From 6a8ca04eea8c6c279da0030b5857b2c2a13da24d Mon Sep 17 00:00:00 2001 From: JimmyZhang12 <67203904+JimmyZhang12@users.noreply.github.com> Date: Mon, 16 Sep 2024 22:49:40 -0700 Subject: [PATCH] Update cudagraphs.md Signed-off-by: JimmyZhang12 <67203904+JimmyZhang12@users.noreply.github.com> --- docs/source/performance/cudagraphs.md | 14 ++++++++------ 1 file changed, 8 insertions(+), 6 deletions(-) diff --git a/docs/source/performance/cudagraphs.md b/docs/source/performance/cudagraphs.md index aaceb8c0c08f..e7d69cc5c447 100644 --- a/docs/source/performance/cudagraphs.md +++ b/docs/source/performance/cudagraphs.md @@ -1,14 +1,17 @@ # Cudagraphs for NeMo ### Background and Motivation -Cudagraphs are a performance feature used mainly to eliminate reduce the presence of GPU idling due to host overhead and jitter. Commonly, if the host cannot issue kernels fast enough, the GPU may empty its work queue and begin idling. Cudagraphs solve this by encapsulating a sequence of many kernels into single unit. Thus, the host can launch multiple kernels by calling a single cudagraph, which greatly reduces the amount of work required by the host in comparison to eagerly launching each kernel individually. -The current cudagraph implementation in NeMo maps each transformer layer’s forward and backpasses to a cudagraph. On the first step, `CudaGraphManager` intercepts the forward pass to each transformer layer and uses it to trace the transformer layer via stream capture. On subsequent steps, `CudagraphManager`, rather than of executing the layer eagerly, reroutes calls into a single cudagraph, +Cudagraphs are a performance feature used mainly to eliminate or reduce the presence of GPU idling due to host overhead and jitter. Commonly, if the host cannot issue kernels fast enough, the GPU may empty its work queue and begin idling. Cudagraphs solve this by encapsulating a sequence of many kernels into single unit. Thus, the host can launch multiple kernels by calling a single cudagraph, which greatly reduces the amount of work required by the host in comparison to eagerly launching each kernel individually. -Currently, cudagraphs increases the memory usage of activations by ~20%. End to end, this is roughly an increase of about 10-20GB. The reason for this increase is due to cudagraphs preallocating any memory used by temporary tensors in the forward pass. As a result, temporary tensors that normally are deallocated when dereferenced are kept allocated for the lifespan of the cudagraph. The implementation of cudagraphs for NeMo can be found in [ megatron/core/transformer/cuda_graphs.py](https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/core/transformer/cuda_graphs.py). +The current implementation in NeMo captures all kernels per transformer layer in each of the forward and backward passes as a single graph. On the first step, `CudaGraphManager` intercepts the forward pass to each transformer layer and uses it to trace the transformer layer via stream capture. On subsequent steps, `CudagraphManager`, rather than executing the layer eagerly, reroutes calls into a single cudagraph, + +Currently, cudagraphs increases the memory usage of activations by ~20%. End to end, this is roughly an increase of about 10-20GB. The reason for this increase is due to cudagraphs preallocating any memory used by temporary tensors in the forward pass. As a result, temporary tensors that normally are deallocated when dereferenced are kept allocated for the lifespan of the cudagraph. These unnecessary allocations will be removed in future versions, in which cudagraphs are expected to not meaningfully increase memory usage. + +The implementation of cudagraphs for NeMo can be found in [ megatron/core/transformer/cuda_graphs.py](https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/core/transformer/cuda_graphs.py). ### When to use -Cudagraphs are recommended to be used when host overheads are expected to be significant, for instance systems with weak single-threaded CPU performance. -As a demonstration we show that with a GPT3 20B model, cudagraphs can improve performance by 11.4%. The increase in memory usage due to cudagraphs was observed to be 15GB per gpu. +Cudagraphs are recommended to be used when host overheads are expected to be significant, for instance systems with weak single-threaded CPU performance, or training workloads with small tensors due to small problem sizes (e.g. a small micro-batch size, sequence length, or hidden size). +As a demonstration we show that with a GPT3 20B model, cudagraphs can improve performance by 11.4%. The increase in memory usage due to cudagraphs was observed to be 15GB per GPU. | Setting | TFLOP/s | | ----- | ------ | | No Cudagraphs | 750 | @@ -23,4 +26,3 @@ To enable please add the following configs: `model.use_te_rng_tracker=True` `model.get_attention_mask_from_fusion=True` -