Support group_mlp offload for TE op fuser

[lhb8125]

What does this PR do ?

⚠️ For major changes (either in lines of code or in its impact), please make sure to first share a design doc with the team. If you're unsure what's the best way to do so, contact the @mcore-oncall.

Contribution process

Pre-checks

• I have added relevant unit tests
• I have added relevant functional tests
• I have added proper typing to my code Typing guidelines
• I have added relevant documentation
• I have run the autoformatter.sh on my PR

Code review

Feel free to message or comment the @mcore-oncall to help accelerate your merge into main. The less complex your PR is, the faster it will be approved and merged!

All PRs start as draft. If you open a non-draft PR, it will be automatically converted to draft.

Step 1: Mark PR as "Ready for Review"

1. When your PR is ready, click Ready for Review.
2. An oncall reviewer is auto-assigned and expert reviewers are notified based on your changes.
   • Some PRs may jump straight to step 2. This is determined by .github/CODEOWNERS.

⚠️ Only mark as ready once merge-conflicts are resolved and the CI is passing.
Final Review might get declined if these requirements are not fulfilled.

Step 2: Final Review

For PRs that change megatron/core, once all expert reviewers have approved, the Final Review label is applied automatically and final reviewers are assigned.

For PRs outside megatron/core, this step is skipped.

Step 3: Approved

Once all required reviewers have approved, the Approved label is applied automatically.

Merge

Any member of mcore-engineers will be able to merge your PR.

For MRs into `dev` branch

The proposed review process for `dev` branch is under active discussion.

MRs are mergable after one approval by either eharper@nvidia.com or zijiey@nvidia.com.

[timmoon10]

I'm not sure this is correct. GroupedTensor.prepare_for_saving removes all the buffers within the GroupedTensor, so we need to call restore_from_saved even if no buffers were moved to CPU.

[timmoon10]

Nit: This implementation is correct, but we could make it a bit nicer by using a consistent structure for offloaded and non-offloaded buffers.

Suggested change

	if not isinstance(buffer, torch.Tensor):
	buffer_states.append(None)
	elif _regular_tensor_needs_offloading(buffer, self.min_offloaded_tensor_size):
	buffer_states.append(
	self._offload_tensor(buffer, pin_memory=pin_memory, use_cpu_pool=use_cpu_pool)
	)
	buffer.record_stream(self.d2h_stream)
	offload_nbytes += buffer.numel() * buffer.element_size()
	else:
	buffer_states.append((_TE_GROUPED_TENSOR_RESIDENT_BUFFER_STATE, buffer))
	is_offloaded = (
	buffer is not None
	and _regular_tensor_needs_offloading(buffer, self.min_offloaded_tensor_size)
	)
	buffer_state = (
	self._offload_tensor(buffer, pin_memory, use_cpu_pool=use_cpu_pool)
	if is_offloaded
	else buffer
	)
	buffer_states.append((is_offloaded, buffer_state))

The reloading function becomes:

    def _reload_te_grouped_tensor(self, state, non_blocking=None):
        """Reload TE GroupedTensor backing buffers and reconstruct the wrapper."""
        debug_rank("------reload TE GroupedTensor")
        _, tensor_obj, buffer_states, _ = state

        buffers = []
        for is_offloaded, buffer_state in buffer_states:
            if is_offloaded:
                buffers.append(self._reload_tensor(buffer_state, non_blocking=non_blocking))
            else:
                buffers.append(buffer_state)

        return te_grouped_tensor_restore_from_saved(tensor_obj, buffers)

[timmoon10]

This looks correct, but I'm curious why we don't have similar logic for QuantizedTensor, which should have similar problems as GroupedTensor. If we already have TE infrastructure for CPU offloading for quantized tensors, then we should use that for grouped tensors instead of implementing custom logic in Mcore.