waste. It front-loads the core action and key detail ('Unsloth optimizations'), making it easy to parse. Every word earns its place without redundancy or unnecessary elaboration.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
Given the tool's complexity (loading models with optimizations), lack of annotations, and no output schema, the description is incomplete. It doesn't cover what happens after loading (e.g., model availability for other tools), potential errors, or the impact of Unsloth optimizations. For a tool with 4 parameters and no structured safety or output info, more context is needed to guide effective use.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 100%, so the schema fully documents all parameters. The description adds no additional meaning beyond what's in the schema (e.g., it doesn't clarify the purpose of 'max_seq_length' or trade-offs for 'load_in_4bit'). Baseline score of 3 is appropriate as the schema handles parameter documentation adequately.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description clearly states the action ('Load') and resource ('a pretrained model'), and specifies the optimization framework ('with Unsloth optimizations'). It distinguishes from siblings like 'finetune_model' (training) and 'list_supported_models' (listing), but doesn't explicitly differentiate from 'export_model' or 'generate_text' in terms of loading vs. using the model.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
No guidance is provided on when to use this tool versus alternatives. It doesn't mention prerequisites (e.g., after checking installation with 'check_installation'), when loading is needed (e.g., before fine-tuning or generation), or what happens if the model is already loaded. The description assumes context without explicit direction.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
check_installationexport_modelfinetune_modelgenerate_textlist_supported_modelsload_modelScored across 6 tools
Each tool has a clearly distinct purpose with no overlap: checking installation, listing models, loading models, fine-tuning, generating text, and exporting models. The descriptions make it unambiguous which tool to use for each step in the workflow.
All tools follow a consistent verb_noun pattern with clear, descriptive names (e.g., check_installation, export_model, finetune_model). There are no deviations in naming conventions or style mixing.
With 6 tools, this server is well-scoped for its purpose of fine-tuning and using models with Unsloth optimizations. Each tool earns its place by covering a distinct, essential operation in the model lifecycle.
The tool set provides complete coverage for the domain: from checking prerequisites and listing models, through loading, fine-tuning, and generating text, to exporting the final model. There are no obvious gaps or dead ends in the workflow.
LLM chat, text tools, image generation, editing, batch image jobs, and asynchronous video generation
Image, video, chat and speech models with one account — local (npx) or hosted with OAuth sign-in.
Turn any LLM multimodal; generate images, voices, videos, 3D models, music, and more.
Provision private AI model endpoints on dedicated GPUs (Llama, Qwen, Mistral). Pay per minute.
