megatron-core 0.16.0rc0.dev127378__tar.gz → 0.16.0rc0.dev127802__tar.gz

{megatron_core-0.16.0rc0.dev127378 → megatron_core-0.16.0rc0.dev127802}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: megatron-core
-Version: 0.16.0rc0.dev127378
+Version: 0.16.0rc0.dev127802
 Summary: Megatron Core - a library for efficient and scalable training of transformer based models
 Author-email: NVIDIA <nemo-toolkit@nvidia.com>
 Maintainer-email: NVIDIA <nemo-toolkit@nvidia.com>

{megatron_core-0.16.0rc0.dev127378 → megatron_core-0.16.0rc0.dev127802}/megatron/core/datasets/megatron_tokenizer.py

@@ -1,11 +1,14 @@
 # Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
 import json
+import logging
 from abc import ABC, abstractmethod
 from collections import OrderedDict
 from typing import Any
 
 import numpy
 
+logger = logging.getLogger(__name__)
+
 
 class MegatronLegacyTokenizer(ABC):
     """Abstract class for tokenizer
@@ -20,6 +23,12 @@ class MegatronLegacyTokenizer(ABC):
     """
 
     def __init__(self, *tokenizer_paths: str, **tokenizer_options: Any):
+        # Deprecation warning
+        logger.warning(
+            "You’re using the legacy tokenizer system, which is deprecated "
+            "and will be removed in a future release. Please migrate to the new tokenizer system "
+            "(`megatron.core.tokenizers.MegatronTokenizer`)."
+        )
         self.unique_identifiers = OrderedDict()
         self.unique_identifiers["class"] = type(self).__name__
         self.unique_identifiers["tokenizer_path"] = list(tokenizer_paths)

{megatron_core-0.16.0rc0.dev127378 → megatron_core-0.16.0rc0.dev127802}/megatron/core/fp8_utils.py

@@ -10,6 +10,12 @@ from typing import List, Optional
 import torch
 
 from megatron.core.enums import Fp8Recipe
+from megatron.core.tensor_parallel import (
+    ColumnParallelLinear,
+    RowParallelLinear,
+    gather_from_sequence_parallel_region,
+    reduce_scatter_to_sequence_parallel_region,
+)
 from megatron.core.transformer.transformer_config import TransformerConfig
 from megatron.core.utils import get_te_version, is_te_min_version
 
@@ -112,6 +118,27 @@ def get_fp8_align_size(fp8_recipe: Fp8Recipe) -> int:
         return 16
 
 
+def is_column_parallel_linear(module):
+    """Returns whether the given module is a ColumnParallelLinear layer."""
+    if HAVE_TE and (
+        isinstance(module, TEColumnParallelLinear)
+        or isinstance(module, TELayerNormColumnParallelLinear)
+    ):
+        return True
+    elif isinstance(module, ColumnParallelLinear):
+        return True
+    return False
+
+
+def is_row_parallel_linear(module):
+    """Returns whether the given module is a RowParallelLinear layer."""
+    if HAVE_TE and isinstance(module, TERowParallelLinear):
+        return True
+    elif isinstance(module, RowParallelLinear):
+        return True
+    return False
+
+
 """
 The code below abstracts the functionalities needed for implementing "--fp8-param-gather" into
 several functions. It provides different implementations for each function based on different
@@ -587,6 +614,18 @@ if HAVE_TE:
             if not FP8GlobalStateManager.is_fp8_enabled():
                 return original_forward(input_tensor, *args, **kwargs)
 
+            # With sequence parallelism we need to all-gather before padding
+            # and reduce-scatter after unpadding
+            if is_sequence_parallel := getattr(module, "sequence_parallel", False):
+                if is_column_parallel_linear(module):
+                    input_tensor = gather_from_sequence_parallel_region(
+                        input_tensor, group=module.tp_group
+                    )
+
+                # Disable sequence parallelism on the module because we are handling the
+                # all-gather and reduce-scatter externally
+                module.sequence_parallel = False
+
             seq_len, batch_size, hidden_size = input_tensor.shape
             # Reshape to (S, B*H) to pad sequence dimension
             input_2d = input_tensor.reshape(seq_len, -1)
@@ -612,6 +651,16 @@ if HAVE_TE:
             unpadded_output_2d = _unpad_func(output_2d, [seq_len])
             unpadded_output = unpadded_output_2d.reshape(seq_len, batch_size, output_hidden_size)
 
+            if is_sequence_parallel:
+                # Reduce-scatter after unpadding
+                if is_row_parallel_linear(module):
+                    unpadded_output = reduce_scatter_to_sequence_parallel_region(
+                        unpadded_output, group=module.tp_group
+                    )
+
+                # Reset sequence parallelism flag on the module
+                module.sequence_parallel = True
+
             if other_outputs:
                 return (unpadded_output,) + other_outputs
             else:

{megatron_core-0.16.0rc0.dev127378 → megatron_core-0.16.0rc0.dev127802}/megatron/core/inference/async_stream.py

@@ -9,6 +9,7 @@ import asyncio
 from typing import Any, AsyncGenerator, Callable, Optional, Type, Union
 
 from megatron.core.inference.inference_request import InferenceRequest
+from megatron.core.utils import get_asyncio_loop
 
 STOP_ITERATION = Exception()
 
@@ -20,12 +21,17 @@ class AsyncStream:
     Adopted from https://github.com/vllm-project/vllm/blob/eb881ed006ca458b052905e33f0d16dbb428063a/vllm/v1/engine/async_stream.py # pylint: disable=line-too-long
     """
 
-    def __init__(self, request_id: int, cancel: Callable[[str], None]) -> None:
+    def __init__(
+        self,
+        request_id: int,
+        cancel: Callable[[str], None],
+        loop: Optional[asyncio.AbstractEventLoop] = None,
+    ) -> None:
         self._request_id = request_id
         self._cancel = cancel
         self._queue: asyncio.Queue = asyncio.Queue()
         self._finished = False
-        self._loop = asyncio.get_running_loop()
+        self._loop = get_asyncio_loop(loop)
 
     def put(self, item: Union[InferenceRequest, Exception]) -> None:
         """Adds a new value to the stream"""

megatron_core-0.16.0rc0.dev127802/megatron/core/inference/contexts/attention_context/mamba_metadata.py

@@ -0,0 +1,106 @@
+# Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+
+import torch
+
+
+class MambaMetadata:
+    """Manages the metadata tensors required for Mamba layers during inference."""
+
+    def __init__(self, max_requests: int):
+        """
+        Initializes the Mamba slot allocator.
+
+        Args:
+            max_requests (int): The maximum number of concurrent requests.
+        """
+        self.max_requests = max_requests
+
+        # Metadata for mapping requests to slots in the static Mamba state buffer
+        self.request_to_mamba_state_idx = torch.full(
+            (self.max_requests,), -1, dtype=torch.int32, device=torch.cuda.current_device()
+        )
+
+        # Separate mapping used only for CUDA graph compatibility
+        self.request_to_mamba_state_idx_cudagraph_only = torch.full(
+            (self.max_requests,), -1, dtype=torch.int32, device=torch.cuda.current_device()
+        )
+
+        # Allocator for Mamba state slots
+        self.mamba_state_free_slots = torch.arange(
+            self.max_requests, dtype=torch.int32, device=torch.cuda.current_device()
+        )
+        self.mamba_state_free_slot_count = self.max_requests
+
+    def reset(self) -> None:
+        """
+        Resets all Mamba states and frees all allocated slots.
+        """
+        self.request_to_mamba_state_idx.fill_(-1)
+        self.request_to_mamba_state_idx_cudagraph_only.fill_(-1)
+
+        # Re-initialize the free slot pool
+        self.mamba_state_free_slots = torch.arange(
+            self.max_requests, dtype=torch.int32, device=torch.cuda.current_device()
+        )
+        self.mamba_state_free_slot_count = self.max_requests
+
+    def reset_cudagraph_mapping(self) -> None:
+        """
+        Resets only the CUDA graph mapping tensor.
+        """
+        self.request_to_mamba_state_idx_cudagraph_only.fill_(-1)
+
+    def update_cudagraph_mapping(
+        self, active_mamba_indices: torch.Tensor, num_active_requests: int
+    ) -> None:
+        """
+        Updates the dedicated CUDA graph mapping tensor with the indices
+        of currently active requests.
+
+        Args:
+            active_mamba_indices (Tensor): Tensor containing the Mamba slot indices
+                                           for active requests.
+            num_active_requests (int): The number of active requests.
+        """
+        self.request_to_mamba_state_idx_cudagraph_only[0:num_active_requests] = active_mamba_indices
+
+    def allocate_slot(self) -> int:
+        """
+        Allocates a new slot for a request in the Mamba state buffers.
+
+        Returns:
+            int: The index of the allocated slot.
+            Returns None if no slots are available.
+        """
+        if self.mamba_state_free_slot_count == 0:
+            return None
+
+        # Get a free slot
+        self.mamba_state_free_slot_count -= 1
+        mamba_idx = self.mamba_state_free_slots[self.mamba_state_free_slot_count]
+
+        return mamba_idx
+
+    def free_slots(self, request_indices: torch.Tensor) -> None:
+        """
+        Frees the Mamba state slots associated with the given request indices.
+
+        Args:
+            request_indices (Tensor): A 1D tensor of request indices to free.
+        """
+        # Get the Mamba state indices for finished requests
+        mamba_indices_to_free = self.request_to_mamba_state_idx[request_indices]
+
+        # Filter out any invalid indices (e.g., -1)
+        mamba_indices_to_free = mamba_indices_to_free[mamba_indices_to_free != -1]
+        num_to_free = len(mamba_indices_to_free)
+
+        if num_to_free > 0:
+            # Add the freed indices back to the free slot pool
+            start_idx = self.mamba_state_free_slot_count
+            end_idx = start_idx + num_to_free
+            self.mamba_state_free_slots[start_idx:end_idx] = mamba_indices_to_free
+            self.mamba_state_free_slot_count = end_idx
+
+        # Invalidate the Mamba state index for the finished requests
+        self.request_to_mamba_state_idx[request_indices] = -1