Open-source SDKs
One unified developer surface for TsugiCinema's distributed-training SDKs.
pip install tsugi TsugiLabs is the research division of TsugiCinema, Inc., a Delaware C-corporation (2026). NVIDIA Inception member.
Two engines
The tsugi meta-package wraps two physically separate SDKs under a unified import namespace. The two engines share zero code and can be installed and used independently.
Cross-rack distributed-training resilience: straggler and fail-slow detection with concurrent recovery. Patent-independent by construction. The default mode is bit-exact against tsugi-mend's own synchronous reference path. Throughput is benchmarked against FSDP baselines.
from tsugi.mend import MendConfig, mend_init, mend_shutdown
# After your model is wrapped (FSDP, TP, etc.):
config = MendConfig(quorum_min_learners=2, grace_window_ms=2000)
mend_init(model, config)
# ... train ...
mend_shutdown(model) K-of-N LoRA continual learning without replay buffers. Practices US App. 64/060,315 and 64/055,093 as distributed.
from tsugi.kpool import KPoolLoraConfig, plesio_init, plesio_shutdown
config = KPoolLoraConfig(
r=16,
lora_alpha=32,
n_adapters=8,
k_active=4,
sideband_enabled=True,
aggregation_mode="buffer_convergence",
)
plesio_init(model, config)
# ... train with K-of-N adapter routing ...
plesio_shutdown(model) Validated hardware
Commodity Ethernet
SXM4
SXM5 over InfiniBand
Benchmark campaign
A pre-registered H100-over-InfiniBand stall-recovery benchmark campaign is prepared. The pre-registration (hypotheses, arms, analysis plan) publishes openly in the tsugi-mend repository on the day the venue is booked, and results publish openly regardless of outcome.
Licensing
All three packages, tsugi, tsugi-mend, and tsugi-kpool, are licensed under the Apache License, Version 2.0.
Links
Contact
research@tsugilabs.ai for research correspondence. partnerships@tsugicinema.com for biz-dev and licensing questions.