Fine-tune LLMs from one YAML. Layer streaming trains an 8B model on a 4 GB laptop GPU.
Python cliconsumer-gpudpofine-tuningggufhuggingfacellmllmopslocal-ailocal-llmloralow-vrammachine-learningollamapeftpythonpytorchqlorasfttransformers
"""Known Kimi K2 resident remote-code boundary; no model or network imports."""
import os
KIMI_K2_RESIDENT_REASON = "Resident Kimi K2 is unsupported on Transformers 5.x"
KIMI_K2_GUARDED_STREAM_REASON = (
KIMI_K2_RESIDENT_REASON + ". Use the guarded training.stream_layers: true route "
"(#2107), with --trust-remote-code for the tokenizer and pip install tiktoken. "
"See docs/performance-and-quantization.md, Kimi K2: "
"CPU-only parity on a tiny K2-like checkpoint; full K2/GPU training is unverified."
)
KIMI_K2_NO_SUPPORTED_ROUTE_REASON = KIMI_K2_RESIDENT_REASON + ". No supported route yet."
KIMI_K2_RESIDENT_REASONS = {
"moonshotai/kimi-k2-base": KIMI_K2_GUARDED_STREAM_REASON,
"moonshotai/kimi-k2-instruct": KIMI_K2_GUARDED_STREAM_REASON,
"moonshotai/kimi-k2-instruct-0905": KIMI_K2_NO_SUPPORTED_ROUTE_REASON,
"moonshotai/kimi-k2-thinking": KIMI_K2_NO_SUPPORTED_ROUTE_REASON,
}
def resident_kimi_k2_refusal(
base: str, *, backend: str, hub: str, stream_layers: bool,
) -> str | None:
"""Only qualify known HF IDs on the resident Transformers path."""
if backend == "transformers" and hub == "hf" and not stream_layers:
reason = KIMI_K2_RESIDENT_REASONS.get(base.casefold())
# A real relative local directory can have the same two-component name
# as a Hub ID. Do not classify its potentially repaired code by its name.
if reason is not None and not os.path.isdir(base):
return reason
return None