unslothai/unsloth · error · ValueError

base_precision must be one of nf4 / bf16 / int8 / fp8 / mxfp

Error message

base_precision must be one of nf4 / bf16 / int8 / fp8 / mxfp8 / auto

What it means

The validator rejected a base_precision value outside ('nf4', 'bf16', 'int8', 'fp8', 'mxfp8', 'auto'). base_precision selects how the frozen base model's weights are quantized during LoRA training (NF4/INT8/FP8 via bitsandbytes, dense bf16, or 'auto'). The value is strip().lower()-ed first, so this error is purely about an unrecognized name.

Source

Thrown at studio/backend/core/training/diffusion_train_common.py:1104

                    f"checkpoint bundle. Leave it at 0; the adapter is still saved at the end."
                )
        try:
            ema_decay = float(self.ema_decay or 0.0)
        except (TypeError, ValueError) as exc:
            raise ValueError(f"ema_decay must be a number, got {self.ema_decay!r}") from exc
        # decay = 1.0 would freeze the shadow at its init forever; the update is shadow * decay + param * (1 - decay), so valid decays live in [0, 1).
        if not 0.0 <= ema_decay < 1.0:
            raise ValueError("ema_decay must be in [0, 1); 0 disables the EMA adapter")
        # A blank cond_cache_dir (the Studio default when unset) means "off", not cwd.
        cond_cache_dir = (
            str(self.cond_cache_dir).strip() if self.cond_cache_dir is not None else ""
        ) or None
        compile_transformer = str(self.compile_transformer or "auto").strip().lower()
        if compile_transformer not in ("off", "on", "auto"):
            raise ValueError("compile_transformer must be one of off / on / auto")
        base_precision = str(self.base_precision or "nf4").strip().lower()
        if base_precision not in ("nf4", "bf16", "int8", "fp8", "mxfp8", "auto"):
            raise ValueError("base_precision must be one of nf4 / bf16 / int8 / fp8 / mxfp8 / auto")
        # base_precision is a DiT-only lever, so the dense-mode gates apply only to the DiT families. The mode-name check above still runs for every family.
        if resolved_family != "sdxl" and base_precision in ("bf16", "int8", "fp8", "mxfp8"):
            if repo_is_prequantized(self.base_model):
                raise ValueError(
                    f"base_precision={base_precision!r} needs a dense base repo, but "
                    f"'{self.base_model}' is already bitsandbytes-quantized. Pick the "
                    f"family's dense (bf16) base repo for this mode, or use nf4/auto."
                )
            if self.mixed_precision != "bf16":
                raise ValueError(
                    f"base_precision={base_precision!r} trains in bf16 compute; set "
                    f"mixed_precision to bf16."
                )
            # Refuse a scheme this family's DiT is known to corrupt, and also one the training bar holds back while
            # inference allows it: qwen-image fp8 now renders inside the accuracy gate, but no one has measured whether a
            # LoRA converges against fp8-frozen linears, so it fails fast here rather than silently training on faith.
            # MiniMax-H3 runs all three modalities through one set of linears, so the
            # per-family activation range the fp8 module filter was measured against does not

View on GitHub (pinned to 203007d190)

Solutions

1. Use one of: nf4, bf16, int8, fp8, mxfp8, auto (case-insensitive).
2. For 4-bit use 'nf4'; for 8-bit use 'int8'; there is no fp16 option by design.
3. If unsure, 'auto' picks a sensible default for the family.

Example fix

# before
config = TrainConfig(base_precision='4bit')

# after
config = TrainConfig(base_precision='nf4')

VALID_BASE_PRECISION = {"nf4", "bf16", "int8", "fp8", "mxfp8", "auto"}

def check_base_precision(v) -> str:
    s = str(v or "nf4").strip().lower()
    alias = {"4bit": "nf4", "nf4": "nf4", "8bit": "int8", "none": "auto"}
    s = alias.get(s, s)
    if s not in VALID_BASE_PRECISION:
        raise ValueError(f"base_precision must be one of {sorted(VALID_BASE_PRECISION)}, got {v!r}")
    return s

def is_valid_base_precision(v) -> bool:
    return str(v or "nf4").strip().lower() in {"nf4", "bf16", "int8", "fp8", "mxfp8", "auto"}

try:
    session.submit_training(config)
except ValueError as e:
    if "base_precision must be one of" in str(e):
        config.base_precision = "auto"
        session.submit_training(config)
    else:
        raise

When it happens

Trigger: Passing base_precision='fp16' (not supported — fp16 is a compute precision, not a base quantization), '4bit', '8bit', 'q4', 'none', or spellings from other tools (GGUF names, 'awq', 'gptq').

Common situations: Copying quantization vocabulary from llama.cpp/GGUF or autogptq configs; confusion between mixed_precision (compute) and base_precision (weight storage); UI free-text entry.

Related errors

• base_precision={base_precision!r} needs a dense base repo, b
• gradient_accumulation_steps must be >= 1
• lora_rank must be >= 1
• lora_alpha must be >= 1 (a zero/negative alpha scales the ad
• resolution must be a multiple of 8 and >= 64

AI-assisted analysis of unslothai/unsloth@203007d190 (2026-08-15). Data as JSON: /api/errors/2fe0e55b0a30c06e. Report an issue: GitHub.