---
license: apache-2.0
metrics:
- perplexity
pipeline_tag: text-generation
---

Train in 30B Byte. Mode size 353M. Table 2 in [MambaByte](https://arxiv.org/abs/2401.13660)

To use

```
import torch
import numpy as np

from mamba_ssm.models.mixer_seq_simple import MambaLMHeadModel

model=MambaLMHeadModel.from_pretrained("JunxiongWang/MambaByte_Arxiv", device='cuda', dtype=torch.bfloat16)

text = "\documentclass[12pt]{article}"
text_byte = np.frombuffer(text.encode('utf-8'), dtype=np.uint8)
input_ids = torch.from_numpy(text_byte[None, :].copy()).long().cuda()

sample = model.generate(
    input_ids=input_ids,
    max_length=2048,
    cg=True,
    return_dict_in_generate=True,
    output_scores=True,
    enable_timing=True,
    temperature=1,
    top_k=256,
    top_p=0.9,
)

print(bytes(sample.sequences[0].tolist()).decode('utf-8'))
```

Output:

```
\documentclass[12pt]{article}}}}^{{\mathbf{P}}\uplus{\mathbf{Q}}}}}}}{}}$ is a symmetric poset. This implies that $$\operatorname{end}({\mathscr{L}}) = \operatorname{end}({\mathscr{L}}\setminus\{\sigma_{{\mathbf{P}}}\}) = \operatorname{end}({\mathscr{L}}\setminus\{\sigma_{{\mathbf{Q}}}\}) = \operatorname{end}({\mathscr{L}}\setminus\{\sigma_{{\mathbf{P}}},\sigma_{{\mathbf{Q}}}\}),$$ i.e., ${\mathscr{L}}$ is $\{\sigma_{{\mathbf{P}}},\sigma_{{\mathbf{Q}}}\}$-bistochastic for any ${\mathbf{P}}\neq{\mathbf{Q}}$. Thus, ${\mathscr{L}}$ is reversible, and is in fact maximal among all $\{\sigma_{{\mathbf{P}}},\sigma_{{\mathbf{Q}}}\}$-bistochastic matrices.

Since ${\mathscr{L}}$ is in the same class as $\sigma_{{\mathbf{P}}}^{{\mathbf{Q}}}$, we have $\operatorname{end}({\mathscr{L}})\subseteq\operatorname{end}({\mathscr{L}})$. Conversely, if $\operatorname{end}({\mathscr{L}})=\operatorname{end}({\mathscr{L}})$, then $\sigma_{{\mathbf{P}}}^{{\mathbf{Q}}}$ is maximal in $\operatorname{end}({\mathscr{L}})$. Since ${\mathbf{P}}\setminus\{\sigma_{{\mathbf{P}}}\}\subseteq\operatorname{end}({\mathscr{L}})$, this implies that ${\mathscr{L}}$ is in the same class as $\sigma_{{\mathbf{P}}}^{{\mathbf{Q}}}$, and hence ${\mathscr{L}}$ is reversible.

We are now ready to show that $\{\sigma_{{\mathbf{P}}},\sigma_{{\mathbf{Q}}}\}$-bistochastic matrices form a symmetric poset of ends.

\[lem:end\_symm\_class\] Let ${\mathbf{P}},{\mathbf{Q}}\in{\mathscr{M}}$. Then $\sigma_{{\mathbf{P}}}^{{\mathbf{Q}}}$ is symmetric if and only if $\operatorname{end}({\mathscr{L}})=\operatorname{end}({\mathscr{L}})$.

Suppose that $\operatorname{end}({\mathscr{L}})=\operatorname{end}({\mathscr{L}})$, and we prove that $\sigma_{{\mathbf{P}}}^{{\mathbf{Q}}}$ is symmetric. Clearly, $\operatorname{end}({\mathscr{L}})$ contains exactly the ends of $\operatorname{end}({\mathscr{L}})$ by definition, and the only case that survives is when $\operatorname{end}({\mathscr{L}})=\operatorname{end}({\mathscr{L}})$. By construction, this means that $\sigma_{{\mathbf{P}}}
```