메뉴
Navigate model merging strategies and theory. Quick reference for Task Arithmetic, DARE, TIES merging methods, alpha selection, and RTX 4090 validation results.
Codex 또는 Claude로 설치 이 Prompt를 복사해 Codex, Claude 또는 다른 어시스턴트에 붙여 넣으면 Skill 페이지를 검토하고 설치를 진행할 수 있습니다.
Validate GRPO training configs against 36 cross-framework bug avoidance rules (19 MUST DO + 17 MUST NOT). Covers 7 frameworks (DeepSpeed, Megatron, vLLM, verl, SGLang, rLLM, MindIE), 7 pattern classes, and MoE NaN avoidance.
Debug and diagnose GRPO training issues across verl, Megatron, DeepSpeed, and RTX 4090. Symptom-to-bug mapping, NaN debugging, Muon clipping avoidance, and CUDA stream safety analysis.
Navigate SGLang source code. Quick reference for subsystem locations, key classes, and data flow.
Navigate mechanistic interpretability concepts, tools, and SAE architectures. Quick reference for SAE variants, feature steering methods, circuit discovery, and trust calibration.
Navigate RDMA, InfiniBand, RoCE v2, AI cluster networking topology, and NCCL transport internals. Quick reference for Verbs API, QP/CQ/SRQ/MR, fat-tree/rail-aligned topology, DCQCN congestion control, GPUDirect RDMA, and cluster configuration.
Navigate PyTorch torch.distributed, NCCL, and distributed training internals. Quick reference for ProcessGroupNCCL, DDP, FSDP, NCCL channels/proxy/protocol, and training configuration.
SOC 직업 분류 기준
| name | model-merging-nav |
| description | Navigate model merging strategies and theory. Quick reference for Task Arithmetic, DARE, TIES merging methods, alpha selection, and RTX 4090 validation results. |
| triggers | ["model merging","task arithmetic","DARE","TIES","merge models","multi-task merge","alpha parameter","model composition","fine-tune merge"] |
| Strategy | Method | Key Parameter | Success Condition | RTX 4090 Result |
|---|---|---|---|---|
| Task Arithmetic | base + α*Σ(task_vectors) | α=0.75 optimal | α<1.0 safe, α≥1.0 risky | ✓ 100% bench-tested |
| DARE | Random drop p% + rescale 1/(1-p) | drop_rate=0.9 | Works on large models | ✓ 90% drop OK on large |
| TIES | Trim top-k% + sign majority + merge | top_k=20% | Minimal interference | ✓ best interference score |
α=0.1 → cos_sim=0.90 ✓ GOOD (low signal preservation)
α=0.25 → cos_sim=0.92 ✓ GOOD
α=0.5 → cos_sim=0.91 ✓ GOOD
α=0.75 → cos_sim=0.88 ✓ GOOD ← OPTIMAL (best preservation-safety balance)
α=1.0 → cos_sim=0.83 ✗ RISKY (norm explosion starts)
α=1.5 → cos_sim=0.73 ✗ RISKY
α=2.0 → cos_sim=0.66 ✗ RISKY (severe norm explosion)
Rule: α<1.0 is safe. α≥1.0 causes weight norm ratio >1.5 → unstable.
drop=0.5 → cos_sim=0.69 ✗ RISKY (too many active → noise)
drop=0.7 → cos_sim=0.57 ✗ RISKY
drop=0.9 → ✓ GOOD on real large models (synthetic small models show different behavior)
drop=0.95 → ✓ OK but sparse → needs more real data
drop=0.99 → ✗ RISKY (only 1% kept → unstable)
Key insight: DARE works well on real large models (many redundant parameters) but poorly on small synthetic matrices. Our RTX 4090 bench confirms 90% drop works in practice.
Result: Lowest interference (-0.04 vs +0.42 for arithmetic) → best for conflicting tasks.
tools/model_merging_simulator.py — NumPy-based validation (no GPU)results/model_merging_simulator_results.jsonnotebook/fundamentals/model-merging-theory.md — theoretical foundationsmemory/model-merging-experiment-rtx4090.md — RTX 4090 bench resultsNeed to merge N fine-tuned models into one?
→ Tasks similar/compatible? → Task Arithmetic α=0.75 (simplest, proven)
→ Tasks conflicting/different directions? → TIES (best interference control)
→ Very large models (7B+)? → DARE 90% drop (reduces compute, proven on large)
→ All methods? → Try all three, compare cos_sim + preservation + interference