Improve compiler toolkit logging

Summary:
1. When an input is not a DTensor, print out its index and type.
2. Graph dump spams the initialization log, make it to debug level and output the graphs to the dump folder.

ghstack-source-id: fb2239f
Pull-Request: #2028

Author: fegin

torchtitan/experiments/compiler_toolkit/deepseek_v3/parallelize.py

@@ -80,14 +80,17 @@ def parallelize_deepseekv3(
     compiler_passes = get_compiler_passes_from_config(job_config)
 
     # Create compilers with specified passes (defaults to no passes)
-    fw_compiler, bw_compiler = make_compiler_with_passes(compiler_passes)
+    fw_compiler, bw_compiler = make_compiler_with_passes(
+        compiler_passes, dump_folder=job_config.job.dump_folder
+    )
 
     # Create custom joint_graph_builder with deepseekv3-specific compilers
     deepseekv3_joint_graph_builder = functools.partial(
         joint_graph_builder,
         fw_compiler=fw_compiler,
         bw_compiler=bw_compiler,
         joint_custom_pass=validate_flex_attention_annotation,
+        dump_folder=job_config.job.dump_folder,
     )
 
     # TODO: CompiledModule should take sample input as well, so that we can

torchtitan/experiments/compiler_toolkit/graph_utils.py

@@ -5,6 +5,7 @@
 # LICENSE file in the root directory of this source tree.
 
 import contextlib
+from pathlib import Path
 from typing import Callable, List, Optional
 
 import torch
@@ -21,8 +22,18 @@
 from torchtitan.tools.logging import logger
 
 
+def _dump_gm(dump_folder: str | None, gm: torch.fx.GraphModule, name: str) -> None:
+    # TODO: make the dump rank configurable
+    if not dump_folder or torch.distributed.get_rank() != 0:
+        return
+
+    output_path = Path(dump_folder) / "compiler" / f"{name}.txt"
+    output_path.parent.mkdir(parents=True, exist_ok=True)
+    output_path.write_text(gm.print_readable(print_output=False))
+
+
 def export_joint(
-    model, args, kwargs=None
+    model, args, kwargs=None, dump_folder: str | None = None
 ) -> tuple[JointWithDescriptors, TracingContext]:
     if kwargs is None:
         kwargs = {}
@@ -35,8 +46,10 @@ def export_joint(
         torch.fx.traceback.preserve_node_meta(),
     ):
         gm = dynamo_graph_capture_for_export(model)(*args, **kwargs)
-        logger.info("Dynamo gm:")
-        logger.info(gm.print_readable(print_output=False))
+        logger.debug("Dynamo gm:")
+        logger.debug(gm.print_readable(print_output=False))
+        _dump_gm(dump_folder, gm, "dynamo_gm")
+
         tracing_context = gm.meta["tracing_context"]
 
     with tracing(tracing_context):
@@ -68,6 +81,7 @@ def joint_graph_builder(
     fw_compiler: Optional[Callable] = None,
     bw_compiler: Optional[Callable] = None,
     joint_custom_pass: Optional[Callable] = None,
+    dump_folder: str | None = None,
 ):
     """
     Build a joint forward-backward graph for the model with optional custom compilers.
@@ -79,16 +93,17 @@ def joint_graph_builder(
         fw_compiler: Optional custom forward compiler function
         bw_compiler: Optional custom backward compiler function
         joint_custom_pass: Optional custom pass to run on the joint graph
+        dump_folder: Optional folder to dump the graph to
     """
     assert isinstance(model_args, tuple)
-    for arg in model_args:
-        assert isinstance(arg, DTensor)
+    for idx, arg in enumerate(model_args):
+        assert isinstance(arg, DTensor), f"Argument {idx} is of type {type(arg)}"
 
     # get joint graph
     (
         joint_with_descriptors,
         tracing_context,
-    ) = export_joint(model, model_args, model_kwargs)
+    ) = export_joint(model, model_args, model_kwargs, dump_folder=dump_folder)
 
     # Optional validation
     if joint_custom_pass is not None:
@@ -179,6 +194,7 @@ def compiler(
     gm: torch.fx.GraphModule,
     example_inputs,
     passes: List[Callable] = None,
+    dump_folder: str | None = None,
 ):
     """
     Compile a graph module by applying a sequence of compiler passes.
@@ -194,19 +210,23 @@ def compiler(
     if passes is None:
         passes = DEFAULT_COMPILER_PASSES
 
-    logger.info(f"{name} before compiler:")
-    logger.info(gm.print_readable(print_output=False))
+    logger.debug(f"{name} before compiler:")
+    logger.debug(gm.print_readable(print_output=False))
+    _dump_gm(dump_folder, gm, f"{name}_before_compiler")
 
     for pass_fn in passes:
         logger.info(f"Applying pass: {pass_fn.__name__}")
         gm = pass_fn(gm, example_inputs)
 
-    logger.info(f"{name} after compiler:")
-    logger.info(gm.print_readable(print_output=False))
+    logger.debug(f"{name} after compiler:")
+    logger.debug(gm.print_readable(print_output=False))
+    _dump_gm(dump_folder, gm, f"{name}_after_compiler")
     return gm
 
 
-def make_compiler_with_passes(passes: List[Callable] = None):
+def make_compiler_with_passes(
+    passes: List[Callable] = None, dump_folder: str | None = None
+):
     """
     Create forward and backward compilers with specified passes.
 
@@ -218,10 +238,14 @@ def make_compiler_with_passes(passes: List[Callable] = None):
     """
 
     def fw_compiler(gm: torch.fx.GraphModule, example_inputs) -> None:
-        return compiler("fwd_gm", gm, example_inputs, passes=passes)
+        return compiler(
+            "fwd_gm", gm, example_inputs, passes=passes, dump_folder=dump_folder
+        )
 
     def bw_compiler(gm: torch.fx.GraphModule, example_inputs) -> None:
-        return compiler("bwd_gm", gm, example_inputs, passes=passes)
+        return compiler(
+            "bwd_gm", gm, example_inputs, passes=passes, dump_folder=dump_folder
+        )
 
     return fw_compiler, bw_compiler
 

torchtitan/experiments/compiler_toolkit/llama3/parallelize.py

@@ -67,14 +67,17 @@ def parallelize_llama(
     compiler_passes = get_compiler_passes_from_config(job_config)
 
     # Create compilers with specified passes (defaults to no passes)
-    fw_compiler, bw_compiler = make_compiler_with_passes(compiler_passes)
+    fw_compiler, bw_compiler = make_compiler_with_passes(
+        compiler_passes, dump_folder=job_config.job.dump_folder
+    )
 
     # Create custom joint_graph_builder with llama-specific compilers and validation
     llama_joint_graph_builder = functools.partial(
         joint_graph_builder,
         fw_compiler=fw_compiler,
         bw_compiler=bw_compiler,
         joint_custom_pass=validate_flex_attention_annotation,
+        dump_folder=job_config.job.dump_folder,
     )
 
     # TODO: CompiledModule should take sample input as well, so that we can