I'm running roberta on huggingface language_modeling.py. After doing 400 steps I suddenly get a CUDA out of memory issue. Don't know how to deal with it. Can you please help? Thanks
This can have multiple reasons. If you only get it after a few iterations, it might be that you don't free the computational graphs. Do you use loss.backward(retain_graph=True) or something similar?
Also, when you're running inference, be sure to use
with torch.no_grad():
model.forward(...)
Otherwise the computational graphs are saved there as well and potentially never freed since you never call backward() on them.
My problem was that I didn't check the size of my GPU memory with comparison to the sizes of samples. I had a lot of pretty small samples and after many iterations a large one. My bad.
Thank you and remember to check these things if it happens to you to.
Related
I have a large MILP that I build with cvxpy and want to solve with GUROBI. When I give use the solve() function of cvxpy it take a really really really long time to setup and does not start solving for hours. Whilest doing that only 1 core of my cluster is being used. It is used for 100%. I would like to use multiple cores to build the model so that the process of building the model does not take so long. Running grbprobe also shows that gurobi knows about the other cores and for solving the problem it uses multiple cores.
I have tried to run with different flags i.e. turning presolve off and on or giving the number of Threads to be used (this seemed like i didn't even for the solving.
I also have reduce the number of constraints in the problem and it start solving much faster which means that this is definitively not a problem of the model itself.
The problem in it's normal state should have 2200 constraints i reduce it to 150 and it took a couple of seconds until it started to search for a solution.
The problem is that I don't see anything since it takes so long to get the ""set username parameters"" flag and I don't get any information on what the computer does in the mean time.
Is there a way to tell GUROBI or CVXPY that it can take more cpus for the build-up?
Is there another way to solve this problem?
Sorry. The first part of the solve (cvxpy model generation, setup, presolving, scaling, solving the root, preprocessing) is almost completely serial. The parallel part is when it really starts working on the branch-and-bound tree. For many problems, the parallel part is by far the most expensive, but not for all.
This is not only the case for Gurobi. Other high-end solvers have the same behavior.
There are options to do less presolving and preprocessing. That may get you earlier in the B&B. However, usually, it is better not to touch these options.
Running things with verbose=True may give you more information. If you have more detailed questions, you may want to share the log.
I have found this post but am still unclear on what the redzone_checker kernel is doing and why. Specifically, should it be taking > 90% of my application's runtime? TensorBoard reports that it is taking the vast majority of the runtime of my JAX code, and I'd like to know
Is it actually the case that this kernel is taking way too much time, or is this a side effect of profiling JAX with TensorBoard (i.e., the output is misleading in some way)?
Is there a way to reduce the amount of time taken by the redzone_checker kernel? Is that even a good idea?
Thanks in advance for any insights.
make sure warmup before profiling.
it maybe jit compiling time.
I profiled a model that I am running and the vast majority of the time in each step (295 of 320ms) is being taken up by "device-to-device" operations (see image). I assume this means loading data from my cpu onto my gpu and back is the bottleneck.
I am running this on a single machine. The data is stored on an SSD and being fed into a GPU.
I am using tensorflow's tf.data.Dataset API and doing all the recommended things like prefetching and num_parallel_calls=tf.data.experimental.AUTOTUNE
My questions are:
(1) Is my assumption correct?
(2) How do I reduce this huge burden on my model?
Tensorboard Profiling Overview
Not a proper answer but it's something; by using tensorflow's mixed precision training I was able to reduce the "device-to-device" time to ~ 145ms. This is still an immense burden compared to everything else profiled and I'd love to be able to reduce it further.
I don't know why this helped either. I assume that mp-training means smaller numbers of bytes are being passed around so maybe that helps.
While GPUs speed math calculations there is a fixed overhead for moving a kernel out to the GPU for execution that is high.
I'm using cupy and numba. THe first time I execute a function call that is using cupy's GPU version of numpy it is quite slow. But the second time it is fast.
I've realized I don't understand how the kernel, or GPU code, gets out to the GPU to run. Operationally I want to understand this better so that I can know when the things I do will accidentally create a slow step due to some kernel transfer. So I need some sorts of rules or rules of thumb understand the concept.
For example, if I multiply two cupy arrays that are stashed on the GPU already I might write C= A*B
At some point the cupy overload on * multiplication has to be coded out on the GPU, and it automagically needs will also get wrapped by loops that break it down into blocks and threads. So presumably this code is some kernel that gets transported out to the GPU. I'm guessing that the next time I call C*D that the GPU no longer needs to be taught what * means and so it will be fast.
But at some point I would imagine the GPU needs to clear out old code so * or other operations not being used at that moment might get flushed from memory, and so later on when the call for A*B happens again there's going to be a penalty in time to recompile it out on the GPU.
Or so I imagine. If I'm right how do I know when these kernels are going to stick around or disappear?
If I'm wrong and this isn't how it works or there is some other slow step (I'm assuming the data is already transported to arrays on the GPU) then what is this slow step and how does organize things so one pay it as little as possible?
I'm trying to avoid writing explicit numba thread managment kernels as one does in cuda++ but just use the standard numba #njit, #vectorize, #stencil decorators. Likewise in Cupy I want to just work at the level of the numpy syntax not dive into thread management.
I've read a lot of docs on this but they just refer to overheads of kernels, not when these get paid and how one controls that so I'm confused.
I don't have a full answer to this yet. But so far the biggest clue I've gotten has come from reading up on the currently undocumented function #cupy.fuse() which makes it more clear than the #numba.jit documents where the kernel launch costs are paid. I have not found the connection to Contexts yet as recommended by #talonmies.
see https://gist.github.com/unnonouno/877f314870d1e3a2f3f45d84de78d56c
The key example is this
c = cupy.arange(4)
##cupy.fuse()
def foo(x):
return x+x+x+x+x
foo(.) will be three times slower with #cupy.fuse() commented out because each "+" involves a kernel load and a kernel free. Fusion merges all the adds into a single kernel so those the launch and free are paid onces. FOr matricies less than 1 million in size on a typical 2018 GPU, the add() is so fast that the launch and free are the dominate times.
I wish I could find some documentation on #fuse. FOr example, does it unroll internal functions the way #jit does. Could I achieve that by stacking #jit and #fuse?
I'm still however largely in the dark about when the costs are getting paid in numba.
I'm using Tensorflow-GPU 1.8 API on Windows 10. For many projects I use the tf.Estimator's, which really work great. It takes care of a bunch of steps including writting summaries for Tensorboard. But right now the 'events.out.tfevents' file getting way to big and I am running into "out of space" errors. For that reason I want to disable the summary writting or at least reduce the amount of summaries written.
Going along with that mission I found out about the RunConfig you can pass over at construction of tf.Estimator. Apparently the parameter 'save_summary_steps' (which by default is 200) controls the way summaries are wrtitten out. Unfortunately changing this parameter seems to have no effect at all. It won't disable (using None value) the summary or reducing (choosing higher values, e.g. 3000) the file size of 'events.out.tfevents'.
I hope you guys can help me out here. Any help is appreciated.
Cheers,
Tobs.
I've observed the following behavior. It doesn't make sense to me so I hope we get a better answer:
When the input_fn gets data from tf.data.TFRecordDataset then the number of steps between saving events is the minimum of save_summary_steps and (number of training examples divided by batch size). That means it does it a minimum of once per epoch.
When the input_fn gets data from tf.TextLineReader, it follows save_summary_steps as you'd expect and I can give it a large value for infrequent updates.