README.md

---
library_name: easydel
pipeline_tag: text-generation
tags:
  - easydel
  - jax
  - "kimi_vl"
  - "CausalLM"
  - "ragged_page_attention_v3"
---

<p align="center">
  <img alt="easydel" src="https://raw.githubusercontent.com/erfanzar/easydel/main/images/easydel-logo-with-text.png">
</p>

<h1 align="center">Kimi-VL-A3B-Instruct</h1>

<div align="center">
  A model compatible with the EasyDeL JAX stack.
</div>

## Overview

This checkpoint is intended to be loaded with EasyDeL on JAX (CPU/GPU/TPU). It supports sharded loading with `auto_shard_model=True` and configurable precision via `dtype`, `param_dtype`, and `precision`.

## Quickstart

```python
import easydel as ed
from jax import numpy as jnp, lax

repo_id = "EasyDeL/Kimi-VL-A3B-Instruct"

dtype = jnp.bfloat16  # try jnp.float16 on many GPUs

model = ed.AutoEasyDeLModelForCausalLM.from_pretrained(
    repo_id,
    dtype=dtype,
    param_dtype=dtype,
    precision=lax.Precision("fastest"),
    sharding_axis_names=("dp", "fsdp", "ep", "tp", "sp"),
    sharding_axis_dims=(1, -1, 1, 1, 1),
    config_kwargs=ed.EasyDeLBaseConfigDict(
        attn_dtype=dtype,
        attn_mechanism=ed.AttentionMechanisms.RAGGED_PAGE_ATTENTION_V3,
        fsdp_is_ep_bound=True,
        sp_is_ep_bound=True,
        moe_method=ed.MoEMethods.FUSED_MOE,
    ),
    auto_shard_model=True,
    partition_axis=ed.PartitionAxis(),
)
```

If the repository only provides PyTorch weights, pass `from_torch=True` to `from_pretrained(...)`.

## Sharding & Parallelism (Multi-Device)

EasyDeL can scale to multiple devices by creating a logical device mesh. Most EasyDeL loaders use a 5D mesh:

- `dp`: data parallel (replicated parameters, different batch shards)
- `fsdp`: parameter sharding (memory saver; often the biggest axis)
- `ep`: expert parallel (MoE; keep `1` for non-MoE models)
- `tp`: tensor parallel (splits large matmuls)
- `sp`: sequence parallel (splits sequence dimension)

Use `sharding_axis_names=("dp","fsdp","ep","tp","sp")` and choose `sharding_axis_dims` so that their product equals your device count.
You can use `-1` in `sharding_axis_dims` to let EasyDeL infer the remaining dimension.

<details>
<summary>Example sharding configs</summary>

```python
# 8 devices, pure FSDP
sharding_axis_dims = (1, 8, 1, 1, 1)

# 8 devices, 2-way DP x 4-way FSDP
sharding_axis_dims = (2, 4, 1, 1, 1)

# 8 devices, 4-way FSDP x 2-way TP
sharding_axis_dims = (1, 4, 1, 2, 1)
```
</details>

## Using via `eLargeModel` (ELM)

`eLargeModel` is a higher-level interface that wires together loading, sharding, training, and eSurge inference from a single config.

```python
from easydel import eLargeModel

repo_id = "EasyDeL/Kimi-VL-A3B-Instruct"

elm = eLargeModel.from_pretrained(repo_id)  # task is auto-detected
elm.set_dtype("bf16")
elm.set_sharding(axis_names=("dp", "fsdp", "ep", "tp", "sp"), axis_dims=(1, -1, 1, 1, 1))

model = elm.build_model()
# Optional: build an inference engine
# engine = elm.build_esurge()
```

<details>
<summary>ELM YAML config example</summary>

```yaml
model:
  name_or_path: "EasyDeL/Kimi-VL-A3B-Instruct"

loader:
  dtype: bf16
  param_dtype: bf16

sharding:
  axis_dims: [1, -1, 1, 1, 1]
  auto_shard_model: true
```
</details>

## Features

**EasyDeL:**
- JAX native implementation and sharded execution
- Configurable attention backends via `AttentionMechanisms.*`
- Precision control via `dtype`, `param_dtype`, and `precision`

## Installation

```bash
pip install easydel
```

## Links

- EasyDeL GitHub: https://github.com/erfanzar/EasyDeL
- Docs: https://easydel.readthedocs.io/en/latest/

## Supported Tasks

- CausalLM

## Limitations

- Refer to the original model card for training data, evaluation, and intended use.

## License

EasyDeL is released under the Apache-2.0 license. The license for this model's weights may differ; please consult the original repository.

## Citation

```bibtex
@misc{Zare Chavoshi_2023,
    title={EasyDeL: An open-source library for enhancing and streamlining the training process of machine learning models},
    url={https://github.com/erfanzar/EasyDeL},
    author={Zare Chavoshi, Erfan},
    year={2023}
}
```