New mcore transformer block spec

nemo/collections/multimodal/models/multimodal_llm/neva/neva_model.py

@@ -728,8 +728,7 @@ def dummy():
                 config=self.transformer_config,
                 transformer_layer_spec=get_specs(
                     self.spec_name,
-                    self.transformer_config.num_moe_experts,
-                    self.transformer_config.moe_grouped_gemm,
+                    self.transformer_config,
                     self.transformer_engine,
                 ),
                 vocab_size=self.cfg.get('override_vocab_size', self.padded_vocab_size),

nemo/collections/multimodal/models/vision_language_foundation/clip/megatron_clip_models.py

@@ -501,8 +501,7 @@ def __init__(
                 add_class_token = True
             vision_layer_spec = get_specs(
                 model_cfg.text.get('name', ''),
-                vision_transformer_config.num_moe_experts,
-                vision_transformer_config.moe_grouped_gemm,
+                vision_transformer_config,
                 model_cfg.get('transformer_engine', True),
             )
             vision_layer_spec.submodules.self_attention.params['attn_mask_type'] = MCoreAttnMaskType.no_mask
@@ -527,8 +526,7 @@ def __init__(
                 config=text_transformer_config,
                 transformer_layer_spec=get_specs(
                     model_cfg.text.get('name', ''),
-                    text_transformer_config.num_moe_experts,
-                    text_transformer_config.moe_grouped_gemm,
+                    text_transformer_config,
                     model_cfg.get('transformer_engine', True),
                 ),
                 vocab_size=model_cfg.text.get('override_vocab_size', padded_vocab_size),

nemo/collections/nlp/models/language_modeling/megatron/gpt_full_te_layer_autocast_spec.py

@@ -35,7 +35,9 @@
 
 try:
     from megatron.core import parallel_state, tensor_parallel
+    from megatron.core.fusions.fused_layer_norm import FusedLayerNorm
     from megatron.core.transformer.spec_utils import ModuleSpec
+    from megatron.core.transformer.transformer_block import TransformerBlockSubmodules, get_num_layers_to_build
     from megatron.core.transformer.transformer_layer import BaseTransformerLayer
     from megatron.core.transformer.utils import make_sharded_tensors_for_checkpoint
 
@@ -322,8 +324,10 @@
 
 
 # Use this spec to use the full Transformer layer from Transformer Engine
-def get_gpt_full_te_layer_autocast_spec() -> ModuleSpec:
+def get_gpt_full_te_layer_autocast_spec(transformer_config) -> ModuleSpec:
     if not HAVE_MEGATRON_CORE or not HAVE_TE:
         raise ImportError(IMPORT_ERROR)
-
-    return ModuleSpec(module=TETransformerLayerAutocast)
+    num_layers = get_num_layers_to_build(transformer_config)
+    return TransformerBlockSubmodules(
+        layer_specs=[ModuleSpec(module=TETransformerLayerAutocast)] * num_layers, layer_norm=FusedLayerNorm
+    )

nemo/collections/nlp/models/language_modeling/megatron_gpt_model.py

@@ -138,7 +138,11 @@ def mcore_supports_moe() -> bool:
 
 
 ## TODO: This function will not work if TE is not installed
-def get_specs(spec_name, num_experts=None, moe_grouped_gemm=False, use_te=True, hyena_cfg: Dict = None):
+def get_specs(spec_name, transformer_config=None, use_te=True, hyena_cfg: Dict = None):
+    # else cases for backwards compatibility with neva
+    num_experts = transformer_config.num_moe_experts if transformer_config else None
+    moe_grouped_gemm = transformer_config.moe_grouped_gemm if transformer_config else False
+
     if num_experts is not None:
         assert mcore_supports_moe(), "Megatron-core >= v0.5.0 is required for MoE"
 
@@ -148,7 +152,7 @@ def get_specs(spec_name, num_experts=None, moe_grouped_gemm=False, use_te=True,
         "": get_gpt_layer_local_spec(num_experts, moe_grouped_gemm),
         "te_gpt": get_gpt_layer_with_transformer_engine_spec(num_experts, moe_grouped_gemm),
         "megatron_falcon_gpt": get_falcon_layer_spec(),
-        "megatron_gpt_full_te_layer_autocast": get_gpt_full_te_layer_autocast_spec(),
+        "megatron_gpt_full_te_layer_autocast": get_gpt_full_te_layer_autocast_spec(transformer_config),
         "modelopt": get_gpt_layer_modelopt_spec(num_experts),
         "te_gpt_hyena": get_gpt_layer_with_te_and_hyena_spec(hyena_cfg),
     }
@@ -415,8 +419,7 @@ def model_provider_func(self, pre_process, post_process):
                 config=self.transformer_config,
                 transformer_layer_spec=get_specs(
                     self.spec_name,
-                    self.transformer_config.num_moe_experts,
-                    self.transformer_config.moe_grouped_gemm,
+                    self.transformer_config,
                     self.transformer_engine,
                     self.cfg.get('hyena', None),
                 ),