pyds4 1.0.0

Python bindings for the DS4 native inference engine

pyds4

Python bindings for the DS4 DeepSeek V4 Flash inference engine

pyds4 is a Python package for running DS4-supported DeepSeek V4 Flash GGUF models from Python. It wraps the DS4 native engine with synchronous and asyncio APIs, token streaming, chat prompt helpers, token logprobs, DSML tool-call helpers, session snapshots, and payload-backed disk KV cache helpers.

This is not a generic GGUF runner. It targets the model files and native API supported by DS4. Avalan uses pyds4 for native DS4 inference, but pyds4 is usable directly from any Python application.

Contents

• Install
• Model Files
• Async Streaming Quickstart
• Runnable Examples
• Tool Use With DSML
• Advanced APIs
• Build From Source
• Benchmark
• Test
• Release

Install

Install the published package from PyPI:

python -m pip install -U pyds4

Check which native backend the installed package was built for:

python - <<'PY'
import pyds4

backend = pyds4.__ds4_native_backend__
print("pyds4:", pyds4.__version__)
print("backend:", backend)
print("available:", pyds4.is_backend_available(backend))
if not pyds4.is_backend_available(backend):
    print(pyds4.backend_unavailable_reason(backend))
PY

pyds4 requires Python 3.11 or newer. Production targets are macOS arm64 with Metal and Linux with CUDA. CPU builds exist for diagnostics and tests only.

pyds4 wheels are built for one selected native backend. If a wheel for your platform is not available, build from source against a DS4 checkout as shown in Build From Source.

Model Files

DS4 opens a local GGUF file directly from the filesystem. Use DS4 to download one of the supported DeepSeek V4 Flash GGUFs:

git clone https://github.com/antirez/ds4.git /path/to/ds4
cd /path/to/ds4
./download_model.sh q2-imatrix

The DS4 repository documents the available quantizations, memory expectations, optional MTP model, and current engine limitations. The examples below assume /path/to/ds4/ds4flash.gguf.

Async Streaming Quickstart

AsyncEngine owns DS4 on a single worker thread and serializes native calls there. For most applications, AsyncSession.stream_text() is the right high-level API: it advances the session, suppresses EOS text, buffers stop strings, and handles incremental UTF-8 decoding.

import asyncio

import pyds4

MODEL_PATH = "/path/to/ds4/ds4flash.gguf"
CTX_SIZE = 4096


async def main() -> None:
    backend = pyds4.Backend(pyds4.__ds4_native_backend__)
    if not pyds4.is_backend_available(backend.value):
        raise RuntimeError(pyds4.backend_unavailable_reason(backend.value))

    options = pyds4.EngineOptions(
        model_path=MODEL_PATH,
        backend=backend,
        native_log=False,
    )

    async with pyds4.AsyncEngine(options) as engine:
        prompt_tokens = await engine.encode_chat_prompt(
            system="You are concise.",
            prompt="Explain why KV caches matter in one paragraph.",
            think_mode=pyds4.think_mode_for_context(
                pyds4.ThinkMode.NONE,
                CTX_SIZE,
            ),
        )

        async with await engine.create_session(CTX_SIZE) as session:
            await session.sync(prompt_tokens)

            generation = pyds4.GenerationOptions(max_new_tokens=128)
            async for chunk in session.stream_text(generation):
                print(chunk, end="", flush=True)
            print()


asyncio.run(main())

Use sampling by passing SamplingOptions into GenerationOptions:

generation = pyds4.GenerationOptions(
    max_new_tokens=128,
    sampling=pyds4.SamplingOptions(
        temperature=0.7,
        top_k=40,
        top_p=0.95,
        seed=1,
    ),
)

If you build your own token loop but want the same stop-string behavior, StopStringBuffer exposes the generic buffering used by stream_text():

buffer = pyds4.StopStringBuffer(["</s>", "STOP"])
for chunk in buffer.push(decoded_token_text):
    print(chunk, end="")
for chunk in buffer.flush():
    print(chunk, end="")

Runnable Examples

The async example exposes the common knobs for backend, context size, sampling, streaming, and thread count:

python examples/generate_text_async.py \
  --model /path/to/ds4/ds4flash.gguf \
  --backend metal \
  --ctx-size 4096 \
  --max-new-tokens 128 \
  --temperature 0 \
  "Explain LLM distillation in one paragraph."

For lower-level control, the sync example shows the synchronous session loop: sync() the prompt, pick argmax() or sample(), call eval() to advance, then decode engine.token_text(token_id).

Tool Use With DSML

DeepSeek V4 Flash tool calls use DSML text. pyds4 can render prompts with tool schemas, tokenize rendered DSML chat prompts, parse generated tool calls, and stream argument-value deltas from a growing tool block. It does not execute tools; your application owns dispatching the parsed call and appending the tool result on the next turn.

import pyds4
from pyds4.dsml import (
    DsmlMessage,
    DsmlParseStatus,
    DsmlPrompt,
    DsmlToolCallBufferStatus,
    parse_generated_message,
    render_prompt,
    tool_call_buffer_status,
)

tool_schema = {
    "type": "function",
    "function": {
        "name": "math.calculator",
        "description": "Evaluate a small arithmetic expression.",
        "parameters": {
            "type": "object",
            "properties": {"expression": {"type": "string"}},
            "required": ["expression"],
        },
    },
}

rendered = render_prompt(
    DsmlPrompt(
        system_content="Use tools for arithmetic.",
        messages=[DsmlMessage(role="user", content="What is 4 * 7?")],
        tool_schemas=[tool_schema],
    ),
    think_mode=pyds4.ThinkMode.NONE,
)

async def collect_tool_call_text(engine: pyds4.AsyncEngine) -> str:
    prompt_tokens = await engine.tokenize_rendered_chat(rendered)
    generated = ""

    async with await engine.create_session(4096) as session:
        await session.sync(prompt_tokens)
        async for chunk in session.stream_text(
            pyds4.GenerationOptions(max_new_tokens=512),
        ):
            generated += chunk
            if (
                tool_call_buffer_status(generated)
                is DsmlToolCallBufferStatus.CLOSED
            ):
                break

    return generated


async def run_tool_prompt(engine: pyds4.AsyncEngine) -> None:
    generated = await collect_tool_call_text(engine)
    parsed = parse_generated_message(generated)
    if parsed.status is DsmlParseStatus.COMPLETE:
        for call in parsed.calls:
            print(call.name, call.arguments)

When continuing after a tool result, render the next prompt with the prior assistant tool call and a DsmlMessage(role="tool", content=...) result so the DSML transcript stays aligned with DS4's native prompt format. String parameter rendering escapes every accepted DSML parameter close-marker variant before it reaches the generated block, while preserving literal entity text such as </parameter> and </parameter> when parsed back.

Advanced APIs

Token Scores

Use AsyncSession.next_token() when you need token-level metadata instead of plain text chunks:

step = await session.next_token(
    decode=True,
    scores=pyds4.GenerationScoreOptions(
        mode=pyds4.TokenScoreMode.TOKEN_LOGPROB_AND_TOP_LOGPROBS,
        top_k=5,
    ),
)

print(step.decoded_text, step.token_logprob)
for score in step.top_logprobs:
    print(score.token_id, score.logprob)

The synchronous Session exposes the same primitives directly with argmax(), sample(), token_logprob(), top_logprobs(), and eval().

Snapshots And Disk KV Cache

Sessions can save and restore in-memory snapshots and serialized payloads. Ds4DiskKvCache uses payloads to cache a prompt prefix on disk.

from pathlib import Path

from pyds4.kv_cache import Ds4DiskKvCache

cache = Ds4DiskKvCache(
    Path("~/.cache/pyds4/kv").expanduser(),
    model_namespace="deepseek-v4-flash-q2-imatrix",
    backend=backend,
)

async with await engine.create_session(CTX_SIZE) as session:
    restored = await cache.arestore(session, prompt_tokens, CTX_SIZE)
    if restored.status == "miss" and restored.synced:
        await cache.astore(
            session,
            prompt_tokens,
            CTX_SIZE,
            size_budget_bytes=2_000_000_000,
        )

    text = await session.generate_text(
        pyds4.GenerationOptions(max_new_tokens=128),
    )

Store immediately after syncing or restoring the prefix you want to cache. Ds4DiskKvCache caches payload bytes, not snapshots. On a hit, it loads the payload with load_payload(); on a miss or corrupt entry, it calls sync(prompt_tokens) by default. model_namespace is caller-defined and should identify the exact model and configuration. Cache metadata can include prompt text, and payload bytes are session state, so treat cache directories as sensitive. size_budget_bytes is opt-in and enforced after a payload is written.

Progress And Cancellation

AsyncSession.progress is an asyncio.Queue[pyds4.ProgressEvent] populated when the native backend reports long-running progress. Cancelling a mutating async operation poisons that session and closes the native session during cleanup; create a fresh session after cancellation.

MTP And Speculative Evaluation

EngineOptions accepts mtp_path, mtp_draft_tokens, and mtp_margin for DS4's optional MTP path. Engine metadata exposes has_mtp and mtp_draft_tokens, while sessions expose eval_speculative_argmax(). Treat this as an advanced DS4-specific path and validate it with your target model.

Build From Source

Source builds need a DS4 checkout from the repository's default branch, plus CMake and a platform C/C++ toolchain:

git clone https://github.com/antirez/ds4.git /path/to/ds4

Build a Metal package on macOS arm64:

DS4_SOURCE_DIR=/path/to/ds4 \
PYDS4_BACKEND=metal \
python -m pip install --no-binary=pyds4 pyds4

Build a CUDA package on Linux:

DS4_SOURCE_DIR=/path/to/ds4 \
PYDS4_BACKEND=cuda \
CUDA_ARCH=90 \
python -m pip install --no-binary=pyds4 pyds4

To build this checkout and install it into a specific project's virtual environment, point PYTHON at that project's interpreter:

DS4_SOURCE_DIR=.local/ds4 \
PYDS4_BACKEND=metal \
PYTHON=/path/to/project/.venv/bin/python \
make ds4-bridge

If DS4_SOURCE_DIR is omitted during a source build, the package remains import-safe but native inference is unavailable.

For local wrapper development without a real GGUF or GPU, use the deterministic fake DS4 shim:

PYDS4_USE_FAKE_DS4=1 PYDS4_BACKEND=cpu \
python -m pip install -e ".[test,dev]"

Benchmark

Benchmark the pyds4 sync path, async primitive path, and async next_token() path:

python scripts/benchmark_generation.py \
  --model /path/to/ds4/ds4flash.gguf \
  --backend metal \
  --ctx-size 4096 \
  --max-new-tokens 128 \
  --mode pyds4 \
  --json-output /tmp/pyds4-bench.json

The benchmark reports open, prompt, warmup, sync, generation and total time, time to first token, tokens per second, event-loop latency, queue round-trip latency, and output preview.

Test

Run the fake-native test suite:

PYDS4_USE_FAKE_DS4=1 PYDS4_BACKEND=cpu python -m pip install -e ".[test,dev]"
make test
make test-cpp-sanitizers

Run real-model integration tests when a supported DS4 GGUF is available:

PYDS4_MODEL=/path/to/ds4/ds4flash.gguf \
PYDS4_BACKEND=metal \
PYDS4_CTX=4096 \
python -m pytest -q \
  tests/test_real_ds4_integration.py \
  tests/test_async_real_ds4_integration.py

Build and smoke-test a wheel:

DS4_SOURCE_DIR=/path/to/ds4 PYDS4_BACKEND=metal make wheel

WHEEL="dist/pyds4-*.whl" \
SMOKE_BACKEND=metal \
SMOKE_EXPECT_AVAILABLE=true \
SMOKE_MODEL=/path/to/ds4/ds4flash.gguf \
SMOKE_CTX=4096 \
make wheel-smoke

Release

pyproject.toml is the version source. For a new release, bump it and commit:

make version VERSION=1.0.1

Build and publish one backend wheel plus the sdist:

DS4_SOURCE_DIR=/path/to/ds4 PYDS4_BACKEND=metal make release

Use PYDS4_BACKEND=cuda on a CUDA Linux build host to produce the NVIDIA wheel. Metal and CUDA wheels can be uploaded for the same pyds4 version because they have different platform tags. The GitHub Release workflow builds the sdist, macOS arm64 Metal wheels, and Linux x86_64 CUDA wheels. The cuda_arch workflow input controls the NVIDIA architecture passed to CMake.

The CUDA wheel is audited in the release workflow but keeps CUDA runtime and cuBLAS libraries external, so it is published as a linux_x86_64 wheel. A strict auditwheel repair currently bundles those NVIDIA libraries and creates an artifact too large for PyPI's default file limit.