WIP: add nkigen-lite as a standalone IR-based kernel generation backend - #59

Draft

ymwangg wants to merge 107 commits into

mainfrom

Commits on Jun 2, 2026

feat: add nkigen-lite as a standalone IR-based kernel generation backend
ymwangg
committed
ac1ba42

Commits on Jun 3, 2026

feat: integrate nkigen-lite as a third backend for nkipy
ymwangg
committed
934f863
test: run shared test suites with both hlo and nkigen-lite backends
ymwangg
committed
1ed3441
fix: support keepdims=False reduce and scalar outputs in nkigen-lite
ymwangg
committed
1d93cbb
fix: improve nkigen-lite op coverage and input validation
ymwangg
committed
b4b6aa3
fix: add expand_dims validation and skip HLO-specific error tests
ymwangg
committed
69c7f1b
feat: support cross-lane MIN reduction via negate-max-negate
ymwangg
committed
2643fc9
test: use on_device_test in alias tests for backend-agnostic execution
ymwangg
committed
176ee95
fix: handle scalar broadcast and f16 tensor_scalar_arith
ymwangg
committed
b4e23cc
feat: add native bitwise ops (AND, OR, XOR) via tensor_tensor_bitvec
ymwangg
committed
82feeec
feat: add comparison and scalar bitvec primitives; rewrite floor
ymwangg
committed
770b679

Commits on Jun 4, 2026

feat: support strided slicing and fix numpy array in dynamic_update_slice
ymwangg
committed
8cdd829
feat: add float8_e4m3 (IEEE) dtype support
ymwangg
committed
a36c7d0
fix: set NEURON_RT_VISIBLE_CORES for xdist parallel test isolation
ymwangg
committed
d7f3102
feat: add comparison ops and where lowering for nkigen-lite
ymwangg
committed
652a484
feat: add cos decomposition via sin(x + π/2)
ymwangg
committed
f5ec24b
feat: add dot op via composed_impl routing to matmul
ymwangg
committed
1defcad

Commits on Jun 5, 2026

Commits on Jun 23, 2026

feat: support float8_e4m3fn on nkigen-lite backend
ymwangg
committed
ea53687
feat: add iota primitive to nkigen-lite tensor_ir
ymwangg
committed
ab68be6

Commits on Jun 24, 2026

feat: add tril/triu/diag/trace for nkigen-lite
ymwangg
committed
a9e5886
feat: add pad/flip/roll/tile/diff for nkigen-lite
ymwangg
committed
75c7dae
feat: add argmax/argmin for nkigen-lite
ymwangg
committed
79289c7
feat: support non-uniform constants + cumsum for nkigen-lite
ymwangg
committed
b9d4c0a
feat: add conv2d/conv3d for nkigen-lite via im2col
ymwangg
committed
26cdc0b
feat: add repeat + split-with-indices for nkigen-lite
ymwangg
committed
3412eb2
fix: range-reduce sin/cos arguments for nkigen-lite
ymwangg
committed
e771269
feat: add topk for nkigen-lite via iterative max-extraction
ymwangg
committed
9198d74
feat: use hardware max8 primitive for nkigen-lite topk values
ymwangg
committed
7e8662c
feat: topk via canonical max8 + match_replace8 scan (k>8 support)
ymwangg
committed
eaf0650
test: fix nkigen-lite xdist failures (core isolation + spike import clash)
ymwangg
committed
f6ec7f1
test: skip slow/hanging conv2d and conv3d tests on nkigen-lite
ymwangg
committed
a1f10a3

Commits on Jun 25, 2026

Commits on Jun 26, 2026

perf: collapse adjacent in-order axes in transpose lowering (8x on Qwen conv3d weight)
ymwangg
committed
146d745

Commits on Jun 27, 2026

Commits on Jun 29, 2026

Commits on Jun 30, 2026

Commits on Jul 1, 2026

Commits on Jul 6, 2026

Commits on Jul 7, 2026

Commits on Jul 8, 2026

Commits on Jul 16, 2026