Ocannl (neural_nets

module Row = Ocannl_tensor.Row

module Shape = Ocannl_tensor.Shape

module Tensor = Ocannl_tensor.Tensor

module Operation = Ocannl_tensor.Operation

module PrintBox_utils = Ocannl_tensor.PrintBox_utils

module Train : sig ... end

User-facing modules

module Calibrate : sig ... end

Memory-bandwidth calibration (gh-ocannl-578): STREAM-style streaming kernels whose byte counts are exact by construction — one access site per tensor node and direction, no guards — so the calibration rows they append under config autotune_calibration_file satisfy the fitter's per-leg exactness rule and can constrain the envelope's memory leg (model_peak_memory_bandwidth). Matmul-family tuning data cannot: its rows are compute-bound (their bytes-over-time understates achievable bandwidth no matter how exactly counted) and typically carry guards-taken/union upper bounds, so the fit of the memory leg needs exactly this kind of calibration diversity — see tools/calibrate_bandwidth.ml for the command-line entry point and Ir.Cost_model.Calibration for the fit semantics.

module Parallel : sig ... end

Data-parallel training over independent per-shard backend contexts, and the host-level sharding helpers it is built from.

module Nn_blocks : sig ... end

module Demo_models : sig ... end

module Persistence : sig ... end

module Checkpoint : sig ... end

module Precision_policy : sig ... end

Precision-assignment policy over a model (gh-ocannl-492 task 1): assigns storage precisions across a tensor graph instead of the user hand-annotating every tensor. This is the storage-precision sibling of the compute-precision policy in Ir.Numerics — the two are the "one roof" gh-ocannl-478 asks for.

module Mixed_prec : sig ... end

Mixed-precision training recipe (gh-ocannl-492): master weights via cast twins, and dynamic loss scaling. This is the torch-AMP / JAX-policy recipe translated to OCANNL's structures; it composes with Precision_policy (storage-precision assignment over the rest of the graph) and with Ir.Numerics (compute-precision policy, gh-ocannl-478).

module Safetensors : sig ... end