gpt-oss implementation

[sbhavani]

Description

This outlines the current status of gpt-oss features that need to be implemented in Megatron Core, leveraging Transformer Engine.

✅ UPDATE: All core GPT-OSS functionality is now available in Megatron Core (training) and Megatron Bridge (checkpoint conversion).

MoE Layer

Enabled Bias

• Status: ✅ Supported
• Implementation: Available in main branch: Fixes for gpt-oss #2038

Attention Mechanisms

Alternating Sliding-Window Attention Pattern

• Status: ✅ Supported - Infrastructure exists for per-layer patterns and sliding window attention using TE

Attention Sinks

• Status: ✅ Implemented - in Transformer Engine and cuDNN
• Reference: Streaming LLM
• Related Transformer Engine PR: [PyTorch] Add sink attention support from cuDNN TransformerEngine#2148

Activation Functions

Custom SwiGLU with Clamping

• Status: ✅ Supported
• Implementation: Megatron Core added partially fused version as "custom quick GeGLU"

FP8-aware fused kernel merged into Transformer Engine

• Related Transformer Engine PR: [Pytorch] Support for Swiglu Activation used in GPT OSS TransformerEngine#2161

Positional Encodings

YaRN RoPE Scaling

• Status: ✅ Fully Supported
• Implementation:
  • YaRN scaling to 128k+ context
  • Integration with existing RoPE
  • YaRN for general RoPE/GPT models
  • Convergence validation
• Usage: --position-embedding-type yarn with YaRN configuration parameters
• Reference: arXiv:2309.00071

Megatron Bridge Support

Megatron Bridge provides full GPT-OSS integration:

• ✅ Checkpoint Conversion: Hugging Face ↔ Megatron format
• ✅ Pre-configured Providers: GPTOSSProvider20B and GPTOSSProvider120B
• ✅ Quantization Support: Handles MXFP4 weight dequantization

Megatron Bridge + Megatron-LM Example

PR: #2383 provides end-to-end example scripts covering checkpoint conversion (convert_mcore_bf16_checkpoint_from_hf.py) and training/fine-tuning (training_gptoss_20b_h100_bf16_fp8.sh)

Credits: @cuichenx for core implementation, @yiakwy-xpu-ml-framework-team for example scripts

[Pikachu1412]

So how can I train gptoss by this branch?

[sbhavani]

We have a guide in NeMo Framework (using Megatron Core): https://docs.nvidia.com/nemo-framework/user-guide/latest/llms/gpt_oss.html.

Megatron-LM training scripts will be added soon

[liuqianchao]

how is the progress of support gpt-oss

[jgcb00]

Hi, is there any plan to implement intra-doc masking (Attention being only compute on a document not on the full sequence length) ?

Best regards

[quantLm14]

@cuichenx Any update on long seq optimisations?

Or for this branch to be merged in main?

[JimmyAwoe]

Thanks for contribution first. So has this implementation already finished or closing? I haven't see any related argument about sliding attention window in megatron/training/argument.py.

[yiakwy-xpu-ml-framework-team]

@JimmyAwoe GPTOSS model is now provided in megatron bridge.

They use megatron bridge to provide model variants now.

[yiakwy-xpu-ml-framework-team]

I have given GPT OSS training support here :

#2383

@JimmyAwoe simply try it out

[radna0]

@sbhavani is gpt-oss now supported in megatron bridge? I'm getting
Model architecture 'GptOssForCausalLM' is not yet supported
this is in megatron-bridge==0.2.0rc6 and megatron-lm==core_r0.15.0