xai-org/x-algorithm · error · ValueError

Expected {k.shape=} to be {kv_shape} (inferred from q)

Error message

Expected {k.shape=} to be {kv_shape} (inferred from q)

What it means

After unpacking q's shape, _attention_forward infers the expected kv shape (batch_size, kv_seq_len, num_kv_heads, head_dim) from q and k, and verifies k matches exactly: batch, head_dim and kv-head count must agree with q; only seq_len may differ.

Source

Thrown at phoenix/xrex/pallas/ranker_attention_fa3.py:95

def _attention_forward(
    q,
    k,
    v,
    config: TuningConfig,
    save_residuals: bool = False,
    bound=None,
    sm_scale: float = 1.0,
    cap: float = -1.0,
    cap_method: str = "tanh",
):
    if q.ndim != 4 or k.ndim != 4 or v.ndim != 4:
        raise ValueError(f"q, k, and v should all be 4D, got: {q.ndim=}, {k.ndim=}, {v.ndim=}")
    batch_size, q_seq_len, num_q_heads, head_dim = q.shape
    _, kv_seq_len, num_kv_heads, _ = k.shape
    kv_shape = (batch_size, kv_seq_len, num_kv_heads, head_dim)
    if k.shape != kv_shape:
        raise ValueError(f"Expected {k.shape=} to be {kv_shape} (inferred from q)")
    if v.shape != kv_shape:
        raise ValueError(f"Expected {v.shape=} to be {kv_shape} (inferred from q)")
    if (dtype := q.dtype) != k.dtype or dtype != v.dtype:
        raise ValueError(
            f"q, k, and v should all have the same dtype, got: {q.dtype}, {k.dtype}, {v.dtype}"
        )
    if num_q_heads % num_kv_heads:
        raise ValueError(f"{num_q_heads=} must be divisible by and {num_kv_heads=}")
    q_heads_per_kv_head = num_q_heads // num_kv_heads
    if head_dim % 64:
        raise ValueError(f"{head_dim=} must be divisible by 64")
    if jnp.dtype(dtype) not in map(jnp.dtype, [jnp.float16, jnp.bfloat16]):
        raise NotImplementedError(f"Only f16 and bf16 are supported, got dtype: {dtype}")

    max_concurrent_steps = min(config.max_concurrent_steps, kv_seq_len // config.block_kv)
    block_q, block_kv = config.block_q, config.block_kv
    if kv_seq_len % block_kv:
        raise ValueError(f"{kv_seq_len=} must be a multiple of {block_kv=}")

View on GitHub (pinned to 24c60942c5)

Solutions

  1. Group kv heads correctly: num_kv_heads = num_q_heads / group_size
  2. Ensure batch size and head_dim match q exactly
  3. Verify with an assert/print of q.shape and k.shape before the call

Example fix

# before
k = k  # (B, S_kv, num_q_heads, D) — ungrouped
# after
k = k[:, :, :num_kv_heads, :]  # or repeat correctly for GQA: (B, S_kv, num_kv_heads, D)
Defensive patterns

Strategy: validation

Validate before calling

B, _, Hq, D = q.shape
assert k.shape == (B, k.shape[1], k.shape[2], D) and k.shape[2] in divisors(Hq)

Prevention

When it happens

Trigger: k with a different batch size, head_dim, or num_kv_heads than q (e.g. k reshaped to (B, S_kv, H_q, D) instead of grouping kv heads for GQA).

Common situations: GQA setups where k/v still carry the full query-head count; mixed dtypes/layouts after rearrange; off-by-one head grouping in MQA.

Understand the failure class

Background: Tensor shape mismatch errors ("must have shape", "expected shape ... got ..."): when tensor dimensions disagree with what an op or layer was told to expect — this error's family across 6 libraries.

Related errors

AI-assisted analysis of xai-org/x-algorithm@24c60942c5 (2026-08-28). Data as JSON: /api/errors/d8b12f4331555373. Report an issue: GitHub.