Havoc1904/edge-moe-logits

Edge-MoE Logits Pipeline

Overview

Multi-teacher distillation logit generation pipeline for training a 7-9B Edge-MoE model. Generates teacher logits across 6 domains using Hugging Face models served via vLLM on AMD MI300X (ROCm, 192GB VRAM).

Teacher Model Chain

Attempts in order:

1. Qwen/Qwen3-Omni-30B-A3B-Instruct (primary multimodal) — FAILED: cu_seqlens_q must be on CUDA (ROCm incompatibility)
2. nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16 (multimodal fallback) — SUCCESS: 61.59 GiB. Vision via image_url (base64). No audio.
3. Qwen/Qwen3.5-27B (text-only fallback) — works but text-only.

Domains & Datasets

Domain	Source	Size	Status
coding	Llama-Nemotron SFT code split	40,000	✅
math	Llama-Nemotron SFT math split	40,000	✅
dialogue	Llama-Nemotron SFT chat + Tulu-3-SFT	40,000	✅
agentic	OpenHermes-2.5 + Tulu-3-SFT filtered + Nemotron RL	40,000	✅
knowledge	Llama-Nemotron SFT science split	40,000	✅
vision	LLaVA-Instruct-150K (COCO train2014)	40,000	✅ (auto-graded)
audio	Menlo/instruction-speech-encodec-v1.5	—	❌ Dropped (no audio support)

GPU Budget

• Total: ~50h credit
• Used: ~2.5h (smoke tests + validation + benchmarks)
• Remaining: ~47.5h
• Full 200K run (256 concurrency): estimated ~17-20h

Key Findings

Concurrency Benchmarking (Qwen3.5-27B, 768 tokens)

Concurrency	Ex/min	vs baseline
1	12	1x
64	136	11.3x
128	169.5	14.1x
256	193.7	16.1x
512	plateaus	~0 gain

Model Compatibility on ROCm

• Qwen3-Omni: ❌ cu_seqlens_q must be on CUDA — ROCm kernel issue
• Nemotron-3-Nano-Omni: ✅ Vision works. Needs --trust-remote-code. ~62GB download, 61.59 GiB VRAM. No audio.
• Qwen3.5-27B: ✅ Text-only. 52GB download, ~28 GiB VRAM at FP8. Fastest.

Nemotron-Nano Vision Quality

Auto-graded 30 examples vs LLaVA refs (keyword overlap F1):

• Mean F1: 0.114 (low due to meta-style framing, not capability)
• Model correctly identifies image content but wraps in "The user wants to know..."
• Vision capability is real and grounded

How to Resume

Prerequisites

hf download nvidia/Nemotron-3-Nano-Omni-30B-A3B-Reasoning-BF16
cd /root/parshu && bash scripts/launch_vllm_teachers.sh

Launch Full Generation

bash scripts/run_all_domains.sh

Auto-resumes from checkpoints in data/logits/.checkpoints/.

Monitor

tail -f logs/gen_*.log
rocm-smi
wc -l data/logits/*.jsonl

Push to HF

python3 scripts/push_to_hub.py --repo-id Havoc1904/edge-moe-logits

Download on Laptop

hf download Havoc1904/edge-moe-logits --repo-type dataset --local-dir ./edge-moe-data

File Layout

• scripts/ — Pipeline scripts
• data/prompts/ — 40K prompts per domain (6 domains)
• data/logits/ — Generated logit JSONL files
• data/logits/.checkpoints/ — Resume checkpoints
• logs/ — Server and generation logs