Add Tealite: pure-Python TransformerEngine for ROCm/AMD GPUs#581

Open

jayfurmanek wants to merge 103 commits intodevfrom

furmanek/dev-lite

Commits on Apr 6, 2026

Add TE Lite: pure-Python replacement for C++ extensions (NVTE_LITE=1)

jayfurmanek
and
claude
committed
Fix TE Lite Phase 0: GEMM, norms, and attention backend for GPU verification

jayfurmanek
and
claude
committed
Add Phase 0 test suite for TE Lite mode

jayfurmanek
and
claude
committed

Commits on Apr 7, 2026

TE Lite Phase 2: Wire Triton norm kernels into _lite backend

jayfurmanek
and
claude
committed
Wire Triton cast kernels into _lite/quantize.py, fix FP8 recursion bug

jayfurmanek
and
claude
committed
Vectorize fp8_block_scaling functions in _lite/quantize.py

jayfurmanek
and
claude
committed
Add Triton kernels for fp8_block_scaling in _lite/quantize.py

jayfurmanek
and
claude
committed
Document bgrad_quantize fusion trade-off in _lite/quantize.py

jayfurmanek
and
claude
committed
Add AITER integration for GEMM, activations, and RoPE in _lite

jayfurmanek
and
claude
committed
Add lite-only wheel build mode (NVTE_LITE_ONLY=1)

jayfurmanek
and
claude
committed

Commits on Apr 8, 2026

Commits on Apr 9, 2026

Commits on Apr 10, 2026

Commits on Apr 16, 2026

Commits on Apr 17, 2026

Commits on Apr 20, 2026

Remove CPU-GPU syncs from lite FP8 amax/scale updates

jayfurmanek
and
claude
committed

Commits on Apr 21, 2026

Commits on Apr 22, 2026

Commits on Apr 23, 2026

Commits on Apr 24, 2026

Commits on Apr 28, 2026

Commits on Apr 29, 2026

Commits on Apr 30, 2026

Add NVTE_CONTIG_DIAG harness for full vs lite materialize attribution

jayfurmanek
and
claude
committed

Commits on May 5, 2026

Commits on May 6, 2026

Add tealite SKILLS.md: operational notes complementing the README

jayfurmanek
and
claude
committed

Commits on May 7, 2026

Commits on May 8, 2026