[RFC][WIP][CP] Enable FlexAttention CP for llama3

This PR uses the latest CP APIs to enable FlexAttention + CP for llama3.  This PR removes the usage of context_paralle() context manager and use `_context_parallel_shard()` to shard the input data.

ghstack-source-id: 5d04d61
Pull-Request: #1857

Author: fegin

torchtitan/distributed/utils.py

@@ -19,6 +19,7 @@
 
 from torchtitan.config import Comm as CommConfig, TORCH_DTYPE_MAP
 from torchtitan.distributed.parallel_dims import ParallelDims
+from torchtitan.protocols.model import AttentionMasksType
 from torchtitan.tools.logging import logger
 from torchtitan.tools.utils import device_module, device_type
 
@@ -200,9 +201,6 @@ def context(cp_context: Generator[None, None, None] | None = None):
             if enable_loss_parallel:
                 stack.enter_context(torch.distributed.tensor.parallel.loss_parallel())
 
-            if cp_context:
-                stack.enter_context(cp_context)
-
             yield
 
     return context
@@ -443,3 +441,53 @@ def _clip_grad_norm_with_ep(
     torch.nn.utils.clip_grads_with_norm_(non_ep_params, max_norm, total_norm, foreach)
 
     return total_norm
+
+
+def cp_shard(
+    cp_mesh: DeviceMesh,
+    inputs: torch.Tensor,
+    labels: torch.Tensor,
+    attention_masks: AttentionMasksType | None,
+    order_sensitive_buffers: dict[str, torch.Tensor],
+    order_sensitive_buffers_seq_dims: dict[str, int],
+):
+    from torch.distributed.tensor.experimental._attention import _context_parallel_shard
+    from torch.distributed.tensor.experimental._load_balancer import _PTRRLoadBalancer
+    from torch.nn.attention.flex_attention import BlockMask
+
+    load_balancer = _PTRRLoadBalancer(attention_masks, cp_mesh.size(0))
+    inputs, labels = _context_parallel_shard(
+        mesh=cp_mesh,
+        buffers=(inputs, labels),
+        seq_dims=(1, 1),
+        load_balancer=load_balancer,
+    )
+
+    order_sensitive_buffers = _context_parallel_shard(
+        mesh=cp_mesh,
+        buffers=order_sensitive_buffers,
+        seq_dims=order_sensitive_buffers_seq_dims,
+        load_balancer=load_balancer,
+    )
+
+    if attention_masks is None:
+        return inputs, labels, None, order_sensitive_buffers
+
+    masks = (
+        [attention_masks]
+        if isinstance(attention_masks, BlockMask)
+        else list(attention_masks.values())
+    )
+    masks = _context_parallel_shard(
+        mesh=cp_mesh,
+        buffers=masks,
+        seq_dims=(2,) * len(masks),
+        load_balancer=load_balancer,
+    )
+    attention_masks = (
+        masks[0]
+        if isinstance(attention_masks, BlockMask)
+        else {k: v for k, v in zip(attention_masks.keys(), masks)}
+    )
+
+    return inputs, labels, attention_masks, order_sensitive_buffers

torchtitan/models/attention.py

@@ -32,6 +32,22 @@
 ]
 
 
+class FlexAttentionKernel(torch.nn.Module):
+    """Wrapper to enable FlexCP"""
+
+    _compiled_flex_attn: ClassVar[Callable] = torch.compile(
+        flex_attention, mode="max-autotune-no-cudagraphs"
+    )
+
+    def forward(self, *args, **kwargs):
+        # 1. _compiled_flex_attn has to be a class variable, otherwise there will
+        #    be multiple compiled flex_attention instances, which can be slow.
+        # 2. `self._compiled_flex_attn` is not correct, `self` will be passed in
+        #    as the first argument, which will cause an error.
+        #    `FlexAttentionKernel._compiled_flex_attn` is correct.
+        return FlexAttentionKernel._compiled_flex_attn(*args, **kwargs)
+
+
 class FlexAttentionWrapper(torch.nn.Module):
     """Wrapper around `flex_attention` to make it torch.compile and CP compatible.
 
@@ -45,9 +61,11 @@ class FlexAttentionWrapper(torch.nn.Module):
         block_mask as a keyword argument to be compatible with _ContextParallel.
     """
 
-    _compiled_flex_attn: ClassVar[Callable] = torch.compile(
-        flex_attention, mode="max-autotune-no-cudagraphs"
-    )
+    def __init__(self) -> None:
+        super().__init__()
+        # TODO: remove this wrapper once FlexAttentionWrapper.forward() has the
+        # same signature as flex_attention() and is compatible with _ContextParallel.
+        self._flex_attention_kernel = FlexAttentionKernel()
 
     def forward(
         self,
@@ -59,15 +77,10 @@ def forward(
         scale: float | None = None,
         return_lse: bool = False,
     ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
-        # 1. _compiled_flex_attn has to be a class variable, otherwise there will
-        #    be multiple compiled flex_attention instances, which can be slow.
-        # 2. `self._compiled_flex_attn` is not correct, `self` will be passed in
-        #    as the first argument, which will cause an error.
-        #    `FlexAttentionWrapper._compiled_flex_attn` is correct.
-        # 3. Used `return_lse` instead of `return_aux` because of easier TP module notation
-        #    to convert `lse` to be DTensor.
+        # Used `return_lse` instead of `return_aux` because of easier TP module notation
+        # to convert `lse` to be DTensor.
 
-        return FlexAttentionWrapper._compiled_flex_attn(
+        return self._flex_attention_kernel(
             q,
             k,
             v,

torchtitan/models/llama3/infra/parallelize.py

@@ -27,6 +27,7 @@
 from torchtitan.distributed import ParallelDims
 from torchtitan.distributed.activation_checkpoint import apply_ac
 from torchtitan.distributed.tensor_parallel import maybe_enable_async_tp
+from torchtitan.protocols.model import AttentionMasksType
 from torchtitan.tools.logging import logger
 
 
@@ -67,10 +68,6 @@ def parallelize_llama(
         ({parallel_dims.tp}) and 2 * CP degree ({parallel_dims.cp}).
         """
 
-    use_flex_attn = getattr(model.model_args, "use_flex_attn", False)
-    if job_config.parallelism.context_parallel_degree > 1 and use_flex_attn:
-        raise NotImplementedError("CP support for FlexAttention is still in progress.")
-
     if parallel_dims.tp_enabled:
         enable_float8_linear = "float8" in job_config.model.converters
         float8_is_rowwise = job_config.quantize.linear.float8.recipe_name in (
@@ -91,6 +88,11 @@ def parallelize_llama(
         )
         maybe_enable_async_tp(job_config, world_mesh["tp"])
 
+    use_flex_attn = getattr(model.model_args, "use_flex_attn", False)
+    if parallel_dims.cp_enabled:
+        logger.info("Applied Context Parallel to the model")
+        apply_cp(model, world_mesh["cp"], use_flex_attn)
+
     model_compile_enabled = (
         job_config.compile.enable and "model" in job_config.compile.components
     )
@@ -131,9 +133,6 @@ def parallelize_llama(
         else:
             logger.info("Applied FSDP to the model")
 
-        if parallel_dims.cp_enabled:
-            logger.info("Applied Context Parallel to the model")
-
         if job_config.training.enable_cpu_offload:
             logger.info("Applied CPU Offloading to the model")
     elif parallel_dims.dp_replicate_enabled:
@@ -328,3 +327,91 @@ def apply_ddp(
     replicate(model, device_mesh=dp_mesh, bucket_cap_mb=100)
 
     logger.info("Applied DDP to the model")
+
+
+def apply_cp(
+    model: nn.Module,
+    cp_mesh: DeviceMesh,
+    use_flex_attn: bool,
+) -> None:
+    """
+    Apply context parallelism to the model.
+    """
+    from torch.distributed.tensor.experimental._attention import (
+        _ContextParallel,
+        _enable_context_parallel_dispatcher,
+    )
+
+    # Apply context parallelism to every transformer block
+    # TODO: make seq_sim configurable once the implementation doesn't assume 2
+    # internally.
+    if use_flex_attn:
+        cp_plan = _ContextParallel(
+            seq_dim=2, attention_type=_ContextParallel.AttentionType.FLEX
+        )
+    else:
+        # This is currently required as DTensor dispatcher is not enabled to
+        # dispatch SDPA to CP implementation. We don't disable the CP
+        # dispatching in TorchTitan as it is not needed. But there is a
+        # corresponding API, _disable_context_parallel_dispatcher to do
+        # that if users have this use case.
+        _enable_context_parallel_dispatcher()
+        cp_plan = _ContextParallel(
+            seq_dim=2, attention_type=_ContextParallel.AttentionType.SDPA
+        )
+
+    for transformer_block in model.layers.values():
+        module = transformer_block.attention.inner_attention
+        if use_flex_attn:
+            module = module._flex_attention_kernel
+
+        parallelize_module(
+            module=module,
+            device_mesh=cp_mesh,
+            parallelize_plan=cp_plan,
+        )
+
+
+def cp_shard(
+    cp_mesh: DeviceMesh,
+    inputs: torch.Tensor,
+    labels: torch.Tensor,
+    attention_masks: AttentionMasksType,
+    order_sensitive_buffers: dict[str, torch.Tensor],
+    order_sensitive_buffers_seq_dims: dict[str, int],
+):
+    from torch.distributed.tensor.experimental._attention import _context_parallel_shard
+    from torch.nn.attention.flex_attention import BlockMask
+
+    load_balancer = None
+    inputs, labels = _context_parallel_shard(
+        mesh=cp_mesh,
+        buffers=(inputs, labels),
+        seq_dims=(1, 1),
+        load_balancer=load_balancer,
+    )
+
+    masks = (
+        [attention_masks]
+        if isinstance(attention_masks, BlockMask)
+        else list(attention_masks.values())
+    )
+    masks = _context_parallel_shard(
+        mesh=cp_mesh,
+        buffers=masks,
+        seq_dims=(2,) * len(masks),
+        load_balancer=load_balancer,
+    )
+    attention_masks = (
+        masks[0]
+        if isinstance(attention_masks, BlockMask)
+        else {k: v for k, v in zip(attention_masks.keys(), masks)}
+    )
+
+    order_sensitive_buffers = _context_parallel_shard(
+        mesh=cp_mesh,
+        buffers=order_sensitive_buffers,
+        seq_dims=order_sensitive_buffers_seq_dims,
+        load_balancer=load_balancer,
+    )
+    return inputs, labels, attention_masks, order_sensitive_buffers

torchtitan/models/llama3/model/args.py

@@ -55,11 +55,6 @@ def update_from_config(self, job_config: JobConfig, **kwargs) -> None:
             )
         self.max_seq_len = seq_len
 
-        if job_config.parallelism.context_parallel_degree > 1 and self.use_flex_attn:
-            raise NotImplementedError(
-                "CP support for FlexAttention is still in progress."
-            )
-
     def get_nparams_and_flops(
         self, model: nn.Module, seq_len: int
     ) -> tuple[int, float]:

torchtitan/train.py

@@ -409,12 +409,10 @@ def batch_generator(
 
             yield input_dict, labels
 
-    def forward_backward_step(
+    def post_dataloader_step(
         self, input_dict: dict[str, torch.Tensor], labels: torch.Tensor
-    ) -> torch.Tensor:
-        model_parts = self.model_parts
-        parallel_dims = self.parallel_dims
-
+    ) -> tuple[torch.Tensor, torch.Tensor, dict[str, Any], dict[str, Any],]:
+        """Post processing of the batch and label after being loaded from the dataloader."""
         inputs = input_dict["input"]
         extra_inputs = {k: v for k, v in input_dict.items() if k != "input"}
         # For arguments, like attention_masks, we have to put them in a separate
@@ -423,38 +421,53 @@ def forward_backward_step(
         extra_kwargs = {}
 
         if getattr(self.model_args, "use_flex_attn", False):
-            extra_kwargs["attention_masks"] = model_parts[0].get_attention_masks(
+            extra_kwargs["attention_masks"] = self.model_parts[0].get_attention_masks(
                 input_batch=inputs,
                 tokenizer=self.tokenizer,
                 extra_inputs=extra_inputs,
             )
+        else:
+            extra_kwargs["attention_masks"] = None
 
         # Get the order sensitive buffers
-        order_sensitive_buffers = model_parts[0].get_order_sensitive_buffers(
+        order_sensitive_buffers = self.model_parts[0].get_order_sensitive_buffers(
             inputs.size(0), inputs.size(1)
         )
-        extra_args.update(order_sensitive_buffers[0])
-
-        # apply context parallelism if cp is enabled
-        # ensure CP handles the separate freqs_cis buffer for each pp stage
-        optional_context_parallel_ctx = (
-            dist_utils.create_context_parallel_ctx(
-                cp_mesh=parallel_dims.world_mesh["cp"],
-                cp_buffers=[inputs, labels] + [m.freqs_cis for m in model_parts],
-                cp_seq_dims=[1, 1] + [0 for _ in model_parts],
-                cp_no_restore_buffers={inputs, labels},
-                cp_rotate_method=self.job_config.parallelism.context_parallel_rotate_method,
-            )
-            if parallel_dims.cp_enabled
+        cp_mesh = (
+            self.parallel_dims.world_mesh["cp"]
+            if self.parallel_dims.cp_enabled
             else None
         )
+        if cp_mesh:
+            (
+                inputs,
+                labels,
+                extra_kwargs["attention_masks"],
+                *order_sensitive_buffers,
+            ) = dist_utils.cp_shard(
+                cp_mesh,
+                inputs,
+                labels,
+                extra_kwargs["attention_masks"],
+                *order_sensitive_buffers,
+            )
+        extra_kwargs.update(order_sensitive_buffers[0])
+        return inputs, labels, extra_inputs, extra_kwargs
+
+    def forward_backward_step(
+        self, input_dict: dict[str, torch.Tensor], labels: torch.Tensor
+    ) -> torch.Tensor:
+        model_parts = self.model_parts
+        parallel_dims = self.parallel_dims
+
+        inputs, labels, extra_inputs, extra_kwargs = self.post_dataloader_step(
+            input_dict, labels
+        )
 
         if parallel_dims.pp_enabled:
             # Pipeline Parallel forward / backward inside step() call
-            with self.train_context(optional_context_parallel_ctx):
-                targets, losses = (
-                    (labels, []) if self.pp_has_last_stage else (None, None)
-                )
+            targets, losses = (labels, []) if self.pp_has_last_stage else (None, None)
+            with self.train_context():
                 if self.pp_has_first_stage:
                     self.pp_schedule.step(
                         inputs,
@@ -484,7 +497,7 @@ def forward_backward_step(
             )
         else:
             # Non-PP forward / backward
-            with self.train_context(optional_context_parallel_ctx):
+            with self.train_context():
                 assert len(model_parts) == 1
                 with self.maybe_enable_amp:
                     pred = model_parts[0](inputs, **extra_inputs, **extra_kwargs)