- Add Quick Start section with random initialization example
- Show actual generation code that works
- Include example output (random tokens)
- Update model configuration to reflect 24 layers
- Add note that config matches gpt-oss-20b
- Show ~2.4B parameters
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Update config to mirror gpt-oss-20b (excluding MoE params)
- Changed from 201088 to 199998 to match openai/gpt-oss-20b tokenizer
- Ensures model and tokenizer are compatible
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Inherit from GenerationMixin to add generate() method
- Add @can_return_tuple decorator
- Add logits_to_keep parameter for efficient generation
- Use self.loss_function instead of manual loss computation
- Import Union type for logits_to_keep
This ensures GptOssDense has the same API as GptOss.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Add badges for license, Python, and transformers compatibility
- Add verification note that model works with trust_remote_code=True
- Tested on transformers 4.40+
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Add rope_scaling attribute for stable transformers compatibility
- Set both rope_scaling and rope_parameters
- rope_scaling is expected by stable transformers YARN implementation
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Remove check_model_inputs decorator for stable transformers compatibility
- Remove check_model_inputs import and decorator
- This decorator API changed between versions
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Fix rope_parameters compatibility with stable transformers
- Remove RopeParameters import (not available in stable)
- Use Dict type hint instead
- Remove standardize_rope_params call (not available in stable)
- Add rope_theta attribute for compatibility
- Make rope validation optional
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Fix PretrainedConfig import for compatibility with stable transformers
- Change PreTrainedConfig to PretrainedConfig (lowercase 't')
- This matches the stable transformers API
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Update README to reference marksverdhei/transformers fork
- Replace generic "if model is registered" with specific fork reference
- Add installation instructions for the fork
- Link to github.com/marksverdhei/transformers
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Add test script and update README with usage instructions
- Add test_model.py to demonstrate loading and using the model
- Update README with clear instructions for:
- Loading config and initializing model
- Loading pre-trained weights (when available)
- Using with transformers library directly
- Clarify that repo currently has architecture but no weights
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
- Change relative imports to absolute transformers imports
- This allows the model to be loaded with trust_remote_code=True
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
Add custom modeling and configuration files for trust_remote_code support
- Add modeling_gpt_oss_dense.py with GptOssDense model implementation
- Add configuration_gpt_oss_dense.py with GptOssDenseConfig
- Add config.json with auto_map for remote code loading
- Add README.md with model documentation
This allows the model to be loaded with trust_remote_code=True from the Hub.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude <noreply@anthropic.com>
