wip

Signed-off-by: Alexandros Koumparoulis <[email protected]>

Author: akoumpa

nemo_automodel/_transformers/auto_model.py

@@ -17,7 +17,7 @@
 import inspect
 import logging
 import types
-from typing import List, Optional, Union
+from typing import Any, Dict, List, Optional, Union
 
 import torch
 from torch.nn.attention import SDPBackend, sdpa_kernel
@@ -35,6 +35,9 @@
 from nemo_automodel._transformers.registry import ModelRegistry
 from nemo_automodel.shared.import_utils import safe_import
 from nemo_automodel.shared.utils import dtype_from_str
+from nemo_automodel.components.distributed.fsdp2 import FSDP2Manager
+from nemo_automodel.components.distributed.ddp import DDPManager
+from torch.distributed.device_mesh import DeviceMesh
 
 HAS_LIGER_KERNEL, liger_kernel_trf = safe_import("liger_kernel.transformers")
 logger = logging.getLogger(__name__)
@@ -148,6 +151,29 @@ def _get_next_fallback_attn(attn_implementation: str) -> str:
         return priorities[0]
 
 
+def _choose_device(device: Optional[torch.device]) -> torch.device:
+    if device is not None:
+        return device
+    if torch.cuda.is_available():
+        import os
+
+        local_rank = int(os.environ.get("LOCAL_RANK", 0))
+        return torch.device("cuda", local_rank)
+    return torch.device("cpu")
+
+
+def _move_module_to_device(module: torch.nn.Module, device: torch.device, torch_dtype: Any) -> None:
+    # torch_dtype can be "auto", torch.dtype, or string
+    if torch_dtype == "auto":
+        dtype = None
+    else:
+        dtype = dtype_from_str(torch_dtype) if isinstance(torch_dtype, str) else torch_dtype
+    if dtype is not None:
+        module.to(device=device, dtype=dtype)
+    else:
+        module.to(device=device)
+
+
 class _BaseNeMoAutoModelClass(_BaseAutoModelClass):
     """
     Drop-in replacement for ``_BaseAutoModelClass`` that includes custom-kernels.
@@ -182,6 +208,10 @@ def from_pretrained(
         attn_implementation: str = "flash_attention_2",
         quantization_config=None,
         force_hf: bool = False,
+        device_mesh: Optional[DeviceMesh] = None,
+        distributed: Optional[Dict[str, Any]] = None,
+        device: Optional[torch.device] = None,
+        move_to_device: bool = False,
         **kwargs,
     ) -> PreTrainedModel:
         """
@@ -248,6 +278,10 @@ def _retry(**override):
                 use_liger_kernel=override.get("use_liger_kernel", use_liger_kernel),
                 use_sdpa_patching=override.get("use_sdpa_patching", use_sdpa_patching),
                 sdpa_method=sdpa_method,
+                device_mesh=device_mesh,
+                distributed=distributed,
+                device=device,
+                move_to_device=move_to_device,
                 **kwargs,
             )
 
@@ -291,6 +325,47 @@ def _retry(**override):
                 return _retry(attn_implementation=attn_implementation)
             raise e
 
+        # Decide kernel patching based on requested parallelism
+        if distributed is not None:
+            conf = distributed if isinstance(distributed, dict) else {}
+            tp = int(conf.get("tp_size", 1))
+            cp = int(conf.get("cp_size", 1))
+            if tp > 1 or cp > 1:
+                use_liger_kernel = False
+
+        # Build internal wrapper (FSDP2 or DDP) using device_mesh or config
+        internal_wrapper = None
+        if torch.distributed.is_initialized():
+            conf = distributed if isinstance(distributed, dict) else {}
+            tp = int(conf.get("tp_size", 1))
+            cp = int(conf.get("cp_size", 1))
+            pp = int(conf.get("pp_size", 1))
+            dp = int(conf.get("dp_size", 0))
+            world_size = conf.get("world_size", torch.distributed.get_world_size())
+            backend = conf.get("backend", "nccl" if torch.cuda.is_available() else "gloo")
+
+            if tp > 1 or cp > 1 or pp > 1:
+                internal_wrapper = FSDP2Manager(
+                    dp_size=dp if dp > 0 else None,
+                    tp_size=tp,
+                    cp_size=cp,
+                    pp_size=pp,
+                    backend=backend,
+                    world_size=world_size,
+                    use_hf_tp_plan=bool(conf.get("use_hf_tp_plan", False)),
+                    sequence_parallel=bool(conf.get("sequence_parallel", False)),
+                    activation_checkpointing=bool(conf.get("activation_checkpointing", False)),
+                )
+                # If caller provided a prebuilt DeviceMesh, inject it
+                if device_mesh is not None:
+                    internal_wrapper.device_mesh = device_mesh
+            elif conf:
+                internal_wrapper = DDPManager(
+                    backend=backend,
+                    activation_checkpointing=bool(conf.get("activation_checkpointing", False)),
+                )
+            # no-op: model_wrapper removed; liger disabled via conf above
+
         # Kernel patching
         try:
             if use_liger_kernel:
@@ -309,6 +384,22 @@ def _retry(**override):
             logging.warning("Retrying without SDPA patching.")
             return _retry(use_sdpa_patching=False)
 
+        # Optional: parallelize model according to parallel_scheme decision
+        # Priority: FSDP2 when tp_size>1 or cp_size>1; else DDP when dp_size>1; else single GPU
+        if internal_wrapper is not None:
+            if torch.distributed.is_initialized():
+                try:
+                    model = internal_wrapper.parallelize(model)
+                except Exception as e:
+                    logger.warning("internal parallelize failed: %s", e)
+            else:
+                logger.warning("distributed requested but torch.distributed is not initialized; skipping parallelize")
+
+        # Finally move the model to the chosen device (post-parallelization), if requested
+        if move_to_device:
+            dev = _choose_device(device)
+            _move_module_to_device(model, dev, torch_dtype)
+
         model.config.update({"nemo_version": __version__})
         return model
 
@@ -324,6 +415,10 @@ def from_config(
         attn_implementation: str = "flash_attention_2",
         quantization_config=None,
         force_hf: bool = False,
+        device_mesh: Optional[DeviceMesh] = None,
+        distributed: Optional[Dict[str, Any]] = None,
+        device: Optional[torch.device] = None,
+        move_to_device: bool = False,
         **kwargs,
     ) -> PreTrainedModel:
         """
@@ -382,6 +477,10 @@ def _retry(**override):
                 use_liger_kernel=override.get("use_liger_kernel", use_liger_kernel),
                 use_sdpa_patching=override.get("use_sdpa_patching", use_sdpa_patching),
                 sdpa_method=sdpa_method,
+                device_mesh=device_mesh,
+                distributed=distributed,
+                device=device,
+                move_to_device=move_to_device,
                 **kwargs,
             )
 
@@ -419,6 +518,31 @@ def _retry(**override):
                 return _retry(attn_implementation="eager")
             raise e
 
+        # Initialize model wrapper from `distributed` if provided
+        if distributed is not None and model_wrapper is None:
+            if isinstance(distributed, FSDP2Manager):
+                model_wrapper = distributed
+            elif isinstance(distributed, dict):
+                init_kwargs = dict(distributed)
+                try:
+                    import torch.distributed as dist
+
+                    if "world_size" not in init_kwargs and dist.is_available() and dist.is_initialized():
+                        init_kwargs["world_size"] = dist.get_world_size()
+                except Exception:
+                    pass
+                if "backend" not in init_kwargs:
+                    init_kwargs["backend"] = "nccl" if torch.cuda.is_available() else "gloo"
+                model_wrapper = FSDP2Manager(**init_kwargs)
+
+        # If distributed tensor/context parallelism is requested, avoid Liger patch like in train_ft
+        if model_wrapper is not None:
+            try:
+                if getattr(model_wrapper, "tp_size", 1) > 1 or getattr(model_wrapper, "cp_size", 1) > 1:
+                    use_liger_kernel = False
+            except Exception:
+                pass
+
         # Kernel patching
         try:
             if use_liger_kernel:
@@ -437,6 +561,54 @@ def _retry(**override):
             logging.warning("Retrying without SDPA patching.")
             return _retry(use_sdpa_patching=False)
 
+        # Optional: parallelize model according to parallel_scheme decision
+        # Priority: FSDP2 when tp_size>1 or cp_size>1; else DDP when dp_size>1; else single GPU
+        if model_wrapper is not None:
+            if torch.distributed.is_initialized():
+                try:
+                    model = model_wrapper.parallelize(model)
+                except Exception as e:
+                    logger.warning("model_wrapper.parallelize failed: %s", e)
+            else:
+                logger.warning("model_wrapper provided but torch.distributed is not initialized; skipping parallelize")
+        elif distributed is not None and torch.distributed.is_initialized():
+            conf = distributed if isinstance(distributed, dict) else {}
+            tp = int(conf.get("tp_size", 1))
+            cp = int(conf.get("cp_size", 1))
+            pp = int(conf.get("pp_size", 1))
+            dp = int(conf.get("dp_size", 0))
+            world_size = conf.get("world_size", torch.distributed.get_world_size())
+            backend = conf.get("backend", "nccl" if torch.cuda.is_available() else "gloo")
+
+            if tp > 1 or cp > 1 or pp > 1:
+                manager = FSDP2Manager(
+                    dp_size=dp if dp > 0 else None,
+                    tp_size=tp,
+                    cp_size=cp,
+                    pp_size=pp,
+                    backend=backend,
+                    world_size=world_size,
+                    use_hf_tp_plan=bool(conf.get("use_hf_tp_plan", False)),
+                    sequence_parallel=bool(conf.get("sequence_parallel", False)),
+                    activation_checkpointing=bool(conf.get("activation_checkpointing", False)),
+                )
+                try:
+                    model = manager.parallelize(model)
+                except Exception as e:
+                    logger.warning("FSDP2Manager.parallelize failed: %s", e)
+            else:
+                ddp_backend = backend
+                try:
+                    ddp_manager = DDPManager(backend=ddp_backend, activation_checkpointing=bool(conf.get("activation_checkpointing", False)))
+                    model = ddp_manager.parallelize(model)
+                except Exception as e:
+                    logger.warning("DDPManager.parallelize failed: %s", e)
+
+        # Finally move the model to the chosen device (post-parallelization), if requested
+        if move_to_device:
+            dev = _choose_device(device)
+            _move_module_to_device(model, dev, torch_dtype)
+
         model.config.update({"nemo_version": __version__})
         return model
 

nemo_automodel/recipes/llm/train_ft.py

@@ -201,6 +201,32 @@ def build_model_and_optimizer(
 
             kwargs["quantization_config"] = create_bnb_config(cfg_quantization)
 
+        # Optionally pass internal parallel scheme for HF models (non-pipeline) to let the model self-parallelize
+        used_internal_parallel = False
+        if (
+            is_hf_model
+            and autopipeline is None
+            and get_world_size_safe() > 1
+            and not isinstance(model_wrapper, MegatronFSDPManager)
+        ):
+            internal_scheme = {
+                "tp_size": int(tp_size) if tp_size is not None else 1,
+                "cp_size": int(cp_size) if cp_size is not None else 1,
+                "pp_size": 1,
+            }
+            try:
+                world_size = get_world_size_safe()
+                denom = max(1, internal_scheme["tp_size"] * internal_scheme["cp_size"])
+                if world_size > 1 and world_size % denom == 0:
+                    internal_scheme["dp_size"] = world_size // denom
+            except Exception:
+                pass
+            internal_scheme["backend"] = "nccl" if torch.cuda.is_available() else "gloo"
+            kwargs["distributed"] = internal_scheme
+            if getattr(model_wrapper, "device_mesh", None) is not None:
+                kwargs["device_mesh"] = model_wrapper.device_mesh
+            used_internal_parallel = True
+
         # Instantiate the model in meta device to avoid OOM
         with init_ctx:
             if is_hf_model and (tp_size > 1 or cp_size > 1):
@@ -263,7 +289,7 @@ def build_model_and_optimizer(
             model = autopipeline
     else:
         load_weights = False
-        if parallelize_fn is not None and get_world_size_safe() > 1:
+        if not (is_hf_model and autopipeline is None and get_world_size_safe() > 1 and 'distributed' in kwargs) and parallelize_fn is not None and get_world_size_safe() > 1:
             parallelize_fn(
                 model,
                 world_mesh=model_wrapper.device_mesh,
@@ -281,7 +307,7 @@ def build_model_and_optimizer(
                 ep_shard_axis_names=("ep_shard",),
             )
             load_weights = True
-        elif callable(getattr(model_wrapper, "parallelize", None)):
+        elif not (is_hf_model and autopipeline is None and get_world_size_safe() > 1 and 'distributed' in kwargs) and callable(getattr(model_wrapper, "parallelize", None)):
             # FSDP2 and MegatronFSDP should already be on the correct device
             if isinstance(model_wrapper, MegatronFSDPManager):
                 # MegatronFSDP instantiate optimizer inside parallelize_function