Spaces:
Sleeping
Sleeping
Commit History
-
Fix default temp=0.8, top_k=20, fix fallback tokenizer to 38 chars
DavinciDreams443b747 -
Switch to Python/FastAPI server (RandyGPT pattern)
DavinciDreams +1: Claude Opus 4.6Replace Julia HTTP.jl server with FastAPI/uvicorn for reliable HF Spaces streaming. Loads Flux JLD2 weights via h5py into PyTorch. - server.py: FastAPI + StreamingResponse, JLD2→PyTorch weight loader - GPT-2 architecture: LayerNorm, GELU, combined QKV, 6L×384D, 10.7M params - Downloads best_model.jld2 + vocab.json from HF Hub at startup - Dockerfile: python:3.11-slim + uvicorn - Remove Julia files Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>492f84f -
Aggressive sampling: temp=0.1, top_k=8, rep_penalty=1.3
DavinciDreamsba27b12 -
Fix inference: top-k sampling, repetition penalty, prompt encoding, vocab download
DavinciDreams- model.jl: Added top-k=40 sampling, repetition penalty (1.2), lowered default temperature from 0.8 to 0.3, accept seed_ids for prompt context - server.jl: Download vocab.json from HF model repo, encode user prompt as char-level seed tokens, expose top_k and repetition_penalty API params, fix fallback charset (removed stray comma), handle vocab size mismatches6e446b9
-
Upgrade to juliaflux v1: Flux.jl GPT-2 architecture with HF model download
Verified signature3a9d7d1
-
Upload checkpoints/best_model.json with huggingface_hub
Verified signature50ac038 -
Upload server.jl with huggingface_hub
Verified signature319e3ec -
Upload checkpoint.jl with huggingface_hub
Verified signature9439897 -
Upload model.jl with huggingface_hub
Verified signature5ba8ddd -
Upload Project.toml with huggingface_hub
Verified signaturefaed5ac -
Upload Dockerfile with huggingface_hub
Verified signaturea4ce50d -
Upload README.md with huggingface_hub
Verified signature5076bf9 -
initial commit
Verified signatureadb4c3f