Instructions to use Sayankotor/qwen35-122b-a10b-k4e2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Sayankotor/qwen35-122b-a10b-k4e2 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Sayankotor/qwen35-122b-a10b-k4e2", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoProcessor, AutoModel

processor = AutoProcessor.from_pretrained("Sayankotor/qwen35-122b-a10b-k4e2", trust_remote_code=True)
model = AutoModel.from_pretrained("Sayankotor/qwen35-122b-a10b-k4e2", trust_remote_code=True, device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Trellis

How to use Sayankotor/qwen35-122b-a10b-k4e2 with Trellis:

# No code snippets available yet for this library.

# To use this model, check the repository files and the library's documentation.

# Want to help? PRs adding snippets are welcome at:
# https://github.com/huggingface/huggingface.js

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Sayankotor/qwen35-122b-a10b-k4e2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Sayankotor/qwen35-122b-a10b-k4e2"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Sayankotor/qwen35-122b-a10b-k4e2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Sayankotor/qwen35-122b-a10b-k4e2

SGLang

How to use Sayankotor/qwen35-122b-a10b-k4e2 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Sayankotor/qwen35-122b-a10b-k4e2" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Sayankotor/qwen35-122b-a10b-k4e2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Sayankotor/qwen35-122b-a10b-k4e2" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Sayankotor/qwen35-122b-a10b-k4e2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Sayankotor/qwen35-122b-a10b-k4e2 with Docker Model Runner:
```
docker model run hf.co/Sayankotor/qwen35-122b-a10b-k4e2
```

qwen35-122b-a10b-k4e2 / qtip_kernel_decompress.py

Sayankotor

Upload QTIP K4/K2 quantized Qwen3.5-122B-A10B (weights + self-loading code)

5a15552 verified 10 days ago

Raw

History Blame Contribute Delete

2.24 kB

	import torch


	#@torch.compile
	def decode_compressed(L, S, R, V, m, k, compressed, expanded_lut):
	if compressed.dtype != torch.uint16:
	compressed = compressed.view(torch.uint16)

	assert compressed.shape == (R * m * k // 16, )

	BITS_PER_BLOCK = R * 16 * 16 # R bits * f16 mma tile A size

	# unswizzle interleaved blocks

	BLOCK_SIZE = 16 * 16

	BITS_PER_BLOCK = R * 16 * 16 # R bits * f16 mma tile A size

	compressed = (compressed.view(torch.uint8).reshape(
	m // 16 // 2, k // 16 // 2, BLOCK_SIZE // 8, 2, 2,
	R).permute(0, -2, 1, -3, 2, -1).flip(
	(-1, )).reshape(m // 16, k // 16, BITS_PER_BLOCK // 16, 2).flip(
	(-1, )).view(torch.uint16).reshape(m // 16, k // 16,
	BITS_PER_BLOCK // 16))
	# decode block

	assert L <= 16

	blocked = compressed.reshape(R * m * k // BITS_PER_BLOCK,
	BITS_PER_BLOCK // 16, 1)
	blocked_roll = torch.roll(blocked.to(torch.int32), -1,
	-2).to(blocked.dtype)
	blocked32 = torch.cat((blocked_roll, blocked),
	dim=-1).reshape(blocked.shape[0],
	-1).contiguous().view(torch.uint32)
	# blocked32 is 16bits[-1]\|\|16bits[0] 16bits[0]\|\|16bits[1] ... 16bits[-2]\|\|16bits[-1]

	expanded32 = blocked32.reshape(*blocked32.shape,
	1).expand(*blocked32.shape,
	16).view(torch.int32)
	shifts = (torch.arange(0, 16, dtype=torch.int32,
	device=blocked.device)).to(torch.int32).reshape(
	1, 1, -1).expand(expanded32.shape)
	shifted = expanded32 >> (16 - shifts)
	indices = torch.bitwise_and(
	shifted.reshape(shifted.shape[0], -1)[:, 16 - L::R << V], (1 << L) - 1)

	# decode lut
	mma_swizzled = expanded_lut[indices]

	# deswizzle m16n8k16 mma pattern
	decompressed = (mma_swizzled.reshape(m // 16, k // 16, 16, 16).reshape(
	m // 16, k // 16, 8, 4, 2, 2, 2).permute(0, -2, 2, 1, -3, 3,
	-1).reshape(m, k))
	return decompressed