Crate spark_runtime
Source - buffers
- Pre-allocated GPU buffer arena for intermediate tensors.
- cublaslt
- Minimal cuBLASLt FFI for the high-efficiency GEMM path (
ATLAS_CUBLAS_GEMM). - cuda_backend
- Real CUDA GPU backend using AtlasRegistry.
- cutlass
- Optional CUTLASS host-wrapper FFI for de-risking GB10 GEMM replacements.
- fast_weights
- Fast safetensors loader (InstantTensor-style) — pure Rust.
- flashinfer
- Host-callable FlashInfer ragged/varlen prefill attention FFI (GB10/sm_121).
- gpu
- GPU backend abstraction (SBIO IORouter for GPU operations).
- kernel_args
- Type-safe kernel argument builder for CUDA + Metal kernel launches.
- kernel_audit
- Startup kernel-resolution audit + embedded-kernel-set table.
- kv_cache
- Paged KV cache block allocator.
- kv_dequant
- Host-side dequantization of paged KV cache blocks → BF16.
- kv_spill
- File-backed KV cache spill manager.
- launch_trace
- ANOMALIES A56 diagnostic: record every GPU op a step enqueues, so two
consecutive steps can be diffed.
- op_cache
OpCache — per-backend kernel handles and scratch buffers.- pinned_hosts
- Which host buffers are page-locked, and therefore which
copy_h2d_async
calls are genuinely asynchronous. - prefix_cache
- Prefix caching trait for KV block reuse (SDD).
- progress
- Structured startup-progress events for the Atlas TUI.
- radix_tree
- Radix tree prefix cache for KV block reuse.
- run_metrics
- Run mailboxes — the observability surfaces that stay process-global on
purpose, and the one call that keeps them honest across a model swap.
- sampler
- Token sampling strategies.
- weights
- Weight loading from safetensors files (SBIO IORouter for filesystem I/O).
- set_ssm_tail_midchunk
- Default-ON switch for MID-CHUNK tail SSM capture (opt-out
ATLAS_SSM_TAIL_MIDCHUNK=0). - ssm_tail_boundary
- Last paged-KV block boundary strictly below
total_tokens. - ssm_tail_ckpt_enabled
- OPT-IN switch for the tail checkpoint (
ATLAS_SSM_TAIL_CKPT=1). - ssm_tail_midchunk_enabled