Compatibility

What runs where, what converts in which direction, and what is missing, all tested under python/ and tests/.

Backends

Target	Call	Notes
PyTorch	linnet.torch.load	interpreted, or generated source (compile=True); trainable=True for autograd
JAX	linnet.jax.load, load_source, load_nnx	XLA (no VJP), differentiable generated jnp code, or either as Flax NNX
StableHLO	linnet stablehlo --bind ...	static shapes
ONNX	linnet onnx --bind ...	opset 20 text format
ONNX, packaged	linnet.onnx.export_model	checkpoint embedded; runs in any ONNX runtime
vLLM, SGLang, TGI, transformers	linnet.hf.export	llama, qwen2, qwen3, phi3, and gpt2 families
llama.cpp, Ollama	linnet.gguf.export	the same families
Triton Inference Server	linnet.triton.export	packaged ONNX, or a Python backend for stateful entries
ComfyUI	linnet-comfyui	node pack

Importers

Source	Call	Recovers
PyTorch modules	linnet.torch.export_linnet	module tree as blocks; common layers as library calls; matmul and einsum as index notation
JAX functions	linnet.jax.export_linnet	parameter pytree as blocks; softmax, sigmoid, silu, gelu, and RMS norm
StableHLO text	linnet.jax.import_stablehlo	the same, for modules from elsewhere
ONNX models	linnet.onnx.import_onnx	initializer names as the hierarchy, dim_params as generics, the common operators, PyTorch's RMS norm, SiLU, and softmax

Importers never guess: an unmapped operation stops the import by name.

Real checkpoints

The Llama and GPT-2 examples with TinyLlama-1.1B's and GPT-2's weights match transformers logits within 2e-2 in f32, with identical argmax, through KV-cache decoding. LINNET_HF_TESTS=1 runs python/linnet/tests/torch/test_hf_checkpoints.py.

Language features by backend

Feature	PyTorch	JAX and StableHLO	ONNX
Symbolic shapes	bound at load and from inputs	bound per compile	bound per export
static for	yes	unrolled	unrolled
while	yes, also in generated source	stablehlo.while, lax.while_loop	Loop
Index notation and reductions	yes	yes	yes
Optional parameters	yes	--optionals	yes
Enum constants and match	yes	folded	folded
Tuple results	yes	yes	one output per element
state members	buffers, reset_state()	threaded in and out	threaded in and out
std.random	yes	yes	yes (right shifts are logical)
Training	linnet.train: SFT, DPO, GRPO, LoRA, FSDP; transformers' Trainer and TRL	linnet.jax.train: SFT, DPO, GRPO, LoRA, FSDP over a mesh, remat; load_source with jax.grad	no
dtypes	f16, bf16, f32, f64, i8 to i64, u8, bool	as the backend supports	as opset 20 supports

Not yet

• A scan that collects per-iteration outputs; data-dependent shapes.
• Importers: erf-based GELU, softmax over an axis other than the last, general gather and scatter, custom calls.
• Activation quantization. Weights can be int8, int4, group-wise 4-bit (GPTQ, AWQ), or MXFP4.