Support CUDA Graph for MoE models (#1233)

* Align RNG tracker with megatron

Signed-off-by: Robin Zhang <[email protected]>
Co-authored-by: Yifei Song <[email protected]>

* Fix module_params order and warmup bug in cudagraph

Signed-off-by: Robin Zhang <[email protected]>
Co-authored-by: Yifei Song <[email protected]>

* Add fp8_group argument and fix fp8 accuracy issue for cudagraph

Signed-off-by: Robin Zhang <[email protected]>
Co-authored-by: Yifei Song <[email protected]>

* Add TE modules and weights filters to support MoE models

Signed-off-by: Robin Zhang <[email protected]>
Co-authored-by: Yifei Song <[email protected]>

* Revert self.fp8

Signed-off-by: Robin Zhang <[email protected]>

* Use hooks to filter module params

Signed-off-by: Robin Zhang <[email protected]>

* Filter all TE modules in hooks

Signed-off-by: Robin Zhang <[email protected]>
Co-authored-by: Yifei Song <[email protected]>

* Format code

Signed-off-by: Robin Zhang <[email protected]>

* [pre-commit.ci] auto fixes from pre-commit.com hooks

for more information, see https://pre-commit.ci

* Update graph.py

Signed-off-by: Xin Yao <[email protected]>

* Revert CudaRNGStatesTracker

Signed-off-by: Robin Zhang <[email protected]>

* Format Update

Signed-off-by: Yifei Song <[email protected]>

* Revert "Use hooks to filter module params"

This reverts commit 73a22e2.

Signed-off-by: Yifei Song <[email protected]>

* Remove filtering module params

Signed-off-by: Robin Zhang <[email protected]>

---------

Signed-off-by: Robin Zhang <[email protected]>
Signed-off-by: Xin Yao <[email protected]>
Signed-off-by: Yifei Song <[email protected]>
Co-authored-by: Yifei Song <[email protected]>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Xin Yao <[email protected]>
Co-authored-by: Xin Yao <[email protected]>
Co-authored-by: Tim Moon <[email protected]>

transformer_engine/pytorch/fp8.py

@@ -442,16 +442,16 @@ def get_old_fp8_meta_tensors_for_recompute(cls, fp8_meta: Dict[str, Any]) -> Non
         stashed_fp8_meta = cls.fp8_tensors_recompute_buffer[fp8_meta[buffer_position_key]].popleft()
 
         # Replace amaxes and scales with stashed values for phase 2 forward
-        fp8_meta["scaling_fwd"].amax_history = stashed_fp8_meta[0]
-        fp8_meta["scaling_fwd"].scale = stashed_fp8_meta[1]
-        fp8_meta["scaling_fwd"].scale_inv = stashed_fp8_meta[2]
+        fp8_meta["scaling_fwd"].amax_history.copy_(stashed_fp8_meta[0])
+        fp8_meta["scaling_fwd"].scale.copy_(stashed_fp8_meta[1])
+        fp8_meta["scaling_fwd"].scale_inv.copy_(stashed_fp8_meta[2])
 
     @staticmethod
     def restore_fp8_meta_tensors(fp8_meta: Dict[str, Any]) -> None:
         """Restore latest scaling factors and amaxes after recompute forward run."""
-        fp8_meta["scaling_fwd"].amax_history = fp8_meta["updated_amax_history_fwd"]
-        fp8_meta["scaling_fwd"].scale = fp8_meta["updated_scale_fwd"]
-        fp8_meta["scaling_fwd"].scale_inv = fp8_meta["updated_scale_inv_fwd"]
+        fp8_meta["scaling_fwd"].amax_history.copy_(fp8_meta["updated_amax_history_fwd"])
+        fp8_meta["scaling_fwd"].scale.copy_(fp8_meta["updated_scale_fwd"])
+        fp8_meta["scaling_fwd"].scale_inv.copy_(fp8_meta["updated_scale_inv_fwd"])
 
 
 @contextmanager

transformer_engine/pytorch/graph.py

@@ -12,6 +12,7 @@
 from torch._C import _graph_pool_handle
 
 from transformer_engine.common.recipe import DelayedScaling
+from transformer_engine.pytorch.constants import dist_group_type
 from .fp8 import (
     fp8_autocast,
     FP8GlobalStateManager,
@@ -173,11 +174,14 @@ def _make_graphed_callables(
         ]
     else:
         per_callable_module_params = []
-        for c in callables:
-            for i in range(num_microbatches):
-                per_callable_module_params.append(
-                    tuple(c.parameters()) if isinstance(c, torch.nn.Module) else ()
-                )
+        for m_chunk in range(num_model_chunks):
+            for _ in range(num_microbatches):
+                for l_no in range(num_layers):
+                    per_callable_module_params.append(
+                        tuple(callables[m_chunk * num_layers + l_no].parameters())
+                        if isinstance(callables[m_chunk * num_layers + l_no], torch.nn.Module)
+                        else ()
+                    )
         assert len(per_callable_module_params) == len(flatten_sample_args)
         per_callable_static_input_surfaces = [
             flatten_sample_args[i] + per_callable_module_params[i]
@@ -201,13 +205,55 @@ def _make_graphed_callables(
     # Hopefully prevents cudnn benchmarking and other lazy-initialization cuda work
     # from ending up in any captures.
     torch.cuda.synchronize()
-    with torch.cuda.stream(torch.cuda.Stream()):
+
+    # Get warmup func and func_idx.
+    warmup_func_idx = []
+    warmup_func = []
+    if _order is None:
         for func_idx, func in enumerate(callables):
+            warmup_func_idx.append(func_idx)
+            warmup_func.append(func)
+    else:
+        fwd_idx = [0] * num_model_chunks
+        for c_id in _order:
+            if c_id > 0:
+                m_chunk = c_id - 1
+                for l_no in range(num_layers):
+                    func = callables[m_chunk * num_layers + l_no]
+                    func_idx = (m_chunk * num_microbatches * num_layers) + (
+                        fwd_idx[m_chunk] * num_layers + l_no
+                    )
+                    warmup_func_idx.append(func_idx)
+                    warmup_func.append(func)
+                fwd_idx[m_chunk] += 1
+    assert len(warmup_func) == len(
+        sample_args
+    ), f"Warmup runs {len(warmup_func)} don't match args {len(sample_args)}."
+    assert len(warmup_func_idx) == len(
+        set(warmup_func_idx)
+    ), f"Warmup runs {len(warmup_func)} but only {len(set(warmup_func_idx))} are unique."
+
+    # Filter the TE modules that cudagraph can access.
+    visited_te_modules = set()
+
+    def hook_fn(module, inputs, outputs):  # pylint: disable=unused-argument
+        if isinstance(module, TransformerEngineBaseModule):
+            visited_te_modules.add(module)
+
+    # Run warmup and do the above filtering.
+    with torch.cuda.stream(torch.cuda.Stream()):
+        for func_idx, func in zip(warmup_func_idx, warmup_func):
             args = sample_args[func_idx]
             kwargs = sample_kwargs[func_idx]
             static_input_surface = per_callable_static_input_surfaces[func_idx]
             for _ in range(num_warmup_iters):
+                hooks = []
+                for module in func.modules():
+                    hook = module.register_forward_hook(hook_fn)
+                    hooks.append(hook)
                 outputs, _ = _tree_flatten(func(*args, **kwargs))
+                for hook in hooks:
+                    hook.remove()
                 grad_inputs = torch.autograd.grad(
                     outputs=tuple(o for o in outputs if o.requires_grad),
                     inputs=tuple(i for i in static_input_surface if i.requires_grad),
@@ -216,6 +262,11 @@ def _make_graphed_callables(
                     allow_unused=allow_unused_input,
                 )
                 del outputs, grad_inputs
+            # The following code is added specifically for MCore's special requirements,
+            # aimed at preventing warmup from altering the control flow.
+            for module in func.modules():
+                if hasattr(module, "is_first_microbatch"):
+                    module.is_first_microbatch = True
     torch.cuda.synchronize()
 
     # All captures here share a mempool. To avoid replays corrupting each other's memory,
@@ -462,6 +513,19 @@ def new_fwd(*user_args, **user_kwargs):
                                 isinstance(m, TransformerEngineBaseModule)
                                 and FP8GlobalStateManager.is_fp8_enabled()
                             ):
+                                if m not in visited_te_modules:
+                                    # Only Set the FP8 meta for the modules included by forward
+                                    continue
+                                fp8_recipe = FP8GlobalStateManager.get_fp8_recipe()
+                                from transformer_engine.pytorch.attention import DotProductAttention
+
+                                if (
+                                    isinstance(m, DotProductAttention)
+                                    and not fp8_recipe.fp8_mha
+                                    and not fp8_recipe.fp8_dpa
+                                ):
+                                    # Don't need to update FP8 meta for non-FP8 DPA
+                                    continue
                                 m.fp8_meta["fp8_group"] = FP8GlobalStateManager.get_fp8_group()
                                 m.fp8_meta["recipe"] = FP8GlobalStateManager.get_fp8_recipe()
                                 FP8GlobalStateManager.add_fp8_tensors_to_global_buffer(
@@ -538,6 +602,7 @@ def make_graphed_callables(
     fp8_enabled: bool = False,
     fp8_calibrating: bool = False,
     fp8_recipe: Optional[DelayedScaling] = None,
+    fp8_group: Optional[dist_group_type] = None,
     fp8_weight_caching: bool = False,
     _order: Optional[List[int]] = None,
     pool: Optional[Tuple[int, ...]] = None,
@@ -579,6 +644,9 @@ def make_graphed_callables(
                      using a higher precision.
     fp8_recipe: recipe.DelayedScaling, default = `None`
                 recipe used for FP8 training.
+    fp8_group: torch._C._distributed_c10d.ProcessGroup, default = `None`
+               distributed group over which amaxes for the fp8 tensors
+               are reduced at the end of each training step.
     fp8_weight_caching: bool, default = `False`
                         Whether or not to cache FP8 weights across microbatches. if set to `True`,
                         the `is_first_microbatch` boolean argument must be passed into the forward
@@ -607,7 +675,11 @@ def wrap_autocast(block):
 
         def forward_func(*args, **kwargs):
             with fp8_autocast(
-                enabled=fp8_enabled, calibrating=fp8_calibrating, fp8_recipe=fp8_recipe, _graph=True
+                enabled=fp8_enabled,
+                calibrating=fp8_calibrating,
+                fp8_recipe=fp8_recipe,
+                fp8_group=fp8_group,
+                _graph=True,
             ):
                 outputs = old_forward(*args, **kwargs)
             return outputs

transformer_engine/pytorch/module/layernorm_linear.py

@@ -1152,7 +1152,10 @@ def forward(
                                produced)
         """
 
-        skip_fp8_weight_update = FP8GlobalStateManager.get_skip_fp8_weight_update_tensor()
+        if FP8GlobalStateManager.fp8_graph_capturing():
+            skip_fp8_weight_update = FP8GlobalStateManager.get_skip_fp8_weight_update_tensor()
+        else:
+            skip_fp8_weight_update = None
         if skip_fp8_weight_update is not None:
             is_first_microbatch = False
 

transformer_engine/pytorch/module/layernorm_mlp.py

@@ -1484,7 +1484,10 @@ def forward(
                                produced)
         """
 
-        skip_fp8_weight_update = FP8GlobalStateManager.get_skip_fp8_weight_update_tensor()
+        if FP8GlobalStateManager.fp8_graph_capturing():
+            skip_fp8_weight_update = FP8GlobalStateManager.get_skip_fp8_weight_update_tensor()
+        else:
+            skip_fp8_weight_update = None
         if skip_fp8_weight_update is not None:
             is_first_microbatch = False
 

transformer_engine/pytorch/module/linear.py

@@ -938,8 +938,10 @@ def forward(
                                first microbatch (since it is the first gradient being
                                produced)
         """
-
-        skip_fp8_weight_update = FP8GlobalStateManager.get_skip_fp8_weight_update_tensor()
+        if FP8GlobalStateManager.fp8_graph_capturing():
+            skip_fp8_weight_update = FP8GlobalStateManager.get_skip_fp8_weight_update_tensor()
+        else:
+            skip_fp8_weight_update = None
         if skip_fp8_weight_update is not None:
             is_first_microbatch = False