megatron-core 0.16.0rc0.dev123469__tar.gz → 0.16.0rc0.dev123790__tar.gz

{megatron_core-0.16.0rc0.dev123469 → megatron_core-0.16.0rc0.dev123790}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: megatron-core
-Version: 0.16.0rc0.dev123469
+Version: 0.16.0rc0.dev123790
 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>
@@ -41,7 +41,7 @@ Requires-Dist: transformers; extra == "mlm"
 Provides-Extra: dev
 Requires-Dist: nvidia-modelopt[torch]; sys_platform != "darwin" and extra == "dev"
 Requires-Dist: transformer-engine[pytorch]<2.10.0,>=2.9.0a0; extra == "dev"
-Requires-Dist: nvidia-resiliency-ext<0.5.0,>=0.4.0a0; extra == "dev"
+Requires-Dist: nvidia-resiliency-ext; extra == "dev"
 Requires-Dist: tqdm; extra == "dev"
 Requires-Dist: einops~=0.8; extra == "dev"
 Requires-Dist: tensorstore!=0.1.46,!=0.1.72,~=0.1; extra == "dev"
@@ -59,13 +59,20 @@ Requires-Dist: wget; extra == "dev"
 Requires-Dist: onnxscript; extra == "dev"
 Provides-Extra: lts
 Requires-Dist: tqdm; extra == "lts"
-Requires-Dist: einops; extra == "lts"
-Requires-Dist: tensorstore!=0.1.46,!=0.1.72; extra == "lts"
-Requires-Dist: nvtx; extra == "lts"
-Requires-Dist: transformers; extra == "lts"
-Requires-Dist: zarr; extra == "lts"
+Requires-Dist: einops~=0.8; extra == "lts"
+Requires-Dist: tensorstore!=0.1.46,!=0.1.72,~=0.1; extra == "lts"
+Requires-Dist: nvtx~=0.2; extra == "lts"
+Requires-Dist: multi-storage-client~=0.27; extra == "lts"
+Requires-Dist: opentelemetry-api~=1.33.1; extra == "lts"
 Requires-Dist: setuptools<80.0.0; extra == "lts"
+Requires-Dist: mamba-ssm~=2.2; extra == "lts"
+Requires-Dist: causal-conv1d~=1.5; extra == "lts"
+Requires-Dist: nv-grouped-gemm~=1.1; extra == "lts"
+Requires-Dist: megatron-energon[av_decode]~=6.0; extra == "lts"
+Requires-Dist: av<16.0.0; extra == "lts"
+Requires-Dist: flashinfer-python; extra == "lts"
 Requires-Dist: wget; extra == "lts"
+Requires-Dist: onnxscript; extra == "lts"
 
 <div align="center">
 

{megatron_core-0.16.0rc0.dev123469 → megatron_core-0.16.0rc0.dev123790}/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.dev123469 → megatron_core-0.16.0rc0.dev123790}/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.dev123790/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

megatron_core-0.16.0rc0.dev123790/megatron/core/inference/contexts/attention_context/metadata_base.py

@@ -0,0 +1,72 @@
+# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
+
+
+class MetadataBase:
+    """
+    Base class for attention metadata.
+    High-performance attention kernels often require input metadata in specific
+    formats—such as cumulative query lengths, cumulative key/value lengths,
+    and similar structures. Moreover, when using CUDA Graphs, these metadata
+    buffers must be statically allocated. This class serves as a unified container
+    that manages all such metadata in one place.
+    """
+
+    def __init__(self):
+        """
+        Initialize the metadata.
+        """
+        self.state_data = {}
+
+    def update(self, *args, **kwargs):
+        """
+        Construct the metadata from request states.
+        """
+        pass
+
+    def reset(self):
+        """
+        Reset the metadata.
+        """
+        pass
+
+    def tensor_copy_and_pad(
+        self,
+        tensor_buf,
+        unpadded_tensor,
+        real_batch_size,
+        padded_batch_size,
+        is_cumulative_tensor=False,
+        pad_value=0,
+    ):
+        """
+        Copy the unpadded tensor to the tensor_buf,
+        pad the tensor_buf with zero or the last value of the tensor,
+        depending on whether the tensor is cumulative.
+        Args:
+            tensor_buf: The destination tensor, at least padded_batch_size long.
+            unpadded_tensor: The tensor to copy, at least real_batch_size long.
+            real_batch_size: The real batch size.
+            padded_batch_size: Padded boundary of the tensor.
+            is_cumulative_tensor: Whether the tensor is cumulative.
+                If True, we pad the tensor_buf with the last value of the unpadded_tensor.
+            pad_value: The value to pad the tensor_buf with when the tensor is not cumulative.
+        """
+        assert real_batch_size <= padded_batch_size
+        assert tensor_buf.shape[0] >= padded_batch_size
+        assert unpadded_tensor.shape[0] >= real_batch_size
+        if is_cumulative_tensor:
+            if real_batch_size == 0:
+                value = pad_value
+            else:
+                value = unpadded_tensor[real_batch_size - 1]
+        else:
+            value = pad_value
+        tensor_buf[0:real_batch_size] = unpadded_tensor[:real_batch_size]
+        tensor_buf[real_batch_size:padded_batch_size] = value
+        return tensor_buf
+
+    def __str__(self):
+        """
+        Return a string representation of the metadata.
+        """
+        return "\n".join([f"{key}: {value}" for key, value in self.state_data.items()])

megatron_core-0.16.0rc0.dev123790/megatron/core/inference/contexts/attention_context/mha_metadata.py

@@ -0,0 +1,220 @@
+# Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
+
+from typing import Optional
+
+import torch
+
+from .metadata_base import MetadataBase
+
+
+class MHAMetadata(MetadataBase):
+    """
+    Metadata for MHA layer using flash-attention.
+    """
+
+    def __init__(
+        self, block_count_total, max_kv_block_count, max_requests, block_size_tokens, max_seqlen
+    ):
+        super().__init__()
+        device = torch.cuda.current_device()
+        self.device = device
+        self.max_blocks = block_count_total
+        self.max_kv_blocks = max_kv_block_count
+        self.max_bs = max_requests
+        self.max_seqlen = max_seqlen
+        self._query_lengths_buf = torch.zeros(self.max_bs, dtype=torch.int32, device=device)
+        self._cu_query_seq_lengths_buf = torch.zeros(
+            self.max_bs + 1, dtype=torch.int32, device=device
+        )
+        self._cu_kv_seq_lengths_buf = torch.zeros(self.max_bs + 1, dtype=torch.int32, device=device)
+        self._kv_seq_lengths_buf = torch.zeros(self.max_bs, dtype=torch.int32, device=device)
+        self._block_table_buf = torch.zeros(
+            (self.max_bs, self.max_kv_blocks), dtype=torch.int32, device=device
+        )
+        self._max_seqlen_q = 0
+        self._max_seqlen_k = 0
+        self.state_data = {}
+
+    def update(
+        self,
+        request_query_lengths: torch.Tensor,
+        request_kv_length_offsets: torch.Tensor,
+        request_to_kv_block_ids: torch.Tensor,
+        padded_active_token_count: int,
+        real_batch_size: int,
+        padded_active_request_count: Optional[int] = None,
+        decode_only: bool = False,
+    ):
+        """
+        Args:
+            request_query_lengths: (>real_batch_size,)
+            request_kv_length_offsets: (>real_batch_size,)
+            request_to_kv_block_ids: (>real_batch_size, max_kv_blocks)
+            padded_active_token_count: int
+            real_batch_size: int
+            padded_active_request_count: Optional[int]
+            decode_only: bool
+        """
+        if padded_active_request_count is None:
+            padded_active_request_count = real_batch_size
+
+        assert real_batch_size <= padded_active_request_count <= self.max_bs
+        assert request_query_lengths.shape[0] == real_batch_size
+        assert request_kv_length_offsets.shape[0] == real_batch_size
+        assert request_to_kv_block_ids.shape[0] == real_batch_size
+
+        self.tensor_copy_and_pad(
+            self._query_lengths_buf,
+            request_query_lengths,
+            real_batch_size,
+            padded_active_request_count,
+        )
+        self._cu_query_seq_lengths_buf[0] = 0
+        self.tensor_copy_and_pad(
+            self._cu_query_seq_lengths_buf[1:],
+            torch.cumsum(request_query_lengths, dim=0),
+            real_batch_size,
+            padded_active_request_count,
+            is_cumulative_tensor=True,
+        )
+        self.tensor_copy_and_pad(
+            self._kv_seq_lengths_buf,
+            request_kv_length_offsets + request_query_lengths,
+            real_batch_size,
+            padded_active_request_count,
+        )
+        self.tensor_copy_and_pad(
+            self._block_table_buf,
+            request_to_kv_block_ids,
+            real_batch_size,
+            padded_active_request_count,
+            pad_value=torch.tensor(self.max_kv_blocks, dtype=torch.int32, device=self.device).fill_(
+                -1
+            ),
+        )
+        self._cu_kv_seq_lengths_buf[0] = 0
+        self.tensor_copy_and_pad(
+            self._cu_kv_seq_lengths_buf[1:],
+            torch.cumsum(self._kv_seq_lengths_buf, dim=0),
+            real_batch_size,
+            padded_active_request_count,
+            is_cumulative_tensor=True,
+        )
+
+        if decode_only:
+            self._max_seqlen_q = 1
+        else:
+            self._max_seqlen_q = max(2, padded_active_token_count)
+        self._max_seqlen_k = self.max_seqlen
+
+        self.state_data = {
+            "query_lengths": self._query_lengths_buf[:padded_active_request_count],
+            "cu_query_seq_lengths": self._cu_query_seq_lengths_buf[
+                : padded_active_request_count + 1
+            ],
+            "cu_kv_seq_lengths": self._cu_kv_seq_lengths_buf[: padded_active_request_count + 1],
+            "kv_seq_lengths": self._kv_seq_lengths_buf[:padded_active_request_count],
+            "block_table": self._block_table_buf[0:padded_active_request_count, :],
+            "max_seqlen_q": self._max_seqlen_q,
+            "max_seqlen_k": self._max_seqlen_k,
+        }
+
+    def reset(self):
+        """
+        Reset the metadata for the next batch.
+        """
+        self._query_lengths_buf.fill_(0)
+        self._cu_query_seq_lengths_buf.fill_(0)
+        self._cu_kv_seq_lengths_buf.fill_(0)
+        self._kv_seq_lengths_buf.fill_(0)
+        self._block_table_buf.fill_(0)
+        self._max_seqlen_q = 0
+        self._max_seqlen_k = 0
+
+
+class GraphedMHAMetadata(MHAMetadata):
+    """
+    Metadata for MHA layer using flash-attention with CUDA graphs.
+    """
+
+    def __init__(
+        self, block_count_total, max_kv_block_count, max_requests, block_size_tokens, max_seqlen
+    ):
+        super().__init__(
+            block_count_total, max_kv_block_count, max_requests, block_size_tokens, max_seqlen
+        )
+
+    def update(
+        self,
+        request_query_lengths: torch.Tensor,
+        request_kv_length_offsets: torch.Tensor,
+        request_to_kv_block_ids: torch.Tensor,
+        padded_active_token_count: int,
+        real_batch_size: int,
+        padded_active_request_count: Optional[int] = None,
+        decode_only: bool = False,
+    ):
+        """
+        Args:
+            request_query_lengths: (>real_batch_size,)
+            request_kv_length_offsets: (>real_batch_size,)
+            request_to_kv_block_ids: (>real_batch_size, max_kv_blocks)
+            padded_active_token_count: int
+            real_batch_size: int
+            padded_active_request_count: Optional[int]
+            decode_only: bool
+        """
+        super().update(
+            request_query_lengths,
+            request_kv_length_offsets,
+            request_to_kv_block_ids,
+            padded_active_token_count,
+            real_batch_size,
+            padded_active_request_count,
+            decode_only,
+        )
+
+    def reset(self):
+        super().reset()
+
+
+class NonGraphedMHAMetadata(MHAMetadata):
+    """
+    Metadata for MHA layer using flash-attention without CUDA graphs.
+    """
+
+    def update(
+        self,
+        request_query_lengths: torch.Tensor,
+        request_kv_length_offsets: torch.Tensor,
+        request_to_kv_block_ids: torch.Tensor,
+        padded_active_token_count: int,
+        real_batch_size: int,
+        padded_active_request_count: Optional[int] = None,
+        decode_only: bool = False,
+    ):
+        """
+        Args:
+            request_query_lengths: (>real_batch_size,)
+            request_kv_length_offsets: (>real_batch_size,)
+            request_to_kv_block_ids: (>real_batch_size, max_kv_blocks)
+            padded_active_token_count: int
+            real_batch_size: int
+            padded_active_request_count: Optional[int]
+            decode_only: bool
+        """
+        super().update(
+            request_query_lengths,
+            request_kv_length_offsets,
+            request_to_kv_block_ids,
+            padded_active_token_count,
+            real_batch_size,
+            padded_active_request_count,
+            decode_only,
+        )
+        if len(self.state_data["query_lengths"]) > 0:
+            self.state_data["max_seqlen_q"] = torch.max(self.state_data["query_lengths"]).item()
+            self.state_data["max_seqlen_k"] = torch.max(self.state_data["kv_seq_lengths"]).item()
+        else:
+            self.state_data["max_seqlen_q"] = 1
+            self.state_data["max_seqlen_k"] = 1