[DSL] Expose F32_FP8 conversion from packed_i32

[big-yellow-duck]

Summary

• add a typed fx.rocdl.cvt_f32_fp8 wrapper for the scalar ROCDL conversion
• export it through the public ROCDL expression API
• add a backend-agnostic test for the emitted operation and result type

Motivation

RDNA4 SplitKV needs scalar FP8 conversion for irregular key/value loads where a packed conversion is not applicable. The underlying MLIR ROCDL operation already exists, but FlyDSL does not currently expose it.

API stability

This is an additive stable API change. The stable API catalog gains flydsl.expr.rocdl.cvt_f32_fp8; no existing entries are removed or changed.

Validation

• bash scripts/check_python_style.sh
• python3 -m pytest tests/unit/test_rocdl_conversions.py -q
• stable API catalog comparison against main

[coderfeli]

why not Float32() directly?

[big-yellow-duck]

Float32() works when its operand is already a typed FP8 value, but this API is needed for packed raw data.

The consumer is the SplitKV kernel in vLLM PR #55996. It performs a 64-bit copy into an eight-element Uint8 register fragment, then loads that fragment and dequantizes its eight FP8 payloads.

The helper bitcasts those eight bytes into two packed i32 words and decodes bytes 0–3 from each word. This matches the ROCDL/hardware interface: an i32 source plus a byte selector.

Calling Float32(packed_i32) would instead perform a numerical integer-to-float conversion of the entire word; it would not select and interpret one byte as FP8. We could conceptually bitcast the bytes to Float8E4M3FN first and then convert to Float32, but the current FlyDSL ROCm pipeline does not lower that FP8 arith.extf path to this ROCDL instruction.

IF there was an api to do Float32(packged_i32) and compile down to cvt_f32_fp8 instuction that would be best.

[big-yellow-duck]

Hi @coderfeli is it all ok from your side? hope we can get this merged into the next release