xai-org/x-algorithm · error · ValueError

feature_prep_enabled (candidate project-then-sum) and enable

Error message

feature_prep_enabled (candidate project-then-sum) and enable_linear_proj are mutually exclusive; enable at most one candidate-tower combine mode (or neither for mean-pool).

What it means

RecsysCandidateTower.make offers three candidate combine modes: feature_prep project-then-sum (use_project_then_sum=True), a shared linear projection (enable_linear_proj on the parent config), and the default mean-pool. Enabling both projections at once is ambiguous, so make raises ValueError telling you to pick at most one (or neither).

Source

Thrown at phoenix/xrex/models/recsys_two_tower_model.py:253

@configclass
class RecsysCandidateModelConfig(Config):
    hash_table: HashTable
    enable_linear_proj: bool = False
    emb_table_width: int = 128
    scale_config: ScaleConfig = ScaleConfig()
    max_posts: int = 10_240_000
    num_candidate_heads: int = 1

    def make(
        self,
        sharding_context: ShardingContext,
        *,
        use_post_embedding: bool = True,
        use_post_sid: bool = False,
        use_project_then_sum: bool = False,
    ):
        if use_project_then_sum and self.enable_linear_proj:
            raise ValueError(
                "feature_prep_enabled (candidate project-then-sum) and "
                "enable_linear_proj are mutually exclusive; enable at most one "
                "candidate-tower combine mode (or neither for mean-pool)."
            )

        return RecsysCandidateTower(
            self,
            sharding_context,
            use_post_embedding=use_post_embedding,
            use_post_sid=use_post_sid,
            use_project_then_sum=use_project_then_sum,
        )

    def make_post_embeddings(self):
        all_post_ids = jnp.arange(self.max_posts * 2).reshape(-1, 2).astype(jnp.int32)
        all_author_ids = jnp.arange(self.max_posts * 2).reshape(-1, 2).astype(jnp.int32)
        all_dataset_types = jnp.full(
            (self.max_posts, 1), RetrievalDataset.PAD.value, dtype=jnp.int32

View on GitHub (pinned to 24c60942c5)

Solutions

Set enable_linear_proj=false when using project-then-sum (feature_prep).
Or keep enable_linear_proj=true and pass use_project_then_sum=False.
Set both off for the default mean-pool combine.

Example fix

# before
config.enable_linear_proj = True
tower = Tower.make(..., use_project_then_sum=True)

# after
config.enable_linear_proj = False
tower = Tower.make(..., use_project_then_sum=True)

Defensive patterns

Strategy: validation

Validate before calling

assert not (use_project_then_sum and config.enable_linear_proj), "mutually exclusive combine modes"

Prevention

Represent combine mode as a single enum field instead of two booleans.
Validate tower construction args in a shared factory.

When it happens

Trigger: Constructing the two-tower model with feature_prep_enabled=true in config (which sets use_project_then_sum) while enable_linear_proj is also true, or passing use_project_then_sum=True explicitly alongside a config with enable_linear_proj.

Common situations: Turning on the newer project-then-sum path while an old config still sets enable_linear_proj; merging config overlays that each enable one mode.

Related errors

AI-assisted analysis of xai-org/x-algorithm@24c60942c5 (2026-08-28). Data as JSON: /api/errors/a2d30b865a65ddb3. Report an issue: GitHub.