Systems

Learning systems and where they run.

A system is a model, a learner, a run policy and a pairing with a device, compiled whole into one executable. Each state below links to the record behind it. As of build 2877a0c1.

Runs on hardware

A recorded run on the device: train, resume from a checkpoint, predict.

Verified, run pending

The executable builds and verifies; a fresh run on the device is still to come.

In development

Maintained source that does not yet run end to end on a device.

Systems

Every system in the repository

Each is a program under systems/. Model code is hardware-independent; the pairing chooses the device.

Trains on hardware

Coppelius

A 57.7M-parameter transformer: 16 blocks, width 512, rotary attention, gated GELU, bfloat16 matrix products and AdamW, trained on WikiText-103.

Runs on: x86-64 or AArch64 host → sm_86 (RTX 3070, 3090) and sm_121 (GB10); sm_89 envelope for the RTX 4090
Status: Trains, resumes from a checkpoint and predicts on the RTX 3070, RTX 3090 and DGX Spark GB10. On the RTX 3090 it ran 10,000 steps and trained 1.51× faster than PyTorch with CUDA graphs in matched runs. A fresh RTX 4090 run is pending.
Namespaces: Coppelius · Training

System README RTX 3090 vs PyTorch GB10 training record System definition Open in catalogue

Trains on hardware

Bob

The smallest system: a next-byte predictor with one 256 × 256 FP32 logits table, exact full-vocabulary cross-entropy and AdamW.

Runs on: x86-64 host → sm_86 (RTX 3070, 3090); sm_89 envelope for the RTX 4090
Status: Trains, resumes and predicts on the RTX 3070 and 3090. Its RTX 4090 executable builds and passes envelope verification; a fresh run on that card is pending.
Namespaces: Bob

System README System definition Physical qualification Open in catalogue

In development

Baguette

A 321M-parameter grouped-query transformer: 80 blocks, width 576, 9 query and 3 key/value heads, 4,096-token context.

Runs on: x86-64 host → sm_86 (RTX 3090)
Status: The forward plan and the first optimizer update compile as a whole-program plan. A complete training executable has not been built yet; memory figures are plans until a run confirms them.
Namespaces: Baguette

Goal and plan System definition Open in catalogue

In development

KickDelta

A small harmonic language model: eight oscillator groups feeding a delta-rule memory whose read is the next-byte prediction (7,227 parameters).

Runs on: x86-64 host
Status: The complete forward pass is checked Alpha source and matches the reference model within its declared tolerance. Gradients through every component and an AdamW cell exist; batching and the full parameter update are still open.
Namespaces: KickDelta

System README Open in catalogue

In development

Polf

A port of the parameter-golf GPT baseline: 17.1M parameters, nine blocks, width 512, grouped-query attention, tied embeddings and the Muon optimizer.

Runs on: Not yet paired with a device
Status: The model geometry, parameter roles and the Muon and Adam update equations are checked in Alpha. The forward and backward GPU realization, and so training, are still to come.
Namespaces: Polf · Model.ParameterGolf

System README Open in catalogue

Paused

Vathek

A pretrained reader and decoder with a trainable bridge between them, for structured extraction from text.

Runs on: AArch64 host → sm_121 (GB10)
Status: Short training runs on synthetic documents pass on the GB10. Work is paused; the handoff records the verified results and how to resume.
Namespaces: Vathek

System README Handoff Open in catalogue

Historical

Alpha-ER

The original Alpha model family, kept with its model-specific training and inference code.

Runs on: x86-64 host → sm_86
Status: Maintained source, not currently qualified on a device. Its sm_86 pipeline has a known parse defect and no importers.
Namespaces: Model · Training · Inference · Accelerator · Domain · Pipeline

System README Open in catalogue

Targets

Hosts, GPU architectures and cards

Emitting code for an architecture and running a model on a particular card are different claims, so they have separate rows. Hardware has a page for each card.

Target	Layer	State	What it does	Evidence
x86-64 LinuxELF64 / x86-64 / Linux	Host output	Emits and runs	The default host target: static executables with no libc, including the host side of every RTX training executable. A host target says nothing about a GPU. GPU execution also needs a device image the card admits.	Sources Target implementation
AArch64 LinuxELF64 / AArch64 / Linux	Host output	Emits and runs	The host of the DGX Spark: the Coppelius and Vathek GB10 executables run their host code here. Direct lowering only. The evaluator-carrying artifact and native object output remain x86-64 only.	Sources AArch64 ELF owner Bring-up record
sm_86 (Ampere)NVIDIA Ampere / compute capability 8.6	Accelerator architecture	Trains models	The primary GPU architecture: instruction encoding, scoreboard checking, launch descriptors and command streams for the RTX 3070 and 3090. Built through each system's build roots rather than the generic --target flag.	Sources Architecture owner Target contract Linux/NVIDIA platform
sm_121 (Blackwell)NVIDIA Blackwell / compute capability 12.1	Accelerator architecture	Trains models	A compiler back end lowers the sm_86 kernels to sm_121 and emits the Blackwell launch descriptors, so the same model source runs on the GB10. Kernels are written for sm_86 and lowered; nothing is yet tuned specifically for sm_121.	Sources sm_121 lowering GB10 training record
sm_89 envelope (Ada)NVIDIA Ada / compute capability 8.9	Accelerator architecture	Verified · run pending	Runs the sm_86 instructions on Ada cards under sm_89 launch descriptors and driver contracts. The current RTX 4090 executables pass compiler and independent envelope verification. A fresh physical run is pending.	Sources sm_89 platform profile RTX 4090 integration report
GeForce RTX 3070GA104 / sm_86 / 8 GiB	Physical device	Qualified	Bob and Coppelius train, resume and predict here. The smallest-memory qualified card.	Sources Physical qualification
GeForce RTX 3090GA102 / sm_86 / 24 GiB	Physical device	Qualified	Bob and Coppelius train, resume and predict here. Coppelius ran 10,000 steps and trained 1.51× faster than PyTorch with CUDA graphs. Performance work happens on this card; Baguette's first runs will too.	Sources RTX 3090 vs PyTorch Physical qualification
GeForce RTX 4090AD102 / sm_89 / 24 GiB	Physical device	Verified · run pending	Coppelius was brought up on this card before the current refactor; the current Bob and Coppelius executables pass local verification. A fresh train, resume and predict run is pending: the rented cards available so far failed to open /dev/nvidia-uvm.	Sources RTX 4090 evidence
DGX Spark (GB10)20-core AArch64 / sm_121 / 128 GB unified memory	Physical device	Qualified	Coppelius trains, resumes and predicts here, including 10,000-step runs; Vathek's short training runs pass. The GPU shares the host's memory, so bandwidth and capacity are shared with the CPU.	Sources GB10 training record GB10 performance record Measured profile

Compatibility

Which system runs where

Derived from the rows above: the host each system emits, the GPU architectures it targets and the cards with a record for it.

System	Host	GPU architecture	Cards
CoppeliusTrains on hardware	x86-64 Linux AArch64 Linux	sm_86 (Ampere)sm_121 (Blackwell)sm_89 envelope (Ada)	GeForce RTX 3070 GeForce RTX 3090 GeForce RTX 4090 DGX Spark (GB10)
BobTrains on hardware	x86-64 Linux	sm_86 (Ampere)sm_89 envelope (Ada)	GeForce RTX 3070 GeForce RTX 3090 GeForce RTX 4090
BaguetteIn development	x86-64 Linux	sm_86 (Ampere)	GeForce RTX 3090
KickDeltaIn development	x86-64 Linux	None yet	None yet
PolfIn development	None yet	None yet	None yet
VathekPaused	AArch64 Linux	sm_121 (Blackwell)	DGX Spark (GB10)
Alpha-ERHistorical	x86-64 Linux	sm_86 (Ampere)	None yet