Enable CP

This PR adds experimental flags and functions to enable context parallelism. We currently support only FSDP + CP and CP only. CP + TP is being tested.

ghstack-source-id: 20b8844
Pull Request resolved: #433

Author: fegin

estimation.py

@@ -67,6 +67,7 @@ def estimate_memory(job_config: JobConfig):
 
     parallel_dims = ParallelDims(
         dp=job_config.training.data_parallel_degree,
+        cp=job_config.experimental.context_parallel_degree,
         tp=job_config.training.tensor_parallel_degree,
         pp=job_config.experimental.pipeline_parallel_degree,
         world_size=world_size,

torchtitan/config_manager.py

@@ -323,6 +323,12 @@ def __init__(self):
             action="store_true",
             help="Enable CompiledAutograd to compile the backward.",
         )
+        self.parser.add_argument(
+            "--experimental.context_parallel_degree",
+            type=int,
+            default=1,
+            help="Context parallelism degree. 1 means disabled.",
+        )
         self.parser.add_argument(
             "--training.mixed_precision_param",
             type=str,

torchtitan/parallelisms/__init__.py

@@ -24,6 +24,7 @@
 @dataclass
 class ParallelDims:
     dp: int
+    cp: int
     tp: int
     pp: int
     world_size: int
@@ -35,22 +36,24 @@ def __post_init__(self):
         self._validate()
 
     def _validate(self):
-        dp, tp, pp = self.dp, self.tp, self.pp
+        dp, cp, tp, pp = self.dp, self.cp, self.tp, self.pp
         if dp == -1:
-            self.dp = dp = self.world_size // (tp * pp)
+            self.dp = dp = self.world_size // (cp * tp * pp)
         assert dp >= 1, dp
+        assert cp >= 1, cp
         assert tp >= 1, tp
         assert pp >= 1, pp
-        assert (
-            dp * tp * pp == self.world_size
-        ), f"Invalid parallel dims: dp({dp}) * tp({tp}) * pp({pp}) != WORLD_SIZE({self.world_size})"
+        assert dp * cp * tp * pp == self.world_size, (
+            f"Invalid parallel dims: dp({dp}) * cp ({cp}) * tp({tp}) * pp({pp}) "
+            f"!= WORLD_SIZE({self.world_size})"
+        )
         assert self.dp_type in ("fsdp", "ddp")
 
     def build_mesh(self, device_type):
         dims = []
         names = []
         for d, name in zip(
-            [self.pp, self.dp, self.tp], ["pp", "dp", "tp"], strict=True
+            [self.pp, self.dp, self.cp, self.tp], ["pp", "dp", "cp", "tp"], strict=True
         ):
             if d > 1:
                 dims.append(d)
@@ -63,6 +66,10 @@ def build_mesh(self, device_type):
     def dp_enabled(self):
         return self.dp > 1
 
+    @property
+    def cp_enabled(self):
+        return self.cp > 1
+
     @property
     def tp_enabled(self):
         return self.tp > 1

torchtitan/parallelisms/parallelize_llama.py

@@ -19,9 +19,15 @@
 
 from torch.distributed._composable.replicate import replicate
 from torch.distributed._tensor import Replicate, Shard
+
+try:
+    from torch.distributed._tensor.experimental.attention import enable_context_parallel
+except ImportError:
+    print("The PyTorch version does not include the experimental CP APIs.")
 from torch.distributed.algorithms._checkpoint.checkpoint_wrapper import (
     checkpoint_wrapper as ptd_checkpoint_wrapper,
 )
+from torch.distributed.device_mesh import init_device_mesh
 from torch.distributed.pipelining import pipeline, PipelineStage, SplitPoint
 from torch.distributed.tensor.parallel import (
     ColwiseParallel,
@@ -455,17 +461,43 @@ def apply_compile(model: nn.Module, job_config: JobConfig):
     return model
 
 
+def apply_cp(model, world_mesh, parallel_dims, job_config: JobConfig):
+    """
+    Apply context parallelism to the model. This is an experimental feature.
+    """
+    if parallel_dims.tp_enabled or parallel_dims.pp_enabled:
+        raise NotImplementedError("CP + TP or CP + PP are not supported yet.")
+    cp_mesh = world_mesh["cp"]
+    callers = []
+    for layer_id, transformer_block in model.layers.items():
+        callers.append(transformer_block.attention)
+    enable_context_parallel(seq_dim=2, callers=callers, device_mesh=cp_mesh)
+    logger.info("Applied CP to the model")
+
+    return model
+
+
 def apply_fsdp(
     model: nn.Module,
     world_mesh: DeviceMesh,
     parallel_dims: "ParallelDims",
     job_config: JobConfig,
 ):
+
     """
     Apply data parallelism to the model. FSDP2 is used here.
     """
 
-    dp_mesh = world_mesh["dp"] if world_mesh.ndim > 1 else world_mesh
+    if parallel_dims.cp_enabled:
+        # Temporary solution to enable FSDP + CP
+        dp_mesh = init_device_mesh(
+            world_mesh.device_type,
+            (parallel_dims.dp * parallel_dims.cp,),
+            mesh_dim_names=["dp"],
+        )
+    else:
+        dp_mesh = world_mesh["dp"]
+
     assert dp_mesh.mesh_dim_names == ("dp",), dp_mesh.mesh_dim_names
 
     mp_policy = MixedPrecisionPolicy(
@@ -542,6 +574,9 @@ def parallelize_llama(
     if job_config.training.compile:
         model = apply_compile(model, job_config)
 
+    if parallel_dims.cp_enabled:
+        model = apply_cp(model, world_mesh, parallel_dims, job_config)
+
     if parallel_dims.dp_enabled:
         if parallel_dims.dp_type == "fsdp":
             model = apply_fsdp(model, world_mesh, parallel_dims, job_config)

train.py

@@ -11,6 +11,7 @@
 
 from dataclasses import dataclass, field
 from datetime import timedelta
+from functools import partial
 from io import BytesIO
 from timeit import default_timer as timer
 from typing import Any, Dict, List
@@ -20,6 +21,7 @@
 import torch
 import torch.nn.functional as F
 from torch.distributed import destroy_process_group
+from torch.distributed._tensor.experimental.attention import context_parallel_buffers
 from torch.distributed.checkpoint.stateful import Stateful
 from torch.distributed.elastic.multiprocessing.errors import record
 from torch.distributed.tensor.parallel import loss_parallel
@@ -172,6 +174,7 @@ def main(job_config: JobConfig):
     world_size = int(os.environ["WORLD_SIZE"])
     parallel_dims = ParallelDims(
         dp=job_config.training.data_parallel_degree,
+        cp=job_config.experimental.context_parallel_degree,
         tp=job_config.training.tensor_parallel_degree,
         pp=job_config.experimental.pipeline_parallel_degree,
         world_size=world_size,
@@ -216,6 +219,20 @@ def main(job_config: JobConfig):
         job_config.experimental.enable_compiled_autograd,
     )
 
+    if parallel_dims.cp_enabled:
+        cp_mesh = world_mesh["cp"]
+        context_parallel_ctx = partial(
+            context_parallel_buffers,
+            cp_rank=cp_mesh.get_local_rank(),
+            cp_world_size=cp_mesh.size(),
+        )
+    else:
+        context_parallel_ctx = partial(
+            context_parallel_buffers,
+            cp_rank=0,
+            cp_world_size=1,
+        )
+
     # 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))
@@ -373,38 +390,43 @@ def loss_fn(pred, labels):
             ntokens_since_last_log += labels.numel()
             data_loading_times.append(timer() - data_load_start)
 
-            input_ids = input_ids.cuda()
-            labels = labels.cuda()
             optimizers.zero_grad()
 
-            if parallel_dims.pp_enabled:
-                # pipeline parallel forward / backward inside step() call
-                is_last_stage = pp_mesh.get_local_rank() == pp_mesh.size() - 1
-
-                with train_context():
-                    if pp_mesh.get_local_rank() == 0:
-                        pp_schedule.step(input_ids)
-                    elif is_last_stage:
-                        losses = []
-                        pp_schedule.step(target=labels, losses=losses)
-                    else:
-                        pp_schedule.step()
-
-                # accumulate losses across pipeline microbatches
-                loss = (
-                    torch.mean(torch.stack(losses))
-                    if is_last_stage
-                    else torch.Tensor([-1.0])
-                )
-            else:
-                # Non-PP forward / backward
-                with train_context():
-                    pred = model(input_ids)
-                    loss = loss_fn(pred, labels)
-                    # pred.shape=(bs, seq_len, vocab_size)
-                    # need to free to before bwd to avoid peaking memory
-                    del pred
-                    loss.backward()
+            with context_parallel_ctx(
+                buffers=[input_ids, labels, model.freqs_cis],
+                seq_dims=[1, 1, 0],
+                keep_orig_buffers=[False, False, True],
+            ):
+                input_ids = input_ids.cuda()
+                labels = labels.cuda()
+                if parallel_dims.pp_enabled:
+                    # pipeline parallel forward / backward inside step() call
+                    is_last_stage = pp_mesh.get_local_rank() == pp_mesh.size() - 1
+
+                    with train_context():
+                        if pp_mesh.get_local_rank() == 0:
+                            pp_schedule.step(input_ids)
+                        elif is_last_stage:
+                            losses = []
+                            pp_schedule.step(target=labels, losses=losses)
+                        else:
+                            pp_schedule.step()
+
+                    # accumulate losses across pipeline microbatches
+                    loss = (
+                        torch.mean(torch.stack(losses))
+                        if is_last_stage
+                        else torch.Tensor([-1.0])
+                    )
+                else:
+                    # Non-PP forward / backward
+                    with train_context():
+                        pred = model(input_ids)
+                        loss = loss_fn(pred, labels)
+                        # pred.shape=(bs, seq_len, vocab_size)
+                        # need to free to before bwd to avoid peaking memory
+                        del pred
+                        loss.backward()
 
             # clip gradients
             for model in model_parts: