chidori-gpu 0.1.5

Run any PyTorch script on a remote GPU. Change zero lines of code.

Chidori GPU

Use any GPU, from anywhere. Zero code changes.

---

Chidori makes remote GPUs feel local. Your PyTorch training scripts, HuggingFace models, and CUDA applications run on remote GPU servers without changing a single line of code. Just pip install and go.

pip install chidori-gpu

30-Second Demo

On a machine with a GPU:

pip install chidori-gpu
chidori serve

On your laptop (no GPU needed):

pip install chidori-gpu
chidori run --server gpu-box:43211 python train.py

That's it. Your train.py runs on the remote GPU. No Docker, no SSH tunneling, no code changes.

Why Chidori?

Without Chidori	With Chidori
SSH into GPU server, set up env, scp files	chidori run -s host python train.py
Manage CUDA versions across machines	Auto-detects and translates between versions
One user per machine	One user per GPU, 8 clients on 8 GPUs
Port-forward Jupyter, sync files	Run locally, compute remotely

Performance

Tested over WAN with 70ms round-trip latency:

Workload	GPU	Throughput
Matrix multiply (4096x4096)	A100 80GB	96 TFLOPS
Training (23M params, FP32)	A100 80GB	32,600 tok/s
Training (23M params, FP16)	A100 80GB	59,000 tok/s
Inference (server-side decode)	A100 80GB	100+ tok/s

4 parallel clients on 4 A100s: 384 TFLOPS combined.

Getting Started

1. Install

pip install chidori-gpu

Both client and server. One package, two roles.

2. Start a GPU Server

On any machine with an NVIDIA GPU:

chidori serve

First run auto-compiles the CUDA worker from bundled source (needs gcc + CUDA toolkit). Subsequent runs use the cached binary.

3. Run Your Code Remotely

chidori run --server 10.0.0.5:43211 python train.py

Your script sees a GPU. All CUDA calls are transparently routed over TCP to the remote server.

Multi-GPU

Got a server with 8 GPUs? Start one server per GPU:

for i in $(seq 0 7); do
    CUDA_VISIBLE_DEVICES=$i chidori serve --port $((43211 + i)) &
done

Each client connects to its own GPU:

chidori run -s gpu-box:43211 python job_a.py &   # GPU 0
chidori run -s gpu-box:43212 python job_b.py &   # GPU 1
chidori run -s gpu-box:43213 python job_c.py &   # GPU 2
# ...

8 researchers, 8 GPUs, zero contention.

Fast Inference

Standard model.generate() works out of the box at ~1 tok/s (limited by per-token RPC overhead). For 100x speedup, enable server-side graph decode:

import chidori
chidori.optimize()

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

tokenizer = AutoTokenizer.from_pretrained("TinyLlama/TinyLlama-1.1B-Chat-v1.0")
model = AutoModelForCausalLM.from_pretrained(
    "TinyLlama/TinyLlama-1.1B-Chat-v1.0",
    dtype=torch.float16
).cuda()

# 100+ tok/s — server runs entire decode loop in a single RPC
output = model.generate(
    tokenizer("Hello", return_tensors="pt").input_ids.cuda(),
    max_new_tokens=200,
    do_sample=False,
    cache_implementation="static",
)
print(tokenizer.decode(output[0]))

Or use the direct API:

from chidori import server_decode
text = server_decode(model, tokenizer, "Explain quantum computing:", max_new_tokens=200)
# Returns generated text at 100+ tok/s

How It Works

┌──────────────────┐          TCP/IP          ┌──────────────────┐
│   Your Laptop    │  ◄──────────────────────► │   GPU Server     │
│                  │                           │                  │
│  python train.py │                           │  chidori serve   │
│       │          │                           │       │          │
│  LD_PRELOAD      │                           │  Real CUDA       │
│  libchidori.so   │   CUDA calls over wire    │  Driver + GPU    │
│       │          │  ─────────────────────►   │       │          │
│  intercepts      │                           │  executes on     │
│  280+ CUDA calls │  ◄─────────────────────   │  real hardware   │
│                  │   results back             │                  │
└──────────────────┘                           └──────────────────┘

Chidori uses LD_PRELOAD to intercept CUDA API calls at the driver level before they reach the (non-existent) local GPU. Each call is serialized into a compact wire protocol and sent over TCP. The remote server executes it on a real GPU and returns the result.

Key technical details:

• 280+ CUDA functions intercepted — Driver API, Runtime API, cuBLAS, cuBLASLt, cuDNN
• Async RPC batching — kernel launches are fire-and-forget, batched via TCP_CORK into single TCP segments
• Server-side graph decode — captures a CUDA graph of the forward pass, then runs the entire autoregressive loop server-side (one RPC instead of hundreds)
• Version-aware struct mapping — automatically translates cudaDeviceProp between CUDA 11.8 through 13.1
• Zero-copy device-to-device — D2D copies send only 24 bytes over the wire (data stays on GPU)

Supported CUDA Versions

PyTorch Build	CUDA Version	Status
cu118	CUDA 11.8	Supported
cu121	CUDA 12.1	Supported
cu126	CUDA 12.6	Supported
cu128	CUDA 12.8	Supported
cu130	CUDA 13.0	Supported
cu131	CUDA 13.1	Supported

Client and server can run different CUDA versions. Chidori handles the translation.

CLI Reference

# Start a GPU server (auto-builds on first run)
chidori serve [--port 43211]

# Run a command on a remote GPU
chidori run --server HOST[:PORT] [--verbose] [--trace] COMMAND...

# Examples
chidori run -s 10.0.0.5 python train.py
chidori run -s gpu1:43211 python -c "import torch; print(torch.cuda.get_device_name(0))"
chidori run -s gpu1:43211 --trace python benchmark.py   # show RPC trace

Requirements

Client (your machine):

• Linux x86_64
• Python 3.9+
• PyTorch (any cu11x/cu12x/cu13x build)

Server (GPU machine):

• Linux x86_64
• NVIDIA GPU with drivers installed
• CUDA toolkit (nvcc, headers, libraries)
• gcc and libzstd-dev

Contributing

Interested in contributing? Reach out to us.

License

Proprietary. All rights reserved.