davidfoss commited on
Commit
354bc76
·
verified ·
1 Parent(s): 843ee73

Fix inline math rendering in model card

Browse files
Files changed (1) hide show
  1. README.md +9 -9
README.md CHANGED
@@ -34,7 +34,7 @@ tags:
34
 
35
  [Paper](paper/Wirkung_2026_Origami_HCGM.pdf) · [DOI](https://doi.org/10.13140/RG.2.2.31826.72647) · [Runtime (PyPI)](https://pypi.org/project/wirkung/) · [Wirkung](https://wirku.ng) · [Base weights](https://huggingface.co/Qwen/Qwen3.5-27B) · [Contact](mailto:contact@wirku.ng) · [License](LICENSE)
36
 
37
- **`Origami-27B` is a 27.80B-parameter Holographic Causal Graph Machine (HCGM) Preview compiled in closed form from `Qwen/Qwen3.5-27B` with zero retraining, zero fine-tuning, and zero gradient steps. It folds the 55.56 GB source checkpoint into 7.44 GB (7.46x compression), executes directly from NVMe/SSD storage via OS `mmap` with `0 GB` GPU VRAM and `0 GB` dedicated weight-RAM, and keeps live state strictly constant ($O(1)$) across arbitrary horizons on CPU, NPU, `sGPU`, Metal, and standalone PTX.**
38
 
39
  `Origami-27B` preserves the full `Qwen/Qwen3.5-27B` architecture (48 Gated DeltaNet layers, 16 full-attention layers (64 KV groups), rescued native Multi-Token Prediction (`MTP`), and a 248,320-entry vocabulary) while turning a passive weight checkpoint into a self-contained binary file executed under one cross-silicon contract. It succeeds our first-generation constant-state releases published under [`huggingface.co/tfwnotops`](https://huggingface.co/tfwnotops) by removing `vLLM`, `MLX`, and `Transformers` runtime dependencies altogether. This repository is a **Preview** release: structural byte footprints, fold ratios, and constant-memory invariants below are fixed properties of the compiled artifact, while hardware-dependent throughput figures are omitted from the card and documented in the [Origami Family Paper](paper/Wirkung_2026_Origami_HCGM.pdf).
40
 
@@ -52,37 +52,37 @@ wkng run ./Origami-27B -p "Summarize how zero-retraining spectral folding preser
52
  | Checkpoint footprint on disk | 55.56 GB BF16 | **7.44 GB** (**7.46x** smaller; 410 `.o3f` operator files (including rescued `MTP`) + `exact_core.safetensors`) |
53
  | Required GPU VRAM | Multi-tens of GB HBM | **0 GB** (runs on plain CPU / NPU / `sGPU` or optional GPU) |
54
  | Dedicated weight-RAM | Pinned in host / device RAM | **0 GB** (read-only OS `mmap` directly from SSD/NVMe) |
55
- | Live session / horizon state | Linearly growing $O(T)$ KV cache (68.72 GB at 1M tokens; 1.18 GiB in Gen-1 `ConstantKV`) | **Constant 34.24 MB $O(1)$ session state at any length (35x smaller than `ConstantKV`)** |
56
  | Operator & readout path | Dense per-step matrix scan | Certified Cauchy-Schwarz `lm_head` gate + rescued native `MTP` multi-row verification (0 greedy drift) |
57
  | Runtime dependencies | PyTorch / `vLLM` / `MLX` / CUDA | **`wirkung` + `numpy`** (standalone C11, NPU, `sGPU`, Metal, PTX) |
58
 
59
  ![Physical byte anatomy and sequence-horizon state scaling of Origami-27B](assets/architecture.png)
60
 
61
- Figure 1. (a) Physical byte scale contrasting the 55.56 GB BF16 `Qwen/Qwen3.5-27B` checkpoint against the 7.44 GB `Origami-27B` `.o3f` bundle mapped directly from SSD (`0 GB` GPU VRAM, `0 GB` dedicated weight-RAM). (b) Live conversation state across 4k to 1M tokens: linear $O(T)$ KV growth against the flat 34.24 MB `Origami-27B` carrier.
62
 
63
  ## Zero-retraining compilation and constant-state architecture
64
 
65
  Every Origami model is produced by the same four-part compilation pipeline without gradient descent:
66
 
67
  1. **The file is the machine (disk-resident `mmap` execution).** An HCGM model is an immutable binary artifact on disk. The operating system maps operator pages read-only on demand from SSD or NVMe storage, requiring neither GPU VRAM nor weight-deserialization buffers in host RAM. Under our `v3-exakt` dyadic integer reduction contract, CPU (`AVX2` / `NEON`), general NPUs, embedded `sGPU` (`cpu_cuda`), Apple Metal, and driver-API PTX evaluate identical canonical reductions rather than drifting floating-point kernel approximations.
68
- 2. **Closed-form spectral folding (`.o3f`, 7.46x without retraining) plus rescued native `MTP`.** All 48 Gated DeltaNet blocks, 16 full-attention blocks, and the native Multi-Token Prediction (`MTP`) head are compiled directly from the frozen 55.56 GB BF16 checkpoint into 7.44 GB across 410 `.o3f` files (7.33 GB and 7.57x without `MTP`) via Rademacher-scrambled `FWHT_128` bitplanes and rank-$r$ projective crystals ($U_r V_r^\top$), fitting comfortably on 16 GB consumer nodes without swap.
69
- 3. **Constant 34.24 MB $O(1)$ state synthesized via `Wirkung`.** Using `Wirkung`, our internal exact structural-law and closed-form operator synthesis engine, all 48 Gated DeltaNet layers execute via associative WY monoid scans while all 64 full-attention KV groups across the 16 global layers are bounded by a pinned Token-0 sink, an exact 1,024-token frontier, and key-centered Ward centroids, cutting live state from 1.18 GiB in our first-generation `ConstantKV` build to 34.24 MB at every sequence horizon.
70
- 4. **Certified readout and online trajectory decomposition ($X = X_{\mathrm{known}} + R$).** Rather than re-evaluating every parameter blindly for isolated single tokens, HCGM decomposes live hidden trajectories into known low-rank structure $X_{\mathrm{known}}$ (cosine trajectory banks, multi-order suffix tables, and rank-1 Sherman-Morrison-Woodbury updates) plus novel residual $R$, verifying multiple candidate steps in one weight pass with zero greedy output drift.
71
 
72
  ![Constant-memory read geometry and folded operator execution in Origami-27B](assets/mechanism.png)
73
 
74
- Figure 2. (a) Constant-memory attention read over token coordinate $j$: pinned sink $s_0$, bounded Ward + WY Gated DeltaNet carrier, and exact local frontier ($W = 1024$). (b) Single-pass `.o3f` contraction pairing 4-class Walsh-Hadamard bitplanes with a rank-$r$ projective crystal ($U_r V_r^\top$ + rescued `MTP`) and Cauchy-Schwarz pruning.
75
 
76
  ## Shared trajectory assimilation (M-Rail), GPU-less server fleets and multi-model routing (HMoE)
77
 
78
  Beyond single-node execution, `Origami-27B` attaches natively to **Shared `M-Rail`** (`.mrail` v3), a content-addressed associative trajectory manifold shared across processes or across an enterprise network:
79
 
80
- - **Warm up once on one GPU, run permanently on disk-only CPU/NPU servers.** Because HCGM models require only SSD storage and megabytes of live state, an organization can run a single GPU (or a short temporary cloud GPU lease) until its recurring domain workflows are crystallized into a shared `.mrail` file, then shut the GPU down and serve frontier models across commodity CPU/NPU VMs and CPU-only servers. Known trajectories execute via `Track` copy morphisms or resume in $O(1)$ space from `Station` (`ANCHOR` + `DELTA`) snapshots; whenever any server encounters an unseen task ($R \neq 0$), computing and committing that residual once makes the solution immediately available to every other node on the rail.
81
  - **Heavy Mixture of Experts (`HMoE`).** Because idle disk-mapped models consume zero weight-RAM, `M-Rail` routes across complete compiled Origami models (`Origami-9B`, `Origami-27B`, `Origami-31B`, `Origami-78B`, `Origami-H3`), zero-recompile static-graph NPU specialist pools, and sub-2 KB exact algebraic group organs on a synchronous station bus (`cstar`) under deterministic check gates.
82
 
83
  ![Shared M-Rail GPU-less server fleet topology and multi-session memory scaling for Origami-27B](assets/mrail.png)
84
 
85
- Figure 3. (a) Shared `M-Rail` (`.mrail` v3) fleet architecture: optional temporary GPU warmup feeds a permanent zero-GPU CPU/NPU/`sGPU` server fleet where any novel residual $+R$ solved once commits back to all nodes alongside `HMoE` routing on `cstar`. (b) Total resident RAM/VRAM across $N = 1..32$ concurrent 32k-token sessions: pinned weights plus $N \cdot \mathrm{KV}$ against shared read-only `mmap` plus $N \cdot O(1)$ constant state.
86
 
87
  ## How to run
88
 
 
34
 
35
  [Paper](paper/Wirkung_2026_Origami_HCGM.pdf) · [DOI](https://doi.org/10.13140/RG.2.2.31826.72647) · [Runtime (PyPI)](https://pypi.org/project/wirkung/) · [Wirkung](https://wirku.ng) · [Base weights](https://huggingface.co/Qwen/Qwen3.5-27B) · [Contact](mailto:contact@wirku.ng) · [License](LICENSE)
36
 
37
+ **`Origami-27B` is a 27.80B-parameter Holographic Causal Graph Machine (HCGM) Preview compiled in closed form from `Qwen/Qwen3.5-27B` with zero retraining, zero fine-tuning, and zero gradient steps. It folds the 55.56 GB source checkpoint into 7.44 GB (7.46x compression), executes directly from NVMe/SSD storage via OS `mmap` with `0 GB` GPU VRAM and `0 GB` dedicated weight-RAM, and keeps live state strictly constant (O(1)) across arbitrary horizons on CPU, NPU, `sGPU`, Metal, and standalone PTX.**
38
 
39
  `Origami-27B` preserves the full `Qwen/Qwen3.5-27B` architecture (48 Gated DeltaNet layers, 16 full-attention layers (64 KV groups), rescued native Multi-Token Prediction (`MTP`), and a 248,320-entry vocabulary) while turning a passive weight checkpoint into a self-contained binary file executed under one cross-silicon contract. It succeeds our first-generation constant-state releases published under [`huggingface.co/tfwnotops`](https://huggingface.co/tfwnotops) by removing `vLLM`, `MLX`, and `Transformers` runtime dependencies altogether. This repository is a **Preview** release: structural byte footprints, fold ratios, and constant-memory invariants below are fixed properties of the compiled artifact, while hardware-dependent throughput figures are omitted from the card and documented in the [Origami Family Paper](paper/Wirkung_2026_Origami_HCGM.pdf).
40
 
 
52
  | Checkpoint footprint on disk | 55.56 GB BF16 | **7.44 GB** (**7.46x** smaller; 410 `.o3f` operator files (including rescued `MTP`) + `exact_core.safetensors`) |
53
  | Required GPU VRAM | Multi-tens of GB HBM | **0 GB** (runs on plain CPU / NPU / `sGPU` or optional GPU) |
54
  | Dedicated weight-RAM | Pinned in host / device RAM | **0 GB** (read-only OS `mmap` directly from SSD/NVMe) |
55
+ | Live session / horizon state | Linearly growing O(<i>T</i>) KV cache (68.72 GB at 1M tokens; 1.18 GiB in Gen-1 `ConstantKV`) | **Constant 34.24 MB O(1) session state at any length (35x smaller than `ConstantKV`)** |
56
  | Operator & readout path | Dense per-step matrix scan | Certified Cauchy-Schwarz `lm_head` gate + rescued native `MTP` multi-row verification (0 greedy drift) |
57
  | Runtime dependencies | PyTorch / `vLLM` / `MLX` / CUDA | **`wirkung` + `numpy`** (standalone C11, NPU, `sGPU`, Metal, PTX) |
58
 
59
  ![Physical byte anatomy and sequence-horizon state scaling of Origami-27B](assets/architecture.png)
60
 
61
+ Figure 1. (a) Physical byte scale contrasting the 55.56 GB BF16 `Qwen/Qwen3.5-27B` checkpoint against the 7.44 GB `Origami-27B` `.o3f` bundle mapped directly from SSD (`0 GB` GPU VRAM, `0 GB` dedicated weight-RAM). (b) Live conversation state across 4k to 1M tokens: linear O(<i>T</i>) KV growth against the flat 34.24 MB `Origami-27B` carrier.
62
 
63
  ## Zero-retraining compilation and constant-state architecture
64
 
65
  Every Origami model is produced by the same four-part compilation pipeline without gradient descent:
66
 
67
  1. **The file is the machine (disk-resident `mmap` execution).** An HCGM model is an immutable binary artifact on disk. The operating system maps operator pages read-only on demand from SSD or NVMe storage, requiring neither GPU VRAM nor weight-deserialization buffers in host RAM. Under our `v3-exakt` dyadic integer reduction contract, CPU (`AVX2` / `NEON`), general NPUs, embedded `sGPU` (`cpu_cuda`), Apple Metal, and driver-API PTX evaluate identical canonical reductions rather than drifting floating-point kernel approximations.
68
+ 2. **Closed-form spectral folding (`.o3f`, 7.46x without retraining) plus rescued native `MTP`.** All 48 Gated DeltaNet blocks, 16 full-attention blocks, and the native Multi-Token Prediction (`MTP`) head are compiled directly from the frozen 55.56 GB BF16 checkpoint into 7.44 GB across 410 `.o3f` files (7.33 GB and 7.57x without `MTP`) via Rademacher-scrambled `FWHT_128` bitplanes and rank-<i>r</i> projective crystals (<i>U<sub>r</sub>V<sub>r</sub></i><sup>⊤</sup>), fitting comfortably on 16 GB consumer nodes without swap.
69
+ 3. **Constant 34.24 MB O(1) state synthesized via `Wirkung`.** Using `Wirkung`, our internal exact structural-law and closed-form operator synthesis engine, all 48 Gated DeltaNet layers execute via associative WY monoid scans while all 64 full-attention KV groups across the 16 global layers are bounded by a pinned Token-0 sink, an exact 1,024-token frontier, and key-centered Ward centroids, cutting live state from 1.18 GiB in our first-generation `ConstantKV` build to 34.24 MB at every sequence horizon.
70
+ 4. **Certified readout and online trajectory decomposition (<i>X</i> = <i>X</i><sub>known</sub> + <i>R</i>).** Rather than re-evaluating every parameter blindly for isolated single tokens, HCGM decomposes live hidden trajectories into known low-rank structure <i>X</i><sub>known</sub> (cosine trajectory banks, multi-order suffix tables, and rank-1 Sherman-Morrison-Woodbury updates) plus novel residual <i>R</i>, verifying multiple candidate steps in one weight pass with zero greedy output drift.
71
 
72
  ![Constant-memory read geometry and folded operator execution in Origami-27B](assets/mechanism.png)
73
 
74
+ Figure 2. (a) Constant-memory attention read over token coordinate <i>j</i>: pinned sink <i>s</i><sub>0</sub>, bounded Ward + WY Gated DeltaNet carrier, and exact local frontier (<i>W</i> = 1024). (b) Single-pass `.o3f` contraction pairing 4-class Walsh-Hadamard bitplanes with a rank-<i>r</i> projective crystal (<i>U<sub>r</sub>V<sub>r</sub></i><sup>⊤</sup> + rescued `MTP`) and Cauchy-Schwarz pruning.
75
 
76
  ## Shared trajectory assimilation (M-Rail), GPU-less server fleets and multi-model routing (HMoE)
77
 
78
  Beyond single-node execution, `Origami-27B` attaches natively to **Shared `M-Rail`** (`.mrail` v3), a content-addressed associative trajectory manifold shared across processes or across an enterprise network:
79
 
80
+ - **Warm up once on one GPU, run permanently on disk-only CPU/NPU servers.** Because HCGM models require only SSD storage and megabytes of live state, an organization can run a single GPU (or a short temporary cloud GPU lease) until its recurring domain workflows are crystallized into a shared `.mrail` file, then shut the GPU down and serve frontier models across commodity CPU/NPU VMs and CPU-only servers. Known trajectories execute via `Track` copy morphisms or resume in O(1) space from `Station` (`ANCHOR` + `DELTA`) snapshots; whenever any server encounters an unseen task (<i>R</i> ≠ 0), computing and committing that residual once makes the solution immediately available to every other node on the rail.
81
  - **Heavy Mixture of Experts (`HMoE`).** Because idle disk-mapped models consume zero weight-RAM, `M-Rail` routes across complete compiled Origami models (`Origami-9B`, `Origami-27B`, `Origami-31B`, `Origami-78B`, `Origami-H3`), zero-recompile static-graph NPU specialist pools, and sub-2 KB exact algebraic group organs on a synchronous station bus (`cstar`) under deterministic check gates.
82
 
83
  ![Shared M-Rail GPU-less server fleet topology and multi-session memory scaling for Origami-27B](assets/mrail.png)
84
 
85
+ Figure 3. (a) Shared `M-Rail` (`.mrail` v3) fleet architecture: optional temporary GPU warmup feeds a permanent zero-GPU CPU/NPU/`sGPU` server fleet where any novel residual +<i>R</i> solved once commits back to all nodes alongside `HMoE` routing on `cstar`. (b) Total resident RAM/VRAM across <i>N</i> = 1..32 concurrent 32k-token sessions: pinned weights plus <i>N</i> · KV against shared read-only `mmap` plus <i>N</i> · O(1) constant state.
86
 
87
  ## How to run
88