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: 5d4f276bcff9ff53c2caadd161161a9b7a33142a
Pull Request resolved: #433

estimation.py

@@ -70,6 +70,7 @@ 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/models/llama/model.py

@@ -410,7 +410,7 @@ def _precompute_freqs_cis(self) -> torch.Tensor:
             self.model_args.dim // self.model_args.n_heads,
             # Need to compute until at least the max token limit for generation
             # (use 2x max sequence length to be safe)
-            self.model_args.max_seq_len * 2,
+            self.model_args.max_seq_len,
             self.model_args.rope_theta,
         )
 

torchtitan/parallelisms/__init__.py

@@ -24,6 +24,7 @@
 @dataclass
 class ParallelDims:
     dp: int
+    cp: int
     tp: int
     pp: int
     world_size: int
@@ -35,34 +36,41 @@ 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)
                 names.append(name)
         logger.info(f"Building {len(dims)}-D device mesh with {names}, {dims}")
         names = tuple(names)
-        return init_device_mesh(device_type, dims, mesh_dim_names=names)
+        world_mesh = init_device_mesh(device_type, dims, mesh_dim_names=names)
+        return world_mesh
 
     @property
     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

@@ -19,6 +19,11 @@
 
 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,
 )
@@ -472,6 +477,22 @@ 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.pp_enabled:
+        raise NotImplementedError("CP + PP is 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,
@@ -482,8 +503,14 @@ def apply_fsdp(
     Apply data parallelism to the model. FSDP2 is used here.
     """
 
-    dp_mesh = world_mesh["dp"] if world_mesh.ndim > 1 else world_mesh
-    assert dp_mesh.mesh_dim_names == ("dp",), dp_mesh.mesh_dim_names
+    if parallel_dims.cp_enabled:
+        # Temporary solution to enable FSDP + CP
+        if parallel_dims.dp_enabled:
+            dp_mesh = world_mesh["dp", "cp"]._flatten()
+        else:
+            dp_mesh = world_mesh["cp"]
+    else:
+        dp_mesh = world_mesh["dp"]
 
     mp_policy = MixedPrecisionPolicy(
         param_dtype=TORCH_DTYPE_MAP[job_config.training.mixed_precision_param],
@@ -570,7 +597,10 @@ def parallelize_llama(
     if job_config.training.compile:
         model = apply_compile(model, job_config)
 
-    if parallel_dims.dp_enabled:
+    if parallel_dims.cp_enabled:
+        model = apply_cp(model, world_mesh, parallel_dims, job_config)
+
+    if parallel_dims.dp_enabled or parallel_dims.cp_enabled:
         if parallel_dims.dp_type == "fsdp":
             model = apply_fsdp(model, world_mesh, parallel_dims, job_config)
         else:

train.py

@@ -11,18 +11,21 @@
 
 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
+from typing import Any, Callable, Dict, List, Optional
 
 import numpy as np
 
 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
+from torch.distributed.device_mesh import DeviceMesh
 
 from torchtitan.checkpoint import CheckpointManager
 from torchtitan.config_manager import JobConfig
@@ -139,18 +142,40 @@ def zero_grad(self):
     return OptimizersContainer([_build_optimizer(model) for model in model_parts])
 
 
-def get_train_context(enable_loss_parallel: bool, enable_compiled_autograd: bool):
+def get_train_context(
+    enable_loss_parallel: bool,
+    enable_compiled_autograd: bool,
+    cp_mesh: Optional[DeviceMesh],
+):
+    if cp_mesh is not None:
+        context_parallel_ctx = partial(context_parallel_buffers, mesh=cp_mesh)
+    else:
+        context_parallel_ctx = partial(context_parallel_buffers, mesh=None)
+
     @contextlib.contextmanager
-    def context():
+    def context(
+        cp_buffers: List[torch.Tensor],
+        cp_seq_dims: List[int],
+        cp_restore_funcs: List[Optional[Callable]],
+    ):
         with contextlib.ExitStack() as stack:
             if enable_loss_parallel:
                 stack.enter_context(loss_parallel())
+
             if enable_compiled_autograd:
                 stack.enter_context(
                     torch._dynamo.utils.maybe_enable_compiled_autograd(True)
                 )
 
-            yield
+            buffers = stack.enter_context(
+                context_parallel_ctx(
+                    buffers=cp_buffers,
+                    seq_dims=cp_seq_dims,
+                    restore_funcs=cp_restore_funcs,
+                )
+            )
+
+            yield buffers
 
     return context
 
@@ -173,6 +198,7 @@ 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,
@@ -215,6 +241,7 @@ def main(job_config: JobConfig):
     train_context = get_train_context(
         parallel_dims.loss_parallel_enabled,
         job_config.experimental.enable_compiled_autograd,
+        world_mesh["cp"] if parallel_dims.cp_enabled else None,
     )
 
     # loss fn can be shared by pipeline-parallel or non-pp execution
@@ -376,18 +403,28 @@ def loss_fn(pred, labels):
             data_load_start = timer()
             batch = next(data_iterator)
             input_ids, labels = batch
-            ntokens_since_last_log += labels.numel()
+            ntokens_since_last_log += labels.numel() // parallel_dims.cp
             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(
+                cp_buffers=[input_ids, labels, model.freqs_cis],
+                cp_seq_dims=[1, 1, 0],
+                cp_restore_funcs=[
+                    None,
+                    None,
+                    lambda buf, m=model: setattr(m, "freqs_cis", buf),
+                ],
+            ) as cp_buffers:
+                input_ids = cp_buffers[0].cuda()
+                labels = cp_buffers[1].cuda()
+                model.freqs_cis = cp_buffers[2]
+
+                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:
@@ -396,15 +433,13 @@ def loss_fn(pred, labels):
                     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():
+                    loss = (
+                        torch.mean(torch.stack(losses))
+                        if is_last_stage
+                        else torch.Tensor([-1.0])
+                    )
+                else:
+                    # Non-PP forward / backward
                     pred = model(input_ids)
                     loss = loss_fn(pred, labels)
                     # pred.shape=(bs, seq_len, vocab_size)