Fix fake_initialize_model_parallel for MoE models

* Make the use of RankGenerator consistent with recent Mcore changes !1940
( NVIDIA/Megatron-LM@7f22e21)
  - use ep=1 for decoder_rank_generator, making it treat EP as part of DP
  - define a new expert_decoder_rank_generator to handle EP groups/ranks only

Signed-off-by: Guyue Huang <[email protected]>

nemo/collections/nlp/modules/common/megatron/megatron_init.py

@@ -349,23 +349,42 @@ def fake_initialize_model_parallel(
 
     decoder_rank_generator = RankGenerator(
         tp=tensor_model_parallel_size,
-        ep=expert_model_parallel_size_,
+        ep=1,
         dp=data_parallel_size,
         pp=pipeline_model_parallel_size,
         cp=context_parallel_size,
         order='tp-pp-dp' if use_tp_pp_dp_mapping else 'tp-cp-ep-dp-pp',
         rank_offset=encoder_world_size,
     )
+
+    # the default setting uses DEP (expert-parallel ranks for FFN are data-parallel ranks for Attention. This definition follows that rule.)
+    expert_decoder_rank_generator = RankGenerator(
+        tp=tensor_model_parallel_size, # the same as Attention part
+        ep=expert_model_parallel_size_,
+        dp=(decoder_world_size // (expert_model_parallel_size_ * tensor_model_parallel_size * pipeline_model_parallel_size)),
+        pp=pipeline_model_parallel_size,
+        cp=1,
+        order='tp-pp-dp' if use_tp_pp_dp_mapping else 'tp-cp-ep-dp-pp',
+        rank_offset=encoder_world_size,
+    )
 
-    def generator_wrapper(group_type, **kwargs):
+    assert decoder_rank_generator.get_ranks("pp") == expert_decoder_rank_generator.get_ranks(
+        "pp"
+    ), f"Pipeline parallel groups are expected to be the same for Non-Expert and Expert part, \
+    but got {decoder_rank_generator.get_ranks('pp')} and {expert_decoder_rank_generator.get_ranks('pp')}"
+
+    def generator_wrapper(group_type, is_expert=False, **kwargs):
         from itertools import cycle
 
         """The `RankGenerator` class produces a hyper-rectangle for a given set of
         tensor, pipeline, data, expert, and context parallelism. If we have an encoder,
         in addition to the default decoder, we essentially instantiate two `RankGenerator`
         classes to construct the parallelism for each module separately, and we then have
         to stitch them together for the right groups. For now, this means pp and tp-pp."""
-        d_ranks = decoder_rank_generator.get_ranks(group_type, **kwargs)
+        if is_expert:
+            d_ranks = expert_decoder_rank_generator.get_ranks(group_type, **kwargs)
+        else:
+            d_ranks = decoder_rank_generator.get_ranks(group_type, **kwargs)
         if encoder_rank_generator is None:
             for x in d_ranks:
                 yield x
@@ -446,7 +465,7 @@ def generator_wrapper(group_type, **kwargs):
     # EP rank
     expert_model_parallel_rank = 0
     if expert_model_parallel_size_ is not None and expert_model_parallel_size_ > 1:
-        for ranks in generator_wrapper('ep', independent_ep=True):
+        for ranks in generator_wrapper('ep', is_expert=True):
             if rank in ranks:
                 expert_model_parallel_rank = list(ranks).index(rank)