enable Context Parallel

ghstack-source-id: 5a8900c
Pull Request resolved: #592

Author: XilunWu

estimation.py

@@ -14,14 +14,14 @@
 from torch.distributed._tools.fsdp2_mem_tracker import FSDPMemTracker
 from torch.testing._internal.distributed.fake_pg import FakeStore
 
+from torchtitan import utils
 from torchtitan.config_manager import JobConfig
 from torchtitan.datasets import build_tokenizer
 from torchtitan.float8 import Float8Handler
 from torchtitan.logging import init_logger, logger
 from torchtitan.models import model_name_to_cls, model_name_to_tokenizer, models_config
 from torchtitan.optimizer import build_lr_schedulers, build_optimizers
 from torchtitan.parallelisms import models_parallelize_fns, ParallelDims
-from train import get_train_context
 
 
 def estimate_memory(job_config: JobConfig):
@@ -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,
@@ -94,7 +95,7 @@ def estimate_memory(job_config: JobConfig):
     tokenizer_type = model_name_to_tokenizer[model_name]
     tokenizer = build_tokenizer(tokenizer_type, job_config.model.tokenizer_path)
 
-    train_context = get_train_context(
+    train_context = utils.get_train_context(
         parallel_dims.loss_parallel_enabled,
         job_config.experimental.enable_compiled_autograd,
     )

test_runner.py

@@ -306,6 +306,41 @@ 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(
+            [
+                [
+                    "--training.data_parallel_shard_degree=2",
+                    "--training.tensor_parallel_degree=2",
+                    "--experimental.context_parallel_degree=2",
+                ]
+            ],
+            "FSDP+TP+CP",
+            "fsdp+tp+cp",
+            ngpu=8,
+        ),
         OverrideDefinitions(
             [
                 [

torchtitan/config_manager.py

@@ -325,6 +325,12 @@ 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 +415,9 @@ 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,
+            # TODO: explain in docs/composability.md why we removed the 2x
+            # relaxing in our CP enablement PR
+            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"],
         ):
             if d > 1:
                 dims.append(d)
@@ -71,6 +74,13 @@ def build_mesh(self, device_type):
         # initialized
         if self.dp_replicate > 1 and self.dp_shard > 1:
             mesh["dp_replicate", "dp_shard"]._flatten(mesh_dim_name="dp")
+
+        if self.cp > 1:
+            if self.dp_replicate > 1 and self.dp_shard > 1:
+                mesh["dp_replicate", "dp_shard", "cp"]._flatten(mesh_dim_name="dp_cp")
+            else:
+                mesh["dp", "cp"]._flatten(mesh_dim_name="dp_cp")
+
         return mesh
 
     @property
@@ -85,6 +95,10 @@ 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
@@ -98,5 +112,5 @@ def loss_parallel_enabled(self):
         return self.tp > 1 and self.enable_loss_parallel
 
     @cached_property
-    def model_parallel_size(self):
-        return self.tp * self.pp
+    def non_data_parallel_size(self):
+        return self.cp * self.tp * self.pp

torchtitan/parallelisms/parallelize_llama.py

@@ -11,6 +11,7 @@
 
 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
@@ -72,36 +73,44 @@ def parallelize_llama(
             )
         apply_compile(model)
 
-    if parallel_dims.dp_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"]
-
-            apply_fsdp(
-                model,
-                dp_mesh,
-                param_dtype=TORCH_DTYPE_MAP[job_config.training.mixed_precision_param],
-                reduce_dtype=TORCH_DTYPE_MAP[
-                    job_config.training.mixed_precision_reduce
-                ],
-                tp_enabled=parallel_dims.tp_enabled,
-                pp_enabled=parallel_dims.pp_enabled,
-            )
-            if parallel_dims.dp_replicate_enabled:
-                logger.info("Applied HSDP to the model")
-            else:
-                logger.info("Applied FSDP to the model")
+    if (
+        parallel_dims.dp_shard_enabled
+    ):  # apply FSDP or HSDP, potentially with Context Parallel
+        dp_mesh_dim_names = (
+            ("dp_replicate", "dp_shard")
+            if parallel_dims.dp_replicate_enabled
+            else ("dp",)
+        )
+        dp_mesh = (
+            world_mesh[(*dp_mesh_dim_names, "cp")]._flatten("dp_cp")
+            if parallel_dims.cp_enabled
+            else world_mesh[dp_mesh_dim_names]
+        )
+        apply_fsdp(
+            model,
+            dp_mesh,
+            param_dtype=TORCH_DTYPE_MAP[job_config.training.mixed_precision_param],
+            reduce_dtype=TORCH_DTYPE_MAP[job_config.training.mixed_precision_reduce],
+            tp_enabled=parallel_dims.tp_enabled,
+            pp_enabled=parallel_dims.pp_enabled,
+        )
+
+        if parallel_dims.dp_replicate_enabled:
+            logger.info("Applied HSDP to the model")
         else:
-            if world_mesh.ndim > 1:
-                raise RuntimeError("DDP has not supported > 1D parallelism")
-            apply_ddp(
-                model,
-                world_mesh,
-                enable_compile=job_config.training.compile,
-                enable_compiled_autograd=job_config.experimental.enable_compiled_autograd,
-            )
+            logger.info("Applied FSDP to the model")
+
+        if parallel_dims.cp_enabled:
+            logger.info("Applied Context Parallel to the model")
+    elif parallel_dims.dp_replicate_enabled:
+        if world_mesh.ndim > 1:
+            raise RuntimeError("DDP has not supported > 1D parallelism")
+        apply_ddp(
+            model,
+            world_mesh,
+            enable_compile=job_config.training.compile,
+            enable_compiled_autograd=job_config.experimental.enable_compiled_autograd,
+        )
 
 
 def apply_tp(

torchtitan/utils.py

@@ -4,12 +4,13 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+import contextlib
 import gc
 import os
 import subprocess
 from dataclasses import dataclass
 from datetime import timedelta
-from typing import Optional, Union
+from typing import Generator, List, Optional, Set, Union
 
 import torch
 import torch.distributed._functional_collectives as funcol
@@ -101,6 +102,57 @@ def run(self, step_count):
 SKIP_CLEANUP = "3"
 
 
+def create_context_parallel_ctx(
+    cp_mesh: DeviceMesh,
+    cp_buffers: List[torch.Tensor],
+    cp_seq_dims: List[int],
+    cp_no_restore_buffers: Set[torch.Tensor],
+):
+    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."
+        )
+
+    return context_parallel(
+        cp_mesh,
+        buffers=cp_buffers,
+        buffer_seq_dims=cp_seq_dims,
+        no_restore_buffers=cp_no_restore_buffers,
+    )
+
+
+def get_train_context(enable_loss_parallel: bool, enable_compiled_autograd: bool):
+    @contextlib.contextmanager
+    def context(cp_context: Optional[Generator[None, None, None]] = None):
+        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_context is not None:
+                from torch.nn.attention import sdpa_kernel, SDPBackend
+
+                # 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(cp_context)
+
+            yield
+
+    return context
+
+
 def init_distributed(job_config):
     # FlightRecorder is incompatible with =1 mode where watchdog aborts work, must use =3 (skipcleanup)
     # to get flight recorder dumps. See https://github.com/pytorch/pytorch/issues/121055