MotifCL is a Vulkan-first native neural runtime for legacy AMD GPUs. It is maintained as a
supporting systems project for
memory-native, which is the canonical research
implementation of the finite-state optimizer-in-weight method.
The repositories answer different questions:
| Repository | Primary question |
|---|---|
memory-native |
Does the training method learn, scale, and reproduce in a familiar PyTorch/MLX environment? |
| MotifCL | Can the relevant operators and compact state execute efficiently in a native Vulkan runtime on constrained hardware? |
Claims must identify the repository, commit, backend, device, command, and witness that produced them.
- A PyTorch/T4 result is not a MotifCL Vulkan result.
- A MotifCL kernel parity test is not convergence evidence for the training method.
- A skip-capable GPU test is not backend evidence unless strict mode proves that the intended device path executed.
- Historical reports are not substitutes for a fresh release gate.
- Modeled memory and throughput are labeled separately from measured values.
For Vulkan tests, use MOTIFCL_REQUIRE_VULKAN_COMPUTE=1 where supported so an unavailable device
cannot turn a skipped path into a false green result.
The RX 580/Polaris target drives the default training design:
- Vulkan FP32 is the supported primary training path.
- BF16 training is not a target capability of Polaris.
- Full FP16 backward is not a project milestone for the RX 580 profile because the hardware lacks modern mixed-precision matrix acceleration and previous FP16 paths did not establish a useful end-to-end advantage.
- Q4/Q8/K-quant paths are intended for inference.
- Compact counter state and reversible execution are the memory-oriented training research paths.
Mixed-precision infrastructure may remain useful for portability experiments on other devices, but it must not displace correctness and measured performance on the named target.
MotifCL is suitable for:
- Vulkan runtime and kernel research;
- compact Transformer inference;
- small-model FP32 training experiments;
- operator parity and backend studies;
- memory-native state/update experiments;
- legacy-GPU portability work.
It is not presented as:
- a full replacement for PyTorch;
- a production distributed-training platform;
- proof of large-model convergence;
- a universal loader for every HF architecture;
- a hardened service for untrusted model artifacts.
- preserve strict Vulkan training witnesses;
- keep HF/GGUF compatibility claims tied to executable tests;
- separate inference support from backward/training support;
- keep generated SPIR-V synchronized with shader sources;
- improve install/export and consumer-build reliability;
- document unsupported architectures and formats explicitly.