enable Context Parallel

ghstack-source-id: f9bc24f
Pull Request resolved: #592

Author: XilunWu

estimation.py

@@ -66,6 +66,7 @@ def estimate_memory(job_config: JobConfig):
     parallel_dims = ParallelDims(
         dp_shard=job_config.training.data_parallel_shard_degree,
         dp_replicate=job_config.training.data_parallel_replicate_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,

test_runner.py

@@ -316,6 +316,29 @@ def build_test_list():
             "hsdp+tp",
             ngpu=8,
         ),
+        OverrideDefinitions(
+            [
+                [
+                    "--training.data_parallel_shard_degree=2",
+                    "--experimental.context_parallel_degree=2",
+                ]
+            ],
+            "FSDP+CP",
+            "fsdp+cp",
+            ngpu=4,
+        ),
+        OverrideDefinitions(
+            [
+                [
+                    "--training.data_parallel_shard_degree=2",
+                    "--training.data_parallel_replicate_degree=2",
+                    "--experimental.context_parallel_degree=2",
+                ]
+            ],
+            "HSDP+CP",
+            "hsdp+cp",
+            ngpu=8,
+        ),
         OverrideDefinitions(
             [
                 [

torchtitan/config_manager.py

@@ -326,6 +326,12 @@ 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

@@ -415,8 +415,8 @@ def _precompute_freqs_cis(self) -> torch.Tensor:
         return precompute_freqs_cis(
             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,
+            # Note: removed the 2x relaxing in CP enablement
+            self.model_args.max_seq_len,
             self.model_args.rope_theta,
         )
 

torchtitan/parallelisms/parallel_dims.py

@@ -15,6 +15,7 @@
 class ParallelDims:
     dp_replicate: int
     dp_shard: int
+    cp: int
     tp: int
     pp: int
     world_size: int
@@ -24,36 +25,38 @@ def __post_init__(self):
         self._validate()
 
     def _validate(self):
-        dp_replicate, dp_shard, tp, pp = (
+        dp_replicate, dp_shard, cp, tp, pp = (
             self.dp_replicate,
             self.dp_shard,
+            self.cp,
             self.tp,
             self.pp,
         )
-        for d in (dp_replicate, tp, pp):
+        for d in (dp_replicate, cp, tp, pp):
             assert d >= 1, "Parallelism degree should be >= 1, except for dp_shard"
         assert dp_shard == -1 or dp_shard >= 1, " dp_shard must -1 or >=1."
 
         dp = dp_replicate * dp_shard
         if dp < 0:
-            dp = self.world_size // (tp * pp)
+            dp = self.world_size // (cp * tp * pp)
             self.dp_shard = dp_shard = dp // dp_replicate
 
         assert dp_replicate >= 1
         assert dp_shard >= 1
+        assert cp >= 1, cp
         assert tp >= 1, tp
         assert pp >= 1, pp
-        assert dp_replicate * dp_shard * tp * pp == self.world_size, (
+        assert dp_replicate * dp_shard * cp * tp * pp == self.world_size, (
             f"Invalid parallel dims: dp_replicate({dp_replicate}) * dp_shard({dp_shard}) * "
-            f"tp({tp}) * pp({pp}) != WORLD_SIZE({self.world_size})"
+            f"cp({cp}) * tp({tp}) * pp({pp}) != WORLD_SIZE({self.world_size})"
         )
 
     def build_mesh(self, device_type):
         dims = []
         names = []
         for d, name in zip(
-            [self.pp, self.dp_replicate, self.dp_shard, self.tp],
-            ["pp", "dp_replicate", "dp_shard", "tp"],
+            [self.pp, self.dp_replicate, self.dp_shard, self.cp, self.tp],
+            ["pp", "dp_replicate", "dp_shard", "cp", "tp"],
             strict=True,
         ):
             if d > 1:
@@ -86,6 +89,10 @@ def dp_replicate_enabled(self):
     def dp_shard_enabled(self):
         return self.dp_shard > 1
 
+    @property
+    def cp_enabled(self):
+        return self.cp > 1
+
     @property
     def tp_enabled(self):
         return self.tp > 1

torchtitan/parallelisms/parallelize_llama.py

@@ -8,9 +8,11 @@
 # training techniques (e.g. activation checkpointing and compile) to the Llama model.
 
 from collections import defaultdict
+from typing import Tuple
 
 import torch
 import torch.nn as nn
+
 from torch.distributed import DeviceMesh
 from torch.distributed._composable.fsdp import fully_shard, MixedPrecisionPolicy
 from torch.distributed._composable.replicate import replicate
@@ -72,13 +74,18 @@ def parallelize_llama(
             )
         apply_compile(model)
 
-    if parallel_dims.dp_enabled:
+    if parallel_dims.dp_enabled or parallel_dims.cp_enabled:
         if parallel_dims.dp_shard_enabled:
             if parallel_dims.dp_replicate_enabled:
                 dp_mesh = world_mesh["dp_replicate", "dp_shard"]
             else:
                 dp_mesh = world_mesh["dp"]
 
+            if parallel_dims.cp_enabled:
+                dp_dim_names = dp_mesh.mesh_dim_names
+                assert isinstance(dp_dim_names, Tuple)
+                dp_mesh = world_mesh[(*dp_dim_names, "cp")]._flatten()
+
             apply_fsdp(
                 model,
                 dp_mesh,

train.py

@@ -10,7 +10,13 @@
 from datetime import timedelta
 
 import torch
+
+from typing import List, Optional, Set
+from functools import partial
+
+from torch.distributed.device_mesh import DeviceMesh
 from torch.distributed.elastic.multiprocessing.errors import record
+from torch.nn.attention import SDPBackend, sdpa_kernel
 
 from torchtitan import utils
 from torchtitan.checkpoint import CheckpointManager, TrainState
@@ -28,17 +34,52 @@
 )
 from torchtitan.profiling import maybe_enable_memory_snapshot, maybe_enable_profiling
 
+try:
+    from torch.distributed.tensor.experimental import context_parallel
+except ImportError:
+    print(
+        f"PyTorch version {torch.__version__} does not include the experimental "
+        "Context Parallel API. Please update to a newer version."
+    )
+
+
+def get_train_context(
+    enable_loss_parallel: bool,
+    enable_compiled_autograd: bool,
+    cp_mesh: Optional[DeviceMesh] = None,
+):
+    if cp_mesh is not None:
+        context_parallel_ctx = partial(context_parallel, mesh=cp_mesh)
 
-def get_train_context(enable_loss_parallel: bool, enable_compiled_autograd: bool):
     @contextlib.contextmanager
-    def context():
+    def context(
+        cp_buffers: List[torch.Tensor],
+        cp_seq_dims: List[int],
+        cp_no_restore_buffers: Set[torch.Tensor],
+    ):
         with contextlib.ExitStack() as stack:
             if enable_loss_parallel:
                 stack.enter_context(torch.distributed.tensor.parallel.loss_parallel())
+
             if enable_compiled_autograd:
                 stack.enter_context(
                     torch._dynamo.utils.maybe_enable_compiled_autograd(True)
                 )
+
+            if cp_mesh is not None:
+                # currently we only support these two SDP backends.
+                # TODO (xilunwu): support cuDNN backend
+                stack.enter_context(
+                    sdpa_kernel([SDPBackend.FLASH_ATTENTION, SDPBackend.EFFICIENT_ATTENTION])
+                )
+                stack.enter_context(
+                    context_parallel_ctx(
+                        buffers=cp_buffers,
+                        buffer_seq_dims=cp_seq_dims,
+                        no_restore_buffers=cp_no_restore_buffers,
+                    )
+                )
+
             yield
 
     return context
@@ -61,6 +102,7 @@ def main(job_config: JobConfig):
     parallel_dims = ParallelDims(
         dp_shard=job_config.training.data_parallel_shard_degree,
         dp_replicate=job_config.training.data_parallel_replicate_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,
@@ -226,6 +268,7 @@ def loss_fn(pred, labels):
     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,
     )
 
     # variables used to keep info for metrics logging
@@ -259,18 +302,24 @@ def loss_fn(pred, labels):
             data_load_start = time.perf_counter()
             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(time.perf_counter() - data_load_start)
 
             input_ids = input_ids.cuda()
             labels = labels.cuda()
             optimizers.zero_grad()
 
+            training_context = train_context(
+                cp_buffers=[input_ids, labels, model.freqs_cis],
+                cp_seq_dims=[1, 1, 0],
+                cp_no_restore_buffers={input_ids, labels},
+            )
+
             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():
+                with training_context:
                     if pp_mesh.get_local_rank() == 0:
                         pp_schedule.step(input_ids)
                     elif is_last_stage:
@@ -287,7 +336,7 @@ def loss_fn(pred, labels):
                 )
             else:
                 # Non-PP forward / backward
-                with train_context():
+                with training_context:
                     pred = model(input_ids)
                     loss = loss_fn(pred, labels)
                     # pred.shape=(bs, seq_len, vocab_size)

train_configs/debug_model.toml

@@ -42,6 +42,7 @@ compile = false
 dataset = "c4_test"  # supported datasets: c4_test (2K), c4 (177M)
 
 [experimental]
+context_parallel_degree = 1
 pipeline_parallel_degree = 1
 enable_async_tensor_parallel = false