NeMo2.0 llama3 perf scripts

nemo/lightning/run/plugins.py

@@ -314,7 +314,7 @@ class PerfEnvPlugin(run.Plugin):
     enable_layernorm_sm_margin: bool = True
     layernorm_sm_margin: int = 16
     enable_vboost: bool = False
-    nccl_pp_comm_chunksize: int = None
+    nccl_pp_comm_chunksize: Optional[int] = None
 
     def get_vboost_srun_cmd(self, nodes, job_dir):
         "Create the vboost `sudo nvidia-smi boost-slider --vboost 1` command"

scripts/llm/performance/README.md

@@ -0,0 +1,27 @@
+# Performance Recipes
+
+- Scripts defined in `scripts/llm/performance` are recipes optimized for performance. These scripts can launch pre-training experiments on Slurm based clusters.
+- You will need a virtual environemnt with NeMo and Nemo-Run related dependencies installed as the experiment configuration is resolved before launching it inside NeMo container.
+
+## Example
+
+The following line shows an example of how you can launch a pre-training experiment-
+
+`python3 scripts/llm/performance/llama3_8b.py --account <your_slurm_account> -partition <your_slurm_partition>`
+
+## Configuration Options
+
+- Slurm account and partition are mandatory arguments for launching the experiment.
+- You can use the following optional arguments as needed-
+  - -l/--log_dir: Location to store your experiment artifacts and logs. 
+    - Make sure the environemnt variable `NEMORUN_HOME=<log_dir>` is accessible and set correctly in your virtual environment. 
+    - You can run `export NEMORUN_HOME=<log_dir>` in your terminal. You can add it your bashrc file (or equivalent for your OS/Linux distro) for setting it permanently.
+  - -t/--time_limit: Maximum time limit for your experiment. Your slurm job will be cancelled after this. Default is 30 minutes.
+  - -i/--container_image: The NeMo container you want to use. Defaults to latest dev container- 'nvcr.io/nvidia/nemo:dev'.
+  - -c/--compute_dtype: Specifies whether you want to use bf16 or fp8 precision for training. Defaults to 'bf16'. You can choose to use 'fp8'.
+  - -ep/--enable_profiling: Enable nsys profiling. It is disabled by default. When enabled, profiling will be enabled for 1 step from step 5 to step 6. You can change the step in the respective recipe script. 
+  - -tb/--tensorboard: Enable tensorboard logging. It is disabled by default. 
+    - CAUTION: Tensorboard logging may cause performance overhead. 
+  - -d/--dryrun: Using this argument will not launch the experiment. It will simply print the sbatch script to stdout. This can be helpful to verify you have set your experiment correctly as needed.
+- You don't need to set any value for `--enable_profiling`, `--tensorboard` and `--dryrun`. See the below example for reference-
+  `python3 scripts/llm/performance/llama3_8b.py --account <your_slurm_account> -p <your_slurm_partition> -ep --tensorboard -d`

scripts/llm/performance/llama3_405b.py

@@ -0,0 +1,179 @@
+# Copyright (c) 2024, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional
+
+import nemo_run as run
+from nemo_run.config import NEMORUN_HOME
+from utils import get_comm_overlap_callback_idx, hf_tokenizer, parse_cli_args, slurm_executor
+
+from nemo.collections.llm.recipes.llama31_405b import pretrain_recipe
+from nemo.collections.llm.recipes.precision.mixed_precision import bf16_with_fp8_mixed
+from nemo.lightning.pytorch.callbacks.garbage_collection import GarbageCollectionCallback
+from nemo.lightning.run.plugins import NsysPlugin, PerfEnvPlugin
+from nemo.utils import logging
+
+NUM_NODES = 72
+NUM_GPUS_PER_NODE = 8
+MICRO_BATCH_SIZE = 1
+GLOBAL_BATCH_SIZE = 252
+TP_SIZE = 8
+PP_SIZE = 9
+CP_SIZE = 2
+VP_SIZE = 7
+MAX_STEPS = 100
+
+
+def llama3_405b_performance_recipe(
+    compute_dtype: str,
+    num_nodes: int,
+    num_gpus_per_node: int,
+    mbs: int,
+    gbs: int,
+    tp_size: int,
+    pp_size: int,
+    cp_size: int,
+    vp_size: Optional[int],
+    max_steps: int,
+):
+    """
+    llama3 405b pre-train recipe aimed at achieving best possible performance.
+
+    NOTE: Use fp8 precision training with caution. It might not give desirable results.
+    """
+    recipe = pretrain_recipe(performance_mode=True)
+
+    # data module configs
+    recipe.data.micro_batch_size = mbs
+    recipe.data.global_batch_size = gbs
+    recipe.data.num_train_samples = max_steps * gbs  # ensure only 1 epoch for whole run
+    recipe.data.tokenizer = hf_tokenizer("meta-llama/Llama-3.1-405B")
+
+    recipe.trainer.max_steps = max_steps
+    recipe.trainer.num_nodes = num_nodes
+    recipe.trainer.devices = num_gpus_per_node
+
+    # parallelism configs
+    recipe.trainer.strategy.tensor_model_parallel_size = tp_size
+    recipe.trainer.strategy.pipeline_model_parallel_size = pp_size
+    recipe.trainer.strategy.context_parallel_size = cp_size
+    recipe.trainer.strategy.virtual_pipeline_model_parallel_size = vp_size
+    if tp_size > 1:
+        recipe.trainer.strategy.sequence_parallel = True
+    else:
+        recipe.trainer.strategy.sequence_parallel = False
+
+    comm_overlap_callback_idx = get_comm_overlap_callback_idx(recipe.trainer.callbacks)
+
+    # compute dtype configs
+    if compute_dtype.lower() == "fp8":
+        recipe.trainer.plugins = bf16_with_fp8_mixed()
+        recipe.trainer.callbacks[comm_overlap_callback_idx].tp_comm_overlap_cfg.proj_fprop.fp8_buf = True
+        recipe.trainer.callbacks[comm_overlap_callback_idx].tp_comm_overlap_cfg.fc2_fprop.fp8_buf = True
+
+    recipe.trainer.plugins.grad_reduce_in_fp32 = False  # bf16 grad dtype
+
+    # callback configs
+    garbage_collection_callback = run.Config(
+        GarbageCollectionCallback,
+        gc_interval_train=100,
+        gc_interval_val=500,
+    )
+    recipe.trainer.callbacks.extend(
+        [
+            garbage_collection_callback,
+        ]
+    )
+    dp_size = (num_nodes * num_gpus_per_node) / (tp_size * pp_size * cp_size)
+    if dp_size > 1 and pp_size > 1 and vp_size and vp_size > 1:
+        if comm_overlap_callback_idx >= 0:
+            recipe.trainer.callbacks[comm_overlap_callback_idx].overlap_param_gather_with_optimizer_step = True
+
+    # Misc. for overall faster experiment runtime
+    recipe.log.ckpt = None
+    recipe.trainer.enable_checkpointing = False
+    recipe.trainer.val_check_interval = max_steps * gbs / dp_size
+    recipe.trainer.log_every_n_steps = 1
+
+    return recipe
+
+
+if __name__ == "__main__":
+    args = parse_cli_args().parse_args()
+    if args.log_dir != NEMORUN_HOME:
+        import sys
+
+        logging.error(f"Run `export NEMORUN_HOME={args.log_dir}` in your shell environment and rerun this script.")
+        sys.exit(1)
+
+    exp_name = "_".join(
+        [
+            f"llama3_405b",
+            args.compute_dtype,
+            f"{NUM_NODES}nodes",
+            f"tp{TP_SIZE}_pp{PP_SIZE}_cp{CP_SIZE}_vp{VP_SIZE}",
+            f"{MICRO_BATCH_SIZE}mbs_{GLOBAL_BATCH_SIZE}gbs",
+        ]
+    )
+
+    executor = slurm_executor(
+        args.account,
+        args.partition,
+        args.log_dir,
+        NUM_NODES,
+        NUM_GPUS_PER_NODE,
+        args.time_limit,
+        args.container_image,
+        custom_mounts=[],
+        custom_env_vars={},
+        retries=0,
+    )
+
+    recipe = llama3_405b_performance_recipe(
+        args.compute_dtype,
+        NUM_NODES,
+        NUM_GPUS_PER_NODE,
+        MICRO_BATCH_SIZE,
+        GLOBAL_BATCH_SIZE,
+        TP_SIZE,
+        PP_SIZE,
+        CP_SIZE,
+        VP_SIZE,
+        MAX_STEPS,
+    )
+
+    if not args.tensorboard:  # tensorboard adds performance overhead.
+        recipe.log.tensorboard = None
+        recipe.trainer.logger = False
+    else:
+        # default path is NOT intuitive- `<log_dir>/code/nemo_experiments/tb_logs/default/<tfevents_file>`
+        # following line ensures file is at- `<log_dir>/lightning_logs/tb_logs/default/<tfevents_file>`
+        recipe.log.log_dir = "/nemo_run/lightning_logs"
+
+    plugins = [PerfEnvPlugin(enable_vboost=True, nccl_pp_comm_chunksize=2097152)]
+    if args.enable_profiling:
+        plugins.append(NsysPlugin(start_step=5, end_step=6))
+
+    with run.Experiment(exp_name) as exp:
+        exp.add(
+            recipe,
+            executor=executor,
+            name=exp_name,
+            plugins=plugins,
+        )
+
+        if not args.dryrun:
+            exp.run(sequential=True, detach=True)
+        else:
+            exp.dryrun()

scripts/llm/performance/llama3_70b.py

@@ -0,0 +1,179 @@
+# Copyright (c) 2024, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from typing import Optional
+
+import nemo_run as run
+from nemo_run.config import NEMORUN_HOME
+from utils import get_comm_overlap_callback_idx, hf_tokenizer, parse_cli_args, slurm_executor
+
+from nemo.collections.llm.recipes.llama3_70b import pretrain_recipe
+from nemo.collections.llm.recipes.precision.mixed_precision import bf16_with_fp8_mixed
+from nemo.lightning.pytorch.callbacks.garbage_collection import GarbageCollectionCallback
+from nemo.lightning.run.plugins import NsysPlugin, PerfEnvPlugin
+from nemo.utils import logging
+
+NUM_NODES = 8
+NUM_GPUS_PER_NODE = 8
+MICRO_BATCH_SIZE = 1
+GLOBAL_BATCH_SIZE = 128
+TP_SIZE = 4
+PP_SIZE = 4
+CP_SIZE = 2
+VP_SIZE = 5
+MAX_STEPS = 100
+
+
+def llama3_70b_performance_recipe(
+    compute_dtype: str,
+    num_nodes: int,
+    num_gpus_per_node: int,
+    mbs: int,
+    gbs: int,
+    tp_size: int,
+    pp_size: int,
+    cp_size: int,
+    vp_size: Optional[int],
+    max_steps: int,
+):
+    """
+    llama3 70b pre-train recipe aimed at achieving best possible performance.
+
+    NOTE: Use fp8 precision training with caution. It might not give desirable results.
+    """
+    recipe = pretrain_recipe(performance_mode=True)
+
+    # data module configs
+    recipe.data.micro_batch_size = mbs
+    recipe.data.global_batch_size = gbs
+    recipe.data.num_train_samples = max_steps * gbs  # ensure only 1 epoch for whole run
+    recipe.data.tokenizer = hf_tokenizer("meta-llama/Meta-Llama-3-70B")
+
+    recipe.trainer.max_steps = max_steps
+    recipe.trainer.num_nodes = num_nodes
+    recipe.trainer.devices = num_gpus_per_node
+
+    # parallelism configs
+    recipe.trainer.strategy.tensor_model_parallel_size = tp_size
+    recipe.trainer.strategy.pipeline_model_parallel_size = pp_size
+    recipe.trainer.strategy.context_parallel_size = cp_size
+    recipe.trainer.strategy.virtual_pipeline_model_parallel_size = vp_size
+    if tp_size > 1:
+        recipe.trainer.strategy.sequence_parallel = True
+    else:
+        recipe.trainer.strategy.sequence_parallel = False
+
+    comm_overlap_callback_idx = get_comm_overlap_callback_idx(recipe.trainer.callbacks)
+
+    # compute dtype configs
+    if compute_dtype.lower() == "fp8":
+        recipe.trainer.plugins = bf16_with_fp8_mixed()
+        recipe.trainer.callbacks[comm_overlap_callback_idx].tp_comm_overlap_cfg.proj_fprop.fp8_buf = True
+        recipe.trainer.callbacks[comm_overlap_callback_idx].tp_comm_overlap_cfg.fc2_fprop.fp8_buf = True
+
+    recipe.trainer.plugins.grad_reduce_in_fp32 = False  # bf16 grad dtype
+
+    # callback configs
+    garbage_collection_callback = run.Config(
+        GarbageCollectionCallback,
+        gc_interval_train=100,
+        gc_interval_val=500,
+    )
+    recipe.trainer.callbacks.extend(
+        [
+            garbage_collection_callback,
+        ]
+    )
+    dp_size = (num_nodes * num_gpus_per_node) / (tp_size * pp_size * cp_size)
+    if dp_size > 1 and pp_size > 1 and vp_size and vp_size > 1:
+        if comm_overlap_callback_idx >= 0:
+            recipe.trainer.callbacks[comm_overlap_callback_idx].overlap_param_gather_with_optimizer_step = True
+
+    # Misc. for overall faster experiment runtime
+    recipe.log.ckpt = None
+    recipe.trainer.enable_checkpointing = False
+    recipe.trainer.val_check_interval = max_steps * gbs / dp_size
+    recipe.trainer.log_every_n_steps = 1
+
+    return recipe
+
+
+if __name__ == "__main__":
+    args = parse_cli_args().parse_args()
+    if args.log_dir != NEMORUN_HOME:
+        import sys
+
+        logging.error(f"Run `export NEMORUN_HOME={args.log_dir}` in your shell environment and rerun this script.")
+        sys.exit(1)
+
+    exp_name = "_".join(
+        [
+            f"llama3_70b",
+            args.compute_dtype,
+            f"{NUM_NODES}nodes",
+            f"tp{TP_SIZE}_pp{PP_SIZE}_cp{CP_SIZE}_vp{VP_SIZE}",
+            f"{MICRO_BATCH_SIZE}mbs_{GLOBAL_BATCH_SIZE}gbs",
+        ]
+    )
+
+    executor = slurm_executor(
+        args.account,
+        args.partition,
+        args.log_dir,
+        NUM_NODES,
+        NUM_GPUS_PER_NODE,
+        args.time_limit,
+        args.container_image,
+        custom_mounts=[],
+        custom_env_vars={},
+        retries=0,
+    )
+
+    recipe = llama3_70b_performance_recipe(
+        args.compute_dtype,
+        NUM_NODES,
+        NUM_GPUS_PER_NODE,
+        MICRO_BATCH_SIZE,
+        GLOBAL_BATCH_SIZE,
+        TP_SIZE,
+        PP_SIZE,
+        CP_SIZE,
+        VP_SIZE,
+        MAX_STEPS,
+    )
+
+    if not args.tensorboard:  # tensorboard adds performance overhead.
+        recipe.log.tensorboard = None
+        recipe.trainer.logger = False
+    else:
+        # default path is NOT intuitive- `<log_dir>/code/nemo_experiments/tb_logs/default/<tfevents_file>`
+        # following line ensures file is at- `<log_dir>/lightning_logs/tb_logs/default/<tfevents_file>`
+        recipe.log.log_dir = "/nemo_run/lightning_logs"
+
+    plugins = [PerfEnvPlugin(enable_vboost=True, nccl_pp_comm_chunksize=2097152)]
+    if args.enable_profiling:
+        plugins.append(NsysPlugin(start_step=5, end_step=6))
+
+    with run.Experiment(exp_name) as exp:
+        exp.add(
+            recipe,
+            executor=executor,
+            name=exp_name,
+            plugins=plugins,
+        )
+
+        if not args.dryrun:
+            exp.run(sequential=True, detach=True)
+        else:
+            exp.dryrun()