Distributed llama3 example

Author: apbose

examples/distributed_inference/llama3_model.py

@@ -84,20 +84,20 @@ def parallel_rotary_block(rotary_block, tp_mesh):
         "wk": ColwiseParallel(),
         "wo": RowwiseParallel(output_layouts=Shard(0)),
     }
-    rotary_block.n_parallel = 1  # this is for single GPU, to do remove this hardcode
+    rotary_block.n_parallel = tp_mesh.size()
 
     parallelize_module(rotary_block, tp_mesh, plan)
 
 
 class RotaryAttention(nn.Module):
-    def __init__(self, dim: int, seq_len: int):
+    def __init__(self, dim: int, seq_len: int, n_parallel: int = 1):
         super().__init__()
         self.dim = dim
         self.wq = nn.Linear(dim, dim)
         self.wk = nn.Linear(dim, dim)
         self.wo = nn.Linear(dim, dim)
         self.seq_len = seq_len
-        self.n_parallel = 1
+        self.n_parallel = n_parallel
         self.register_buffer("freqs_cis", self._precompute_freqs_cis(), persistent=True)
         self.init_weights()
 

examples/distributed_inference/rotary_embedding.py

@@ -0,0 +1,68 @@
+# Taken and modified pytorch lightening
+# https://lightning.ai/lightning-ai/studios/tensor-parallelism-supercharging-large-model-training-with-pytorch-lightning
+import logging
+import os
+import time
+
+import torch
+import torch_tensorrt
+from llama3_model import ModelArgs, ParallelTransformer
+from torch.distributed._composable.fsdp import MixedPrecisionPolicy
+from torch.distributed._composable.fsdp.fully_shard import fully_shard
+from torch.distributed._tensor import Replicate, Shard
+from torch.distributed.algorithms._checkpoint.checkpoint_wrapper import (
+    checkpoint_wrapper,
+)
+from torch_tensorrt.dynamo.distributed.utils import (
+    cleanup_distributed_env,
+    get_tensor_parallel_device_mesh,
+    initialize_distributed_env,
+    initialize_logger,
+)
+
+if not dist.is_initialized():
+    initialize_distributed_env()
+
+device_mesh, _world_size, _rank = get_tensor_parallel_device_mesh()
+logger = initialize_logger(_rank, "tensor_parallel_simple_example")
+
+logger.info(f"Starting PyTorch TP example on rank {_rank}.")
+assert (
+    _world_size % 2 == 0
+), f"TP examples require even number of GPUs, but got {_world_size} gpus"
+
+model_args = ModelArgs(
+    vocab_size=32000,
+    dim=1024,
+    n_layers=4,
+    n_heads=8,
+    rope_theta=500000.0,
+    n_kv_heads=8,
+    device="cuda",
+)
+
+with torch.no_grad():
+    model = ParallelTransformer(model_args, device_mesh)
+    torch.manual_seed(0)
+    inp = torch.randint(32000, (8, 256), device="cuda")
+    python_result = model(inp)
+    torch_tensorrt.runtime.set_multi_device_safe_mode(True)
+    model = torch.compile(
+        model,
+        fullgraph=True,
+        backend="torch_tensorrt",
+        options={
+            "use_python_runtime": True,
+            "use_distributed_mode_trace": True,
+            "debug": True,
+        },
+        dynamic=False,
+    )
+
+    start = time.time()
+    output = model(inp)
+    end = time.time()
+    logger.info(f"Compilation time is {end-start}")
+    assert (python_result - output).std() < 0.01, "Compilation result is not correct."
+
+    cleanup_distributed_env()

examples/distributed_inference/tensor_parallel_llama3.py

@@ -45,7 +45,7 @@
 DIM = 128
 
 with torch.no_grad():
-    model = RotaryAttention(DIM, SEQ_LEN)
+    model = RotaryAttention(DIM, SEQ_LEN, device_mesh.size())
     parallel_rotary_block(model, device_mesh)
     device = torch.device("cuda", device_mesh.get_rank())
     model.to(device)

examples/distributed_inference/tensor_parallel_rotary_embedding.py

@@ -13,7 +13,6 @@
 from torch_tensorrt._Device import Device
 from torch_tensorrt._enums import EngineCapability, dtype
 from torch_tensorrt._features import needs_cross_compile
-from torch_tensorrt._Input import Input
 from torch_tensorrt.dynamo import _defaults, partitioning
 from torch_tensorrt.dynamo._DryRunTracker import (
     DryRunTracker,
@@ -287,7 +286,6 @@ def cross_compile_for_windows(
         arg_inputs = [arg_inputs]  # type: ignore
 
     # Prepare torch_trt inputs
-    trt_arg_inputs: Sequence[Input] = prepare_inputs(arg_inputs)
     trt_kwarg_inputs: Optional[dict[Any, Any]] = prepare_inputs(kwarg_inputs)
     device = to_torch_tensorrt_device(device)
     enabled_precisions = {dtype._from(p) for p in enabled_precisions}
@@ -377,7 +375,6 @@ def cross_compile_for_windows(
             )
     trt_gm = compile_module(
         gm,
-        trt_arg_inputs,
         trt_kwarg_inputs,
         settings,
     )
@@ -623,7 +620,6 @@ def compile(
         arg_inputs = [arg_inputs]  # type: ignore
 
     # Prepare torch_trt inputs
-    trt_arg_inputs: Sequence[Input] = prepare_inputs(arg_inputs)
     trt_kwarg_inputs: Optional[dict[Any, Any]] = prepare_inputs(kwarg_inputs)
     device = to_torch_tensorrt_device(device)
     enabled_precisions = {dtype._from(p) for p in enabled_precisions}
@@ -709,16 +705,13 @@ def compile(
             logger.warning(
                 "Remaining GPU memory may not be enough to compile the TensorRT engine for this model resulting in an OOM error, Consider setting offload_module_to_cpu=True"
             )
-    trt_gm = compile_module(
-        gm, trt_arg_inputs, trt_kwarg_inputs, settings, engine_cache
-    )
+    trt_gm = compile_module(gm, trt_kwarg_inputs, settings, engine_cache)
     return trt_gm
 
 
 @fn_supports_debugger  # type: ignore[misc]
 def compile_module(
     gm: torch.fx.GraphModule,
-    sample_arg_inputs: Sequence[Input],
     sample_kwarg_inputs: Optional[dict[Any, Any]] = None,
     settings: CompilationSettings = CompilationSettings(),
     engine_cache: Optional[BaseEngineCache] = None,

py/torch_tensorrt/dynamo/_compiler.py

@@ -22,7 +22,6 @@
 from torch_tensorrt.dynamo.utils import (
     is_tegra_platform,
     parse_dynamo_kwargs,
-    prepare_inputs,
     set_log_level,
 )
 
@@ -150,9 +149,6 @@ def _pretraced_backend(
 
             logger.debug("Lowered Input graph:\n " + str(gm.graph))
 
-            torchtrt_inputs = prepare_inputs(
-                torch_inputs, disable_memory_format_check=True
-            )
             if settings.require_full_compilation:
                 logger.warning(
                     "require_full_compilation arg is not applicable for torch.compile with backend='torch_tensorrt"
@@ -163,7 +159,6 @@ def _pretraced_backend(
                 )
             trt_compiled = compile_module(
                 gm,
-                torchtrt_inputs,
                 settings=settings,
                 engine_cache=engine_cache,
             )