support flux for mtp

dcu_megatron/core/distributed/finalize_model_grads.py

-# Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
-# Copyright (c) Huawei Technologies Co., Ltd. 2025-2025. All rights reserved.
-
 from typing import List
 
 import torch

dcu_megatron/core/models/common/embeddings/language_model_embedding.py

-# Copyright (c) 2024, NVIDIA CORPORATION. All rights reserved.
-#  Copyright (c) Huawei Technologies Co., Ltd. 2025-2025. All rights reserved.
 from typing import Literal
 
 import torch

dcu_megatron/core/transformer/mtp/multi_token_predictor.py

-#  Copyright (c) Huawei Technologies Co., Ltd. 2025-2025. All rights reserved.
+import os
 import logging
 from dataclasses import dataclass
 from typing import Union, Optional, Literal
@@ -137,12 +137,16 @@ class MultiTokenPredictor(MegatronModule):
             self.embedding_activation_buffer = None
             self.grad_output_buffer = None
 
-        self.output_layer = tensor_parallel.ColumnParallelLinear(
-            config.hidden_size,
+        if int(os.getenv("USE_FLUX_OVERLAP", "0")):
+            column_parallel_linear_impl = FluxColumnParallelLinear
+        else:
+            column_parallel_linear_impl = tensor_parallel.ColumnParallelLinear
+        self.output_layer = column_parallel_linear_impl(
+                self.config.hidden_size,
                 self.vocab_size,
-            config=config,
-            init_method=config.init_method,
-            bias=self.add_output_layer_bias,
+                config=self.config,
+                init_method=self.config.init_method,
+                bias=False,
                 skip_bias_add=False,
                 gather_output=not self.parallel_output,
                 skip_weight_param_allocation=self.share_mtp_embedding_and_output_weight,