trainer subclass for compiler toolkit

Author: yiming0416

.github/workflows/integration_test_8gpu_compiler_toolkit.yaml

@@ -50,4 +50,4 @@ jobs:
         python -m pip install --force-reinstall --pre torch --index-url https://download.pytorch.org/whl/nightly/cu126
 
         mkdir artifacts-to-be-uploaded
-        python -m torchtitan.experiments.compiler_toolkit.tests.integration_tests artifacts-to-be-uploaded --ngpu 4
+        TRAIN_FILE=torchtitan.experiments.compiler_toolkit.train python -m torchtitan.experiments.compiler_toolkit.tests.integration_tests artifacts-to-be-uploaded --ngpu 4

torchtitan/experiments/compiler_toolkit/README.md

@@ -14,44 +14,44 @@ Joint Graph based Training Prototype:
 
 **SimpleFSDP + TP + EP**
 ```shell
-NGPU=4 CONFIG_FILE=./torchtitan/models/deepseek_v3/train_configs/debug_model.toml ./run_train.sh --model.name compiler_toolkit.deepseek_v3 --parallelism.data_parallel_shard_degree=2 --parallelism.tensor_parallel_degree=2 --parallelism.expert_parallel_degree=2 --activation_checkpoint.mode none
+NGPU=4 TRAIN_FILE=torchtitan.experiments.compiler_toolkit.train CONFIG_FILE=./torchtitan/models/deepseek_v3/train_configs/debug_model.toml ./run_train.sh --model.name compiler_toolkit.deepseek_v3 --parallelism.data_parallel_shard_degree=2 --parallelism.tensor_parallel_degree=2 --parallelism.expert_parallel_degree=2 --activation_checkpoint.mode none
 ```
 
 **SimpleFSDP + TP + EP + FlexAttention**
 ```shell
-NGPU=4 CONFIG_FILE=./torchtitan/models/deepseek_v3/train_configs/debug_model.toml ./run_train.sh --model.name compiler_toolkit.deepseek_v3 --parallelism.data_parallel_shard_degree=2 --parallelism.tensor_parallel_degree=2 --parallelism.expert_parallel_degree=2 --activation_checkpoint.mode none --model.flavor=debugmodel_flex_attn
+NGPU=4 TRAIN_FILE=torchtitan.experiments.compiler_toolkit.train CONFIG_FILE=./torchtitan/models/deepseek_v3/train_configs/debug_model.toml ./run_train.sh --model.name compiler_toolkit.deepseek_v3 --parallelism.data_parallel_shard_degree=2 --parallelism.tensor_parallel_degree=2 --parallelism.expert_parallel_degree=2 --activation_checkpoint.mode none --model.flavor=debugmodel_flex_attn
 ```
 
 ## llama3
 
 **SimpleFSDP + TP**
 ```shell
-NGPU=8 CONFIG_FILE=./torchtitan/models/llama3/train_configs/debug_model.toml ./run_train.sh --model.name compiler_toolkit.llama3 --parallelism.data_parallel_shard_degree=2 --parallelism.tensor_parallel_degree=4
+NGPU=8 TRAIN_FILE=torchtitan.experiments.compiler_toolkit.train CONFIG_FILE=./torchtitan/models/llama3/train_configs/debug_model.toml ./run_train.sh --model.name compiler_toolkit.llama3 --parallelism.data_parallel_shard_degree=2 --parallelism.tensor_parallel_degree=4
 ```
 
 **SimpleFSDP + TP + auto-bucketing**
 ```shell
-NGPU=8 CONFIG_FILE=./torchtitan/models/llama3/train_configs/debug_model.toml ./run_train.sh --model.name compiler_toolkit.llama3 --parallelism.data_parallel_shard_degree=2 --parallelism.tensor_parallel_degree=4 --job.custom_config_module=torchtitan.experiments.compiler_toolkit.job_config --compile.passes autobucketing_reordering
+NGPU=8 TRAIN_FILE=torchtitan.experiments.compiler_toolkit.train CONFIG_FILE=./torchtitan/models/llama3/train_configs/debug_model.toml ./run_train.sh --model.name compiler_toolkit.llama3 --parallelism.data_parallel_shard_degree=2 --parallelism.tensor_parallel_degree=4 --job.custom_config_module=torchtitan.experiments.compiler_toolkit.job_config --compile.passes autobucketing_reordering
 ```
 
 **SimpleFSDP + TP + transformer-block-bucketing**
 ```shell
-NGPU=8 CONFIG_FILE=./torchtitan/models/llama3/train_configs/debug_model.toml ./run_train.sh --model.name compiler_toolkit.llama3 --parallelism.data_parallel_shard_degree=2 --parallelism.tensor_parallel_degree=4 --job.custom_config_module=torchtitan.experiments.compiler_toolkit.job_config --compile.passes transformer_block_bucketing
+NGPU=8 TRAIN_FILE=torchtitan.experiments.compiler_toolkit.train CONFIG_FILE=./torchtitan/models/llama3/train_configs/debug_model.toml ./run_train.sh --model.name compiler_toolkit.llama3 --parallelism.data_parallel_shard_degree=2 --parallelism.tensor_parallel_degree=4 --job.custom_config_module=torchtitan.experiments.compiler_toolkit.job_config --compile.passes transformer_block_bucketing
 ```
 
 **SimpleFSDP + TP + FlexAttention**
 ```shell
-NGPU=8 CONFIG_FILE=./torchtitan/models/llama3/train_configs/debug_model.toml ./run_train.sh --model.name compiler_toolkit.llama3 --parallelism.data_parallel_shard_degree=2 --parallelism.tensor_parallel_degree=4 --model.flavor=debugmodel_flex_attn
+NGPU=8 TRAIN_FILE=torchtitan.experiments.compiler_toolkit.train CONFIG_FILE=./torchtitan/models/llama3/train_configs/debug_model.toml ./run_train.sh --model.name compiler_toolkit.llama3 --parallelism.data_parallel_shard_degree=2 --parallelism.tensor_parallel_degree=4 --model.flavor=debugmodel_flex_attn
 ```
 
 **SimpleFSDP + TP + FlexAttention + auto-bucketing + regional-inductor**
 
 ```shell
-NGPU=8 CONFIG_FILE=./torchtitan/models/llama3/train_configs/debug_model.toml ./run_train.sh --model.name compiler_toolkit.llama3 --parallelism.data_parallel_shard_degree=2 --parallelism.tensor_parallel_degree=4 --job.custom_config_module=torchtitan.experiments.compiler_toolkit.job_config --compile.passes autobucketing_reordering,regional_inductor
+NGPU=8 TRAIN_FILE=torchtitan.experiments.compiler_toolkit.train CONFIG_FILE=./torchtitan/models/llama3/train_configs/debug_model.toml ./run_train.sh --model.name compiler_toolkit.llama3 --parallelism.data_parallel_shard_degree=2 --parallelism.tensor_parallel_degree=4 --job.custom_config_module=torchtitan.experiments.compiler_toolkit.job_config --compile.passes autobucketing_reordering,regional_inductor
 ```
 
 **SimpleFSDP + TP + FlexAttention + transformer-block-bucketing + regional-inductor**
 
 ```shell
-NGPU=8 CONFIG_FILE=./torchtitan/models/llama3/train_configs/debug_model.toml ./run_train.sh --model.name compiler_toolkit.llama3 --parallelism.data_parallel_shard_degree=2 --parallelism.tensor_parallel_degree=4 --job.custom_config_module=torchtitan.experiments.compiler_toolkit.job_config --compile.passes transformer_block_bucketing,regional_inductor
+NGPU=8 TRAIN_FILE=torchtitan.experiments.compiler_toolkit.train CONFIG_FILE=./torchtitan/models/llama3/train_configs/debug_model.toml ./run_train.sh --model.name compiler_toolkit.llama3 --parallelism.data_parallel_shard_degree=2 --parallelism.tensor_parallel_degree=4 --job.custom_config_module=torchtitan.experiments.compiler_toolkit.job_config --compile.passes transformer_block_bucketing,regional_inductor
 ```

torchtitan/experiments/compiler_toolkit/train.py

@@ -0,0 +1,28 @@
+# 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.
+
+import gc
+
+from torchtitan.train import main, Trainer
+
+
+class CompilerToolkitTrainer(Trainer):
+    def close(self) -> None:
+        super().close()
+
+        # Note [explicit cudagraph close]
+        # cudagraph holds reference to nccl which prevents destroy nccl
+        # group. so we need to explicitly delete cudagraph which is held
+        # in joint_graph_module. An explicit gc.collect() is necessary
+        # to clean up reference cycles.
+        for part in self.model_parts:
+            if hasattr(part, "joint_graph_module"):
+                part.joint_graph_module = None
+        gc.collect()
+
+
+if __name__ == "__main__":
+    main(CompilerToolkitTrainer)