tsugi-kpool 0.1.0

K-Pool LoRA SDK. software analog of TsugiCinema's Infinity provisional at LoRA adapter granularity. Productized from the K-Pool LoRA provisional (US App. 64/060,315) and the Infinity provisional (US App. 64/055,093).

tsugi-kpool

K-Pool LoRA SDK. Software productization of TsugiCinema Inc.'s K-Pool LoRA provisional (US App. 64/060,315, filed 2026-05-07) and Infinity provisional (US App. 64/055,093, filed 2026-05-01), packaged as a drop-in extension to PyTorch + PEFT for distributed LoRA fine-tuning with measurable straggler-tax recovery on cross-rack training clusters.

What this is

A Python package that wraps PyTorch distributed and PEFT to implement, at LoRA adapter granularity:

1. K-out-of-N adapter routing. K-Pool LoRA's claim element, selecting K of N adapter modules per step.
2. Adapter-gradient elastic buffer. software analog of Infinity claim 2, FIFO buffer of adapter gradients prior to aggregation.
3. Buffer-convergence aggregation. Infinity claim 4, triggers aggregation when buffer-variance falls below threshold instead of on iteration count.
4. Phase-correction sideband. Infinity claim 3, low-bandwidth TCP channel between training nodes carrying drift telemetry, parallel to (not displacing) the NCCL gradient data plane.

The public API stays close to peft.LoraConfig + accelerate.Accelerator so adoption friction is minimal.

What this is not

• Not a fork of OpenDiLoCo. The architectural sibling exists but Prime Intellect's open-source orchestration layer is a separate prior-art branch; we go through torch.distributed.ProcessGroup directly to keep IP boundaries clean.
• Not a full-model Infinity instance. This SDK demonstrates the mechanism at adapter granularity. The transport-layer / full-model instantiation is a separate (Phase 2) productization track.
• Not yet a benchmark. The benchmark protocol is in benchmarks/llama3_8b_lora/.

Install

pip install -e ".[dev,benchmark]"

Minimal usage

from tsugi_kpool import KPoolLoraConfig, plesio_init
from transformers import AutoModelForCausalLM
from peft import get_peft_model

model = AutoModelForCausalLM.from_pretrained("meta-llama/Meta-Llama-3-8B")

config = KPoolLoraConfig(
    r=16,                        # standard LoRA rank
    lora_alpha=32,
    target_modules=["q_proj", "v_proj"],
    n_adapters=8,                # N
    k_active=2,                  # K-out-of-N per step
    sideband_addr="tcp://0.0.0.0:51820",  # phase-correction sideband
    buffer_convergence_eps=1e-3, # buffer-variance trigger threshold
)

model = get_peft_model(model, config)
plesio_init(model, config)   # starts the sideband + aggregator threads

# from here, train as you would any peft+accelerate fine-tune

License

Apache License, Version 2.0 with its full automatic patent grant. TsugiCinema, Inc. is the Licensor. The Apache-2.0 patent grant in Section 3 extends to TsugiCinema's K-Pool LoRA (US App. 64/060,315) and Infinity (US App. 64/055,093) patent estates AS PRACTICED BY THE SDK CODE AS DISTRIBUTED. See LICENSE for the NOTICE preamble explaining the doctrine and the full Apache-2.0 license text.

The license posture aligns with TsugiCinema's 2026-05-22 strategic doctrine: hyperscaler-product revenue as primary path; patents as defensive moat (DD scaffolding plus counter-assertion insurance) rather than offensive enforcement; unified pip install tsugi product wrapping this SDK and the companion tsugiai-mend-sdk at the packaging level.

Status

Pre-alpha. APIs will change without notice until v0.1. Repository is private and counsel-and-internal-only pending coordinated public release with the companion tsugiai-mend-sdk and the unified pip install tsugi product surface.