File size: 1,646 Bytes
083a5c8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
0cc7462
083a5c8
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
---
library_name: transformers
license: apache-2.0
license_link: https://huggingface.co/Qwen/Qwen3-30B-A3B/blob/main/LICENSE
pipeline_tag: text-generation
base_model: Qwen/Qwen3-30B-A3B
---

Int8 quant for optimized performance on Ampere.

# usage with sglang

Currently, upstream sglang doesn't load this quant correctly due to a few minor issues. Until upstream is fixed, a working fork is available at https://github.com/nytopop/sglang/tree/qwen-30b-a3b:

```shell
uv venv --python 3.12

# use patched sglang from git
uv pip install "git+https://github.com/nytopop/sglang.git@qwen-30b-a3b#subdirectory=python[all]" --find-links https://flashinfer.ai/whl/cu124/torch2.5/flashinfer-python

# run
uv run python -m sglang.launch_server --model-path nytopop/Qwen3-30B-A3B.w8a8 --quantization w8a8_int8 --reasoning-parser qwen3
```

# creation

```python
from transformers import AutoModelForCausalLM
from llmcompressor import oneshot
from llmcompressor.modifiers.quantization import QuantizationModifier
from llmcompressor.transformers.compression.helpers import calculate_offload_device_map

model_id  = "Qwen/Qwen3-30B-A3B"
model_out = model_id.split("/")[1] + ".w8a8"

device_map = calculate_offload_device_map(
    model_id, reserve_for_hessians=False, num_gpus=1, torch_dtype="bfloat16"
)

for k, v in device_map.items():
  if v == 'disk':
    device_map[k] = 'cpu'

model = AutoModelForCausalLM.from_pretrained(
  model_id,
  device_map=device_map,
  torch_dtype="bfloat16",
)

recipe = QuantizationModifier(targets="Linear", scheme="W8A8", ignore=["lm_head", "re:.*mlp.gate$"])

oneshot(model=model, recipe=recipe, output_dir=model_out)
```