Open xiaoxudaxfcv opened 3 months ago
from mamba_ssm import Mamba2 model = Mamba( #该模块大约使用了3 expand d_model^2 个参数 d_model=64, #模型维度 d_model d_state=64, # SSM状态扩展因子,通常为 64 或 128 d_conv=4, #局部清晰度宽度 expand=2, #块扩展因子 ).to("cuda") NameError: 名称 'Mamba' 未定义。您的意思是: 'Mamba2'? 改成model = Mamba2即可解决上述问题,原来的任务使用Mamba不会报错,但是使用from mamba_ssm import Mamba2 model = Mamba2时,报错信息RuntimeError: causal_conv1d with channel last layout require strides (x.stride(0) and x.stride(2)) to be multiples of 8 改成model = Mamba2,解决上面问题,原来的任务使用Mamba不会报错,但是使用from mamba_ssm import Mamba2 model = Mamba2,会报错RuntimeError: causal_conv1d with channel last layout require strides (x.stride(0) and x.stride(2)) to be multiples of 8
I also encountered the same problem
https://github.com/state-spaces/mamba/issues/345#issuecomment-2145035818 You may try this.
maybe add .contiguous() guess the author forget
same problem Has anyone solve it?
maybe add .contiguous() guess the author forget
Thank you for your tip. I created a PR based on your suggestion #362
maybe add .contiguous() guess the author forget
Thank you for your tip. I created a PR based on your suggestion #362
i remember the backward and the forward both need to fix. hope U miss nothing.XD
still dontwork
You can skip causual_1d_conv (comment out and set the causual_conv_1d=None). It will use the nn.Conv1d function. Finally, you need to make the transposed tensor to be contiguous()
You can skip causual_1d_conv (comment out and set the causual_conv_1d=None). It will use the nn.Conv1d function. Finally, you need to make the transposed tensor to be contiguous()
Indeed, the second commit addresses the latter.
maybe add .contiguous() guess the author forget
Thank you for your tip. I created a PR based on your suggestion #362
i remember the backward and the forward both need to fix. hope U miss nothing.XD
Sadly it still won't work even after making it contiguous in the second part. It would now fail with a different error.
Mamba2 has same problem on my two devices(3080 , 3090).
I also try to run Mamba2 on colab(T4). It doesn't have this issue, but other error emerged.
也许添加 .contigious() 猜作者忘记了
谢谢你的提示。我根据你的建议创建了一个 PR #362
我记得向后和向前都需要修复。希望你没有错过任何事情。XD
遗憾的是,即使在第二部分中使其连续之后,它仍然无法工作。现在它会因不同的错误而失败。
I used the method you mentioned, but I encountered a strange mistake: Attributeerror: Module 'Triton. Language' Has No Attribute 'Cumsum'
也许添加 .contigious() 猜作者忘记了
谢谢你的提示。我根据你的建议创建了一个 PR #362
我记得向后和向前都需要修复。希望你没有错过任何事情。XD
遗憾的是,即使在第二部分中使其连续之后,它仍然无法工作。现在它会因不同的错误而失败。
I used the method you mentioned, but I encountered a strange mistake: Attributeerror: Module 'Triton. Language' Has No Attribute 'Cumsum'
version problem
也许添加 .contigious() 猜作者忘记了
谢谢你的提示。我根据你的建议创建了一个 PR #362
我记得向后和向前都需要修复。希望你没有错过任何事情。XD
遗憾的是,即使在第二部分中使其连续之后,它仍然无法工作。现在它会因不同的错误而失败。
I used the method you mentioned, but I encountered a strange mistake: Attributeerror: Module 'Triton. Language' Has No Attribute 'Cumsum'
update the triton
just make sure d_model * expand / headdim = multiple of 8
i had d_model = 192, expand = 2 and i changed headdim from 64 to 48, everything runs fine after
just make sure d_model * expand / headdim = multiple of 8
i had d_model = 192, expand = 2 and i change headdim from 64 to 48, everything runs fine after
I use your make sure method , d_model= 128 , expand=2, headdim= 4, I works,THANKS
By the way , it sames slow than mamba1
['/usr/bin/gcc', '/tmp/tmp9a7rbp53/main.c', '-O3', '-I/home/cpz/anaconda3/envs/mamba/lib/python3.10/site-packages/triton/common/../third_party/cuda/include', '-I/home/cpz/anaconda3/envs/mamba/include/python3.10', '-I/tmp/tmp9a7rbp53', '-shared', '-fPIC', '-lcuda', '-o', '/tmp/tmp9a7rbp53/cuda_utils.cpython-310-x86_64-linux-gnu.so', '-L/lib64', '-L/lib', '-L/lib64', '-L/lib']' returned non-zero exit status 1.
After soving this question, I continue to meet a new error.
['/usr/bin/gcc', '/tmp/tmp9a7rbp53/main.c', '-O3', '-I/home/cpz/anaconda3/envs/mamba/lib/python3.10/site-packages/triton/common/../third_party/cuda/include', '-I/home/cpz/anaconda3/envs/mamba/include/python3.10', '-I/tmp/tmp9a7rbp53', '-shared', '-fPIC', '-lcuda', '-o', '/tmp/tmp9a7rbp53/cuda_utils.cpython-310-x86_64-linux-gnu.so', '-L/lib64', '-L/lib', '-L/lib64', '-L/lib']' returned non-zero exit status 1.
After soving this question, I continue to meet a new error.
me too, did u solve this problem?
from mamba_ssm import Mamba2 model = Mamba(
This module uses roughly 3 expand d_model^2 parameters
).to("cuda") NameError: name 'Mamba' is not defined. Did you mean: 'Mamba2'? Change to model = Mamba2 to solve the above problem, but using Mamba on the original task will not cause an error, but using from mamba_ssm to import Mamba2 model = Mamba2, the error message RuntimeError: causal_conv1d with channel last layout requires strides (x.stride(0) and x.stride(2)) to be multiples of 8
改成model = Mamba2,解决上面问题,但是在原来的任务上使用Mamba不会报错,但使用from mamba_ssm import Mamba2 model = Mamba2,会报错RuntimeError: causal_conv1d with channel last layout requires strides (x.stride(0) and x.stride(2)) to be multiples of 8