unslothai/unsloth · error · ValueError

llama-server does not read an attached value: write '{flag}'

Error message

llama-server does not read an attached value: write '{flag}' and '{value[:32]}' as two separate arguments, not '{token[:64]}'

What it means

ValueError from llama_server_args.py:315 — a flag token uses GNU '--flag=value' spelling. llama.cpp looks the WHOLE token up in its option map (folding only underscores), so '--top-k=20' is an argument it has never heard of, not --top-k with a value: measured on b10342/b10360, --top-k=20, --ctx-size=4096 and --flash-attn=on each exit 'invalid argument'. Accepting it meant the model switch tore down the resident model and the child then refused to start, so it is refused while it is still a 400. The module will not split it for you because for switches the 'value' is not one, and it cannot know an ordinary flag's arity.

Source

Thrown at studio/backend/core/inference/llama_server_args.py:315

            # naming a flag that looks correct in the log. Only flag-shaped tokens:
            # a VALUE may legitimately end in whitespace, a chat template or a
            # grammar being the obvious ones.
            raise ValueError(
                f"llama-server does not accept the spaces around '{token[:64]}': "
                f"write it as '{flag}'"
            )
        elif "=" in token:
            # llama.cpp looks the WHOLE token up in its option map, folding only the
            # underscore spelling, so "--top-k=20" is not "--top-k" with a value: it
            # is an argument it has never heard of. Measured on b10342 and b10360,
            # where --top-k=20, --ctx-size=4096 and --flash-attn=on each exit with
            # "error: invalid argument". Accepting the GNU spelling here meant the
            # switch tore down the resident model and the child then refused to
            # start, so it is refused while it is still a 400 with somewhere to go.
            # Splitting it here would be a guess: for a switch the value is not one,
            # and this module cannot know an ordinary flag's arity.
            value = token.partition("=")[2]
            raise ValueError(
                f"llama-server does not read an attached value: write '{flag}' and "
                f"'{value[:32]}' as two separate arguments, not '{token[:64]}'"
            )
        else:
            # Its own value when attached, otherwise the tokens that follow.
            attached = _value_is_attached(token, flag)
            if pending_two_value > 0:
                raise ValueError(f"llama-server flag '{two_value_flag}' takes two values")
            # An attached value is ONE of the two, not the whole option:
            # "--control-vector-layer-range=1" still owes its END, and
            # llama-server exits on the incomplete option.
            if flag in _TWO_VALUE_FLAGS:
                pending_values = 1 if attached else 2
                pending_two_value = pending_values
            elif flag in _OPTIONAL_SECOND_VALUE_FLAGS:
                # Allowed, not owed: pending_two_value stays 0, so nothing here
                # insists on the second token.
                pending_values = 1 if attached else 2

View on GitHub (pinned to 203007d190)

Solutions

Rewrite as two tokens: '--ctx-size 4096' instead of '--ctx-size=4096' (the error message names the exact split).
If building args programmatically from a dict, emit flag and value as separate list elements.
Grep your stored config for '=' inside tokens starting with '-' and split them at load time — except _TWO_VALUE_FLAGS and _OPTIONAL_SECOND_VALUE_FLAGS where attached handling differs, letting the validator sort it out is safer.

Example fix

# before
extra_args = ["--ctx-size=4096", "--flash-attn=on"]

# after
extra_args = ["--ctx-size", "4096", "--flash-attn", "on"]

Defensive patterns

Strategy: validation

Validate before calling

extra_args = [
    part
    for tok in raw_args
    for part in ((tok.split('=', 1) if tok.startswith('--') and '=' in tok else [tok]))
]

Type guard

def no_attached_values(tokens) -> bool:
    return not any(t.startswith('-') and '=' in t for t in tokens)

Try / catch

try:
    validate_extra_args(args)
except ValueError as e:
    if "attached value" in str(e):
        raise HTTPException(400, "write flags and values as separate tokens")
    raise

Prevention

Default to space-separated flag/value pairs; llama.cpp does not parse '--flag=value'.
Emit args from a dict as alternating [flag, value] elements.
Lint stored configs for tokens matching ^--.*=.
Remember this validator refuses rather than splits — do the split yourself.

When it happens

Trigger: Passing any '--flag=value' token in extra args: --ctx-size=4096, --top-k=20, --flash-attn=on, --split-mode=row. Every attached-value spelling of a known flag hits this branch.

Common situations: Muscle memory from GNU tools; copy-pasting from llama.cpp docs or blog posts that use '='; config generators emitting key=value style.

Related errors

AI-assisted analysis of unslothai/unsloth@203007d190 (2026-08-15). Data as JSON: /api/errors/4ee3820f63259d7b. Report an issue: GitHub.