Debugging

Training on multiple GPUs can be a tricky endeavor whether you’re running into installation issues or communication problems between your GPUs. This debugging guide covers some issues you may run into and how to resolve them.

DeepSpeed CUDA installation

If you’re using DeepSpeed, you’ve probably already installed it with the following command.

pip install deepspeed

DeepSpeed compiles CUDA C++ code and it can be a potential source of errors when building PyTorch extensions that require CUDA. These errors depend on how CUDA is installed on your system, and this section focuses on PyTorch built with CUDA 10.2.

Non-identical CUDA toolkits

PyTorch comes with its own CUDA toolkit, but to use DeepSpeed with PyTorch, you need to have an identical version of CUDA installed system-wide. For example, if you installed PyTorch with cudatoolkit==10.2 in your Python environment, then you’ll also need to have CUDA 10.2 installed system-wide. If you don’t have CUDA installed system-wide, you should install it first.

The exact location may vary from system to system, but usr/local/cuda-10.2 is the most common location on many Unix systems. When CUDA is correctly setup and added to your PATH environment variable, you can find the installation location with the following command:

which nvcc

Multiple CUDA toolkits

You may also have more than one CUDA toolkit installed system-wide.

/usr/local/cuda-10.2
/usr/local/cuda-11.0

Typically, package installers set the paths to whatever the last version was installed. If the package build fails because it can’t find the right CUDA version (despite it being installed system-wide already), then you need to configure the PATH and LD_LIBRARY_PATH environment variables to point to the correct path.

Take a look at the contents of these environment variables first:

echo $PATH
echo $LD_LIBRARY_PATH

PATH lists the locations of the executables and LD_LIBRARY_PATH lists where to look for shared libraries. Earlier entries are prioritized over later ones, and : is used to separate multiple entries. To tell the build program where to find the specific CUDA toolkit you want, insert the correct path to list first. This command prepends rather than overwrites the existing values.

# adjust the version and full path if needed
export PATH=/usr/local/cuda-10.2/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda-10.2/lib64:$LD_LIBRARY_PATH

In addition, you should also check the directories you assign actually exist. The lib64 sub-directory contains various CUDA .so objects (like libcudart.so) and while it is unlikely your system names them differently, you should check the actual names and change them accordingly.

Older CUDA versions

Sometimes, older CUDA versions may refuse to build with newer compilers. For example, if you have gcc-9 but CUDA wants gcc-7. Usually, installing the latest CUDA toolkit enables support for the newer compiler.

You could also install an older version of the compiler in addition to the one you’re currently using (or it may already be installed but it’s not used by default and the build system can’t see it). To resolve this, you can create a symlink to give the build system visibility to the older compiler.

# adapt the path to your system
sudo ln -s /usr/bin/gcc-7  /usr/local/cuda-10.2/bin/gcc
sudo ln -s /usr/bin/g++-7  /usr/local/cuda-10.2/bin/g++

Multi-GPU Network Issues Debug

When training or inferencing with DistributedDataParallel and multiple GPU, if you run into issue of inter-communication between processes and/or nodes, you can use the following script to diagnose network issues.

wget https://raw.githubusercontent.com/huggingface/transformers/main/scripts/distributed/torch-distributed-gpu-test.py

For example to test how 2 GPUs interact do:

python -m torch.distributed.run --nproc_per_node 2 --nnodes 1 torch-distributed-gpu-test.py

If both processes can talk to each and allocate GPU memory each will print an OK status.

For more GPUs or nodes adjust the arguments in the script.

You will find a lot more details inside the diagnostics script and even a recipe to how you could run it in a SLURM environment.

An additional level of debug is to add NCCL_DEBUG=INFO environment variable as follows:

NCCL_DEBUG=INFO python -m torch.distributed.run --nproc_per_node 2 --nnodes 1 torch-distributed-gpu-test.py

This will dump a lot of NCCL-related debug information, which you can then search online if you find that some problems are reported. Or if you’re not sure how to interpret the output you can share the log file in an Issue.

Underflow and Overflow Detection

If you start getting loss=NaN or the model inhibits some other abnormal behavior due to inf or nan in activations or weights one needs to discover where the first underflow or overflow happens and what led to it. Luckily you can accomplish that easily by activating a special module that will do the detection automatically.

If you’re using Trainer, you just need to add:

--debug underflow_overflow

to the normal command line arguments, or pass debug="underflow_overflow" when creating the TrainingArguments object.

If you’re using your own training loop or another Trainer you can accomplish the same with:

from transformers.debug_utils import DebugUnderflowOverflow

debug_overflow = DebugUnderflowOverflow(model)

DebugUnderflowOverflow inserts hooks into the model that immediately after each forward call will test input and output variables and also the corresponding module’s weights. As soon as inf or nan is detected in at least one element of the activations or weights, the program will assert and print a report like this (this was caught with google/mt5-small under fp16 mixed precision):

Detected inf/nan during batch_number=0
Last 21 forward frames:
abs min  abs max  metadata
                  encoder.block.1.layer.1.DenseReluDense.dropout Dropout
0.00e+00 2.57e+02 input[0]
0.00e+00 2.85e+02 output
[...]
                  encoder.block.2.layer.0 T5LayerSelfAttention
6.78e-04 3.15e+03 input[0]
2.65e-04 3.42e+03 output[0]
             None output[1]
2.25e-01 1.00e+04 output[2]
                  encoder.block.2.layer.1.layer_norm T5LayerNorm
8.69e-02 4.18e-01 weight
2.65e-04 3.42e+03 input[0]
1.79e-06 4.65e+00 output
                  encoder.block.2.layer.1.DenseReluDense.wi_0 Linear
2.17e-07 4.50e+00 weight
1.79e-06 4.65e+00 input[0]
2.68e-06 3.70e+01 output
                  encoder.block.2.layer.1.DenseReluDense.wi_1 Linear
8.08e-07 2.66e+01 weight
1.79e-06 4.65e+00 input[0]
1.27e-04 2.37e+02 output
                  encoder.block.2.layer.1.DenseReluDense.dropout Dropout
0.00e+00 8.76e+03 input[0]
0.00e+00 9.74e+03 output
                  encoder.block.2.layer.1.DenseReluDense.wo Linear
1.01e-06 6.44e+00 weight
0.00e+00 9.74e+03 input[0]
3.18e-04 6.27e+04 output
                  encoder.block.2.layer.1.DenseReluDense T5DenseGatedGeluDense
1.79e-06 4.65e+00 input[0]
3.18e-04 6.27e+04 output
                  encoder.block.2.layer.1.dropout Dropout
3.18e-04 6.27e+04 input[0]
0.00e+00      inf output

The example output has been trimmed in the middle for brevity.

The second column shows the value of the absolute largest element, so if you have a closer look at the last few frames, the inputs and outputs were in the range of 1e4. So when this training was done under fp16 mixed precision the very last step overflowed (since under fp16 the largest number before inf is 64e3). To avoid overflows under fp16 the activations must remain way below 1e4, because 1e4 * 1e4 = 1e8 so any matrix multiplication with large activations is going to lead to a numerical overflow condition.

At the very start of the trace you can discover at which batch number the problem occurred (here Detected inf/nan during batch_number=0 means the problem occurred on the first batch).

Each reported frame starts by declaring the fully qualified entry for the corresponding module this frame is reporting for. If we look just at this frame:

                  encoder.block.2.layer.1.layer_norm T5LayerNorm
8.69e-02 4.18e-01 weight
2.65e-04 3.42e+03 input[0]
1.79e-06 4.65e+00 output

Here, encoder.block.2.layer.1.layer_norm indicates that it was a layer norm for the first layer, of the second block of the encoder. And the specific calls of the forward is T5LayerNorm.

Let’s look at the last few frames of that report:

Detected inf/nan during batch_number=0
Last 21 forward frames:
abs min  abs max  metadata
[...]
                  encoder.block.2.layer.1.DenseReluDense.wi_0 Linear
2.17e-07 4.50e+00 weight
1.79e-06 4.65e+00 input[0]
2.68e-06 3.70e+01 output
                  encoder.block.2.layer.1.DenseReluDense.wi_1 Linear
8.08e-07 2.66e+01 weight
1.79e-06 4.65e+00 input[0]
1.27e-04 2.37e+02 output
                  encoder.block.2.layer.1.DenseReluDense.wo Linear
1.01e-06 6.44e+00 weight
0.00e+00 9.74e+03 input[0]
3.18e-04 6.27e+04 output
                  encoder.block.2.layer.1.DenseReluDense T5DenseGatedGeluDense
1.79e-06 4.65e+00 input[0]
3.18e-04 6.27e+04 output
                  encoder.block.2.layer.1.dropout Dropout
3.18e-04 6.27e+04 input[0]
0.00e+00      inf output

The last frame reports for Dropout.forward function with the first entry for the only input and the second for the only output. You can see that it was called from an attribute dropout inside DenseReluDense class. We can see that it happened during the first layer, of the 2nd block, during the very first batch. Finally, the absolute largest input elements was 6.27e+04 and same for the output was inf.

You can see here, that T5DenseGatedGeluDense.forward resulted in output activations, whose absolute max value was around 62.7K, which is very close to fp16’s top limit of 64K. In the next frame we have Dropout which renormalizes the weights, after it zeroed some of the elements, which pushes the absolute max value to more than 64K, and we get an overflow (inf).

As you can see it’s the previous frames that we need to look into when the numbers start going into very large for fp16 numbers.

Let’s match the report to the code from models/t5/modeling_t5.py:

class T5DenseGatedGeluDense(nn.Module):
    def __init__(self, config):
        super().__init__()
        self.wi_0 = nn.Linear(config.d_model, config.d_ff, bias=False)
        self.wi_1 = nn.Linear(config.d_model, config.d_ff, bias=False)
        self.wo = nn.Linear(config.d_ff, config.d_model, bias=False)
        self.dropout = nn.Dropout(config.dropout_rate)
        self.gelu_act = ACT2FN["gelu_new"]

    def forward(self, hidden_states):
        hidden_gelu = self.gelu_act(self.wi_0(hidden_states))
        hidden_linear = self.wi_1(hidden_states)
        hidden_states = hidden_gelu * hidden_linear
        hidden_states = self.dropout(hidden_states)
        hidden_states = self.wo(hidden_states)
        return hidden_states

Now it’s easy to see the dropout call, and all the previous calls as well.

Since the detection is happening in a forward hook, these reports are printed immediately after each forward returns.

Going back to the full report, to act on it and to fix the problem, we need to go a few frames up where the numbers started to go up and most likely switch to the fp32 mode here, so that the numbers don’t overflow when multiplied or summed up. Of course, there might be other solutions. For example, we could turn off amp temporarily if it’s enabled, after moving the original forward into a helper wrapper, like so:

def _forward(self, hidden_states):
    hidden_gelu = self.gelu_act(self.wi_0(hidden_states))
    hidden_linear = self.wi_1(hidden_states)
    hidden_states = hidden_gelu * hidden_linear
    hidden_states = self.dropout(hidden_states)
    hidden_states = self.wo(hidden_states)
    return hidden_states


import torch


def forward(self, hidden_states):
    if torch.is_autocast_enabled():
        with torch.cuda.amp.autocast(enabled=False):
            return self._forward(hidden_states)
    else:
        return self._forward(hidden_states)

Since the automatic detector only reports on inputs and outputs of full frames, once you know where to look, you may want to analyse the intermediary stages of any specific forward function as well. In such a case you can use the detect_overflow helper function to inject the detector where you want it, for example:

from debug_utils import detect_overflow


class T5LayerFF(nn.Module):
    [...]

    def forward(self, hidden_states):
        forwarded_states = self.layer_norm(hidden_states)
        detect_overflow(forwarded_states, "after layer_norm")
        forwarded_states = self.DenseReluDense(forwarded_states)
        detect_overflow(forwarded_states, "after DenseReluDense")
        return hidden_states + self.dropout(forwarded_states)

You can see that we added 2 of these and now we track if inf or nan for forwarded_states was detected somewhere in between.

Actually, the detector already reports these because each of the calls in the example above is a nn.Module, but let’s say if you had some local direct calculations this is how you’d do that.

Additionally, if you’re instantiating the debugger in your own code, you can adjust the number of frames printed from its default, e.g.:

from transformers.debug_utils import DebugUnderflowOverflow

debug_overflow = DebugUnderflowOverflow(model, max_frames_to_save=100)

Specific batch absolute min and max value tracing

The same debugging class can be used for per-batch tracing with the underflow/overflow detection feature turned off.

Let’s say you want to watch the absolute min and max values for all the ingredients of each forward call of a given batch, and only do that for batches 1 and 3. Then you instantiate this class as:

debug_overflow = DebugUnderflowOverflow(model, trace_batch_nums=[1, 3])

And now full batches 1 and 3 will be traced using the same format as the underflow/overflow detector does.

Batches are 0-indexed.

This is helpful if you know that the program starts misbehaving after a certain batch number, so you can fast-forward right to that area. Here is a sample truncated output for such configuration:

                  *** Starting batch number=1 ***
abs min  abs max  metadata
                  shared Embedding
1.01e-06 7.92e+02 weight
0.00e+00 2.47e+04 input[0]
5.36e-05 7.92e+02 output
[...]
                  decoder.dropout Dropout
1.60e-07 2.27e+01 input[0]
0.00e+00 2.52e+01 output
                  decoder T5Stack
     not a tensor output
                  lm_head Linear
1.01e-06 7.92e+02 weight
0.00e+00 1.11e+00 input[0]
6.06e-02 8.39e+01 output
                   T5ForConditionalGeneration
     not a tensor output

                  *** Starting batch number=3 ***
abs min  abs max  metadata
                  shared Embedding
1.01e-06 7.92e+02 weight
0.00e+00 2.78e+04 input[0]
5.36e-05 7.92e+02 output
[...]

Here you will get a huge number of frames dumped - as many as there were forward calls in your model, so it may or may not what you want, but sometimes it can be easier to use for debugging purposes than a normal debugger. For example, if a problem starts happening at batch number 150. So you can dump traces for batches 149 and 150 and compare where numbers started to diverge.

You can also specify the batch number after which to stop the training, with:

debug_overflow = DebugUnderflowOverflow(model, trace_batch_nums=[1, 3], abort_after_batch_num=3)