Instructions to use staccs/qwencoin with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use staccs/qwencoin with Transformers:
# pip install -U transformers accelerate # Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("staccs/qwencoin", device_map="auto") - Notebooks
- Google Colab
- Kaggle
Download exec/qwork.h from staccs/qwencoin: direct link, hf CLI and curl.
- Browser
- Download file 1.94 kB
-
https://huggingface.co/staccs/qwencoin/resolve/main/exec/qwork.h
- Command line
-
hf download hf://staccs/qwencoin/exec/qwork.h
-
curl -L -o qwork.h https://huggingface.co/staccs/qwencoin/resolve/main/exec/qwork.h
1.94 kB
| /* The work function of the chain: a pass through the MLP matrices of Qwen3.5-0.8B. | |
| * | |
| * Table (qwork.bin, built by `python -m hashformer.work build`): for each of the model's 24 | |
| * layers its gate_proj and up_proj (3584 x 1024) and down_proj (1024 x 3584), every weight w | |
| * as the 16-bit integer rint(4096 w). 528 MB. | |
| * | |
| * seed h0 = SHA256d(block header with nNonce = 0) (the caller's SHA-256) | |
| * x 1024 values of +-1 expanded from (h0, nonce) (qwork_expand) | |
| * layer a = sgn(G x) * sgn(U x) 3584 values; x = sgn(D a) sgn(v) = +1 if v > 0 else -1 | |
| * out the 1024 signs after the last layer, packed, folded to 32 bytes by XOR | |
| * | |
| * The proof-of-work value of a header is SHA256(out XOR h0). All arithmetic is on integers: | |
| * every implementation agrees bit for bit, in any order of summation. | |
| */ | |
| extern "C" { | |
| typedef struct { | |
| int layers, n, m; /* 24, 1024, 3584 */ | |
| const int16_t *data; /* per layer: G (m x n), U (m x n), D (n x m), row-major */ | |
| size_t bytes; /* size of the mapped file */ | |
| uint64_t fnv; /* checksum of the whole file (qwork_fnv64) */ | |
| void *map; | |
| } qwork_table; | |
| /* Maps the table read-only and checks its header. Returns 0, or a negative error. */ | |
| int qwork_open(const char *path, qwork_table *t); | |
| void qwork_close(qwork_table *t); | |
| /* FNV-1a over the file's little-endian 64-bit words. */ | |
| uint64_t qwork_fnv64(const unsigned char *p, size_t bytes); | |
| /* The start state: n values of +1 / -1 from the seed and the nonce. */ | |
| void qwork_expand(const unsigned char h0[32], uint32_t nonce, int16_t *x, int n); | |
| /* The whole pass. out: 32 bytes. Uses up to `layers` layers (0: all). */ | |
| void qwork_eval(const qwork_table *t, const unsigned char h0[32], uint32_t nonce, unsigned char out[32]); | |
| } | |