river-client

Python client for the River ML training API — sampling, LoRA and full-weight fine-tuning, and reinforcement learning against River-hosted models.

Installation

pip install river-client

Requires Python 3.12+.

Quick start

River provides a low-level API for sampling, training, and checkpointing, plus an integrated RL library that manages rollouts and training from your environment and reward function. The example below uses the low-level API.

import river_client as river

client = river.Client(api_key="your-key", endpoint="api.river.ai")

# Stateless sampling from a base model
samples = client.sample(
    "What is 2+2?",
    base_model="Qwen/Qwen3.6-35B-A3B-FP8",
    max_tokens=50,
)
print(samples[0].text)

# Training with a session
with client.session() as session:
    model = session.create_model(
        base_model="Qwen/Qwen3.6-35B-A3B-FP8",
        lora=river.LoraConfig(rank=16),
    )

    # Forward + backward, then an optimizer step
    result = model.forward_backward(data, loss_fn="cross_entropy")
    model.optim_step(lr=1e-4)

    # Sample from the current weights
    sample_groups = model.sample("Continue:", max_tokens=100)

Capacity hints

Capacity hints are disabled by default. Enable them with client.session(capacity_hints=True) to print capacity changes to stderr: workers preparing, additional workers starting, or waits for available capacity. A state must persist for five seconds before a hint appears. Each run and workload prints at most once per 30 seconds, with no repeated messages for an unchanged state. A ready message appears only after a capacity hint was shown.

Structured status remains available with printing disabled:

with client.session() as session:
    # Create models and run your job as usual. In your progress update:
    for status in session.capacity_status:
        print(status.training_run_id, status.workload, status.state)

session.capacity_status returns an immutable tuple of river.CapacityStatus objects from the latest heartbeat without an extra API call. Each object contains training_run_id, workload (training or inference), state, capacity_wait, and observed_at (a UTC datetime when available). States are ready, scaling_up, preparing, waiting, at_limit, and unavailable.

Observations expire after 30 seconds. An empty tuple or unavailable means no fresh status is available; neither means capacity is ready. A scaling limit alone does not imply queuing, and workload capacity pressure does not explain every individual request's latency.

These observations require a server that supplies heartbeat capacity status. They cover the session's training runs and their checkpoint inference, matching Console. Stateless base-model sampling is outside this scope. Set session.capacity_hints = True to enable hints in an already-running session, or set it to False to silence them.

Full-weight training

On an endpoint with the GLM 5.3 Flash BF16 pool enabled, select full_weight=True instead of lora:

model = session.create_model(
    base_model="zai-org/GLM-5.3-Flash-BF16",
    full_weight=True,
)
model.forward_backward(data, loss_fn="cross_entropy")
model.optim_step(lr=1e-5, beta1=0.9, beta2=0.95, weight_decay=0.01)
checkpoint = model.save_weights("step1", mode="training", immutable=True, timeout=3600)
model.load_weights(checkpoint, load_optimizer=True, timeout=3600)

Checkpoints include FP32 master weights, Adam moments, optimizer parameters, and the training step. Full-weight saves require immutable=True; repeated names create separate checkpoint URIs with independent retention. Only an RL model such as zai-org/GLM-5.3-Flash-FW-RL can sample: its run owns a sampler, and river_client.rl refreshes it with save_weights(mode="publish"). See the full-weight training guide for RL, resume semantics and deployment requirements.

For temporary sampling or training checkpoints, use durable=False:

checkpoint = model.save_weights("sampling", mode="inference", durable=False)

These checkpoints stay on shared NFS for 24 hours by default and never upload to cloud storage. Pass ttl=datetime.timedelta(...) to choose another lifetime within the server's retention cap. They cannot be restored from cloud storage after expiry or NFS loss. This option requires server support; the SDK rejects it on an endpoint without temporary_checkpoints_v1. Changing an existing checkpoint's storage policy requires a new name or immutable=True. The RL trainer uses it for per-step sampling snapshots. Recovery and evaluation saves remain durable so unfinished evaluations can resume after a long pause. Update the server before using the new RL client. Custom model implementations must accept the durable keyword on save_weights.

On trainer resume, unfinished rollouts pinned to temporary snapshots restart from their environment's initial state, even after a short pause. Recovery republishes only the current policy; older pins may have expired. Completed rollout data is preserved, and each restarted rollout selects a fresh policy without mixing policies within a trajectory.

Reinforcement learning

river_client.rl manages rollouts and training around your environment, tools, and reward function. It supports multi-turn conversations, images, KV-cache reuse, and synchronous or bounded asynchronous training.

import operator
import os
from typing import Literal

import river_client as river
from river_client import rl
from river_client.renderers import get_renderer, get_text_content

# Tool schemas come from type hints and docstrings; execution stays in your code.
@rl.tool
async def calculator(a: float, operation: Literal["+", "-", "*", "/"], b: float) -> str:
    """Calculate the result of an arithmetic operation on two numbers."""
    operations = {"+": operator.add, "-": operator.sub, "*": operator.mul, "/": operator.truediv}
    return f"{operations[operation](a, b):g}"

class AnswerEnv(rl.Env):
    tools = [calculator]  # The engine runs tool calls and feeds results back.

    async def reset(self, row):
        return [{"role": "user", "content": (
            row["question"] + " Use the calculator if helpful. Return only the answer."
        )}]

    async def reward(self, traj, row):
        # Score the final answer, excluding the model's reasoning.
        answer = get_text_content(traj.messages[-1]).strip()
        return float(answer == row["answer"])

# Replace this toy dataset and reward with your task.
dataset = [
    {"question": "What is 17 * 23?", "answer": "391"},
    {"question": "What is 144 / 12?", "answer": "12"},
]
base_model = "zai-org/GLM-5.3-Flash"
renderer = get_renderer(base_model)
client = river.Client(api_key=os.environ["RIVER_API_KEY"], endpoint="api.river.ai")

with client.session() as session:
    model = session.create_model(
        base_model=base_model,
        tokenizer=renderer.tokenizer,
        lora=river.LoraConfig(rank=16),
    )
    trainer = rl.AsyncTrainer(
        engine=rl.RolloutEngine(
            model, env=AnswerEnv, renderer=renderer,
            budget=rl.Budget(max_turns=4, max_generated_tokens=2048),
            # Sampling stays on one policy for the whole trajectory by default.
        ),
        optimizer=rl.Adam(lr=1e-5),
        advantage=rl.GroupCentered(),  # Center rewards within each prompt group.
        completion=rl.GroupCompletion(mode="wait"),
        normalize="token",
        groups_per_step=2, group_size=8,  # Two prompts, eight rollouts each.
        max_staleness=0,  # Synchronous; >0 allows bounded sampling ahead.
    )
    rl.run(trainer, dataset, steps=10, on_step=lambda step: print(step.n, step.metrics))

See the bundled agent skill for image observations, cache policies, checkpoint/resume, and evaluation.

Dedicated streaming inference

Gated feature — disabled by default. Contact River to enable dedicated deployments for your team and the checkpoint's base model before using these APIs. Use a team API key with that access; personal API keys cannot create deployments.

client.create_deployment(checkpoint, ...) provisions capacity for a checkpoint and returns a base URL that the standard OpenAI client streams from unchanged; list_deployments, get_deployment_usage, scale_on_target and delete_deployment manage it from there. See the deployment walkthrough for a complete example. The bundled agent skill below carries the full workflow: replica roles, scale-to-zero and resume, streaming error handling, and usage accounting.

AI agent skill

The package bundles an agent skill — a SKILL.md that teaches AI coding agents (Claude Code and compatible tools) the current training API: train_step semantics, data formats, RL/SFT/distillation loop patterns, image uploads and handles, and fault-tolerant auto-recovery. Because it ships inside the wheel, the skill always matches the installed client version.

Install it into your agent's skills directory:

python -m river_client.skill --install

This copies the skill into ~/.claude/skills/; pass --dest for a different location (e.g. a project's .claude/skills/). Run without --install to print the bundled skill's path instead.

License

Apache-2.0. See LICENSE.

Export downloaded checkpoints to Hugging Face

The converter is available in river-client 0.13.0 and later. In Console, open an inference checkpoint and choose Download HF export script. Install the script's dependencies and run it with your API key:

pip install -U "river-client[hf-export]>=0.13.0" requests
RIVER_API_KEY=... python download.py

The script downloads to a temporary directory and writes an HF/PEFT LoRA adapter under the checkpoint ID. Load that directory with PEFT and the corresponding base model. Conversion reads the base model's safetensors headers from Hugging Face; it does not download base weights. Set HF_TOKEN for gated base models.

For a checkpoint already on disk, install the export extra and run the converter:

pip install -U "river-client[hf-export]>=0.13.0"
python -m river_client.hf_export ./river-checkpoint ./hf-adapter

The source directory must contain adapter_config.json and either adapter_model.safetensors or its sharded index and referenced files. The output contains adapter_config.json and adapter_model.safetensors in HF/PEFT adapter format. The converter uses base_model_name_or_path from the adapter configuration unless you override it:

python -m river_client.hf_export ./river-checkpoint ./hf-adapter \
  --base-model "<base-model-id>" --revision "<base-model-commit>"

Use --base-model /path/to/local/base to read a local base model's safetensors headers. In Python, call the same converter directly:

from river_client.hf_export import export_hf_adapter

output = export_hf_adapter("./river-checkpoint", "./hf-adapter")
# Optional keyword arguments: base_model="...", revision="...".
print(output)

The converter checks every target name and tensor shape against the base model. It restores text-model prefixes, splits supported fused projections, preserves rank/alpha scaling, and combines indexed adapter shards into PEFT's single safetensors file. It runs on CPU with memory proportional to adapter size. The source is preserved, and an existing output directory is never overwritten. Unsupported layouts fail before an output directory is published. This exports LoRA adapters, not merged base weights or training/optimizer state.

Conversion preserves adapter values and dtype; inference compatibility also depends on the loader. In our tests with Transformers 5.8.1 and PEFT 0.18.1, loading BF16 adapters onto native FP8 layers cast the adapter weights to FP8, changing their values. Use a loader that preserves BF16 adapters alongside the FP8 base. The workaround verified in that test was to load the base with FineGrainedFP8Config(dequantize=True) before attaching the adapter. This requires more memory and does not quantize the adapter.

For unsupported layouts, including NVFP4 expert adapters, choose Download raw script in Console. Run RIVER_API_KEY=... python download-raw.py after installing requests. It saves the original files under <checkpoint-id>-raw without conversion; those files are not guaranteed to load in HF/PEFT.