Fixing Degenerate GPT-2 Model Output | Generated by AI

Home 2026.06 English 中文

---

Comprehensive Summary

The Problem

The SEC-EDGAR GPT-2 124M model on RunPod was producing degenerate output — repetitive loops like “ivableivable…” regardless of serving method, sampling strategy, or prompt. This was a model quality issue, not a serving bug.

Root Cause Analysis

There were two separate issues stacked on top of each other:

1. Old model on RunPod (/workspace/model/hf-model/ with vocab_size=50304):

• Properly converted to HF format (correct weight shapes [768, 2304], all biases present)
• But had padded vocab from nanoGPT: vocab_size=50304 (50257 real + 47 padding to nearest 64)
• server.py had a runtime hack to trim embeddings 50304→50257
• This model produced degenerate output — the trimming hack was a band-aid, but the underlying trained weights were likely corrupted or from a bad checkpoint

2. New model from HF (lzwjava/sec-edgar-gpt-124m-hf):

• Has vocab_size=50257 (correct, no padding) — this is the fix
• But the safetensors was saved directly from nanoGPT without proper conversion:
  • Linear weight shapes are transposed: [2304, 768] instead of [768, 2304]
  • All biases are missing (ln_1, ln_2, c_attn, c_proj, c_fc, c_proj)
• transformers 5.5.0 on RunPod rejects this with RuntimeError: shape mismatch

What We’ve Done So Far

1. Updated local server.py — removed the 18-line embedding trimming hack, simplified to direct load
2. Backed up old model on RunPod → hf-model-old/
3. Downloaded new model from HF to RunPod → hf-model/ (7 files, 650MB safetensors)
4. Server crashed — transformers refuses to load due to weight shape mismatches
5. Wrote conversion script (scripts/convert_nanogpt_to_hf.py) that:
   • Transposes linear weights from [out, in] to [in, out]
   • Adds zero biases for all missing params
   • Saves proper config.json with transformers metadata
   • Copies tokenizer files

What Still Needs To Be Done

1. Run the conversion script on RunPod to produce hf-model-converted/
2. Replace hf-model/ with the converted version
3. Restart the server with the new model
4. Test generation — verify no more degenerate loops
5. Optionally re-upload the corrected model to HF so it’s properly formatted for anyone

Key Architecture Notes

Component	Details
RunPod GPU	RTX 2000 Ada 16GB (EU-RO-1)
Server	FastAPI + transformers on port 8888
Nginx	8081→8888 reverse proxy
Model	GPT-2 124M, 12L, 768d, 12 heads, 1024 context
Vocab	50257 (standard GPT-2, no padding)
API	OpenAI-compatible /v1/completions + /v1/chat/completions

Want me to run the conversion on RunPod and bring the server up?

---

Back Donate