Skip to main content

sageLLM backend provider abstraction (CPU/CUDA/Ascend)

Project description

sagellm-backend

Protocol Compliance (Mandatory)

CI PyPI version Python Version License Code style: ruff

硬件抽象层 - 为 sageLLM 提供统一的硬件接口(CUDA/Ascend/Kunlun)

架构定位

┌─────────────────────────────────────────────────────────────┐
│  sagellm-core (引擎协调层)                                   │
│  • LLMEngine (硬件无关的统一引擎)                           │
│  • 自动选择最佳后端 (cuda > ascend > cpu)                  │
├─────────────────────────────────────────────────────────────┤
│  sagellm-backend (硬件抽象层) ← 本仓库                       │
│  ┌─────────────────────────────────────────────────────┐    │
│  │  BackendProvider Interface                         │    │
│  │  • Stream/Event 异步流                              │    │
│  │  • KVBlock 内存管理                                 │    │
│  │  • Collective 操作(all_reduce/all_gather)        │    │
│  └─────────────────────────────────────────────────────┘    │
│  ┌──────────┐  ┌──────────┐  ┌──────────┐                  │
│  │  CUDA    │  │  Ascend  │  │  Kunlun  │                  │
│  │ Provider │  │ Provider │  │ Provider │                  │
│  └──────────┘  └──────────┘  └──────────┘                  │
├─────────────────────────────────────────────────────────────┤
│  Hardware SDK Layer                                          │
│  CUDA/cuDNN/NCCL │ CANN/HCCL │ XPU SDK │ DCU SDK           │
└─────────────────────────────────────────────────────────────┘

职责分离(v0.2.0 重构)

  • 本仓库负责:硬件抽象、设备管理、内存原语
  • 不再包含:BaseEngine, EngineFactory(已移至 sagellm-core)
  • 🔗 被使用于:sagellm-core 中的引擎实现

Features

  • 统一硬件抽象:单一 API 支持多硬件后端
  • CPU Backend:无 GPU 环境的默认后端
  • CUDA Support:原生 CUDA 后端实现
  • CPU Support:CPU-only 后端实现
  • 能力发现:硬件能力查询与验证

Installation

pip install isagellm-backend

Quick Start

git clone git@github.com:intellistream/sagellm-backend.git
cd sagellm-backend
./quickstart.sh

# Run tests
pytest tests/ -v

Usage Examples

Basic Backend Usage

from sagellm_backend import CPUBackendProvider, DType

# Create backend
backend = CPUBackendProvider()

# Query capabilities
cap = backend.capability()
print(cap.supported_dtypes)

# Allocate KV block
block = backend.kv_block_alloc(128, DType.FP16)

Using with sagellm-core LLMEngine

Backend 现在专注于硬件抽象,引擎使用 sagellm-coreLLMEngine

# LLMEngine 位于 sagellm-core
from sagellm_core import LLMEngine, LLMEngineConfig

# LLMEngine 自动选择最佳后端
config = LLMEngineConfig(
    model_path="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
    backend_type="auto",  # 自动选择: cuda > ascend > cpu
    max_new_tokens=100,
)
engine = LLMEngine(config)
await engine.start()

# 推理
output = await engine.generate("Hello, world!")
print(output)

await engine.stop()

Extending with New Backends

# Create provider in providers/ directory
class AscendBackendProvider:
    def capability(self) -> CapabilityDescriptor:
        return CapabilityDescriptor(
            supported_dtypes=[DType.FP16, DType.BF16, DType.INT8],
            # ...
        )

    # Implement other interface methods...

# Register via entry point in pyproject.toml
[project.entry-points."sagellm.backends"]
ascend_cann = "sagellm_backend.providers.ascend:create_ascend_backend"

Documentation

License

Proprietary

Project details


Release history Release notifications | RSS feed

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

isagellm_backend-0.4.0.6.tar.gz (128.1 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

isagellm_backend-0.4.0.6-py2.py3-none-any.whl (188.9 kB view details)

Uploaded Python 2Python 3

File details

Details for the file isagellm_backend-0.4.0.6.tar.gz.

File metadata

  • Download URL: isagellm_backend-0.4.0.6.tar.gz
  • Upload date:
  • Size: 128.1 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? No
  • Uploaded via: twine/6.2.0 CPython/3.11.14

File hashes

Hashes for isagellm_backend-0.4.0.6.tar.gz
Algorithm Hash digest
SHA256 c370e4a23d7b28d0454c78420eba9592d726659c6556a12303f0aa37ec66fe27
MD5 d35b891fe7eff4c67ee7be8c91cf8acc
BLAKE2b-256 ca76d20e1198536b4782e18f72be6d120181aaec3603360013185a9f2d73472a

See more details on using hashes here.

File details

Details for the file isagellm_backend-0.4.0.6-py2.py3-none-any.whl.

File metadata

File hashes

Hashes for isagellm_backend-0.4.0.6-py2.py3-none-any.whl
Algorithm Hash digest
SHA256 2b05911bdbf2583bc5c3744d78e6e97b1e73f0b12b732a76faa51c9dcc5e21ba
MD5 72cd602a97ad3f2cb8b171a5bbec4a8f
BLAKE2b-256 3bf78ea70f06fa6c4ce205aeef969484a1e13b82b8d9634b7506caacbe2ee2f7

See more details on using hashes here.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page