Instructions to use nvidia/Hymba-1.5B-Instruct with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use nvidia/Hymba-1.5B-Instruct with Transformers:

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

pipe = pipeline("text-generation", model="nvidia/Hymba-1.5B-Instruct", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("nvidia/Hymba-1.5B-Instruct", trust_remote_code=True, dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use nvidia/Hymba-1.5B-Instruct with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "nvidia/Hymba-1.5B-Instruct"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "nvidia/Hymba-1.5B-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/nvidia/Hymba-1.5B-Instruct

SGLang

How to use nvidia/Hymba-1.5B-Instruct 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 "nvidia/Hymba-1.5B-Instruct" \
    --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": "nvidia/Hymba-1.5B-Instruct",
		"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 "nvidia/Hymba-1.5B-Instruct" \
        --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": "nvidia/Hymba-1.5B-Instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use nvidia/Hymba-1.5B-Instruct with Docker Model Runner:
```
docker model run hf.co/nvidia/Hymba-1.5B-Instruct
```

The model only contains mamba and no attention

#11

by AscendingGrass - opened Feb 9, 2025

Discussion

AscendingGrass

Feb 9, 2025

•

edited Feb 9, 2025

Am I doing something wrong? because when I try to download the model and print the named parameters there's only mamba in the layers

from transformers import AutoModelForCausalLM

repo_name = "nvidia/Hymba-1.5B-Instruct"
model = AutoModelForCausalLM.from_pretrained(repo_name, trust_remote_code=True)

for name, param in model.named_parameters():
    print(f"{name}: {param.shape}")

here's the output I got:

model.memory_tokens: torch.Size([128, 1600])
model.embed_tokens.weight: torch.Size([32001, 1600])
model.layers.0.mamba.pre_avg_layernorm1.weight: torch.Size([3200])
model.layers.0.mamba.pre_avg_layernorm2.weight: torch.Size([3200])
model.layers.0.mamba.in_proj.weight: torch.Size([8960, 1600])
model.layers.0.mamba.out_proj.weight: torch.Size([1600, 3200])
model.layers.0.mamba.conv1d.weight: torch.Size([3200, 1, 4])
model.layers.0.mamba.conv1d.bias: torch.Size([3200])
model.layers.0.mamba.x_proj.0.weight: torch.Size([132, 3200])
model.layers.0.mamba.dt_proj.0.weight: torch.Size([3200, 100])
model.layers.0.mamba.dt_proj.0.bias: torch.Size([3200])
model.layers.0.mamba.A_log.0: torch.Size([3200, 16])
model.layers.0.mamba.D.0: torch.Size([3200])
model.layers.0.mamba.dt_layernorm.weight: torch.Size([100])
model.layers.0.mamba.B_layernorm.weight: torch.Size([16])
model.layers.0.mamba.C_layernorm.weight: torch.Size([16])
model.layers.0.input_layernorm.weight: torch.Size([1600])
model.layers.0.moe.experts.0.gate_proj.weight: torch.Size([5504, 1600])
model.layers.0.moe.experts.0.down_proj.weight: torch.Size([1600, 5504])
model.layers.0.moe.experts.0.up_proj.weight: torch.Size([5504, 1600])
model.layers.0.pre_moe_layernorm.weight: torch.Size([1600])
model.layers.1.mamba.pre_avg_layernorm1.weight: torch.Size([3200])
model.layers.1.mamba.pre_avg_layernorm2.weight: torch.Size([3200])
model.layers.1.mamba.in_proj.weight: torch.Size([8960, 1600])
model.layers.1.mamba.out_proj.weight: torch.Size([1600, 3200])
model.layers.1.mamba.conv1d.weight: torch.Size([3200, 1, 4])
model.layers.1.mamba.conv1d.bias: torch.Size([3200])
model.layers.1.mamba.x_proj.0.weight: torch.Size([132, 3200])
model.layers.1.mamba.dt_proj.0.weight: torch.Size([3200, 100])
model.layers.1.mamba.dt_proj.0.bias: torch.Size([3200])
model.layers.1.mamba.A_log.0: torch.Size([3200, 16])
model.layers.1.mamba.D.0: torch.Size([3200])
model.layers.1.mamba.dt_layernorm.weight: torch.Size([100])
model.layers.1.mamba.B_layernorm.weight: torch.Size([16])
model.layers.1.mamba.C_layernorm.weight: torch.Size([16])
model.layers.1.input_layernorm.weight: torch.Size([1600])
model.layers.1.moe.experts.0.gate_proj.weight: torch.Size([5504, 1600])
model.layers.1.moe.experts.0.down_proj.weight: torch.Size([1600, 5504])
model.layers.1.moe.experts.0.up_proj.weight: torch.Size([5504, 1600])
model.layers.1.pre_moe_layernorm.weight: torch.Size([1600])
model.layers.2.mamba.pre_avg_layernorm1.weight: torch.Size([3200])
model.layers.2.mamba.pre_avg_layernorm2.weight: torch.Size([3200])
model.layers.2.mamba.in_proj.weight: torch.Size([8000, 1600])
model.layers.2.mamba.out_proj.weight: torch.Size([1600, 3200])
model.layers.2.mamba.conv1d.weight: torch.Size([3200, 1, 4])
model.layers.2.mamba.conv1d.bias: torch.Size([3200])
model.layers.2.mamba.x_proj.0.weight: torch.Size([132, 3200])
model.layers.2.mamba.dt_proj.0.weight: torch.Size([3200, 100])
model.layers.2.mamba.dt_proj.0.bias: torch.Size([3200])
model.layers.2.mamba.A_log.0: torch.Size([3200, 16])
model.layers.2.mamba.D.0: torch.Size([3200])
model.layers.2.mamba.dt_layernorm.weight: torch.Size([100])
model.layers.2.mamba.B_layernorm.weight: torch.Size([16])
model.layers.2.mamba.C_layernorm.weight: torch.Size([16])
model.layers.2.input_layernorm.weight: torch.Size([1600])
model.layers.2.moe.experts.0.gate_proj.weight: torch.Size([5504, 1600])
model.layers.2.moe.experts.0.down_proj.weight: torch.Size([1600, 5504])
model.layers.2.moe.experts.0.up_proj.weight: torch.Size([5504, 1600])
model.layers.2.pre_moe_layernorm.weight: torch.Size([1600])
...
model.layers.31.mamba.pre_avg_layernorm1.weight: torch.Size([3200])
model.layers.31.mamba.pre_avg_layernorm2.weight: torch.Size([3200])
model.layers.31.mamba.in_proj.weight: torch.Size([8960, 1600])
model.layers.31.mamba.out_proj.weight: torch.Size([1600, 3200])
model.layers.31.mamba.conv1d.weight: torch.Size([3200, 1, 4])
model.layers.31.mamba.conv1d.bias: torch.Size([3200])
model.layers.31.mamba.x_proj.0.weight: torch.Size([132, 3200])
model.layers.31.mamba.dt_proj.0.weight: torch.Size([3200, 100])
model.layers.31.mamba.dt_proj.0.bias: torch.Size([3200])
model.layers.31.mamba.A_log.0: torch.Size([3200, 16])
model.layers.31.mamba.D.0: torch.Size([3200])
model.layers.31.mamba.dt_layernorm.weight: torch.Size([100])
model.layers.31.mamba.B_layernorm.weight: torch.Size([16])
model.layers.31.mamba.C_layernorm.weight: torch.Size([16])
model.layers.31.input_layernorm.weight: torch.Size([1600])
model.layers.31.moe.experts.0.gate_proj.weight: torch.Size([5504, 1600])
model.layers.31.moe.experts.0.down_proj.weight: torch.Size([1600, 5504])
model.layers.31.moe.experts.0.up_proj.weight: torch.Size([5504, 1600])
model.layers.31.pre_moe_layernorm.weight: torch.Size([1600])
model.final_layernorm.weight: torch.Size([1600])

jploski

Sep 19, 2025

The attention query/key/value vectors are computed as projections of Mamba's hidden states - see paper or modeling_hymba.py.

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment