Full BF16 including layernorms by default (minimising number of BF16 atomics) #272
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
I'm assuming this is not meant to merge as-is? Would it make sense to put most of this into dev/cuda and then cherry-pick the layernorm we want to use, as usual, into train_gpt2.cu? |
ngc92
reviewed
Apr 28, 2024
train_gpt2.cu
Outdated
| __global__ void layernorm_backward_kernel3(Tdinp* dinp, Tparams* dweight, Tparams* dbias, | ||
| const Tdout* dout, const Trest* inp, const Tparams* weight, const Trest* mean, const Trest* rstd, | ||
| int B, int T, int C) { | ||
| extern __shared__ float shared[]; // size = 2 * C |
ngc92
reviewed
Apr 28, 2024
train_gpt2.cu
Outdated
| dbias_shared[i] = 0.0f; | ||
| dweight_shared[i] = 0.0f; | ||
| } | ||
| uint *tmp_flag = (uint*)(shared + C*2); |
There was a problem hiding this comment.
is this actually used in this kernel?
ngc92
reviewed
Apr 28, 2024
train_gpt2.cu
Outdated
| dbias_shared[i] = 0.0f; | ||
| dweight_shared[i] = 0.0f; | ||
| } | ||
| uint *tmp_flag = (uint*)(shared + C*2); |
|
Thanks! Fixed those bugs, and only kept Kernel6 in train_gpt2.cu, adding all the kernels to /dev/cuda/ I think this might be the first /dev/cuda change with BF16, so I added support for comparing CPU and GPU data of different types (converting GPU data to CPU type before comparison) in validate_result(). |
ngc92
reviewed
Apr 29, 2024
dev/cuda/common.h
Outdated
| cudaCheck(cudaMemcpy(out_gpu, device_result, num_elements * sizeof(T), cudaMemcpyDeviceToHost)); | ||
| template<class D, class T> | ||
| void validate_result(D* device_result, const T* cpu_reference, const char* name, std::size_t num_elements, T tolerance=1e-4) { | ||
| D* out_gpu = (D*)malloc(num_elements * sizeof(T)); |
There was a problem hiding this comment.
Good catch, fixed!
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
I added 4 different new versions of layernorm_backward_kernel, performance is best for:
We probably just want to integrate Kernel 6 but might want to add all of them to /dev/cuda/ in the future. I haven't fixed the BF16 atomics in encoder_backward yet but the performance penalty of that is much much smaller, it might be worth doing it manually using atomicCAS so we can implement stochastic rounding there as well though.
Performance on my RTX 4090 is a tiny bit faster (potentially noise) than with the previous mixed FP32/BF16.