Instructions to use infosave/cmf with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- cortiq
How to use infosave/cmf with cortiq:
# one Rust binary, no additional dependencies cargo install cortiq-cli # or a prebuilt binary from github.com/infosave2007/cmf/releases hf download infosave/cmf --include "*.cmf" --local-dir . ls *.cmf # some repos ship more than one quantization
cortiq run FILE.cmf --prompt "What is the capital of France?"
cortiq serve FILE.cmf --port 8080 # OpenAI-compatible server
- Notebooks
- Google Colab
- Kaggle
links that only resolve on GitHub now point at GitHub
Browse files- SPEC.md +3 -3
- SPEC.ru.md +3 -3
- SPEC.zh.md +3 -3
SPEC.md
CHANGED
|
@@ -615,7 +615,7 @@ file. Errors: opening a non-first shard directly, a missing sibling, a
|
|
| 615 |
Gate (Qwen3.5-0.8B q8_2f, 5 shards ≤ 0.6 GB): sharded PPL == unsharded
|
| 616 |
byte-exactly on the same binary; `verify` green on every shard alone.
|
| 617 |
|
| 618 |
-
## 11. Defragmentation — physical pruning (USPTO App. 19/452,464, claims 9/10 — [PATENTS.md](.
|
| 619 |
|
| 620 |
A mask (§5) is **virtual sparsity**: pruned neurons are flagged but still
|
| 621 |
stored in full (all tasks share one backbone — you cannot physically cut
|
|
@@ -766,7 +766,7 @@ discrete cards and phones).
|
|
| 766 |
|
| 767 |
---
|
| 768 |
|
| 769 |
-
*Related: [COMPARISON.md](COMPARISON.md) (CMF vs. other model formats),
|
| 770 |
-
[project README](.
|
| 771 |
`python/cmf_reader.py` (standalone reader: stdlib + numpy, reads every
|
| 772 |
dtype, shards, skills, verify).*
|
|
|
|
| 615 |
Gate (Qwen3.5-0.8B q8_2f, 5 shards ≤ 0.6 GB): sharded PPL == unsharded
|
| 616 |
byte-exactly on the same binary; `verify` green on every shard alone.
|
| 617 |
|
| 618 |
+
## 11. Defragmentation — physical pruning (USPTO App. 19/452,464, claims 9/10 — [PATENTS.md](https://github.com/infosave2007/cmf/blob/master/PATENTS.md))
|
| 619 |
|
| 620 |
A mask (§5) is **virtual sparsity**: pruned neurons are flagged but still
|
| 621 |
stored in full (all tasks share one backbone — you cannot physically cut
|
|
|
|
| 766 |
|
| 767 |
---
|
| 768 |
|
| 769 |
+
*Related: [COMPARISON.md](https://github.com/infosave2007/cmf/blob/master/docs/COMPARISON.md) (CMF vs. other model formats),
|
| 770 |
+
[project README](https://github.com/infosave2007/cmf/blob/master/README.md) (overview and quick start),
|
| 771 |
`python/cmf_reader.py` (standalone reader: stdlib + numpy, reads every
|
| 772 |
dtype, shards, skills, verify).*
|
SPEC.ru.md
CHANGED
|
@@ -575,7 +575,7 @@ safetensors). Пользователь открывает ЛЮБОЕ имя; р
|
|
| 575 |
нешардированной байт-в-байт на том же бинарнике; `verify` зелёный на каждом шарде
|
| 576 |
по отдельности.
|
| 577 |
|
| 578 |
-
## 11. Дефрагментация — физический прунинг (USPTO App. 19/452,464, claims 9/10 — [PATENTS.md](.
|
| 579 |
|
| 580 |
Маска (§5) — это **виртуальная разреженность**: вырезанные нейроны
|
| 581 |
помечены, но хранятся целиком (все задачи делят один бэкбон — резать
|
|
@@ -678,7 +678,7 @@ text-to-image стек.
|
|
| 678 |
|
| 679 |
---
|
| 680 |
|
| 681 |
-
*Связанное: [COMPARISON.md](COMPARISON.md) (CMF против других форматов моделей),
|
| 682 |
-
[README проекта](.
|
| 683 |
`python/cmf_reader.py` (автономный ридер: stdlib + numpy, читает каждый
|
| 684 |
dtype, шарды, скиллы, verify).*
|
|
|
|
| 575 |
нешардированной байт-в-байт на том же бинарнике; `verify` зелёный на каждом шарде
|
| 576 |
по отдельности.
|
| 577 |
|
| 578 |
+
## 11. Дефрагментация — физический прунинг (USPTO App. 19/452,464, claims 9/10 — [PATENTS.md](https://github.com/infosave2007/cmf/blob/master/PATENTS.md))
|
| 579 |
|
| 580 |
Маска (§5) — это **виртуальная разреженность**: вырезанные нейроны
|
| 581 |
помечены, но хранятся целиком (все задачи делят один бэкбон — резать
|
|
|
|
| 678 |
|
| 679 |
---
|
| 680 |
|
| 681 |
+
*Связанное: [COMPARISON.md](https://github.com/infosave2007/cmf/blob/master/docs/COMPARISON.md) (CMF против других форматов моделей),
|
| 682 |
+
[README проекта](https://github.com/infosave2007/cmf/blob/master/README.md) (обзор и быстрый старт),
|
| 683 |
`python/cmf_reader.py` (автономный ридер: stdlib + numpy, читает каждый
|
| 684 |
dtype, шарды, скиллы, verify).*
|
SPEC.zh.md
CHANGED
|
@@ -527,7 +527,7 @@ safetensors 兼容)。用户可打开**任意**名称;运行时归一化到
|
|
| 527 |
Gate(Qwen3.5-0.8B q8_2f,5 个分片 ≤ 0.6 GB):分片化 PPL == 未分片化,
|
| 528 |
在同一二进制上逐字节相等;`verify` 在每个单独分片上皆为绿。
|
| 529 |
|
| 530 |
-
## 11. 碎片整理——物理剪枝(USPTO App. 19/452,464,claims 9/10 —— [PATENTS.md](.
|
| 531 |
|
| 532 |
掩码(§5)是**虚拟稀疏**:被剪掉的神经元只是被标记,仍完整存储(所有
|
| 533 |
任务共享一个骨干——在承诺单一任务之前无法物理切除)。碎片整理把虚拟
|
|
@@ -605,7 +605,7 @@ GB(−35%),保留集代码困惑度 +2.8%;在 24 GB 内存机器上—
|
|
| 605 |
|
| 606 |
---
|
| 607 |
|
| 608 |
-
*相关:[COMPARISON.md](COMPARISON.md)(CMF 与其他模型格式的对比)、
|
| 609 |
-
[项目 README](.
|
| 610 |
`python/cmf_reader.py`(独立读取器:标准库 + numpy,读取所有
|
| 611 |
dtype、分片、技能、verify)。*
|
|
|
|
| 527 |
Gate(Qwen3.5-0.8B q8_2f,5 个分片 ≤ 0.6 GB):分片化 PPL == 未分片化,
|
| 528 |
在同一二进制上逐字节相等;`verify` 在每个单独分片上皆为绿。
|
| 529 |
|
| 530 |
+
## 11. 碎片整理——物理剪枝(USPTO App. 19/452,464,claims 9/10 —— [PATENTS.md](https://github.com/infosave2007/cmf/blob/master/PATENTS.md))
|
| 531 |
|
| 532 |
掩码(§5)是**虚拟稀疏**:被剪掉的神经元只是被标记,仍完整存储(所有
|
| 533 |
任务共享一个骨干——在承诺单一任务之前无法物理切除)。碎片整理把虚拟
|
|
|
|
| 605 |
|
| 606 |
---
|
| 607 |
|
| 608 |
+
*相关:[COMPARISON.md](https://github.com/infosave2007/cmf/blob/master/docs/COMPARISON.md)(CMF 与其他模型格式的对比)、
|
| 609 |
+
[项目 README](https://github.com/infosave2007/cmf/blob/master/README.md)(概览与快速上手)、
|
| 610 |
`python/cmf_reader.py`(独立读取器:标准库 + numpy,读取所有
|
| 611 |
dtype、分片、技能、verify)。*
|