Update on "adapt CI tests to use compiled_rmsnorm"

[ghstack-poisoned]

Author: tianyu-l

.github/workflows/integration_test_4gpu.yaml

@@ -39,5 +39,6 @@ jobs:
 
         python -m pip install --force-reinstall --pre torch --index-url https://download.pytorch.org/whl/nightly/cu121
         python -m pip install --pre torchdata --index-url https://download.pytorch.org/whl/nightly/
+        USE_CPP=0 python -m pip install git+https://github.com/pytorch/ao.git
         mkdir artifacts-to-be-uploaded
         python ./test_runner.py artifacts-to-be-uploaded --ngpu 4

README.md

@@ -18,6 +18,14 @@ Our guiding principles when building `torchtitan`:
 
 [![Welcome to torchtitan!](assets/images/titan_play_video.png)](https://youtu.be/ee5DOEqD35I?si=_B94PbVv0V5ZnNKE "Welcome to torchtitan!")
 
+### Dive into the code
+
+You may want to see how the model is defined or how parallelism techniques are applied. For a guided tour, see these files first:
+* [train.py](https://github.com/pytorch/torchtitan/blob/main/train.py) - the main training loop and high-level setup code
+* [torchtitan/parallelisms/parallelize_llama.py](https://github.com/pytorch/torchtitan/blob/main/torchtitan/parallelisms/parallelize_llama.py) - helpers for applying Data / Tensor / Pipeline Parallelisms to the model
+* [torchtitan/checkpoint.py](https://github.com/pytorch/torchtitan/blob/main/torchtitan/checkpoint.py) - utils for saving/loading distributed checkpoints
+* [torchtitan/models/llama/model.py](https://github.com/pytorch/torchtitan/blob/main/torchtitan/models/llama/model.py) - the Llama model definition (shared for Llama2 and Llama3 variants)
+
 ## Pre-Release Updates:
 #### (4/25/2024): `torchtitan` is now public but in a pre-release state and under development.
 Currently we showcase pre-training **Llama 3 and Llama 2** LLMs of various sizes from scratch. `torchtitan` is tested and verified with the PyTorch nightly version `torch-2.4.0.dev20240412`. (We recommend latest PyTorch nightly).
@@ -66,7 +74,7 @@ Once you have confirmed access, you can run the following command to download th
 ```bash
 # Get your HF token from https://huggingface.co/settings/tokens
 
-# llama3 tokenizer.model
+# llama3 or 3.1 tokenizer.model
 python torchtitan/datasets/download_tokenizer.py --repo_id meta-llama/Meta-Llama-3-8B --tokenizer_path "original" --hf_token=...
 
 # llama2 tokenizer.model

create_seed_checkpoint.sh

@@ -18,8 +18,6 @@
 
 set -ex
 
-export USE_LIBUV=1
-TRAINER_DIR=${1:-/home/$USER/local/torchtitan}
 NGPU=1
 LOG_RANK=0
 CONFIG_FILE=${CONFIG_FILE:-"./train_configs/debug_model.toml"}

estimation.py

@@ -9,22 +9,19 @@
 import os
 
 import torch
-import torch.nn.functional as F
 from torch._guards import active_fake_mode
 from torch._subclasses.fake_tensor import FakeTensorMode
-from torch.distributed import destroy_process_group
 from torch.distributed._tools.fsdp2_mem_tracker import FSDPMemTracker
-from torch.distributed.tensor.parallel import loss_parallel
 from torch.testing._internal.distributed.fake_pg import FakeStore
 
 from torchtitan.config_manager import JobConfig
-from torchtitan.datasets import create_tokenizer
-from torchtitan.float8_linear import build_fp8_linear
-from torchtitan.logging_utils import init_logger, logger
-from torchtitan.lr_scheduling import get_lr_schedulers
+from torchtitan.datasets import build_tokenizer
+from torchtitan.float8_linear import Float8Handler
+from torchtitan.logging import init_logger, logger
 from torchtitan.models import model_name_to_cls, model_name_to_tokenizer, models_config
+from torchtitan.optimizer import build_lr_schedulers, build_optimizers
 from torchtitan.parallelisms import models_parallelize_fns, ParallelDims
-from train import build_optimizers
+from train import get_train_context
 
 
 def estimate_memory(job_config: JobConfig):
@@ -61,16 +58,18 @@ def estimate_memory(job_config: JobConfig):
         logger.info("Compiled RMSNorm is not supported yet. Switching to RMSNorm.")
         job_config.model.norm_type = "rmsnorm"
 
-    if job_config.training.compile:
+    if job_config.training.compile or job_config.experimental.enable_compiled_autograd:
         logger.info("Compile mode is not supported yet. Switching to eager mode.")
         job_config.training.compile = False
+        job_config.experimental.enable_compiled_autograd = False
 
     parallel_dims = ParallelDims(
         dp=job_config.training.data_parallel_degree,
         tp=job_config.training.tensor_parallel_degree,
         pp=job_config.experimental.pipeline_parallel_degree,
         world_size=world_size,
         enable_loss_parallel=job_config.training.enable_loss_parallel,
+        dp_type=job_config.training.data_parallel_type,
     )
 
     device = torch.device(f"cuda:{int(os.environ['LOCAL_RANK'])}")
@@ -93,16 +92,18 @@ def estimate_memory(job_config: JobConfig):
 
     # build tokenizer
     tokenizer_type = model_name_to_tokenizer[model_name]
-    tokenizer = create_tokenizer(tokenizer_type, job_config.model.tokenizer_path)
+    tokenizer = build_tokenizer(tokenizer_type, job_config.model.tokenizer_path)
 
-    # loss_parallel enables dispatching to efficient loss operators
-    loss_parallel_ctx = (
-        loss_parallel if parallel_dims.loss_parallel_enabled else contextlib.nullcontext
+    train_context = get_train_context(
+        parallel_dims.loss_parallel_enabled,
+        job_config.experimental.enable_compiled_autograd,
     )
 
     # loss fn can be shared by pipeline-parallel or non-pp execution
     def loss_fn(pred, labels):
-        return F.cross_entropy(pred.flatten(0, 1), labels.flatten(0, 1))
+        return torch.nn.functional.cross_entropy(
+            pred.flatten(0, 1), labels.flatten(0, 1)
+        )
 
     # build model (using meta init)
     model_cls = model_name_to_cls[model_name]
@@ -123,9 +124,10 @@ def loss_fn(pred, labels):
         with torch.device("meta"):
             whole_model = model_cls.from_model_args(model_config)
 
-        # apply fp8 linear module swap
-        if job_config.training.fp8_linear:
-            build_fp8_linear(whole_model, job_config)
+        # a no-op hander if fp8 is not enabled
+        float8_handler = Float8Handler(job_config, parallel_dims)
+        # swap to Float8Linear base on fp8 config
+        float8_handler.convert_to_float8_training(whole_model)
 
         # apply PT-D DP/TP parallelisms and activation checkpointing
         model_parts = [whole_model]
@@ -143,7 +145,7 @@ def loss_fn(pred, labels):
 
         # build optimizer after applying parallelisms to the model
         optimizers = build_optimizers(model_parts, job_config)
-        lr_schedulers = get_lr_schedulers(optimizers.optimizers, job_config)
+        lr_schedulers = build_lr_schedulers(optimizers.optimizers, job_config)
 
         for model in model_parts:
             model.train()
@@ -170,7 +172,7 @@ def loss_fn(pred, labels):
             for iter_idx in range(2):
                 input_ids, labels = batch
                 # train step
-                with loss_parallel_ctx():
+                with train_context():
                     pred = whole_model(input_ids)
                     loss = loss_fn(pred, labels)
                     del pred
@@ -181,9 +183,14 @@ def loss_fn(pred, labels):
                     torch.nn.utils.clip_grad_norm_(
                         model.parameters(), job_config.training.max_norm, foreach=True
                     )
+                # sync float8 amaxes and scales
+                float8_handler.sync_float8_amax_and_scale_history(model)
                 # optimizer step
                 optimizers.step()
                 lr_schedulers.step()
+                # calculate float8 dynamic amax/scale for all-parameter for FSDP2
+                # it issues a single all-reduce for all parameters at once for better performance
+                float8_handler.precompute_fp8_dynamic_scale_for_fsdp(model)
                 optimizers.zero_grad()
                 print(f"Peak Memory at iter: {iter_idx}")
                 fsdp_memtracker.display_snapshot("peak", units="MiB", tabulate=True)
@@ -217,4 +224,4 @@ def loss_fn(pred, labels):
     try:
         estimate_memory(config)
     finally:
-        destroy_process_group()
+        torch.distributed.destroy_process_group()

multinode_trainer.slurm

@@ -53,7 +53,6 @@ export NCCL_SOCKET_IFNAME="eth0,en,eth,em,bond"
 export NCCL_BUFFSIZE=2097152
 #export TORCH_DIST_INIT_BARRIER=1
 export FI_EFA_SET_CUDA_SYNC_MEMOPS=0
-#export USE_LIBUV=1
 CONFIG_FILE=${CONFIG_FILE:-"./train_configs/llama2_13b.toml"}
 
 dcgmi profile --pause

run_llama_train.sh

@@ -7,39 +7,18 @@
 
 set -ex
 
-# libUV is a scalable backend for TCPStore which is used in processGroup
-# rendezvous. This is the recommended backend for distributed training.
-export USE_LIBUV=1
-TRAINER_DIR=${TRAINER_DIR:-/home/$USER/local/torchtitan}
-
 # use envs as local overrides for convenience
 # e.g.
 # LOG_RANK=0,1 NGPU=4 ./run_llama_train.sh
-
 NGPU=${NGPU:-"8"}
-NNODES=${NNODES:-"1"}
-
-# by default log just rank 0 output,
 LOG_RANK=${LOG_RANK:-0}
-
-
 CONFIG_FILE=${CONFIG_FILE:-"./train_configs/debug_model.toml"}
 
 overrides=""
 if [ $# -ne 0 ]; then
     overrides="$*"
 fi
 
-# Check if --estimate.memory=True is in the arguments
-if echo "$overrides" | grep -q -- "--memory_estimation.enabled"; then
-    # Calculate WORLD_SIZE as the product of NGPU and NNODES
-    # Export WORLD_SIZE and LOCAL_RANK
-    export WORLD_SIZE=$((NGPU * NNODES))
-    export LOCAL_RANK=0
-    python estimation.py --job.config_file ${CONFIG_FILE} $overrides
-else
-    # Call train.py if not in estimation mode
-    torchrun --nproc_per_node=${NGPU} --rdzv_backend c10d --rdzv_endpoint="localhost:0" \
-    --local-ranks-filter ${LOG_RANK} --role rank --tee 3 \
-    train.py --job.config_file ${CONFIG_FILE} $overrides
-fi
+torchrun --nproc_per_node=${NGPU} --rdzv_backend c10d --rdzv_endpoint="localhost:0" \
+--local-ranks-filter ${LOG_RANK} --role rank --tee 3 \
+train.py --job.config_file ${CONFIG_FILE} $overrides

run_memory_estimation.sh

@@ -0,0 +1,26 @@
+#!/usr/bin/bash
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the BSD-style license found in the
+# LICENSE file in the root directory of this source tree.
+
+set -ex
+
+# use envs as local overrides for convenience
+# e.g.
+# NGPU=4 ./run_memory_estimation.sh
+NGPU=${NGPU:-"8"}
+NNODES=${NNODES:-"1"}
+CONFIG_FILE=${CONFIG_FILE:-"./train_configs/debug_model.toml"}
+
+overrides=""
+if [ $# -ne 0 ]; then
+    overrides="$*"
+fi
+
+# Calculate WORLD_SIZE as the product of NGPU and NNODES
+# Export WORLD_SIZE and LOCAL_RANK
+export WORLD_SIZE=$((NGPU * NNODES))
+export LOCAL_RANK=0
+python estimation.py --job.config_file ${CONFIG_FILE} --memory_estimation.enabled $overrides

test/datasets/test_checkpoint.py

@@ -6,7 +6,7 @@
 
 import torch
 from torchtitan.datasets.hf_datasets import build_hf_data_loader
-from torchtitan.datasets.tokenizer import create_tokenizer
+from torchtitan.datasets.tokenizer import build_tokenizer
 
 
 class TestCheckpoint:
@@ -42,7 +42,7 @@ def _build_dataloader(
         self, dataset_name, dataset_path, batch_size, seq_len, world_size, rank
     ):
         tokenizer_type = "tiktoken"
-        tokenizer = create_tokenizer("tiktoken", "./test/assets/test_tiktoken.model")
+        tokenizer = build_tokenizer("tiktoken", "./test/assets/test_tiktoken.model")
         return build_hf_data_loader(
             dataset_name=dataset_name,
             dataset_path=dataset_path,

test_runner.py

@@ -46,6 +46,21 @@ def build_test_list():
     """
     integration_tests_flavors = defaultdict(list)
     integration_tests_flavors["debug_model.toml"] = [
+        OverrideDefinitions(
+            [
+                [
+                    "--checkpoint.enable_checkpoint",
+                    "--experimental.pipeline_parallel_degree 4",
+                    "--experimental.pipeline_parallel_split_points layers.1,layers.2,layers.3,layers.4,layers.5,layers.6,layers.7",
+                    "--experimental.pipeline_parallel_schedule flexible_interleaved_1f1b",
+                    "--model.norm_type rmsnorm",  # fused_rmsnorm throws cuda context error with pp
+                ],
+            ],
+            "PP looped flexible 1f1b test",
+            "pp_looped_flexible_1f1b",
+            requires_seed_checkpoint=True,
+            ngpu=4,
+        ),
         OverrideDefinitions(
             [
                 [
@@ -253,7 +268,7 @@ def build_test_list():
                     "--experimental.pipeline_parallel_degree 4",
                     "--experimental.pipeline_parallel_split_points layers.1,layers.2,layers.3,layers.4,layers.5,layers.6,layers.7",
                     "--experimental.pipeline_parallel_schedule interleaved_1f1b",
-                    "--model.norm_type rmsnorm",  # compiled_rmsnorm / fused_rmsnorm throws cuda context error with pp
+                    "--model.norm_type rmsnorm",  # compiled_rmsnorm / fused_rmsnorm crashes with PP
                 ],
             ],
             "PP looped 1f1b test",
@@ -281,6 +296,16 @@ def build_test_list():
             "fsdp2_mem_tracker",
             ngpu=4,
         ),
+        OverrideDefinitions(
+            [
+                [
+                    "--training.data_parallel_type ddp",
+                ]
+            ],
+            "DDP",
+            "ddp",
+            ngpu=4,
+        ),
     ]
     return integration_tests_flavors
 
@@ -312,6 +337,8 @@ def run_test(test_flavor: OverrideDefinitions, full_path: str, output_dir: str):
 
     for override_arg in test_flavor.override_args:
         cmd = f"CONFIG_FILE={full_path} NGPU={test_flavor.ngpu} LOG_RANK={all_ranks} ./run_llama_train.sh"
+        if test_name == "fsdp2_mem_tracker":
+            cmd = f"CONFIG_FILE={full_path} NGPU={test_flavor.ngpu} LOG_RANK={all_ranks} ./run_memory_estimation.sh"
         cmd += " " + dump_folder_arg
         cmd += " " + model_flavor_arg
         if override_arg: