pufferlib – skill Lập trình cho agent AI

PufferLib

Choose the version before choosing an API. Reviewed 2026-10-01:

| Profile | Status | Main use |

|---|---|---|

| Native source 5.0 | Current default branch and live documentation | C/CUDA environments, native trainer; CPU evaluation only |

| pufferlib==3.0.0 | Latest PyPI release, published 2025-06-23; sdist only | Python/Gymnasium/PettingZoo adaptation and Torch PuffeRL |

| Pinned source 4.0 | Historical snapshot | C Ocean interface with an optional Torch fallback |

For 5.0, read [references/native-5.md](references/native-5.md). The reviewed

revision is 6ffa5b10dbbbe4d1e8288367c7d9d3acd3bad4a2. Its CLI is

./puffer train after building an environment, not puffer train ENV_NAME.

There is no 5.0 Python emulation/vector API or --slowly fallback.

The bundled plan schema deliberately supports only 3.0 and pinned 4.0; it does

not launch training. All native/PufferLib training examples are **source-reviewed,

illustrative, and not executed in this review**. Bundled synthetic checks are

executed CPU tests, not evidence of PufferLib installation or learning quality.

The PyPI sdist was hash-verified and its Python sources inspected; the moving

3.0 branch differs, including its load_policy and logger contracts.

Safe defaults

1. Start with bundled synthetic, CPU-only, network-free tools.

2. Do not import an arbitrary environment by dotted path. Bundled tools accept

only allowlisted built-ins and slug identifiers.

3. Do not install or execute an unreviewed environment package, native

extension, ROM, map, checkpoint, or pickle file.

4. Verify official source, immutable revision, licenses, checksums or

attestations, and build hooks. Sandbox native builds and first execution.

5. Cap steps, environments, agents, workers, threads, buffers, memory, disk,

render size, and wall time.

6. Keep training and evaluation environments/seeds separate.

7. Default logging to local/none. External logging requires explicit opt-in,

disclosure acknowledgment, and separate artifact-upload approval.

8. Never pass W&B or Neptune credentials via CLI, INI, JSON, tags, run names, or

logger configuration. Never print them.

9. Never dump all environment variables or recursively search for .env.

10. Hash checkpoint bytes before trusted, sandboxed loading; metadata inspection

is not proof of safety.

First local checks

All bundled CLIs are dependency-free and emit strict JSON:

python3 scripts/env_template.py --help
python3 scripts/env_contract_validator.py
python3 scripts/benchmark_vectorization.py --backend serial
python3 scripts/train_template.py
python3 scripts/validate_plan.py
python3 scripts/repro_plan.py

Defaults are synthetic, deterministic, bounded, local, CPU-only, no-network,

and dry-run where training would otherwise occur.

Installation and provenance

Published 3.0.0

PyPI supplies only pufferlib-3.0.0.tar.gz:

sha256: 7df3a3e3f5f894d78d2a1f5374097890aec01473183e748abefe4f3faa10eaa9
Requires-Python: >=3.9

After source/build review, create a pinned uv project:

uv venv --python 3.11
uv add --exact --no-sync "pufferlib==3.0.0"
uv lock
uv sync --frozen

These installation commands are illustrative and were not executed.

Commit pyproject.toml and uv.lock; verify the archive digest and every

resolved dependency. The source build can compile native code and fetch build

assets, so resolve/build in a sandbox without credentials or sensitive mounts.

The archive declares NumPy <2, Gymnasium <=0.29.1, and PettingZoo <=1.24.1;

latest Gymnasium/NumPy are not valid substitutes for this profile. Its setup

supports Linux/macOS and rejects other systems. Python classifiers alone do

not establish a successful native build. The uploaded metadata does not pin Torch or CUDA; do not claim a supported CUDA

matrix that PyPI does not declare.

Pinned 4.0 source

The reviewed branch head on 2026-07-23 was:

25647630e1b15330bb3153a5a0d3ff8d234c3acf

Pin the commit, not branch 4.0:

uv add --no-sync \
  "pufferlib @ git+https://github.com/PufferAI/PufferLib.git@25647630e1b15330bb3153a5a0d3ff8d234c3acf"
uv lock

The reviewed 4.0 package declares Python >=3.10 and Torch >=2.9. The reviewed

PufferTank snapshot uses Ubuntu 24.04, Python 3.12, and an NVIDIA CUDA

13.0.2/cuDNN development image with the cu130 Torch index, but does not pin

the exact Torch wheel or all system packages. Treat it as a reference, not a

complete lock. Never execute a remote installer directly from a pipe.

Read references/training.md before any installation or build.

Environment workflow

1. Validate the contract

Gymnasium reset returns (observation, info). Step returns:

(observation, reward, terminated, truncated, info)

Validate spaces, shapes, dtypes, finite rewards, booleans, reset-before-step,

reset-after-end, seeding, and cleanup. terminated is an MDP terminal;

truncated is an external cutoff such as a time limit. Preserve the distinction

for bootstrapping and metrics. Both flags can be true in general Gymnasium.

Check autoreset timing and retain the final pre-reset observation; do not

bootstrap from the next episode. The 3.0 trainer has unresolved truncation and

inactive-agent mask handling, described in references/training.md.

python3 scripts/env_contract_validator.py \
  --steps 64 --episodes 8 --seed 42

2. Adapt only after review

Published 3.0 uses explicit wrappers:

import pufferlib.emulation

wrapped = pufferlib.emulation.GymnasiumPufferEnv(reviewed_gymnasium_instance)

For a reviewed PettingZoo Parallel environment:

wrapped = pufferlib.emulation.PettingZooPufferEnv(reviewed_parallel_instance)

There is no supported 3.0 pufferlib.emulate(...) shortcut matching the old

skill. Read references/environments.md and references/integration.md.

3. Native environments

Published 3.0 PufferEnv requires

single_observation_space, single_action_space, and num_agents before

super().__init__(buf). It uses in-place vector buffers and returns separate

terminal/truncation arrays plus a list of info dictionaries.

The reviewed 4.0 source uses C bindings. Start from upstream ocean/squared (single-agent)

or ocean/target (multi-agent), build one environment in local/sanitized mode,

and verify every buffer size/type/index before optimization.

Vectorization workflow

Published 3.0:

import pufferlib.vector

vecenv = pufferlib.vector.make(
    reviewed_creator,
    backend=pufferlib.vector.Serial,
    num_envs=4,
    seed=42,
)

Move to Multiprocessing only after serial traces pass. Record

num_envs, num_workers, batch_size, zero-copy mode, start method, agent

count, masks, and actual returned shapes. For multi-agent environments, batch

length is based on agent slots, not necessarily num_envs.

The reviewed 4.0 config instead uses:

[vec]
total_agents = 4096
num_buffers = 2
num_threads = 16

Read references/vectorization.md. Benchmark fixed work with warmup and at least

three repeats; report simulation and end-to-end training SPS separately. The

bundled benchmark measures only its synthetic harness.

Policy workflow

Published 3.0 policies are Torch modules sized from

single_observation_space/single_action_space. Stable recurrent composition

uses encode_observations and decode_actions; structured emulation uses

pufferlib.pytorch.nativize_dtype and nativize_tensor.

The reviewed 4.0 Torch fallback composes:

pufferlib.models.Policy(encoder=encoder, decoder=decoder, network=network)

It provides MLP, MinGRU, LSTM, and GRU network choices; --slowly selects this

fallback instead of the native backend. Check output/state shapes, masks,

finite values, gradients, and eager-versus-compiled behavior. See

references/policies.md.

Training and evaluation

Published 3.0 trainer import:

from pufferlib import pufferl

# train_config must include the environment name for checkpoint naming.
trainer = pufferl.PuffeRL(train_config, vecenv, policy)

Reviewed 4.0 CLI:

puffer train ENV_NAME
puffer eval ENV_NAME --load-model-path EXACT_TRUSTED_PATH
puffer sweep ENV_NAME

Generate a plan instead of launching by default:

python3 scripts/train_template.py \
  --profile pypi-3.0.0 \
  --environment synthetic \
  --device cpu \
  --total-timesteps 10000

train_template.py emits a report envelope, not a bare plan. Its plan

member is the input to validate_plan.py; passing the whole report is invalid.

For the synthetic environment, command_preview is an empty list because

there is no upstream training command to launch. To save and revalidate:

python3 scripts/train_template.py > training-report.json
python3 -c 'import json; r=json.load(open("training-report.json")); print(json.dumps(r["plan"], allow_nan=False, indent=2))' > plan.json
python3 scripts/vali

… (nội dung đầy đủ có trong gói tải về)