cherrypick #3387

Author: zewenli98

examples/distributed_inference/llama3_model.py

@@ -2,6 +2,7 @@
 
 import tensorrt as trt
 import torch
+import torch.distributed as dist
 import torch.nn as nn
 import torch_tensorrt
 from tensor_parallel_initialize_dist import initialize_distributed_env
@@ -15,7 +16,6 @@
 device_mesh, _world_size, _rank, logger = initialize_distributed_env(
     "./tensor_parallel_simple_example"
 )
-import tensorrt_llm
 
 """
 This example copies some code from https://github.com/pytorch/examples/blob/main/distributed/tensor_parallelism/tensor_parallel_example.py
@@ -65,7 +65,6 @@ def forward(self, x):
 inp = torch.rand(20, 10, device="cuda")
 python_result = tp_model(inp)
 
-
 backend = "torch_tensorrt"
 tp_model = torch.compile(
     tp_model,
@@ -75,23 +74,28 @@ def forward(self, x):
         "enabled_precisions": {torch.float32, torch.float16},
         "use_python_runtime": True,
         "min_block_size": 1,
-        "use_aot_joint_export": False,
+        "use_distributed_mode_trace": True,
     },
-    dynamic=False,
+    dynamic=None,
 )
 
-for i in range(10):
-    # For TP, input needs to be same across all TP ranks.
-    # Setting the random seed is to mimic the behavior of dataloader.
-    torch.manual_seed(i)
-    inp = torch.rand(20, 10, device="cuda")
-    start = time.time()
-    output = tp_model(inp)
-    end = time.time()
-    if i == 0:
-        logger.info(f"Compilation time is {end-start}")
-        assert (
-            python_result - output
-        ).std() < 0.01, "Compilation result is not correct."
-    elif _rank == 0:
-        logger.info(f"Inference time is {end-start}")
+try:
+    for i in range(10):
+        # For TP, input needs to be same across all TP ranks.
+        # Setting the random seed is to mimic the behavior of dataloader.
+        torch.manual_seed(i)
+        inp = torch.rand(20, 10, device="cuda")
+        start = time.time()
+        output = tp_model(inp)
+        end = time.time()
+        if i == 0:
+            logger.info(f"Compilation time is {end-start}")
+            assert (
+                python_result - output
+            ).std() < 0.01, "Compilation result is not correct."
+        elif _rank == 0:
+            logger.info(f"Inference time is {end-start}")
+finally:
+    # This cleans up the distributed process group
+    if dist.is_initialized():
+        dist.destroy_process_group()

examples/distributed_inference/tensor_parallel_llama3.py

@@ -46,9 +46,9 @@
 IMMUTABLE_WEIGHTS = True
 ENABLE_WEIGHT_STREAMING = False
 ENABLE_CROSS_COMPILE_FOR_WINDOWS = False
-USE_AOT_JOINT_EXPORT = True
 TILING_OPTIMIZATION_LEVEL = "none"
 L2_LIMIT_FOR_TILING = -1
+USE_DISTRIBUTED_MODE_TRACE = False
 
 
 def default_device() -> Device:

examples/distributed_inference/tensor_parallel_simple_example.py

@@ -35,7 +35,7 @@
     TILING_OPTIMIZATION_LEVEL,
     TIMING_CACHE_PATH,
     TRUNCATE_DOUBLE,
-    USE_AOT_JOINT_EXPORT,
+    USE_DISTRIBUTED_MODE_TRACE,
     USE_EXPLICIT_TYPING,
     USE_FAST_PARTITIONER,
     USE_FP32_ACC,
@@ -94,9 +94,9 @@ class CompilationSettings:
         enable_weight_streaming (bool): Enable weight streaming.
         enable_cross_compile_for_windows (bool): By default this is False means TensorRT engines can only be executed on the same platform where they were built.
             True will enable cross-platform compatibility which allows the engine to be built on Linux and run on Windows
-        use_aot_joint_export (bool): Use aot_export_joint_simple, else wrap backend with AOT_autograd, required for distributed tensors
         tiling_optimization_level (str): The optimization level of tiling strategies. A higher level allows TensorRT to spend more time searching for better tiling strategy. We currently support ["none", "fast", "moderate", "full"].
         l2_limit_for_tiling (int): The target L2 cache usage limit (in bytes) for tiling optimization (default is -1 which means no limit).
+        use_distributed_mode_trace (bool): Using aot_autograd to trace the graph. This is enabled when DTensors or distributed tensors are present in distributed model
     """
 
     enabled_precisions: Set[dtype] = field(default_factory=lambda: ENABLED_PRECISIONS)
@@ -137,9 +137,9 @@ class CompilationSettings:
     immutable_weights: bool = IMMUTABLE_WEIGHTS
     enable_weight_streaming: bool = ENABLE_WEIGHT_STREAMING
     enable_cross_compile_for_windows: bool = ENABLE_CROSS_COMPILE_FOR_WINDOWS
-    use_aot_joint_export: bool = USE_AOT_JOINT_EXPORT
     tiling_optimization_level: str = TILING_OPTIMIZATION_LEVEL
     l2_limit_for_tiling: int = L2_LIMIT_FOR_TILING
+    use_distributed_mode_trace: bool = USE_DISTRIBUTED_MODE_TRACE
 
 
 _SETTINGS_TO_BE_ENGINE_INVARIANT = (

py/torch_tensorrt/dynamo/_defaults.py

@@ -10,11 +10,11 @@
 from torch._dynamo.backends.common import aot_autograd
 from torch._dynamo.utils import detect_fake_mode
 from torch._functorch.aot_autograd import aot_export_joint_simple
+from torch.distributed.tensor import DTensor
 from torch_tensorrt.dynamo import CompilationSettings
 from torch_tensorrt.dynamo._compiler import compile_module
 from torch_tensorrt.dynamo.lowering import (
     get_decompositions,
-    modify_reshape_complex_nodes,
     post_lowering,
     remove_detach,
     remove_sym_nodes,
@@ -52,25 +52,39 @@ def aot_torch_tensorrt_aten_backend(
     gm: torch.fx.GraphModule, sample_inputs: Sequence[Any], **kwargs: Any
 ) -> torch.nn.Module:
     settings, engine_cache = parse_dynamo_kwargs(kwargs)
-    if settings.use_aot_joint_export:
-        return _pretraced_backend(gm, sample_inputs, settings, engine_cache)
-    logger.debug("Wrapping the backend with aot_autograd\n")
-    _pretraced_backend_autograd = functools.partial(
-        _pretraced_backend, settings=settings, engine_cache=engine_cache
-    )
-    settings_aot_autograd = {}
-    settings_aot_autograd["decompostions"] = get_decompositions(
-        settings.enable_experimental_decompositions
-    )
-    # This is added since detach lowering leads to alias nodes
-    # Error - View operation returned a tensor that is the same as the input base tensor
-    # torch nop_decompositions in torch/_decomp/decompositions.py
-    if aten.detach in settings_aot_autograd["decompositions"]:
-        del settings_aot_autograd["decompositions"][aten.detach]
-    return aot_autograd(
-        fw_compiler=_pretraced_backend_autograd,
-        decompositions=get_decompositions(settings.enable_experimental_decompositions),
-    )(gm, sample_inputs)
+
+    if settings.use_distributed_mode_trace:
+        logger.debug(
+            "Wrapping the backend with aot_autograd for Distributed examples\n"
+        )
+        _pretraced_backend_autograd = functools.partial(
+            _pretraced_backend, settings=settings, engine_cache=engine_cache
+        )
+        settings_aot_autograd = {}
+        settings_aot_autograd["decompositions"] = get_decompositions(
+            settings.enable_experimental_decompositions
+        )
+        # This is added since detach lowering leads to alias nodes
+        # Error - View operation returned a tensor that is the same as the input base tensor
+        # torch nop_decompositions in torch/_decomp/decompositions.py
+        # transpose key deleted since not desirable to lower it to permute
+        to_delete = {
+            key
+            for key in settings_aot_autograd["decompositions"]
+            if "detach" in key._name
+        }
+        for key in to_delete:
+            del settings_aot_autograd["decompositions"][key]
+
+        return aot_autograd(
+            fw_compiler=_pretraced_backend_autograd,
+            decompositions=settings_aot_autograd["decompositions"],
+        )(gm, sample_inputs)
+    if any(isinstance(tensor, DTensor) for tensor in sample_inputs):
+        logger.warning(
+            "It is recommended to run the model with use_distributed_mode_trace = True since there are distributed tensors in the input which is not supported in aot_export_joint_simple"
+        )
+    return _pretraced_backend(gm, sample_inputs, settings, engine_cache)
 
 
 def _pretraced_backend(
@@ -110,18 +124,8 @@ def _pretraced_backend(
             # Remove detach nodes
             remove_detach(gm, settings)
 
-            complexInputIndices = []
-            for i, torch_input in enumerate(torch_inputs):
-                if torch_inputs[i].dtype == torch.complex64:
-                    complexInputIndices.append(i)
-                    torch_input_real = torch_inputs[i].real
-                    torch_input_imaginary = torch_inputs[i].imag
-                    torch_inputs[i] = torch.stack(
-                        (torch_input_real, torch_input_imaginary), dim=-1
-                    )
-
             # Invoke AOTAutograd to translate operators to aten
-            if settings.use_aot_joint_export:
+            if not settings.use_distributed_mode_trace:
                 gm = aot_export_joint_simple(
                     gm,
                     sample_inputs,
@@ -137,12 +141,6 @@ def _pretraced_backend(
 
             logger.debug("Lowered Input graph:\n " + str(gm.graph))
 
-            if complexInputIndices:
-                modify_reshape_complex_nodes(gm, complexInputIndices)
-                logger.debug(
-                    "Input graph after modifying complex nodes:\n " + str(gm.graph)
-                )
-
             torchtrt_inputs = prepare_inputs(
                 torch_inputs, disable_memory_format_check=True
             )

py/torch_tensorrt/dynamo/_settings.py

@@ -3,6 +3,7 @@
 import logging
 from typing import Dict, Sequence, Tuple, Union
 
+import tensorrt as trt
 from torch.fx.node import Argument, Target
 from torch_tensorrt.dynamo._SourceIR import SourceIR
 from torch_tensorrt.dynamo.conversion import impl
@@ -16,8 +17,6 @@
     tensorrt_fused_nccl_reduce_scatter_op,
 )
 
-import tensorrt as trt
-
 _LOGGER: logging.Logger = logging.getLogger(__name__)
 
 if load_tensorrt_llm():
@@ -30,7 +29,7 @@ def fused_nccl_gather(
         kwargs: Dict[str, Argument],
         name: str,
     ) -> Union[trt.ITensor, Sequence[trt.ITensor]]:
-        return impl.distributed.nccl_gather(
+        return impl.nccl_ops.nccl_gather(
             ctx,
             target,
             SourceIR.ATEN,
@@ -46,15 +45,14 @@ def fused_nccl_reduce_scatter(
         kwargs: Dict[str, Argument],
         name: str,
     ) -> Union[trt.ITensor, Sequence[trt.ITensor]]:
-        return impl.distributed.nccl_reduce_scatter(
+        return impl.nccl_ops.nccl_reduce_scatter(
             ctx,
             target,
             SourceIR.ATEN,
             name,
             [args[0]],
         )
 
-    breakpoint()
 else:
     _LOGGER.debug(
         "Did not load torch.distributed converters since TensorRT-LLM is not available"

py/torch_tensorrt/dynamo/backend/backends.py

@@ -3,12 +3,11 @@
 from typing import Optional, Tuple, Union
 
 import numpy as np
+import tensorrt as trt
 from torch.fx.node import Argument, Target
 from torch_tensorrt.dynamo.conversion._ConversionContext import ConversionContext
 from torch_tensorrt.fx.converters.converter_utils import SourceIR, set_layer_name
 
-import tensorrt as trt
-
 
 # class for AllReduce
 class AllReduceStrategy(IntEnum):
@@ -94,7 +93,7 @@ def nccl_reduce_scatter(
         "group", np.array(group, dtype=np.int32), trt.PluginFieldType.INT32
     )
 
-    p_dtype = trt.float16
+    p_dtype = trt.float32
     pf_dtype = trt.PluginField(
         "type_id", np.array([int(p_dtype)], np.int32), trt.PluginFieldType.INT32
     )

py/torch_tensorrt/dynamo/conversion/custom_ops_converters.py

@@ -49,7 +49,6 @@ def fuse_distributed_ops(
             == torch.ops._c10d_functional.wait_tensor.default
         ):
             wait_tensor_node = list(node.users)[0]
-            fused_op = None
             if node.target == torch.ops._c10d_functional.all_gather_into_tensor.default:
                 with gm.graph.inserting_after(wait_tensor_node):
                     fused_node = gm.graph.create_node(
@@ -58,11 +57,12 @@ def fuse_distributed_ops(
                         args=(node.args[0], node.args[1], node.args[2]),
                     )
             else:
-                fused_node = gm.graph.create_node(
-                    op="call_function",
-                    target=tensorrt_fused_nccl_reduce_scatter_op,  # Define your custom fused function
-                    args=(node.args[0], node.args[1], node.args[2], node.args[3]),
-                )
+                with gm.graph.inserting_after(wait_tensor_node):
+                    fused_node = gm.graph.create_node(
+                        op="call_function",
+                        target=tensorrt_fused_nccl_reduce_scatter_op,  # Define your custom fused function
+                        args=(node.args[0], node.args[1], node.args[2], node.args[3]),
+                    )
 
             wait_tensor_node.replace_all_uses_with(fused_node)
             fused_node.meta.update(node.meta)