xai-org/x-algorithm · error · ValueError

beam_width must be >= 1, got {beam_width}

Error message

beam_width must be >= 1, got {beam_width}

What it means

RecsysSIDRetrievalModel.generate validates beam_width before delegating to generate_parallel: beam search requires at least one candidate, so beam_width < 1 raises ValueError immediately. Unlike the runner-side error (sid_retrieval_runner), this is the model-level API check with no flag hint.

Source

Thrown at phoenix/xrex/models/recsys_sid_retrieval_model.py:351

        )
        return ln

    @hk.transparent
    def _decode_hidden(self, emb, mask, level, apply_ln):
        hidden = self._run_transformer(emb, mask, is_training=False)
        h_last = apply_ln(hidden[:, -1:, :])[:, 0, :]
        return self._predict_sid_level(h_last, level)

    @hk.transparent
    def generate(
        self,
        batch: RecsysFeaturesBatch,
        recsys_embeddings: RecsysEmbeddingsParameter,
        beam_width: int = 1,
        decode_levels: int | None = None,
    ) -> tuple[jax.Array, jax.Array]:
        if beam_width < 1:
            raise ValueError(f"beam_width must be >= 1, got {beam_width}")
        return self.generate_parallel(
            batch,
            recsys_embeddings,
            beam_width=beam_width,
            decode_levels=decode_levels,
        )

    @hk.transparent
    def _build_block_causal_mask(self, B: int, H: int, K: int, level: int) -> jax.Array:
        T = H + K * level
        q = jnp.arange(T)[:, None]
        kv = jnp.arange(T)[None, :]
        q_hist, kv_hist = q < H, kv < H
        hist_causal = q_hist & kv_hist & (kv <= q)
        beam_to_hist = (~q_hist) & kv_hist
        same_beam = ((q - H) // level) == ((kv - H) // level)
        beam_causal = (~q_hist) & (~kv_hist) & same_beam & (kv <= q)
        base = hist_causal | beam_to_hist | beam_causal

View on GitHub (pinned to 24c60942c5)

Solutions

1. Pass beam_width=1 for greedy decoding or a larger value for beam search.
2. Clamp sweep/config values: beam_width = max(1, int(beam_width)).
3. Use generate_parallel directly only if you also enforce the same bound.

Example fix

# before
scores, beams = model.generate(batch, emb, beam_width=0)

# after
scores, beams = model.generate(batch, emb, beam_width=1)

beam_width = max(1, int(beam_width))
result = model.generate(batch, emb, beam_width=beam_width)

def is_valid_beam_width(bw) -> bool:
    return isinstance(bw, int) and not isinstance(bw, bool) and bw >= 1

When it happens

Trigger: Calling model.generate(batch, embeddings, beam_width=0) directly, or wiring a config/sweep value that computes to 0 into the generate call.

Common situations: beam_width=0 used to mean 'greedy/off' in older code; programmatic sweeps starting at 0; config default of 0.

Understand the failure class

Background: "Must be a positive integer", "Invalid value", "Unsupported": the invalid-argument-value error family, when a library rejects the value you pass — this error's family across 35 libraries.

Related errors

• beam_width must be >= 1, got {self.beam_width}. Use --beam_w
• k ({k}) must be <= n ({n})
• valid_block_upper and valid_block_lower must be provided tog
• {name}_block_cnt and {name}_block_idx must both be provided
• window_ms must be >= 0, got {window_ms}

AI-assisted analysis of xai-org/x-algorithm@24c60942c5 (2026-08-28). Data as JSON: /api/errors/53a99aa99c71c0b7. Report an issue: GitHub.