Build operator kernels from mlir-aie kernel factories; fix Llama decode; test against kernel tolerance contracts - #220

Draft

hunhoffe wants to merge 23 commits into

develfrom

kernel-factories

Commits on Sep 25, 2026

Build kernels from mlir-aie kernel factories; port transpose and gemv

hunhoffe
and
claude
committed
aecc68a
Port the Llama decode ops and elementwise bases to kernel factories

hunhoffe
and
claude
committed
29fd2a7
Port gemm to the linalg.mm / zero / convert_copy kernel factories

hunhoffe
and
claude
committed
6268a74
Port mha to the linalg.mha / zero / passthrough kernel factories

hunhoffe
and
claude
committed
17624f3
Port dequant and mem_copy to the expand / passthrough kernel factories

hunhoffe
and
claude
committed
13979d3
Port flm/dequant to quant.q4nx_dequant; retire iron/common/kernels.py

hunhoffe
and
claude
committed
774e248
Resolve the factory-kernel output dir before handing it to mlir-aie

hunhoffe
and
claude
committed
07ebe07
Llama decode: mask softmax to the current context, not a running sum

hunhoffe
and
claude
committed
d8cad99
Llama test: check NPU logits against an fp32 CPU reference

hunhoffe
and
claude
committed
480fa95
Llama: flush the prefill KV hand-off to the device, not from it

hunhoffe
and
claude
committed
e4fed44
Llama test: guard run-to-run determinism of NPU logits

hunhoffe
and
claude
committed
263b83b
Flush scratch in the full-ELF sequence callable before each dispatch

hunhoffe
and
claude
committed
3731ae0
Tests: lean on mlir-aie's verify and benchmark utilities

hunhoffe
and
claude
committed
4d3eeea
Tests: judge 1:1 operators by their kernel's tolerance contract

hunhoffe
and
claude
committed
afbc276
Tests: one reference per operator, judged by its kernel's contract

hunhoffe
and
claude
committed
4f4c061
Sequences: judge compare mode by each step's kernel contract

hunhoffe
and
claude
committed
44a5667
Tests and docs: small cleanups

hunhoffe
and
claude
committed
f2f0bc1
Tracing: take the trace buffer from mlir-aie; real tests, no mocks

hunhoffe
and
claude
committed
ff3aaf5
Tracing: write and decode dumps with mlir-aie's TraceConfig

hunhoffe
and
claude
committed
b8a909e
Merge remote-tracking branch 'origin/devel' into kernel-factories

Copilot
and
hunhoffe
authored
44b849d
Make missing Llama weights fail in CI

Copilot
and
hunhoffe
authored
a60d170
Clarify AXPY coefficient rounding and add non-integer regressions

Copilot
and
hunhoffe
authored
a31d9a6
Follow mlir-aie #3801's aie_kernels layout

hunhoffe
and
claude
committed
c4bdfd2

Navigation Menu

Build operator kernels from mlir-aie kernel factories; fix Llama decode; test against kernel tolerance contracts - #220

Build operator kernels from mlir-aie kernel factories; fix Llama decode; test against kernel tolerance contracts#220
hunhoffe wants to merge 23 commits into
develfrom
kernel-factories

Commits on Sep 25, 2026

Build kernels from mlir-aie kernel factories; port transpose and gemv

Port the Llama decode ops and elementwise bases to kernel factories

Port gemm to the linalg.mm / zero / convert_copy kernel factories

Port mha to the linalg.mha / zero / passthrough kernel factories

Port dequant and mem_copy to the expand / passthrough kernel factories

Port flm/dequant to quant.q4nx_dequant; retire iron/common/kernels.py

Resolve the factory-kernel output dir before handing it to mlir-aie

Llama decode: mask softmax to the current context, not a running sum

Llama test: check NPU logits against an fp32 CPU reference

Llama: flush the prefill KV hand-off to the device, not from it

Llama test: guard run-to-run determinism of NPU logits

Flush scratch in the full-ELF sequence callable before each dispatch

Tests: lean on mlir-aie's verify and benchmark utilities

Tests: judge 1:1 operators by their kernel's tolerance contract

Tests: one reference per operator, judged by its kernel's contract

Sequences: judge compare mode by each step's kernel contract

Tests and docs: small cleanups

Tracing: take the trace buffer from mlir-aie; real tests, no mocks

Tracing: write and decode dumps with mlir-aie's TraceConfig

Merge remote-tracking branch 'origin/devel' into kernel-factories

Make missing Llama weights fail in CI

Clarify AXPY coefficient rounding and add non-integer regressions

Follow mlir-aie #3801's aie_kernels layout