Skip to content

Cohere2 moe support#1

Open
csabakecskemeti wants to merge 2 commits into
masterfrom
cohere2-moe-support
Open

Cohere2 moe support#1
csabakecskemeti wants to merge 2 commits into
masterfrom
cohere2-moe-support

Conversation

@csabakecskemeti

Copy link
Copy Markdown
Owner

Overview

Additional information

Requirements

csabakecskemeti and others added 2 commits May 21, 2026 14:02
Adds GGUF conversion and C++ inference for Cohere2MoeForCausalLM /
Cohere2VisionForConditionalGeneration (CohereLabs Command A+, 218B/25B
active). Key design points: parallel attention+FFN block sharing a single
pre-norm, sliding window attention on 3 of every 4 layers with RoPE only
on SWA layers, sigmoid-gated sparse MoE with 128 experts (8 active) plus 4
always-on shared experts averaged before residual addition, and
head_dim=128 explicit (not derived from hidden_size/n_heads).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant