Skip to content

Clarify or restrict tstore_fp legality for A2/A3 vec quantized store #261

@zhoubot

Description

@zhoubot

Summary

pto.tstore_fp currently looks broader at the IR/frontend layer than the backend contract it can legally lower to on A2/A3.

Evidence

  • TStoreFPOp in PTOAS accepts generic PTO shaped-like src/fp/dst operands.
  • The verifier currently only checks:
    • src/fp/dst are PTO-shaped-like
    • src and dst have the same rank
    • src and dst have the same element type
  • The A2/A3 backend contract behind TSTORE_FP is ACC-oriented, not a generic vec-tile quantized store.
  • When trying to use it from a vec rope-cache kernel for f16 vec -> i8 GM, the verifier/backend path is not a legal match.

Impact

Frontend users can reasonably infer that pto.tstore_fp might cover generic tile quantized store, but the current backend contract on A2/A3 does not.

Request

Please either:

  • tighten PTOAS legality/verifier/docs so pto.tstore_fp is clearly ACC-quantized-store only on current A2/A3, or
  • extend lowering/backend support if a true vec quantized-store contract is intended.

Affected kernels

  • rope_quant_kvcache
  • dequant_rope_quant_kvcache
  • future rope/cache ports using int8 cache writeback

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions