[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: 673d743
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 +441,65 @@ 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 (
+        _HeadTailLoadBalancer,
+        _PTRRLoadBalancer,
+    )
+    from torch.nn.attention.flex_attention import BlockMask
+
+    seq_len = inputs.size(1)
+    cp_world_size = cp_mesh.size(0)
+    if isinstance(attention_masks, BlockMask):
+        load_balancer = _PTRRLoadBalancer(attention_masks, cp_world_size)
+    else:
+        # For multiple BlockMasks or SDPA, we use the _HeadTailLoadBalancer.
+        load_balancer = _HeadTailLoadBalancer(
+            seq_len, cp_world_size, cp_mesh.device_type
+        )
+
+    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/models/llama3/model/model.py

@@ -92,8 +92,7 @@ def reshape_for_broadcast(freqs_cis: torch.Tensor, x: torch.Tensor) -> torch.Ten
     This function reshapes the frequency tensor to have the same shape as the target tensor 'x'
     for the purpose of broadcasting the frequency tensor during element-wise operations.
 
-    The input freqs_cis tensor is assumed to be of shape (max_seqlen, dim),
-    and the first seqlen elements will be sliced, but dim must match x.
+    The input freqs_cis tensor is assumed to be of shape (batch_size, seqlen, dim).
 
     Args:
         freqs_cis (torch.Tensor): Frequency tensor to be reshaped.
@@ -104,10 +103,10 @@ def reshape_for_broadcast(freqs_cis: torch.Tensor, x: torch.Tensor) -> torch.Ten
     """
     ndim = x.ndim
     assert ndim > 1
+    batch_size = x.shape[0]
     seqlen = x.shape[1]
-    freqs_cis = freqs_cis[0:seqlen]
-    assert freqs_cis.shape == (seqlen, x.shape[-1])
-    shape = [d if i == 1 or i == ndim - 1 else 1 for i, d in enumerate(x.shape)]
+    assert freqs_cis.shape == (batch_size, seqlen, x.shape[-1])
+    shape = [d if i in (0, 1, ndim - 1) else 1 for i, d in enumerate(x.shape)]
     return freqs_cis.view(*shape)
 
 
@@ -474,9 +473,18 @@ def get_attention_masks(
             and_masks(*mask_mods), B, None, input_batch.shape[1], input_batch.shape[1]
         )
 
+    def get_order_sensitive_buffers(
+        self,
+        batch_size: int,
+        seq_len: int,
+    ) -> tuple[dict[str, torch.Tensor], dict[str, int]]:
+        freqs_cis = self.freqs_cis[:seq_len].repeat(batch_size, 1, 1)
+        return ({"freqs_cis": freqs_cis}, {"freqs_cis": 1})
+
     def forward(
         self,
         tokens: torch.Tensor,
+        freqs_cis: torch.Tensor,
         attention_masks: AttentionMasksType | None = None,
     ):
         """
@@ -496,7 +504,7 @@ def forward(
         h = self.tok_embeddings(tokens) if self.tok_embeddings else tokens
 
         for layer in self.layers.values():
-            h = layer(h, self.freqs_cis, attention_masks=attention_masks)
+            h = layer(h, freqs_cis, attention_masks=attention_masks)
 
         h = self.norm(h) if self.norm else h
         output = self.output(h) if self.output else h

torchtitan/protocols/model.py

@@ -70,3 +70,10 @@ def get_attention_masks(
         raise NotImplementedError(
             "This model does not support attention masking/Flex Attention."
         )
+
+    def get_order_sensitive_buffers(
+        self,
+        batch_size: int,
+        seq_len: int,
+    ) -> tuple[dict[str, torch.Tensor], dict[str, int]]:
+        return ({}, {})