Crate atlas_core
Source - arch
- Does the PTX this binary carries run on the GPU it was handed?
- capabilities
- Model capabilities — config-derived feature flags.
- compute
- Hardware-agnostic compute target abstraction.
- config
- cuda_host
- The process-scoped half of the CUDA runtime: the context, the stream,
and the teardown check that guards a model swap.
- device
- Device-context wrapper + per-hardware constants.
- dtype
- error
- fault
- Process-fatal GPU fault latch.
- kernel
- kimi_k3
- Kimi K3 host graph. This slice is KDA only.
- mxfp4_e8m0
- DeepSeek-V4 native MXFP4 host unpack: packed E2M1 nibbles + E8M0 scales.
- numeric
- Host-side numeric conversions shared by every weight loader: the FP8
E4M3 decode table and the f32 -> BF16 cast.
- registry
- Global kernel registry — load PTX once, cache modules/functions/streams.
- safetensors
- The one place that turns a safetensors
data_offsets pair into a byte span. - scope
- Model teardown — ordered, fallible release of state that owns device memory.
- stream
- target
- Kernel target descriptor for (Hardware, Model_quantization) tuples.
- tensor