DotGeneralConfig lhs_rank/rhs_rank fields become stale, causing runtime RankMismatch panics

[shinaoka]

Summary

DotGeneralConfig.lhs_rank / rhs_rank are redundant metadata that duplicate the actual operand rank. They become stale during compilation, causing runtime panics on complex N-ary einsum contractions (84+ tensors).

Symptoms

When running einsum benchmarks with many tensors, the following panics occur:

• RankMismatch(3, 6) — a Permute instruction has perm.len() == 3 but the input tensor has rank 6
• index out of bounds: the len is 3 but the index is 3 — DotGeneralConfig.lhs_batch_dims contains index 3 but the tensor has rank 3 (indices 0-2)
• Segfaults from cascading invalid memory access

Root Cause

1. Graph construction: DotGeneralConfig stores lhs_rank/rhs_rank set from logical operand shapes at graph-build time
2. Output rank computation: TracedTensor::dot_general (traced.rs:759) computes output rank from these stored ranks → downstream Permute perm lengths are based on this computed rank
3. Runtime divergence: The GEMM backend (gemm/mod.rs) uses actual tensor shape.len() to compute free dims and output shape, which may differ from the stored rank
4. Mismatch: Downstream Permute instructions have perm lengths based on the (wrong) logical output rank, but receive tensors with a different actual rank

The StableHLO IR has no shape table (StableHloInstruction carries no operand shape/rank metadata — stablehlo.rs:111), so compiler passes cannot detect or repair stale rank fields.

Affected Code

Location	Role
tenferro-tensor/src/config.rs:22-24	Stores lhs_rank/rhs_rank
tenferro/src/traced.rs:759	Computes output rank from stored ranks
tenferro-einsum/src/builder.rs:272	Sets rank from LabeledVal.shape.len()
tenferro-tensor/src/cpu/gemm/mod.rs:103	Uses actual tensor rank at runtime
tenferro/src/exec.rs:176	Applies Permute with stale perm length

Reproduction

Run the einsum benchmark with instances having 84+ tensors (e.g., lm_batch_likelihood_brackets_4_4d, str_matrix_chain_multiplication_100).

Recommended Fix

1. Add shape/rank tracking to StableHLO (side table or per-instruction metadata)
2. Remove lhs_rank/rhs_rank from DotGeneralConfig; derive rank from the shape table at every usage site
3. Add validation at rewrite boundaries: perm.len() == producer_rank, all dim indices in range, batch/contract sets disjoint

Why simple removal is insufficient today

lhs_rank/rhs_rank cannot be simply deleted because:

• AD rules (ad/contraction.rs) need operand rank but only receive ValRefs, not shapes
• Compiler passes (compiler.rs) need rank but StableHloInstruction has no shape metadata
• max(dim_indices) + 1 does not recover the rank (e.g., rank 4 with contracting_dims=[1], batch_dims=[] has free dims [0,2,3] but max=1+1=2≠4)

Additional Finding

transpose_folding has a separate correctness issue: it remaps batch/contracting dims through the transpose permutation but does not account for free-dim order changes. This can silently change the DotGeneral output axis order without changing rank. (Not the cause of the RankMismatch panics, but should be fixed.)

Related PRs

• fix(gemm): batch-trailing output + fallback copy for unfusable layouts #657 — batch-trailing output convention
• fix(gemm): use actual tensor rank in canonical_gemm_layout fallback #659 — use actual tensor rank in canonical_gemm_layout
• fix(compiler): disable buggy dot_decomposer pass #661 — disable buggy dot_decomposer
• feat(einsum): accept Unicode symbol labels in notation #662 — Unicode einsum labels

[shinaoka]

Mitigated by #707 (merged): rank validation is now performed at trace/eager time, catching stale metadata before runtime panics. The root cause (redundant lhs_rank/rhs_rank fields in DotGeneralConfig) is documented but not yet fully addressed — keeping open for the structural fix.

[shinaoka]

Status check against main (commit 3676d6c)

• lhs_rank / rhs_rank fields still present at tenferro-tensor/src/config.rs:27-28. Root design unchanged.
• Mitigation added (not a root fix): DotGeneralConfig::validate_ranks() at config.rs:49-60 detects stale rank metadata. It is now called from:
  • tenferro/src/traced.rs:875 (graph-build check)
  • tenferro-tensor/src/cpu/gemm/mod.rs:96-109 (runtime check returning a RankMismatch error instead of letting the mismatch cascade into downstream panics / segfaults)
• Complete removal of the redundant fields is still blocked on the 1-layer IR refactor (docs/superpowers/plans/2026-04-17-1layer-ir-refactor.md) delivering per-instruction shape tracking.

Status: symptoms (cascading panics / segfaults on 84+ tensor contractions) now surface as a clean RankMismatch error early. Root cause (redundant rank metadata in DotGeneralConfig) not yet fixed.

[shinaoka]

Correction to previous status comment

The 2026-04-18T10:17 status comment above is incorrect about the blocker. The 1-layer IR refactor landed in #732 (merged 2026-04-17T23:09), before that status check. Verified on current main:

• tenferro/src/shape_infer.rs provides per-op shape inference
• ExecInstruction.output_shapes: Vec<Vec<DimExpr>> at tenferro/src/exec.rs:144

Revised status: shape tracking is available. Structural fix (removing redundant lhs_rank / rhs_rank from DotGeneralConfig and deriving rank from the shape table instead) is no longer blocked. The validate_ranks() mitigation from #707 remains as defense-in-depth until the fields are removed.