@baimingtao/lbs-agent 1.0.0

package/extensions/skills/optional/mlops/guidance/references/backends.md

@@ -0,0 +1,553 @@
+# Backend 配置指南
+
+配置 Guidance 与不同 LLM backend 的完整指南。
+
+## 目录
+- 基于 API 的模型 (Anthropic, OpenAI)
+- 本地模型 (Transformers, llama.cpp)
+- Backend 对比
+- 性能调优
+- 高级配置
+
+## 基于 API 的模型
+
+### Anthropic Claude
+
+#### 基础配置
+
+```python
+from guidance import models
+
+# Using environment variable
+lm = models.Anthropic("claude-sonnet-4-5-20250929")
+# Reads ANTHROPIC_API_KEY from environment
+
+# Explicit API key
+lm = models.Anthropic(
+    model="claude-sonnet-4-5-20250929",
+    api_key="your-api-key-here"
+)
+```
+
+#### 可用模型
+
+```python
+# Claude 3.5 Sonnet (Latest, recommended)
+lm = models.Anthropic("claude-sonnet-4-5-20250929")
+
+# Claude 3.7 Sonnet (Fast, cost-effective)
+lm = models.Anthropic("claude-sonnet-3.7-20250219")
+
+# Claude 3 Opus (Most capable)
+lm = models.Anthropic("claude-3-opus-20240229")
+
+# Claude 3.5 Haiku (Fastest, cheapest)
+lm = models.Anthropic("claude-3-5-haiku-20241022")
+```
+
+#### 配置选项
+
+```python
+lm = models.Anthropic(
+    model="claude-sonnet-4-5-20250929",
+    api_key="your-api-key",
+    max_tokens=4096,           # Max tokens to generate
+    temperature=0.7,            # Sampling temperature (0-1)
+    top_p=0.9,                  # Nucleus sampling
+    timeout=30,                 # Request timeout (seconds)
+    max_retries=3              # Retry failed requests
+)
+```
+
+#### 结合 Context Managers 使用
+
+```python
+from guidance import models, system, user, assistant, gen
+
+lm = models.Anthropic("claude-sonnet-4-5-20250929")
+
+with system():
+    lm += "You are a helpful assistant."
+
+with user():
+    lm += "What is the capital of France?"
+
+with assistant():
+    lm += gen(max_tokens=50)
+
+print(lm)
+```
+
+### OpenAI
+
+#### 基础配置
+
+```python
+from guidance import models
+
+# Using environment variable
+lm = models.OpenAI("gpt-4o")
+# Reads OPENAI_API_KEY from environment
+
+# Explicit API key
+lm = models.OpenAI(
+    model="gpt-4o",
+    api_key="your-api-key-here"
+)
+```
+
+#### 可用模型
+
+```python
+# GPT-4o (Latest, multimodal)
+lm = models.OpenAI("gpt-4o")
+
+# GPT-4o Mini (Fast, cost-effective)
+lm = models.OpenAI("gpt-4o-mini")
+
+# GPT-4 Turbo
+lm = models.OpenAI("gpt-4-turbo")
+
+# GPT-3.5 Turbo (Cheapest)
+lm = models.OpenAI("gpt-3.5-turbo")
+```
+
+#### 配置选项
+
+```python
+lm = models.OpenAI(
+    model="gpt-4o-mini",
+    api_key="your-api-key",
+    max_tokens=2048,
+    temperature=0.7,
+    top_p=1.0,
+    frequency_penalty=0.0,
+    presence_penalty=0.0,
+    timeout=30
+)
+```
+
+#### Chat Format
+
+```python
+from guidance import models, gen
+
+lm = models.OpenAI("gpt-4o-mini")
+
+# OpenAI uses chat format
+lm += [
+    {"role": "system", "content": "You are a helpful assistant."},
+    {"role": "user", "content": "What is 2+2?"}
+]
+
+# Generate response
+lm += gen(max_tokens=50)
+```
+
+### Azure OpenAI
+
+```python
+from guidance import models
+
+lm = models.AzureOpenAI(
+    model="gpt-4o",
+    azure_endpoint="https://your-resource.openai.azure.com/",
+    api_key="your-azure-api-key",
+    api_version="2024-02-15-preview",
+    deployment_name="your-deployment-name"
+)
+```
+
+## 本地模型
+
+### Transformers (Hugging Face)
+
+#### 基础配置
+
+```python
+from guidance.models import Transformers
+
+# Load model from Hugging Face
+lm = Transformers("microsoft/Phi-4-mini-instruct")
+```
+
+#### GPU 配置
+
+```python
+# Use GPU
+lm = Transformers(
+    "microsoft/Phi-4-mini-instruct",
+    device="cuda"
+)
+
+# Use specific GPU
+lm = Transformers(
+    "microsoft/Phi-4-mini-instruct",
+    device="cuda:0"  # GPU 0
+)
+
+# Use CPU
+lm = Transformers(
+    "microsoft/Phi-4-mini-instruct",
+    device="cpu"
+)
+```
+
+#### 高级配置
+
+```python
+lm = Transformers(
+    "microsoft/Phi-4-mini-instruct",
+    device="cuda",
+    torch_dtype="float16",      # Use FP16 (faster, less memory)
+    load_in_8bit=True,          # 8-bit quantization
+    max_memory={0: "20GB"},     # GPU memory limit
+    offload_folder="./offload"  # Offload to disk if needed
+)
+```
+
+#### 常用模型
+
+```python
+# Phi-4 (Microsoft)
+lm = Transformers("microsoft/Phi-4-mini-instruct")
+lm = Transformers("microsoft/Phi-3-medium-4k-instruct")
+
+# Llama 3 (Meta)
+lm = Transformers("meta-llama/Llama-3.1-8B-Instruct")
+lm = Transformers("meta-llama/Llama-3.1-70B-Instruct")
+
+# Mistral (Mistral AI)
+lm = Transformers("mistralai/Mistral-7B-Instruct-v0.3")
+lm = Transformers("mistralai/Mixtral-8x7B-Instruct-v0.1")
+
+# Qwen (Alibaba)
+lm = Transformers("Qwen/Qwen2.5-7B-Instruct")
+
+# Gemma (Google)
+lm = Transformers("google/gemma-2-9b-it")
+```
+
+#### Generation 配置
+
+```python
+lm = Transformers(
+    "microsoft/Phi-4-mini-instruct",
+    device="cuda"
+)
+
+# Configure generation
+from guidance import gen
+
+result = lm + gen(
+    max_tokens=100,
+    temperature=0.7,
+    top_p=0.9,
+    top_k=50,
+    repetition_penalty=1.1
+)
+```
+
+### llama.cpp
+
+#### 基础配置
+
+```python
+from guidance.models import LlamaCpp
+
+# Load GGUF model
+lm = LlamaCpp(
+    model_path="/path/to/model.gguf",
+    n_ctx=4096  # Context window
+)
+```
+
+#### GPU 配置
+
+```python
+# Use GPU acceleration
+lm = LlamaCpp(
+    model_path="/path/to/model.gguf",
+    n_ctx=4096,
+    n_gpu_layers=35,  # Offload 35 layers to GPU
+    n_threads=8       # CPU threads for remaining layers
+)
+
+# Full GPU offload
+lm = LlamaCpp(
+    model_path="/path/to/model.gguf",
+    n_ctx=4096,
+    n_gpu_layers=-1  # Offload all layers
+)
+```
+
+#### 高级配置
+
+```python
+lm = LlamaCpp(
+    model_path="/path/to/llama-3.1-8b-instruct.Q4_K_M.gguf",
+    n_ctx=8192,          # Context window (tokens)
+    n_gpu_layers=35,     # GPU layers
+    n_threads=8,         # CPU threads
+    n_batch=512,         # Batch size for prompt processing
+    use_mmap=True,       # Memory-map the model file
+    use_mlock=False,     # Lock model in RAM
+    seed=42,             # Random seed
+    verbose=False        # Suppress verbose output
+)
+```
+
+#### 量化模型
+
+```python
+# Q4_K_M (4-bit, recommended for most cases)
+lm = LlamaCpp("/path/to/model.Q4_K_M.gguf")
+
+# Q5_K_M (5-bit, better quality)
+lm = LlamaCpp("/path/to/model.Q5_K_M.gguf")
+
+# Q8_0 (8-bit, high quality)
+lm = LlamaCpp("/path/to/model.Q8_0.gguf")
+
+# F16 (16-bit float, highest quality)
+lm = LlamaCpp("/path/to/model.F16.gguf")
+```
+
+#### 常用 GGUF 模型
+
+```python
+# Llama 3.1
+lm = LlamaCpp("llama-3.1-8b-instruct.Q4_K_M.gguf")
+
+# Mistral
+lm = LlamaCpp("mistral-7b-instruct-v0.3.Q4_K_M.gguf")
+
+# Phi-4
+lm = LlamaCpp("phi-4-mini-instruct.Q4_K_M.gguf")
+```
+
+## Backend 对比
+
+### 特性矩阵
+
+| 特性 | Anthropic | OpenAI | Transformers | llama.cpp |
+|---------|-----------|--------|--------------|-----------|
+| 约束生成 | ✅ 完整 | ✅ 完整 | ✅ 完整 | ✅ 完整 |
+| Token Healing | ✅ 是 | ✅ 是 | ✅ 是 | ✅ 是 |
+| 流式传输 | ✅ 是 | ✅ 是 | ✅ 是 | ✅ 是 |
+| GPU 支持 | N/A | N/A | ✅ 是 | ✅ 是 |
+| 量化 | N/A | N/A | ✅ 是 | ✅ 是 |
+| 成本 | $$$ | $$$ | 免费 | 免费 |
+| 延迟 | 低 | 低 | 中 | 低 |
+| 部署难度 | 简单 | 简单 | 中 | 中 |
+
+### 性能特征
+
+**Anthropic Claude:**
+- **延迟**：200-500ms（API 调用）
+- **吞吐量**：受 API rate limits 限制
+- **成本**：每 1M input tokens $3-15
+- **最适合**：生产系统，高质量输出
+
+**OpenAI:**
+- **延迟**：200-400ms（API 调用）
+- **吞吐量**：受 API rate limits 限制
+- **成本**：每 1M input tokens $0.15-30
+- **最适合**：对成本敏感的生产环境，gpt-4o-mini
+
+**Transformers:**
+- **延迟**：50-200ms（本地推理）
+- **吞吐量**：取决于 GPU（10-100 tokens/秒）
+- **成本**：仅硬件成本
+- **最适合**：隐私敏感场景、大吞吐量场景、实验测试
+
+**llama.cpp:**
+- **延迟**：30-150ms（本地推理）
+- **吞吐量**：取决于硬件（20-150 tokens/秒）
+- **成本**：仅硬件成本
+- **最适合**：边缘部署、Apple Silicon、CPU 推理
+
+### 内存要求
+
+**Transformers (FP16):**
+- 7B 模型：约 14GB GPU VRAM
+- 13B 模型：约 26GB GPU VRAM
+- 70B 模型：约 140GB GPU VRAM（多 GPU）
+
+**llama.cpp (Q4_K_M):**
+- 7B 模型：约 4.5GB RAM
+- 13B 模型：约 8GB RAM
+- 70B 模型：约 40GB RAM
+
+**优化建议：**
+- 使用量化模型 (Q4_K_M) 以降低内存消耗
+- 使用 GPU offloading 加快推理速度
+- 对较小模型 (<7B) 使用 CPU 推理
+
+## 性能调优
+
+### API 模型 (Anthropic, OpenAI)
+
+#### 降低延迟
+
+```python
+from guidance import models, gen
+
+lm = models.Anthropic("claude-sonnet-4-5-20250929")
+
+# Use lower max_tokens (faster response)
+lm += gen(max_tokens=100)  # Instead of 1000
+
+# Use streaming (perceived latency reduction)
+for chunk in lm.stream(gen(max_tokens=500)):
+    print(chunk, end="", flush=True)
+```
+
+#### 降低成本
+
+```python
+# Use cheaper models
+lm = models.Anthropic("claude-3-5-haiku-20241022")  # vs Sonnet
+lm = models.OpenAI("gpt-4o-mini")  # vs gpt-4o
+
+# Reduce context size
+# - Keep prompts concise
+# - Avoid large few-shot examples
+# - Use max_tokens limits
+```
+
+### 本地模型 (Transformers, llama.cpp)
+
+#### 优化 GPU 用法
+
+```python
+from guidance.models import Transformers
+
+# Use FP16 for 2x speedup
+lm = Transformers(
+    "meta-llama/Llama-3.1-8B-Instruct",
+    device="cuda",
+    torch_dtype="float16"
+)
+
+# Use 8-bit quantization for 4x memory reduction
+lm = Transformers(
+    "meta-llama/Llama-3.1-8B-Instruct",
+    device="cuda",
+    load_in_8bit=True
+)
+
+# Use flash attention (requires flash-attn package)
+lm = Transformers(
+    "meta-llama/Llama-3.1-8B-Instruct",
+    device="cuda",
+    use_flash_attention_2=True
+)
+```
+
+#### 优化 llama.cpp
+
+```python
+from guidance.models import LlamaCpp
+
+# Maximize GPU layers
+lm = LlamaCpp(
+    model_path="/path/to/model.Q4_K_M.gguf",
+    n_gpu_layers=-1  # All layers on GPU
+)
+
+# Optimize batch size
+lm = LlamaCpp(
+    model_path="/path/to/model.Q4_K_M.gguf",
+    n_batch=512,     # Larger batch = faster prompt processing
+    n_gpu_layers=-1
+)
+
+# Use Metal (Apple Silicon)
+lm = LlamaCpp(
+    model_path="/path/to/model.Q4_K_M.gguf",
+    n_gpu_layers=-1,  # Use Metal GPU acceleration
+    use_mmap=True
+)
+```
+
+#### 批处理
+
+```python
+# Process multiple requests efficiently
+requests = [
+    "What is 2+2?",
+    "What is the capital of France?",
+    "What is photosynthesis?"
+]
+
+# Bad: Sequential processing
+for req in requests:
+    lm = Transformers("microsoft/Phi-4-mini-instruct")
+    lm += req + gen(max_tokens=50)
+
+# Good: Reuse loaded model
+lm = Transformers("microsoft/Phi-4-mini-instruct")
+for req in requests:
+    lm += req + gen(max_tokens=50)
+```
+
+## 高级配置
+
+### 自定义模型配置
+
+```python
+from transformers import AutoTokenizer, AutoModelForCausalLM
+from guidance.models import Transformers
+
+# Load custom model
+tokenizer = AutoTokenizer.from_pretrained("your-model")
+model = AutoModelForCausalLM.from_pretrained(
+    "your-model",
+    device_map="auto",
+    torch_dtype="float16"
+)
+
+# Use with Guidance
+lm = Transformers(model=model, tokenizer=tokenizer)
+```
+
+### 环境变量
+
+```bash
+# API keys
+export ANTHROPIC_API_KEY="sk-ant-..."
+export OPENAI_API_KEY="sk-..."
+
+# Transformers cache
+export HF_HOME="/path/to/cache"
+export TRANSFORMERS_CACHE="/path/to/cache"
+
+# GPU selection
+export CUDA_VISIBLE_DEVICES=0,1  # Use GPU 0 and 1
+```
+
+### 调试
+
+```python
+# Enable verbose logging
+import logging
+logging.basicConfig(level=logging.DEBUG)
+
+# Check backend info
+lm = models.Anthropic("claude-sonnet-4-5-20250929")
+print(f"Model: {lm.model_name}")
+print(f"Backend: {lm.backend}")
+
+# Check GPU usage (Transformers)
+lm = Transformers("microsoft/Phi-4-mini-instruct", device="cuda")
+print(f"Device: {lm.device}")
+print(f"Memory allocated: {torch.cuda.memory_allocated() / 1e9:.2f} GB")
+```
+
+## 资源
+- **Anthropic 文档**：https://docs.anthropic.com
+- **OpenAI 文档**：https://platform.openai.com/docs
+- **Hugging Face 模型**：https://huggingface.co/models
+- **llama.cpp**：https://github.com/ggerganov/llama.cpp
+- **GGUF 模型**：https://huggingface.co/models?library=gguf