[405B] Add performance data for 405B model

docs/performance.md

@@ -2,9 +2,25 @@ To demonstrate the effectiveness of PyTorch distributed training techniques used
 We report infra metrics achieved by [FSDP2](fsdp.md) (1D parallelism) under various configurations, and loss curves for both 1D parallelism (FSDP2) and 2D parallelism (FSDP2 + Tensor Parallel) training.
 
 
+## Llama 3.1 performance numbers
+
+Below are the WPS (word per second, or more accurately, token per second) and MFU (model FLOPS utilization) results which torchtitan achieves on the 405B model released in [LLaMa 3.1](https://llama.meta.com/docs/model-cards-and-prompt-formats/llama3_1). The way we compute WPS and MFU can be found in `train.py`. Because the model now is larger, we run on 128 H100 GPUs to test both performance and loss curves. Below is the performance result of 405B model with optimizations we have developed. We do see OOM for 1D parallelism (FSDP2), so we only tested 2D parallelism (FSDP2 + Tensor Parallel).
+
+| Model size | Batch size | Activation checkpointing | WPS | MFU | Optimizations |
+| ----- | ----- | ----- | ----- | ----- | ----- |
+| 405B | 2 | full | 118 | 37.1% | None
+| 405B | 2 | full | 177 | 27.77% | Float8
+| 405B | 2 | full | 185 | 29.03% | Float8 + Async TP
+
+Here, we use local batch size 2 (global batch size = local batch size 2 * number of FSDP ranks 16 = 32).
+
+Next, we show the loss curves, all models are trained 3000 steps on the [C4 dataset](https://huggingface.co/datasets/allenai/c4), with global batch size 32. We have to use full AC to save memory usage. The results are shown in the picture (a TensorBoard screenshot) below.
+
+![image](../assets/images/llama3_1_405B_loss_curves.png)
+
 ## Llama 3 performance numbers
 
-Below are the WPS (word per second, or more accurately, token per second) and MFU (model FLOPS utilization) results which torchtitan achieves on Llama 3 models with FSDP2 on 64 A100 (80GB) GPUs. The way we compute WPS and MFU can be found in `train.py`.
+Below are the WPS and MFU results which torchtitan achieves on Llama 3 models with FSDP2 on 64 A100 (80GB) GPUs.
 
 | Model size | Batch size | Activation checkpointing | WPS | MFU |
 | ----- | ----- | ----- | ----- | ----- |
@@ -14,7 +30,7 @@ Below are the WPS (word per second, or more accurately, token per second) and MF
 
 We use local batch size 1 (global batch size = local batch size 1 * number of FSDP ranks 64 = 64), because it mimics the small local batch size in large scaled training, and moreoever allows us to compare 1D (FSDP) and 2D (FSDP + TP) training under the same global batch size on both 8B and 70B Llama 3 models, without the out-of-memory (OOM) issue.
 
-Next we show the loss curves for Llama 3 8B and Llama 3 70B training with both 1D parallelism (FSDP2) and 2D parallelism (FSDP2 + Tensor Parallel). All four models are trained 3000 steps on the [C4 dataset](https://huggingface.co/datasets/allenai/c4), with global batch size 64. In terms of activation checkpointing (AC) configs, the Llama 3 8B training jobs use selective op AC, whereas the Llama 3 70B training jobs use full AC. The results are shown in the picture (a TensorBoard screenshot) below.
+Next we show the loss curves for Llama 3 8B and Llama 3 70B training with both 1D parallelism (FSDP2) and 2D parallelism (FSDP2 + Tensor Parallel). All four models are trained the same way as mentioned above with global batch size 64. In terms of activation checkpointing (AC) configs, the Llama 3 8B training jobs use selective op AC, whereas the Llama 3 70B training jobs use full AC. The results are shown in the picture (a TensorBoard screenshot) below.
 
 ![image](../assets/images/llama3_loss_curves.png)
 

torchtitan/utils.py

@@ -15,6 +15,7 @@
 import torch.distributed.distributed_c10d as c10d
 from torch.distributed.device_mesh import DeviceMesh
 from torchtitan.logging import logger
+from torchtitan.metrics import GPUMemoryMonitor
 
 
 def dist_max(x: Union[int, float], mesh: DeviceMesh) -> float:
@@ -134,7 +135,9 @@ def get_num_flop_per_token(num_params: int, model_config, seq_len) -> int:
 
 
 # hardcoded BF16 type peak flops for NVIDIA A100 and H100 GPU
-def get_peak_flops(device_name: str) -> int:
+def get_peak_flops(gpu_memory_monitor: GPUMemoryMonitor) -> int:
+    device_name = gpu_memory_monitor.device_name
+    device_mem = int(gpu_memory_monitor.device_capacity_gib)
     if "A100" in device_name:
         # data from https://www.nvidia.com/en-us/data-center/a100/
         return 312e12
@@ -145,7 +148,13 @@ def get_peak_flops(device_name: str) -> int:
             return 1979e12
         elif "PCIe" in device_name:
             return 756e12
-        else:  # for SXM and other variants
+        # Sometimes, the device name is just literally "NVIDIA H100", so
+        # we won't hit the two conditions before; we then need
+        # to check the memory size of the device to determine the peak flops.
+        # H100 NVL has 94 GiB memory, source: https://www.nvidia.com/en-us/data-center/h100/
+        elif device_mem >= 94:
+            return 835e12
+        else:  # for H100 SXM
             return 989e12
     else:  # for other GPU types, assume A100
         return 312e12

train.py

@@ -71,7 +71,8 @@ def main(job_config: JobConfig):
     utils.init_distributed(job_config)
     # initialize GPU memory monitor and get peak flops for MFU calculation
     gpu_memory_monitor = build_gpu_memory_monitor()
-    gpu_peak_flops = utils.get_peak_flops(gpu_memory_monitor.device_name)
+    gpu_peak_flops = utils.get_peak_flops(gpu_memory_monitor)
+    logger.info(f"Peak FLOPS used for computing MFU: {gpu_peak_flops:.3e}")
 
     # build meshes
     world_mesh = parallel_dims.build_mesh(device_type="cuda")

train_configs/llama3_405b.toml

@@ -34,11 +34,12 @@ steps = 3000
 data_parallel_degree = -1
 tensor_parallel_degree = 8  # 8-way TP
 enable_float8_linear = false
-compile = false
+compile = true
 dataset = "c4"
 
 [experimental]
 pipeline_parallel_degree = 1
+enable_async_tensor_parallel = true
 
 [checkpoint]
 enable_checkpoint = false
@@ -51,3 +52,8 @@ async_mode = "disabled" # ["disabled", "async", "async_with_pinned_mem"]
 
 [activation_checkpoint]
 mode = 'full' # ['none', 'selective', 'full']
+
+[float8]
+enable_float8_linear = true
+enable_fsdp_float8_all_gather = true
+precompute_float8_dynamic_scale_for_fsdp = true