KTransformers 0.7.0.post4

Kimi K2.5 text LoRA training, checkpoint resume, and SGLang adapter loading with
native RAWINT4 routed experts.

Start with the step-by-step tutorial.
Download the attached kimi-k25-post4-user-kit.tar.gz, verify its SHA256, and
follow the guide in two new environments. It includes the three YAMLs, pinned
installation locks, data preparation/conversion tools, and auxiliary training
wheels. Only the model and output paths need changing.

• Linux x86_64, glibc 2.35+, Python 3.12, CUDA 12.8, Torch 2.9.1.
• Tested: dual AMD EPYC 9355, approximately 1.5 TiB RAM; 8 RTX 5090 GPUs for
  training and 4 for inference. Packing is disabled; this is not a minimum-spec claim.
• Core packages: ktransformers / kt-kernel / sglang-kt 0.7.0.post4,
  transformers-kt 5.6.0.post5, accelerate-kt 1.14.0.post3.
• LF uses the approved fixed KT-installation branch. PEFT/TRL only change
  dependency metadata. These three auxiliary wheels are in the public kit, not on PyPI.

Fresh public-package verification passed: installation via Tsinghua TUNA with
the unchanged published hash locks; 8-GPU Kimi four-step training; exact 2→4
resume across 20 state comparisons; independent 4-GPU SGLang generation with
ordinary LoRA and all 60 expert-LoRA layers loaded. Installed files, original
model files and cached model code were verified unchanged. See
PUBLIC_WHEEL_VERIFICATION.json.

The tutorial's CLI configuration and serving flags were checked against the
actual clean run, and its resume-check code was executed unchanged. TUNA only
mirrors PyPI: hosts without Hugging Face access must first obtain the pinned
model/data snapshots on an online machine, as described in the guide.

Candidate runtime checks passed: Kimi training, exact continuation and independent
adapter loading; Qwen3-30B-A3B BF16 training and base inference; DeepSeek-V3.1
native FP8 training; user-designated GLM next-0808 text inference. This does not
certify untested GPU/CPU architectures, legacy GLM vision, or all sequence lengths.

The separately trained Neko checkpoint-100 learned a clear style, but two strict output-format checks
failed. The recipe is a reproducible style experiment, not a general-quality or
safety guarantee. The complete tutorial states these limits.

All five core wheel versions and bytes are unchanged from the accepted build.
They were built before merge from source trees identical to the frozen official
mains in main-equivalence.json. The independent frozen-main rebuild is still
in progress and will be reported separately; same-version rebuilt wheels will
not replace the published artifacts. K2.5 is the hardware-validated model;
K2.6's shared architecture path was not separately hardware-tested.