ERROR: Failed building wheel for mamba-ssm

[5pace4]

Building wheels for collected packages: mamba-ssm
error: subprocess-exited-with-error

× Building wheel for mamba-ssm (pyproject.toml) did not run successfully.
│ exit code: 1
╰─> See above for output.

note: This error originates from a subprocess, and is likely not a problem with pip.
Building wheel for mamba-ssm (pyproject.toml) ... error
ERROR: Failed building wheel for mamba-ssm
Failed to build mamba-ssm
ERROR: ERROR: Failed to build installable wheels for some pyproject.toml based projects (mamba-ssm)

how to solve this error?

[ahatamiz]

This is most likely due to a mismatch for the right CUDA version.

[Mactor2018]

Maybe my resolution helps #86

[5pace4]

Maybe my resolution helps #86

@Mactor2018 Thanks! Your approach in #86 was really helpful.
I'm here providing a simplified guidelines to set it up on Kaggle or Colab.

MambaVision Installation Guide for Google Colab and Kaggle

Prerequisites

• Google Colab: Linux-based, Python 3.10+, CUDA 12.1 or 12.4.
  Enable GPU runtime: Runtime > Change runtime type > GPU.

• Kaggle: Linux-based, Python 3.10+, CUDA 12.0 or 12.1.
  Enable GPU: Settings > Accelerator > GPU P100 or T4.

• Ensure PyTorch 2.6 is compatible with your CUDA version.

---

Solution 1: Install Compatible mamba-ssm Version

This approach installs mamba-ssm==2.2.3.post2 (or similar) with pre-built wheels, avoiding the need to build from source.

1. Start a Notebook with GPU

2. Check CUDA and PyTorch Versions:

!nvcc --version  # Check CUDA version (e.g., 12.1 or 12.4)
import torch
print(torch.__version__)  # Should be 2.6.0 or compatible
print(torch.cuda.is_available())  # Should print True

3. Install PyTorch (if needed):

!pip install torch==2.6.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121

4. Install mamba-ssm v2.2.3.post2:

!pip install mamba-ssm==2.2.3.post2

If it fails, try:

!pip install mamba-ssm==2.2.3.post2 --no-build-isolation

5. Install causal-conv1d (Optional):

!pip install causal-conv1d==1.4.0

6. Install mambavision:

!pip install mambavision

if it still gives any error, don't worry! run the code, this will works.

7. Test Installation:

from transformers import AutoModel
from PIL import Image
from timm.data.transforms_factory import create_transform
import requests


model = AutoModel.from_pretrained("nvidia/MambaVision-T-1K", trust_remote_code=True)

# eval mode for inference
model.cuda().eval()

# prepare image for the model
url = 'http://images.cocodataset.org/val2017/000000020247.jpg'
image = Image.open(requests.get(url, stream=True).raw)
input_resolution = (3, 224, 224)  # MambaVision supports any input resolutions

transform = create_transform(input_size=input_resolution,
                             is_training=False,
                             mean=model.config.mean,
                             std=model.config.std,
                             crop_mode=model.config.crop_mode,
                             crop_pct=model.config.crop_pct)
inputs = transform(image).unsqueeze(0).cuda()
# model inference
out_avg_pool, features = model(inputs)
print("Size of the averaged pool features:", out_avg_pool.size())  # torch.Size([1, 640])
print("Number of stages in extracted features:", len(features)) # 4 stages
print("Size of extracted features in stage 1:", features[0].size()) # torch.Size([1, 80, 56, 56])
print("Size of extracted features in stage 4:", features[3].size()) # torch.Size([1, 640, 7, 7])

---

Solution 2: Build mambavision from Source

If no compatible mamba-ssm wheel is available, build mambavision from source after modifying its setup.py to remove the strict mamba-ssm==2.2.4 requirement.

1. Clone MambaVision Repository:

!git clone https://github.com/NVlabs/MambaVision.git
%cd MambaVision

2. Modify setup.py:

!sed -i 's/mamba-ssm==2.2.4/mamba-ssm/g' setup.py

3. Install Dependencies:

!pip install mamba-ssm==2.2.3.post2 causal-conv1d==1.4.0

4. Build and Install mambavision:

!pip install -e .

5. Test Installation:

Use the test script from Solution 1.

---

NB: Solution 1 worked for me, so I didn’t need to try Solution 2. However, I had planned to use Solution 2 if Solution 1 failed.

[GiuseppeFrigeni]

I tried both methods used here and still getting errors:

---

RuntimeError Traceback (most recent call last)
/tmp/ipykernel_179/3684884852.py in <cell line: 0>()
23 inputs = transform(image).unsqueeze(0).cuda()
24 # model inference
---> 25 out_avg_pool, features = model(inputs)
26 print("Size of the averaged pool features:", out_avg_pool.size()) # torch.Size([1, 640])
27 print("Number of stages in extracted features:", len(features)) # 4 stages

/usr/local/lib/python3.11/dist-packages/torch/nn/modules/module.py in _wrapped_call_impl(self, *args, **kwargs)
1737 return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc]
1738 else:
-> 1739 return self._call_impl(*args, **kwargs)
1740
1741 # torchrec tests the code consistency with the following code

/usr/local/lib/python3.11/dist-packages/torch/nn/modules/module.py in _call_impl(self, *args, **kwargs)
1748 or _global_backward_pre_hooks or _global_backward_hooks
1749 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1750 return forward_call(*args, **kwargs)
1751
1752 result = None

~/.cache/huggingface/modules/transformers_modules/nvidia/MambaVision-T-1K/b1de77e17599566d98efb701c0231b1095dc3a67/modeling_mambavision.py in forward(self, tensor)
735
736 def forward(self, tensor):
--> 737 return self.model.forward_features(tensor)
738
739

~/.cache/huggingface/modules/transformers_modules/nvidia/MambaVision-T-1K/b1de77e17599566d98efb701c0231b1095dc3a67/modeling_mambavision.py in forward_features(self, x)
700 outs = []
701 for level in self.levels:
--> 702 x, xo = level(x)
703 outs.append(xo)
704 x = self.norm(x)

/usr/local/lib/python3.11/dist-packages/torch/nn/modules/module.py in _wrapped_call_impl(self, *args, **kwargs)
1737 return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc]
1738 else:
-> 1739 return self._call_impl(*args, **kwargs)
1740
1741 # torchrec tests the code consistency with the following code

/usr/local/lib/python3.11/dist-packages/torch/nn/modules/module.py in _call_impl(self, *args, **kwargs)
1748 or _global_backward_pre_hooks or _global_backward_hooks
1749 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1750 return forward_call(*args, **kwargs)
1751
1752 result = None

~/.cache/huggingface/modules/transformers_modules/nvidia/MambaVision-T-1K/b1de77e17599566d98efb701c0231b1095dc3a67/modeling_mambavision.py in forward(self, x)
597
598 for _, blk in enumerate(self.blocks):
--> 599 x = blk(x)
600 if self.transformer_block:
601 x = window_reverse(x, self.window_size, Hp, Wp)

/usr/local/lib/python3.11/dist-packages/torch/nn/modules/module.py in _wrapped_call_impl(self, *args, **kwargs)
1737 return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc]
1738 else:
-> 1739 return self._call_impl(*args, **kwargs)
1740
1741 # torchrec tests the code consistency with the following code

/usr/local/lib/python3.11/dist-packages/torch/nn/modules/module.py in _call_impl(self, *args, **kwargs)
1748 or _global_backward_pre_hooks or _global_backward_hooks
1749 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1750 return forward_call(*args, **kwargs)
1751
1752 result = None

~/.cache/huggingface/modules/transformers_modules/nvidia/MambaVision-T-1K/b1de77e17599566d98efb701c0231b1095dc3a67/modeling_mambavision.py in forward(self, x)
507
508 def forward(self, x):
--> 509 x = x + self.drop_path(self.gamma_1 * self.mixer(self.norm1(x)))
510 x = x + self.drop_path(self.gamma_2 * self.mlp(self.norm2(x)))
511 return x

/usr/local/lib/python3.11/dist-packages/torch/nn/modules/module.py in _wrapped_call_impl(self, *args, **kwargs)
1737 return self._compiled_call_impl(*args, **kwargs) # type: ignore[misc]
1738 else:
-> 1739 return self._call_impl(*args, **kwargs)
1740
1741 # torchrec tests the code consistency with the following code

/usr/local/lib/python3.11/dist-packages/torch/nn/modules/module.py in _call_impl(self, *args, **kwargs)
1748 or _global_backward_pre_hooks or _global_backward_hooks
1749 or _global_forward_hooks or _global_forward_pre_hooks):
-> 1750 return forward_call(*args, **kwargs)
1751
1752 result = None

~/.cache/huggingface/modules/transformers_modules/nvidia/MambaVision-T-1K/b1de77e17599566d98efb701c0231b1095dc3a67/modeling_mambavision.py in forward(self, hidden_states)
395 B = rearrange(B, "(b l) dstate -> b dstate l", l=seqlen).contiguous()
396 C = rearrange(C, "(b l) dstate -> b dstate l", l=seqlen).contiguous()
--> 397 y = selective_scan_fn(x,
398 dt,
399 A,

/usr/local/lib/python3.11/dist-packages/mamba_ssm/ops/selective_scan_interface.py in selective_scan_fn(u, delta, A, B, C, D, z, delta_bias, delta_softplus, return_last_state)
108 not considered in the backward pass.
109 """
--> 110 return SelectiveScanFn.apply(u, delta, A, B, C, D, z, delta_bias, delta_softplus, return_last_state)
111
112

/usr/local/lib/python3.11/dist-packages/torch/autograd/function.py in apply(cls, *args, **kwargs)
573 # See NOTE: [functorch vjp and autograd interaction]
574 args = _functorch.utils.unwrap_dead_wrappers(args)
--> 575 return super().apply(*args, **kwargs) # type: ignore[misc]
576
577 if not is_setup_ctx_defined:

/usr/local/lib/python3.11/dist-packages/mamba_ssm/ops/selective_scan_interface.py in forward(ctx, u, delta, A, B, C, D, z, delta_bias, delta_softplus, return_last_state)
42 C = rearrange(C, "b dstate l -> b 1 dstate l")
43 ctx.squeeze_C = True
---> 44 out, x, *rest = selective_scan_cuda.fwd(u, delta, A, B, C, D, z, delta_bias, delta_softplus)
45 ctx.delta_softplus = delta_softplus
46 ctx.has_z = z is not None

RuntimeError: CUDA error: no kernel image is available for execution on the device
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions.