Skip to main content

EngramDB 的 Python 分发包

Disk-first storage engine for Engram / PLE n-gram memory tables (Rust).

分发名 engramdb-python(PyPI 相似名规避);import 名仍为 engramdb。

当前 v0.2.10 同时包含两条 Python 接入路径:

  1. PyO3 原生扩展(优先):crates/engramdb-pyo3,构建后以 python/engramdb/_engramdb.so 提供 Store / View / PageReader / Linux IoUringPageReader。
  2. ctypes C-ABI 回退:crates/engramdb-python,无 PyO3 构建产物时也能用。

python/engramdb/__init__.py 会自动优先加载 PyO3,失败则回退 ctypes。

安装

已发布到 PyPI,包名 engramdb-python,import 名 engramdb:

# 直接安装发布版
python3 -m pip install --upgrade engramdb-python

# 或使用 uv
uv add engramdb-python

开发/本地构建也可以使用:

# 方式 1:maturin 构建 wheel(推荐,直接产生可安装的 native wheel)
cd python
CARGO_HOME=/tmp/cargo-home RUSTFLAGS="-C link-arg=-undefined -C link-arg=dynamic_lookup" \
  maturin build --release --interpreter python3

# 方式 2:开发期内直接构建并复制到包目录
CARGO_HOME=/tmp/cargo-home RUSTFLAGS="-C link-arg=-undefined -C link-arg=dynamic_lookup" \
  cargo build -p engramdb-pyo3 --release
cp target/release/lib_engramdb.dylib python/engramdb/_engramdb.so

# 在 engram-peft 等项目中用 uv 添加本地开发依赖
cd ~/code/engram-peft
uv add --editable ../EngramDB/python

快速使用

存储 / 视图 / 页读取

import engramdb

# 打开 Store-I:目录内为 shard_000.bin 等定长行文件
store = engramdb.Store("path/to/rows", shards=1, rows_per_shard=100, width=256)
row_bytes = store.fetch([0, 1, 2])   # bytes, 每条 256B
store.close()

# 打开 Store-P 视图
view = engramdb.View("path/to/view.bin")
rec = view.read_record(0)            # 一条 e_t 记录

# SGLang 兼容:从多个 fd/offset 读页(Unix 有 PageReader,Linux 另有 IoUringPageReader)
reader = engramdb.PageReader(page_size=4096)
pages = reader.read_pages([fd0, fd1], [offset0, offset1])

线程安全的 Store 连接池

from engramdb import StorePool, ThreadLocalStore

pool = StorePool("path/to/rows", shards=128, rows_per_shard=2_500_012, width=160, pool_size=4)

with pool as store:          # 借一个句柄,用完自动归还
    data = store.fetch(rowids)

tls = ThreadLocalStore(pool) # 每线程一个句柄
handle = tls.get()
try:
    data = handle.fetch(rowids)
finally:
    tls.release_current()

PLE rowid、自动发现、FP8 scale

from engramdb import rowids_for_seq, discover_ple, load_ple_weight_scale, load_ple_multipliers

# Qwen PLE / Engram 确定性 rowid:[T, 16]
rows = rowids_for_seq([248044, 1000, 99999, 42])

# 从 checkpoint 自动发现 PLE 表元数据、weight_scale 与 rowid multipliers
info = discover_ple("/path/to/Qwen3.8-Flash-Next")
scale = load_ple_weight_scale("/path/to/Qwen3.8-Flash-Next")
mult = load_ple_multipliers("/path/to/Qwen3.8-Flash-Next")

# discovery 返回的 info 已包含 weight_scale 和 multipliers,可直接用于 rowids
rows = rowids_for_seq([248044, 1000, 99999, 42], info=info)

真实 PLE 磁盘 Adapter

from engramdb import Store
from engramdb.ple_adapter import disk_ple_from_discovery

info = discover_ple("/path/to/Qwen3.8-Flash-Next")
store = Store("/path/to/real-ple-rows", shards=128, rows_per_shard=2_500_012, width=160)
ple = disk_ple_from_discovery(store, info)  # 自动使用 weight_scale

快速 e_t tensor 读取(v0.2.9+)

训练/预计算不要再走 Python 逐行 bytes 拼接,使用一次 Store.fetch + torch.frombuffer:

from engramdb import Store, fetch_e_t_tensor

store = Store("/path/to/real-ple-rows", shards=128, rows_per_shard=2_500_012, width=160)

# flat_rowids 是 [T * 16] 的扁平行列表
e_t = fetch_e_t_tensor(
    store,
    flat_rowids,
    scale=0.00019931793212890625,
    num_heads=16,
    head_dim=160,
    dtype=torch.float8_e4m3fn,
    out_dtype=torch.float32,
)
# e_t.shape == (T, 16, 160)

也可以走 PleDiskGather 的 tensor 方法:

from engramdb.vllm import PleDiskGather

gather = PleDiskGather(store, row_bytes=160)
e_t = gather.fetch_tensor(flat_rowids, scale=..., num_heads=16, head_dim=160)

PleDiskGather.fetch 也已改为直接返回 Store.fetch 的连续缓冲区,不再做 Python 去重/切片/join。

流式/带 n-gram history 的 rowid 可使用:

from engramdb import rowids_for_seq_with_history

rows = rowids_for_seq_with_history([eos, eos], [10, 11, 12])

DiskPleEmbedding 预取与运行统计

from engramdb.vllm_plugin import DiskPleEmbedding
from concurrent.futures import ThreadPoolExecutor

shared_executor = ThreadPoolExecutor(max_workers=2)
emb = DiskPleEmbedding(
    store,
    num_embeddings=...,
    embedding_dim=160,
    dtype=torch.float8_e4m3fn,
    cache_size=4096,
    prefetch_executor=shared_executor,
    prefetch_timeout=0.5,
)

emb.prefetch([rowid1, rowid2, ...])
out = emb(torch.tensor([...]))

stats = emb.get_stats()
wait = emb.get_wait_distribution()   # p50 / p90 / p99 / max
emb.close()

后台预取失败会自动回退到同步读取;多个 PLE 模块可以共享同一个 prefetch_executor。

引擎适配层

目标是 不改 vLLM / SGLang 源码,启动前执行一小段 hook 即可把 PLE 表切到 EngramDB。

vLLM

from engramdb import Store
from engramdb.vllm_plugin import install_vllm_ple

store = Store("/path/to/engram-rows", shards=..., rows_per_shard=..., width=...)

install_vllm_ple(
    Qwen3_8FlashNextNGramEmbedding,   # 实际运行的 vLLM 模型类
    store=store,
    attr_name="embed_tokens_per_layer",
    embedding_dim=hidden_size_per_layer_input,
)

from vllm import LLM
llm = LLM(model="...", ...)

如果已经构造好模型实例,可以用:

from engramdb.vllm_plugin import patch_named_embedding
patch_named_embedding(model, "embed_tokens_per_layer", store=store, embedding_dim=...)

SGLang

from engramdb.sglang import install_sglang_ple

install_sglang_ple(
    Gemma4Model,                      # 实际运行的 SGLang 模型类
    store=store,
    attr_name="embed_tokens_per_layer",
    embedding_dim=hidden_size_per_layer_input,
)

低层 reader 替换:

from engramdb.sglang import install_sglang_io_uring_reader
install_sglang_io_uring_reader()

engram-peft 集成

安装 engramdb-python 后,可以直接使用内置的磁盘版 MultiHeadEmbedding:

import engramdb
from engramdb.integrations import install_disk_multi_head_embedding

store = engramdb.Store("path/to/embedding-store", shards=1, rows_per_shard=100, width=256)
install_disk_multi_head_embedding(store)

# 之后再调用 engram-peft 的 get_engram_model(...) 即可让 Engram 层从磁盘读取 embedding

真实 Qwen PLE FP8 Store 使用专用注入:

from engramdb.integrations import install_real_qwen_ple_embedding

# scale 会从 checkpoint 自动读取;也可以显式传 scale=0.0002
install_real_qwen_ple_embedding(
    store,
    model_dir="/path/to/Qwen3.8-Flash-Next",
)

多表 / Arrow / 最小服务原型

多表按目录组织:

from engramdb import Database

db = Database("path/to/tables-root")
print(db.list_tables())  # ["alpha", "beta"]

raw = db.fetch("alpha", [1, 3], shards=1, rows_per_shard=100, width=256)

可选 Arrow 读取(需要 pyarrow):

from engramdb.arrow_utils import store_fetch_arrow, table_to_ipc_bytes

table = store_fetch_arrow(store, [0, 1, 2])
ipc = table_to_ipc_bytes(table)   # Arrow IPC stream bytes

最小服务(当前为原型,提供两种 wire 模式):

JSON 模式:

from engramdb import Database
from engramdb.server import EngramDBServer

server = EngramDBServer(Database("path/to/tables-root"), host="127.0.0.1", port=8765)
server.serve_forever()

二进制模式(长度前缀 + 1-byte kind,fetch_raw 直接返回原始字节, fetch_arrow 直接返回 Arrow IPC stream,不需要 base64 包装):

from engramdb import Database, EngramDBBinaryServer, EngramDBClient

server = EngramDBBinaryServer(
    Database("path/to/tables-root"),
    host="127.0.0.1",
    port=8765,
)
server.serve_forever()

# 客户端
with EngramDBClient("127.0.0.1", 8765) as client:
    tables = client.list_tables()
    raw = client.fetch_raw("alpha", [0, 1, 2], shards=1, rows_per_shard=100, width=256)
    ipc = client.fetch_arrow("alpha", [0, 1, 2], shards=1, rows_per_shard=100, width=256)

服务命令包括:

  • ping
  • list_tables
  • fetch / fetch_raw
  • fetch_arrow(JSON 模式返回 base64 封装的 Arrow IPC;二进制模式返回裸 Arrow IPC stream)
  • view_read

注意:PyO3 Store 已不再是 unsendable,fetch 会释放 GIL,因此同一个 Store 可以从多个 Python 线程并发读取(已有 test_store_concurrent_fetch 冒烟)。 服务端 Database.fetch 仍会在请求线程中按需打开/复用 Store;生产级连接池化还在后续计划中。

定位(一句话)

让"确定性哈希的 n-gram 记忆表"(Qwen PLE、DeepSeek Engram 等)像数据库一样落盘、建索引、预取、服务化——单机 CPU+NVMe 低延迟推理 / 高吞吐训练预处理。

详细文档见上游仓库 docs/(design.md / specifications / roadmap)以及 examples/interop_engram_peft.py。

Metadata

Release files for engramdb-python 0.2.10

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for engramdb-python 0.2.10
File Size Uploaded
engramdb_python-0.2.10.tar.gz 59.9 kB Details

Built distributions (wheels)

Table of built distributions (wheels) for engramdb-python 0.2.10
File
engramdb_python-0.2.10-cp310-abi3-win_amd64.whl CPython 3.10 abi3 Windows x86-64 Details
engramdb_python-0.2.10-cp310-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl CPython 3.10 abi3 Linux glibc 2.17+ x86-64 Details
engramdb_python-0.2.10-cp310-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl CPython 3.10 abi3 Linux glibc 2.17+ ARM64 Details
engramdb_python-0.2.10-cp310-abi3-macosx_11_0_arm64.whl CPython 3.10 abi3 macOS 11.0+ ARM64 Details
engramdb_python-0.2.10-cp310-abi3-macosx_10_12_x86_64.whl CPython 3.10 abi3 macOS 10.12+ x86-64 Details

Total release size: 1.7 MB

Release files / engramdb_python-0.2.10.tar.gz

Download URL engramdb_python-0.2.10.tar.gz
Size 59.9 kB
Tags Source
SHA-256 checksum
How to use checksums
b2481361496371459574f2ea1969df2b5939c2d0f35e0f68056dd5ded762c0ca
BLAKE2b-256 checksum
How to use checksums
f13bf1d73c83e57e221eac776f181785d9f3638c919afaf101e53eeb32b92726
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Aug 31, 2026.

Transparency log

Release files / engramdb_python-0.2.10-cp310-abi3-win_amd64.whl

Download URL engramdb_python-0.2.10-cp310-abi3-win_amd64.whl
Size 242.1 kB
Tags CPython 3.10 Windows x86-64 abi3
SHA-256 checksum
How to use checksums
dc025151d9c3ffdfc3d763ead68cfa26e6bac25faba1027e2fe48aa70c38e8b5
BLAKE2b-256 checksum
How to use checksums
c91f582bb54966da0ad8656ba1ea594dacaa3c91cb02d84abb3522966be061cb
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Aug 31, 2026.

Transparency log

Release files / engramdb_python-0.2.10-cp310-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl

Download URL engramdb_python-0.2.10-cp310-abi3-manylinux_2_17_x86_64.manylinux2014_x86_64.whl
Size 367.2 kB
Tags CPython 3.10 Linux glibc 2.17+ x86-64 abi3
SHA-256 checksum
How to use checksums
059138ea8052c5b0f2882484535c47bc09d7a5efcb03beed5d1196a8c9a65aa7
BLAKE2b-256 checksum
How to use checksums
0aed2a852a9c1693add45a66accfab3090fe4a6925b8678055aa117692d8f485
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Aug 31, 2026.

Transparency log

Release files / engramdb_python-0.2.10-cp310-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl

Download URL engramdb_python-0.2.10-cp310-abi3-manylinux_2_17_aarch64.manylinux2014_aarch64.whl
Size 350.7 kB
Tags CPython 3.10 Linux glibc 2.17+ ARM64 abi3
SHA-256 checksum
How to use checksums
666e8bcb4d864d5619eb93c4ac03f720440ec5c1bc72f0d3a4a729b4bdad65db
BLAKE2b-256 checksum
How to use checksums
fa61928151ff846b814c856eb21107cda49d35f1c3dc570d9cf703878482e4dc
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Aug 31, 2026.

Transparency log

Release files / engramdb_python-0.2.10-cp310-abi3-macosx_11_0_arm64.whl

Download URL engramdb_python-0.2.10-cp310-abi3-macosx_11_0_arm64.whl
Size 348.2 kB
Tags CPython 3.10 abi3 macOS 11.0+ ARM64
SHA-256 checksum
How to use checksums
d16b253a36742484e30961709f71b4f1fb6a1badbf10e48d85121d618da1011d
BLAKE2b-256 checksum
How to use checksums
fb4c7d2fbe6c27f328d8f4607b71ba81d188821c921e6b346b78680cf5cac1f7
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Aug 31, 2026.

Transparency log

Release files / engramdb_python-0.2.10-cp310-abi3-macosx_10_12_x86_64.whl

Download URL engramdb_python-0.2.10-cp310-abi3-macosx_10_12_x86_64.whl
Size 357.0 kB
Tags CPython 3.10 abi3 macOS 10.12+ x86-64
SHA-256 checksum
How to use checksums
e90d8c0452f3f4c40cb5883333c309a5841d083657ce52e64fe89d07b8ba1d2f
BLAKE2b-256 checksum
How to use checksums
c9b76521467c04a0ab5ca8324ffe84ad30f926ffa082d5879d72ecd3dcf0946f
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
Yes
Uploaded via twine/7.0.0 CPython/3.13.14

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

PyPI Publish Attestation

PyPI verified that this artifact, at this checksum, originated from the publisher listed below.

Signed by GitHub Actions, verified by PyPI on Aug 31, 2026.

Transparency log

Release history Release notifications | RSS feed

0.3.0

6 release files

This release

0.2.10 This release

6 release files

0.2.9

6 release files

0.2.8

6 release files

0.2.6

6 release files

0.2.5

6 release files

0.2.4

6 release files

0.2.3

6 release files

0.2.2

6 release files

0.2.1

6 release files

0.2.0

2 release files

0.1.0

2 release files

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page