Prep for refit (NVIDIA#2590)

Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>
Co-authored-by: root <[email protected]>

Author: 9 people

gpt_builders.py

@@ -21,7 +21,7 @@
 # NOTE: Loading `megatron.legacy.model` earlier fails due to circular import
 
 
-def gpt_builder(args, pre_process, post_process, vp_stage=None, config=None):
+def gpt_builder(args, pre_process, post_process, vp_stage=None, config=None, pg_collection=None):
     print_rank_0('building GPT model ...')
     if config is None:
         if args.yaml_cfg is not None:
@@ -93,6 +93,7 @@ def gpt_builder(args, pre_process, post_process, vp_stage=None, config=None):
             rope_scaling=args.use_rope_scaling,
             mtp_block_spec=mtp_block_spec,
             vp_stage=vp_stage,
+            pg_collection=pg_collection,
         )
 
     return model

mamba_builders.py

@@ -8,7 +8,7 @@
 from megatron.training.arguments import core_transformer_config_from_args
 from megatron.core.models.mamba.mamba_layer_specs import mamba_inference_stack_spec
 
-def mamba_builder(args, pre_process, post_process, vp_stage=None, config=None):
+def mamba_builder(args, pre_process, post_process, vp_stage=None, config=None, pg_collection=None):
     print_rank_0('building MAMBA model ...')
     if config is None:
         config = core_transformer_config_from_args(args, TransformerConfig)
@@ -37,6 +37,7 @@ def mamba_builder(args, pre_process, post_process, vp_stage=None, config=None):
         position_embedding_type=args.position_embedding_type,
         rotary_percent=args.rotary_percent,
         rotary_base=args.rotary_base,
+        pg_collection=pg_collection,
     )
 
     for l in range(model.decoder.num_layers_per_pipeline_rank):

megatron/core/inference/engines/dynamic_engine.py

@@ -19,7 +19,6 @@
 from torch import Tensor
 from torch.cuda.nvtx import range_pop, range_push
 
-from megatron.core import parallel_state
 from megatron.core.inference.contexts.dynamic_context import (
     DynamicInferenceContext,
     MaxSequenceLengthOverflowError,
@@ -40,8 +39,16 @@
     TextGenerationController,
 )
 from megatron.core.inference.utils import Counter, await_process_event
+from megatron.core.process_groups_config import ProcessGroupCollection
 from megatron.core.transformer.cuda_graphs import delete_cuda_graphs
-from megatron.core.utils import get_asyncio_loop, internal_api, trace_async_exceptions
+from megatron.core.utils import (
+    get_asyncio_loop,
+    get_pg_rank,
+    get_pg_size,
+    get_pg_src_rank,
+    internal_api,
+    trace_async_exceptions,
+)
 
 try:
     from tqdm import tqdm
@@ -136,6 +143,7 @@ def __init__(
         track_paused_request_events: bool = False,
         enable_chunked_prefill: bool = True,
         inference_logging_step_interval: int = 0,
+        pg_collection: Optional[ProcessGroupCollection] = None,
     ):
 
         assert isinstance(
@@ -159,6 +167,11 @@ def __init__(
                 controller.inference_wrapped_model.model.config.enable_cuda_graph
             )
 
+        if pg_collection is not None:
+            self.pg_collection = pg_collection
+        else:
+            self.pg_collection = ProcessGroupCollection.use_mpu_process_groups()
+
         # Initialization options.
         self.controller = controller
         self.context = context
@@ -378,15 +391,15 @@ async def start_listening_to_data_parallel_coordinator(
         self.zmq_sockets = []  # keep track of all sockets created by this engine
 
         # Get world info.
-        dp_group = parallel_state.get_data_parallel_group()
-        dp_src = parallel_state.get_data_parallel_src_rank()
-        dp_size = parallel_state.get_data_parallel_world_size()
-        dp_rank = parallel_state.get_data_parallel_rank()
+        dp_group = self.pg_collection.dp
+        dp_src = get_pg_src_rank(dp_group)
+        dp_size = get_pg_size(self.pg_collection.dp)
+        dp_rank = get_pg_rank(self.pg_collection.dp)
 
-        mp_group = parallel_state.get_model_parallel_group()
-        mp_src = parallel_state.get_model_parallel_src_rank()
-        tp_rank = parallel_state.get_tensor_model_parallel_rank()
-        pp_rank = parallel_state.get_pipeline_model_parallel_rank()
+        mp_group = self.pg_collection.mp
+        mp_src = get_pg_src_rank(mp_group)
+        tp_rank = get_pg_rank(self.pg_collection.tp)
+        pp_rank = get_pg_rank(self.pg_collection.pp)
 
         self.is_mp_coordinator = tp_rank == 0 and pp_rank == 0
         self.is_dp_coordinator = (dp_rank == 0) and self.is_mp_coordinator
@@ -400,7 +413,7 @@ async def start_listening_to_data_parallel_coordinator(
                 args=(
                     coordinator_ready_event,
                     inference_coordinator_port,
-                    parallel_state.get_data_parallel_world_size(),
+                    get_pg_size(self.pg_collection.dp),
                 ),
             )
             self.inference_coordinator_process.start()

megatron/core/utils.py

@@ -555,6 +555,23 @@ def get_pg_rank(group=None):
     return group.rank()
 
 
+def get_pg_src_rank(group=None):
+    """Calculate the global rank corresponding to the first local rank
+    in the given process group.
+
+    Args:
+        group: Process group to query. If None or distributed is not initialized,
+            returns 0.
+
+    Returns:
+        int: The first (source) global rank in the group.
+    """
+    if not torch.distributed.is_initialized() or group is None:
+        return 0
+    ranks = torch.distributed.get_process_group_ranks(group)
+    return ranks[0]
+
+
 def get_attr_wrapped_model(model, attr, allow_none=True, return_model_obj=False):
     """Get an attribute from a wrapped model.
     If return_model_obj is true, return the object that has the 'attr' attribute;

megatron/rl/inference/megatron.py

@@ -8,6 +8,7 @@
 from pydantic import PrivateAttr
 
 from megatron.core import parallel_state
+from megatron.core.utils import get_attr_wrapped_model
 from megatron.core.inference.contexts.dynamic_context import DynamicInferenceContext
 from megatron.core.inference.engines.abstract_engine import AbstractEngine
 from megatron.core.inference.engines.dynamic_engine import DynamicInferenceEngine
@@ -26,7 +27,11 @@
 from megatron.core.models.gpt.gpt_model import GPTModel
 from megatron.core.ssm.mamba_hybrid_layer_allocation import Symbols
 from megatron.core.transformer.module import MegatronModule
-from megatron.core.utils import get_mamba_inference_state_config_from_model, log_single_rank
+from megatron.core.pipeline_parallel.utils import (
+    is_pp_first_stage,
+    is_pp_last_stage,
+)
+from megatron.core.utils import get_mamba_inference_state_config_from_model, log_single_rank, get_pg_size
 from megatron.training import get_wandb_writer
 from megatron.training.global_vars import get_args, get_tokenizer
 
@@ -109,6 +114,16 @@ def get_dynamic_inference_engine(args: Namespace, model: MegatronModule, inferen
 
     mamba_inference_state_config = get_mamba_inference_state_config_from_model(model)
 
+    # DynamicInferenceContext must use the inference model's TP size, not the
+    # training TP size from global args. The inference model may have a custom
+    # ProcessGroupCollection with a different TP size.
+    pg_collection = get_attr_wrapped_model(model, "pg_collection")
+    tp_group = getattr(pg_collection, 'tp', None) if pg_collection is not None else None
+    if tp_group is not None:
+        inference_tp_size = get_pg_size(tp_group)
+    else:
+        inference_tp_size = args.tensor_model_parallel_size
+
     # Inference context.
     inference_context = DynamicInferenceContext(
         params_dtype=args.params_dtype,
@@ -126,7 +141,7 @@ def get_dynamic_inference_engine(args: Namespace, model: MegatronModule, inferen
         block_size_tokens=args.inference_dynamic_batching_block_size,
         buffer_size_gb=args.inference_dynamic_batching_buffer_size_gb,
         max_tokens=args.inference_dynamic_batching_max_tokens,
-        tensor_model_parallel_size=args.tensor_model_parallel_size,
+        tensor_model_parallel_size=inference_tp_size,
         materialize_only_last_token_logits=True,
         mamba_inference_state_config=mamba_inference_state_config,
         cache_mla_latent=args.multi_latent_attention and args.cache_mla_latents,
@@ -143,7 +158,7 @@ def get_dynamic_inference_engine(args: Namespace, model: MegatronModule, inferen
     inference_wrapped_model = GPTInferenceWrapper(model, args, inference_context)
 
     inference_wrapped_model.model_is_pipeline_parallel = not (
-        parallel_state.is_pipeline_first_stage() and parallel_state.is_pipeline_last_stage()
+        is_pp_first_stage(pg_collection.pp) and is_pp_last_stage(pg_collection.pp)
     )
 
     text_generation_controller = SimpleTextGenerationController(
@@ -156,6 +171,7 @@ def get_dynamic_inference_engine(args: Namespace, model: MegatronModule, inferen
         enable_cuda_graph=enable_cuda_graph,
         random_seed=args.seed,
         inference_logging_step_interval=inference_logging_step_interval,
+        pg_collection=pg_collection,
     )
 
 

megatron/training/training.py

@@ -52,9 +52,18 @@
 from megatron.core.utils import (
     check_param_hashes_across_dp_replicas,
     get_model_config,
+    get_pg_size,
+    get_pg_rank,
     StragglerDetector,
 )
 from megatron.core.fp8_utils import correct_amax_history_if_needed
+from megatron.core.process_groups_config import ProcessGroupCollection
+from megatron.core.pipeline_parallel.utils import (
+    is_pp_first_stage,
+    is_pp_last_stage,
+    is_vp_first_stage,
+    is_vp_last_stage,
+)
 from megatron.training.checkpointing import load_checkpoint
 from megatron.training.checkpointing import save_checkpoint
 from megatron.training.checkpointing import checkpoint_exists
@@ -886,10 +895,12 @@ def update_train_iters(args):
     print_rank_0(f'setting training iterations to {args.train_iters}')
 
 
-def get_model(model_provider_func, model_type=ModelType.encoder_or_decoder, wrap_with_ddp=True):
+def get_model(model_provider_func, model_type=ModelType.encoder_or_decoder, wrap_with_ddp=True, config=None, pg_collection=None):
     """Build the model."""
     args = get_args()
     args.model_type = model_type
+    if pg_collection is None:
+        pg_collection = ProcessGroupCollection.use_mpu_process_groups()
 
     if has_nvidia_modelopt:
         from megatron.post_training.checkpointing import has_modelopt_state
@@ -906,23 +917,38 @@ def get_model(model_provider_func, model_type=ModelType.encoder_or_decoder, wrap
     # Build model.
     def build_model():
         if (
-            mpu.get_pipeline_model_parallel_world_size() > 1
+            get_pg_size(pg_collection.pp) > 1
             and args.virtual_pipeline_model_parallel_size is not None
         ):
             model = []
-            for i in range(args.virtual_pipeline_model_parallel_size):
+            vp_size = args.virtual_pipeline_model_parallel_size
+            for i in range(vp_size):
                 # Set pre_process and post_process only after virtual rank is set.
-                pre_process = mpu.is_pipeline_first_stage(ignore_virtual=False, vp_stage=i)
-                post_process = mpu.is_pipeline_last_stage(ignore_virtual=False, vp_stage=i)
+                pre_process = is_pp_first_stage(pg_collection.pp) and is_vp_first_stage(
+                    vp_stage=i, vp_size=vp_size
+                )
+                post_process = is_pp_last_stage(pg_collection.pp) and is_vp_last_stage(
+                    vp_stage=i, vp_size=vp_size
+                )
                 this_model = model_provider_func(
-                    pre_process=pre_process, post_process=post_process, vp_stage=i)
+                    pre_process=pre_process,
+                    post_process=post_process,
+                    vp_stage=i,
+                    config=config,
+                    pg_collection=pg_collection,
+                )
                 this_model.model_type = model_type
                 this_model.vp_stage = i
                 model.append(this_model)
         else:
-            pre_process = mpu.is_pipeline_first_stage()
-            post_process = mpu.is_pipeline_last_stage()
-            model = model_provider_func(pre_process=pre_process, post_process=post_process)
+            pre_process = is_pp_first_stage(pg_collection.pp)
+            post_process = is_pp_last_stage(pg_collection.pp)
+            model = model_provider_func(
+                pre_process=pre_process,
+                post_process=post_process,
+                config=config,
+                pg_collection=pg_collection,
+            )
             model.model_type = model_type
         return model
 
@@ -947,12 +973,12 @@ def build_model():
     num_parameters = sum(
         [sum([p.nelement() for p in model_module.parameters()]) for model_module in model]
     )
-    if mpu.get_data_parallel_rank() == 0 and mpu.get_context_parallel_rank() == 0:
+    if get_pg_rank(pg_collection.dp) == 0 and get_pg_rank(pg_collection.cp) == 0:
         print(
             ' > number of parameters on (tensor, pipeline) '
             'model parallel rank ({}, {}): {}'.format(
-                mpu.get_tensor_model_parallel_rank(),
-                mpu.get_pipeline_model_parallel_rank(),
+                get_pg_rank(pg_collection.tp),
+                get_pg_rank(pg_collection.pp),
                 num_parameters,
             ),
             flush=True,

model_provider.py

@@ -22,7 +22,7 @@
 
 
 def model_provider(
-    model_builder: Callable, pre_process=True, post_process=True, vp_stage: Optional[int] = None
+    model_builder: Callable, pre_process=True, post_process=True, vp_stage: Optional[int] = None, config=None, pg_collection=None,
 ) -> Union[GPTModel, megatron.legacy.model.GPTModel, MambaModel]:
     """Builds the model.
 
@@ -64,7 +64,7 @@ def oom_observer(device, alloc, device_alloc, device_free):
         # [ModelOpt]: Use custom builder + spec when modelopt is enabled
         model_builder = modelopt_gpt_mamba_builder
 
-    return model_builder(args, pre_process, post_process, vp_stage)
+    return model_builder(args, pre_process, post_process, vp_stage, config=config, pg_collection=pg_collection)
 
 
 def count_parameters_in_layer(model, layer_name):

pretrain_bert.py

@@ -28,13 +28,14 @@
 from megatron.core.tokenizers import MegatronTokenizer
 
 
-def model_provider(pre_process=True, post_process=True, vp_stage=None):
+def model_provider(pre_process=True, post_process=True, vp_stage=None, config=None, pg_collection=None):
     """Build the model."""
 
     print_rank_0('building BERT model ...')
 
     args = get_args()
-    config = core_transformer_config_from_args(args)
+    if config is None:
+        config = core_transformer_config_from_args(args)
     num_tokentypes = 2 if args.bert_binary_head else 0
 
     if args.use_legacy_models:

pretrain_t5.py

@@ -66,7 +66,12 @@
 
 
 def model_provider(
-    pre_process=True, post_process=True, add_encoder=True, add_decoder=True
+    pre_process=True,
+    post_process=True,
+    add_encoder=True,
+    add_decoder=True,
+    config=None,
+    pg_collection=None,
 ) -> Union[megatron.legacy.model.T5Model, T5Model]:
     """Builds the model.
 
@@ -83,7 +88,8 @@ def model_provider(
 
     args = get_args()
 
-    config = core_transformer_config_from_args(args)
+    if config is None:
+        config = core_transformer_config_from_args(args)
     if args.use_legacy_models:
         model = megatron.legacy.model.T5Model(
             config=config,

pretrain_vlm.py

@@ -43,7 +43,13 @@
 
 
 def model_provider(
-    pre_process=True, post_process=True, add_encoder=True, add_decoder=True, parallel_output=True
+    pre_process=True,
+    post_process=True,
+    add_encoder=True,
+    add_decoder=True,
+    parallel_output=True,
+    config=None,
+    pg_collection=None,
 ) -> LLaVAModel:
     """Builds the model.
 
@@ -100,7 +106,10 @@ def model_provider(
     args.max_position_embeddings = max(args.max_position_embeddings, args.decoder_seq_length)
 
     print_rank_0('building a multimodal model ...')
-    language_transformer_config = core_transformer_config_from_args(get_args())
+    if config is None:
+        language_transformer_config = core_transformer_config_from_args(get_args())
+    else:
+        language_transformer_config = config
     if args.decoder_num_layers is not None:
         language_transformer_config.num_layers = args.decoder_num_layers
     else: