HLO: inline dense constants for large tensors — scale / compile-time blocker

[michalharakal]

Module

skainet-compile-hlo / ConstantOperationsConverter

Problem

All weight tensors are emitted as inline stablehlo.constant dense<[[...]]> literals. For Whisper-tiny.en, this produces a 151 MB MLIR file where most of the content is floating-point numbers in text form.

Impact

• Extremely slow to parse (minutes for iree-compile to read).
• SSA type tracking in post-processors fails on multi-megabyte lines.
• Not scalable to larger models — Whisper small/medium would be gigabytes of text MLIR.

Suggested fix options

1. External weight files: emit stablehlo.constant with a reference to an external binary file (IREE supports #util.byte_pattern or resource loading).
2. Splat constants for zeros: when VoidTensorOps produces zero tensors, emit dense<0.0> splat instead of spelling out every element.
3. Separate weight serialization: emit the MLIR structure with placeholder constants and load weights at compile time via IREE's parameter mechanism.

Context

Filed from skainet-whisper IREE GPU bring-up (2026-04-18) on branch feature/iree-vulkan-gpu targeting SKaiNET 0.18.0. The ONNX path (ONNX → iree-import-onnx → MLIR → iree-compile → VMFB) works on device today because iree-import-onnx uses external resources for weights. The native SKaiNET DSL path will need an equivalent mechanism to be scalable beyond tiny models.

Test to reproduce:

./gradlew :SKaiNET-voice:testDebugUnitTest --tests "*WhisperHloExportTest*"
# produces SKaiNET-voice/build/iree/encoder_skainet.mlir
ls -lh SKaiNET-voice/build/iree/encoder_skainet.mlir

[michalharakal]

Superseded by the design in #523, which reframes this as an architectural issue (externalize weights via IREE's util.parameter.load + .irpa archive) rather than a converter-internal serialization problem.

PR #522 lands a cosmetic splat-for-uniform fix on the inline path as partial relief. Full resolution will come from PRs B/C/D/E tracked under #523.