Skip to content

support-matrix.md

Latest commit

 

History

History
62 lines (49 loc) · 2.46 KB

support-matrix.md

File metadata and controls

62 lines (49 loc) · 2.46 KB
Raw

Candle 0.1 Support Matrix (NPU-First)

This matrix defines the effective support scope for Candle 0.1.x.

For NPU installation and runtime prerequisites, see install-npu.md.

GA

  • Ascend 910B single-card training core path on a host where the newest installed CANN toolkit is exposed through /usr/local/Ascend/ascend-toolkit/latest and sourced into the shell.
    • tensor creation,
    • forward ops used by baseline training,
    • autograd backward,
    • optimizer step,
    • checkpoint save/load on the critical path.
  • CPU backend for development and CI baseline.

Experimental

  • Additional NPU ops outside the baseline training path.
  • Transformers compatibility runner (tests/run_test.py) and related patches.
  • Distributed/HCCL collectives: all core collectives implemented; multi-card scenarios validated via the tests/distributed/test_hccl_*_multicard.py suite (requires 2+ Ascend cards).
  • Parallel training stack (single-process contracts fully tested; multi-card requires HCCL hardware):
    • DeviceMesh / init_device_mesh
    • DTensor with Shard / Replicate / Partial placements; Cython fastpath active when the compiled extension is present.
    • Tensor Parallel: parallelize_module, ColwiseParallel, RowwiseParallel for nn.Linear; SequenceParallel deferred.
    • Composable FSDP2: fully_shard with shard bookkeeping in Cython.
    • DDP: bucket reducer with Cython fastpath and comm-hook support.
    • Pipeline parallel: PipelineStage, ScheduleGPipe, TensorChunkSpec, split_microbatches; minimal manual schedule only.
    • Shard-aware distributed checkpoint (save / load with per-rank shards).

Not Supported In 0.1 Scope

  • CUDA / NCCL backends.
  • SequenceParallel, PrepareModuleInput, PrepareModuleOutput.
  • FullyShardedDataParallel (legacy FSDP1 wrapper).
  • transformer_auto_wrap_policy.
  • 2D mesh tensor parallel.
  • True JIT/compile acceleration backends.
  • Full ONNX export compatibility guarantees.

Validation Gates

0.1 release quality is evaluated by:

  • local Ascend 910B NPU gate set:
    • tests/npu/test_npu_golden_training_loop.py,
    • tests/npu/test_npu_training_checkpoint_continuity.py,
    • tests/npu/test_mul_scalar_regression_npu.py,
    • tests/npu/test_no_cpu_fallback_npu.py.
  • CPU + contract CI baseline.
  • MPS CI baseline for cross-backend regression visibility.

Runtime Rule

  • NPU execution paths must remain on NPU.
  • Runtime fallback from NPU kernels to CPU is not allowed.