xai-org/x-algorithm · error · ValueError

Backward block sizes (block_q_dq, block_kv_dq, block_q_dkv,

Error message

Backward block sizes (block_q_dq, block_kv_dq, block_q_dkv, block_kv_dkv) must either all be specified or all be None.

What it means

TuningConfig for varlen ranker attention validates that the four backward block sizes (block_q_dq, block_kv_dq, block_q_dkv, block_kv_dkv) are all-or-nothing. Partially specifying them is ambiguous (autotuning would mix tuned and unset sizes), so __post_init__ rejects it immediately at config construction.

Source

Thrown at phoenix/xrex/pallas/ranker_attention_varlen.py:35

class TuningConfig:
    block_q: int
    block_kv: int
    num_warps: int
    num_stages: int

    block_q_dq: int | None = None
    block_kv_dq: int | None = None
    block_q_dkv: int | None = None
    block_kv_dkv: int | None = None

    bwd_num_warps: int | None = None
    bwd_num_stages: int | None = None

    def __post_init__(self):
        backward_blocks = [self.block_q_dq, self.block_kv_dq, self.block_q_dkv, self.block_kv_dkv]
        block_is_set = [blk is not None for blk in backward_blocks]
        if any(block_is_set) and not all(block_is_set):
            raise ValueError(
                "Backward block sizes (block_q_dq, block_kv_dq, block_q_dkv, "
                "block_kv_dkv) must either all be specified or all be None."
            )


HOPPER_CONFIG = TuningConfig(
    block_q=128,
    block_kv=128,
    num_warps=8,
    num_stages=2,
    block_q_dq=128,
    block_kv_dq=16,
    block_q_dkv=16,
    block_kv_dkv=128,
    bwd_num_warps=8,
    bwd_num_stages=4,
)

View on GitHub (pinned to 24c60942c5)

Solutions

1. Provide all four backward block sizes explicitly, or remove all of them so autotuning/defaults apply.
2. Check dataclass field defaults vs your overrides; a field left as None while siblings are ints triggers this.
3. Build the config programmatically so all four sizes are set together.

Example fix

# before
cfg = TuningConfig(block_q_dq=128)
# after
cfg = TuningConfig(block_q_dq=128, block_kv_dq=128, block_q_dkv=64, block_kv_dkv=128)
# or simply
cfg = TuningConfig()

bwd = [cfg.block_q_dq, cfg.block_kv_dq, cfg.block_q_dkv, cfg.block_kv_dkv]
assert all(b is not None for b in bwd) or all(b is None for b in bwd)

def backward_blocks_consistent(cfg) -> bool:
    blocks = [cfg.block_q_dq, cfg.block_kv_dq, cfg.block_q_dkv, cfg.block_kv_dkv]
    return all(b is None for b in blocks) or all(b is not None for b in blocks)

When it happens

Trigger: Constructing TuningConfig (e.g. HOPPER_CONFIG-style presets or a custom config) where some but not all of block_q_dq, block_kv_dq, block_q_dkv, block_kv_dkv are provided.

Common situations: Copying a preset config and overriding only one block size; merging configs where overrides drop some fields; partial migration to a new config schema.

Related errors

• cap_method must be in [tanh, soft_sign], got {cap_method}
• Invalid backward pass implementation: {backward_pass_impl}
• cap_method must be in [tanh, soft_sign, none], got {cap_meth
• {self.block_q=} must be a multiple of 64
• {self.block_kv=} must be a multiple of 64

AI-assisted analysis of xai-org/x-algorithm@24c60942c5 (2026-08-28). Data as JSON: /api/errors/dd1e882acf828711. Report an issue: GitHub.