Feature request: Provide calibrated OSCAR rotation matrices for Qwen-AgentWorld-35B-A3B (head_dim=256) and deepreinforce-ai/Ornith-1.0-35B

[mingxing0769]

Request

We would like to use OSCAR INT2 KV cache quantization with the following models:

1. Qwen-AgentWorld-35B-A3B (https://huggingface.co/Qwen/Qwen-AgentWorld-35B-A3B)

   • Architecture: Qwen3.5 hybrid MoE (35B total, 3B active)
   • (confirmed from config.json)
   • Currently not available in OSCAR-RotationZoo

2. deepreinforce-ai/Ornith-1.0-35B (https://huggingface.co/deepreinforce-ai/Ornith-1.0-35B)

   • Architecture: Qwen3.5 hybrid MoE (35B total, 3B active)
   • Currently not available in OSCAR-RotationZoo

Context

• The directory in this repo only ships rotation matrices for (head_dim=128).
• The RotationZoo currently does not include matrices for these 35B-A3B models.
• These models have , which is different from the Qwen3-4B/8B/32B models (head_dim=128) already supported.

Why this matters

OSCAR INT2 KV cache is the only way to run 35B-A3B models with long context on consumer hardware (e.g., 16GB VRAM Mac or GPU). Without calibrated rotation matrices, we fall back to degraded data-free INT2 (only ~33% quality on GPQA vs ~70% with proper calibration).

Requested

Please provide calibrated rotation matrices ( and ) for:

• Qwen-AgentWorld-35B-A3B (head_dim=256, 36 layers)
• deepreinforce-ai/Ornith-1.0-35B (head_dim=256, 36 layers)

Or guidance on how to run the OSCAR calibration pipeline ( + ) for these models.

Thank you!

[zzz0906]

Hi I will try to look into it this month. Thank you!