card: why the chip ships beside Qwen
Browse files
README.md
CHANGED
|
@@ -1,68 +1,70 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: apache-2.0
|
| 3 |
-
base_model: Qwen/Qwen3.5-4B
|
| 4 |
-
base_model_relation: adapter
|
| 5 |
-
tags:
|
| 6 |
-
- qwenputer
|
| 7 |
-
- risc-v
|
| 8 |
-
- compiled-weights
|
| 9 |
-
- coding-agent
|
| 10 |
-
- fx
|
| 11 |
-
---
|
| 12 |
-
|
| 13 |
-
<p align="center">
|
| 14 |
-
<img src="qwenfx-banner.svg" alt="QwenFX: a coding agent, running inside the weights of the model it drives." width="100%">
|
| 15 |
-
</p>
|
| 16 |
-
|
| 17 |
-
Code: [GitHub](https://github.com/divagr18/QwenFX)
|
| 18 |
-
Part of [Qwenputer](https://github.com/divagr18/qwenputer), the computer compiled into Qwen3.5's weights.
|
| 19 |
-
|
| 20 |
-
**QwenFX runs [fx](https://github.com/vercel-labs/fx), Vercel's coding agent, as its exact compiled
|
| 21 |
-
binary on a RISC-V CPU compiled into weights that run beside Qwen3.5-4B, with that same Qwen as fx's
|
| 22 |
-
model.** fx solved 15 of 15 small coding tasks, each checked by a hidden test, and opened five pull
|
| 23 |
-
requests on GitHub by itself
|
| 24 |
-
([#6](https://github.com/divagr18/fxputer-sandbox/pull/6)-[#10](https://github.com/divagr18/fxputer-sandbox/pull/10)).
|
| 25 |
-
Every chip token in every run was checked against a reference machine. None was wrong.
|
| 26 |
-
|
| 27 |
-
This repository holds the chip: the CPU's weights, with fx's boot state in them. Qwen3.5-4B is not
|
| 28 |
-
included; use [Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B) unchanged.
|
| 29 |
-
|
| 30 |
-
## Files
|
| 31 |
-
|
| 32 |
-
| file | what | bytes | sha256 |
|
| 33 |
-
|---|---|---|---|
|
| 34 |
-
| `fx-live-u13-f4096.safetensors` | the chip: 162 layers, 4,096 RAM frames, the 13 hot-code units, fx's boot state | 124,960,497 | `48ddb06c鈥 |
|
| 35 |
-
| `fx-live-u13-f4096-text.safetensors` | the same, plus the text bridge that carries Qwen's text through the circuit (`--text`) | 129,060,613 | `fda824ef鈥 |
|
| 36 |
-
| `fx-fastcopy.elf` | fx's RV32IM+Zba ELF, built from fx's unmodified source at `757bdb9` | 75,606,908 | `282e53d4鈥 |
|
| 37 |
-
| `imports.json` | the ELF's import table | | |
|
| 38 |
-
|
| 39 |
-
The chip file is a sparse safetensors file: 19.07M nonzero bf16 weights, plus the machine's state
|
| 40 |
-
after fx's boot. Executor v4 lifts the circuit from these weights at load. With base Qwen beside it,
|
| 41 |
-
a live task peaks at 10-14 GB of GPU memory on an L4.
|
| 42 |
-
|
| 43 |
-
## Use
|
| 44 |
-
|
| 45 |
-
```bash
|
| 46 |
-
git clone https://github.com/divagr18/QwenFX && cd QwenFX && uv sync --extra hf
|
| 47 |
-
uv run python qwenfx.py setup
|
| 48 |
-
uv run hf download divagr1925/QwenFX fx-live-u13-f4096.safetensors --local-dir chip
|
| 49 |
-
uv run python qwenfx.py task t1_bug
|
| 50 |
-
```
|
| 51 |
-
|
| 52 |
-
`qwenfx.py chip` builds the same file from the Python circuit instead (about 3 minutes, 6.4 GB of
|
| 53 |
-
RAM); the arrays are identical. The GitHub README has the issue-to-pull-request flow, what runs on
|
| 54 |
-
the chip and what runs on the host, and the checks.
|
| 55 |
-
|
| 56 |
-
## Not a merged checkpoint
|
| 57 |
-
|
| 58 |
-
The chip and Qwen are two sets of weights in one process on one GPU. Qwen's own weights are
|
| 59 |
-
unchanged.
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
base_model: Qwen/Qwen3.5-4B
|
| 4 |
+
base_model_relation: adapter
|
| 5 |
+
tags:
|
| 6 |
+
- qwenputer
|
| 7 |
+
- risc-v
|
| 8 |
+
- compiled-weights
|
| 9 |
+
- coding-agent
|
| 10 |
+
- fx
|
| 11 |
+
---
|
| 12 |
+
|
| 13 |
+
<p align="center">
|
| 14 |
+
<img src="qwenfx-banner.svg" alt="QwenFX: a coding agent, running inside the weights of the model it drives." width="100%">
|
| 15 |
+
</p>
|
| 16 |
+
|
| 17 |
+
Code: [GitHub](https://github.com/divagr18/QwenFX) 路 Mirror: [ModelScope](https://www.modelscope.ai/models/Spartan1925/QwenFX)
|
| 18 |
+
Part of [Qwenputer](https://github.com/divagr18/qwenputer), the computer compiled into Qwen3.5's weights.
|
| 19 |
+
|
| 20 |
+
**QwenFX runs [fx](https://github.com/vercel-labs/fx), Vercel's coding agent, as its exact compiled
|
| 21 |
+
binary on a RISC-V CPU compiled into weights that run beside Qwen3.5-4B, with that same Qwen as fx's
|
| 22 |
+
model.** fx solved 15 of 15 small coding tasks, each checked by a hidden test, and opened five pull
|
| 23 |
+
requests on GitHub by itself
|
| 24 |
+
([#6](https://github.com/divagr18/fxputer-sandbox/pull/6)-[#10](https://github.com/divagr18/fxputer-sandbox/pull/10)).
|
| 25 |
+
Every chip token in every run was checked against a reference machine. None was wrong.
|
| 26 |
+
|
| 27 |
+
This repository holds the chip: the CPU's weights, with fx's boot state in them. Qwen3.5-4B is not
|
| 28 |
+
included; use [Qwen/Qwen3.5-4B](https://huggingface.co/Qwen/Qwen3.5-4B) unchanged.
|
| 29 |
+
|
| 30 |
+
## Files
|
| 31 |
+
|
| 32 |
+
| file | what | bytes | sha256 |
|
| 33 |
+
|---|---|---|---|
|
| 34 |
+
| `fx-live-u13-f4096.safetensors` | the chip: 162 layers, 4,096 RAM frames, the 13 hot-code units, fx's boot state | 124,960,497 | `48ddb06c鈥 |
|
| 35 |
+
| `fx-live-u13-f4096-text.safetensors` | the same, plus the text bridge that carries Qwen's text through the circuit (`--text`) | 129,060,613 | `fda824ef鈥 |
|
| 36 |
+
| `fx-fastcopy.elf` | fx's RV32IM+Zba ELF, built from fx's unmodified source at `757bdb9` | 75,606,908 | `282e53d4鈥 |
|
| 37 |
+
| `imports.json` | the ELF's import table | | |
|
| 38 |
+
|
| 39 |
+
The chip file is a sparse safetensors file: 19.07M nonzero bf16 weights, plus the machine's state
|
| 40 |
+
after fx's boot. Executor v4 lifts the circuit from these weights at load. With base Qwen beside it,
|
| 41 |
+
a live task peaks at 10-14 GB of GPU memory on an L4.
|
| 42 |
+
|
| 43 |
+
## Use
|
| 44 |
+
|
| 45 |
+
```bash
|
| 46 |
+
git clone https://github.com/divagr18/QwenFX && cd QwenFX && uv sync --extra hf
|
| 47 |
+
uv run python qwenfx.py setup
|
| 48 |
+
uv run hf download divagr1925/QwenFX fx-live-u13-f4096.safetensors --local-dir chip
|
| 49 |
+
uv run python qwenfx.py task t1_bug
|
| 50 |
+
```
|
| 51 |
+
|
| 52 |
+
`qwenfx.py chip` builds the same file from the Python circuit instead (about 3 minutes, 6.4 GB of
|
| 53 |
+
RAM); the arrays are identical. The GitHub README has the issue-to-pull-request flow, what runs on
|
| 54 |
+
the chip and what runs on the host, and the checks.
|
| 55 |
+
|
| 56 |
+
## Not a merged checkpoint
|
| 57 |
+
|
| 58 |
+
The chip and Qwen are two sets of weights in one process on one GPU. Qwen's own weights are
|
| 59 |
+
unchanged. A merged checkpoint is very possible, since the chip uses Qwen3.5's own layer types and
|
| 60 |
+
Qwenputer already ships its CPU merged, but at fx's size it would be about 25-37 GB (estimated),
|
| 61 |
+
mostly zero padding, against 8.5 GB for base Qwen plus this 125 MB file. So I chose this layout. The host still renders the chat template, tokenizes, parses tool calls, keeps the file
|
| 62 |
+
store fx reads and writes, and does the git and GitHub work of a pull request. fx's own code, all of
|
| 63 |
+
it, runs on the chip.
|
| 64 |
+
|
| 65 |
+
## License
|
| 66 |
+
|
| 67 |
+
Apache-2.0. fx is by Vercel, Apache-2.0: `LICENSE` is fx's license, and `THIRD_PARTY_NOTICES.md` is
|
| 68 |
+
fx's, unchanged. The ELF and the chip files' boot state are fx-derived; `NOTICE` in the GitHub
|
| 69 |
+
repository's `third_party/fx/` (and here) says how the ELF was built and what else it contains. Qwen3.5-4B is
|
| 70 |
+
Apache-2.0 and is not redistributed here.
|