Skip to content

Add GPT-Neo-1.3B contrib model port#86

Open
dhwanw wants to merge 4 commits intomainfrom
contrib/gpt-neo-1.3B
Open

Add GPT-Neo-1.3B contrib model port#86
dhwanw wants to merge 4 commits intomainfrom
contrib/gpt-neo-1.3B

Conversation

@dhwanw
Copy link

@dhwanw dhwanw commented Mar 18, 2026

Summary

  • Adds NeuronX Distributed Inference implementation of EleutherAI/gpt-neo-1.3B
  • Validated with 100% greedy token match on all 10 prompts (1185 tokens)
  • Profiled: 81.9 tok/s on trn1.32xlarge (TP=1)

Model Details

  • Architecture: Decoder-only transformer
  • Parameters: 1.3B
  • TP Degree: 1
  • Precision: BF16

Validation

  • Greedy token match: 100% (10/10 prompts, 1185 tokens)
  • Teacher-forced match: 100%
  • Throughput: 81.9 tok/s (TP=1, BS=1, seq_len=128)

Files

  • contrib/models/gpt-neo-1.3B/src/modeling_gpt_neo.py — Model implementation
  • contrib/models/gpt-neo-1.3B/test/ — Integration tests
  • contrib/models/gpt-neo-1.3B/README.md — Documentation

🤖 Generated with Claude Code

dhwanw and others added 2 commits March 18, 2026 23:16
Validated on Trn1: 98.59% teacher-forced accuracy, 81.5 tok/s.
Key: unscaled attention via Q pre-multiplication, absolute position embeddings,
tied lm_head weights.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@dhwanw dhwanw force-pushed the contrib/gpt-neo-1.3B branch from ace831c to e998afe Compare March 18, 2026 23:19
dhwanw and others added 2 commits March 19, 2026 20:40
Not needed — validation instructions are in README.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Previous README had stale 52.97% greedy / 98.59% TF data. Actual package
validation shows 100% greedy and 100% teacher-forced match across all 10
prompts (1185 tokens).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
@dhwanw dhwanw marked this pull request as ready for review March 19, 2026 22:12
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant