Spaces:
Running
Running
Commit History
-
Refresh progress every 5 tokens to reduce bar flicker Running
4d22edc -
Hide token stream from Output box; show complete text only at end (no flicker)
704af0d -
Yield placeholders between progress stages so the bar actually moves
03eb13d -
Real progress: track HF download tqdm + update bar during token streaming
375eb0e -
Tune llama-cpp-python for HF Spaces 2-vCPU: n_ctx=512, n_threads=2, n_batch=8
fe4e741 -
Trigger rebuild to pick up merges-corrected GGUFs
e0118ce -
Switch large models to GGUF via llama-cpp-python 0.2.62
All 4 large model GGUFs were re-quantized using llama.cpp at commit 3b8f1ec (April 17 2024, bundled with llama-cpp-python 0.2.62) to ensure binary-format compatibility with the only available manylinux CPU wheel. - 1B, 2B, 3B-v2, retired-3B all run via Llama.from_pretrained() now - Q4_K_M quantization, ~3-5x faster than fp32 transformers on CPU - Streaming output preserved (llama-cpp-python stream=True) - Progress bar preserved for model load427773a -
Stream output token-by-token + add progress bar for model load
- generate_text is now a generator yielding partial output strings - s1 models yield after each new token; s2_hf models use TextIteratorStreamer - gr.Progress() shows model load stages and "generating..." status - Lower max_tokens default 50->30, slider max 300->150 (faster perceived response) - Subheader notes CPU inference + first-load timing92ce714 -
Drop GGUF path, keep bf16 + preload + new model entries
llama-cpp-python pre-built wheel constraints made the GGUF path unworkable on HF Spaces (manylinux cap at 0.2.62, can't load GGUFs from current llama.cpp). Reverting to transformers loader with bf16 for ~2x speedup vs fp32. Keeps the 3B-Chat-v2 and retired-3B entries and the default-model preload.c8162b7 -
Pin llama-cpp-python to direct manylinux wheel URL (0.2.62 cp311)
7b7f3ca -
Pin llama-cpp-python==0.2.90 with --prefer-binary for cp311 wheels
62f2ee4 -
Use abetlen pre-built CPU wheels for llama-cpp-python
594f0e5 -
Pin Python to 3.11 for llama-cpp-python pre-built wheels
c550b37 -
Add 3B-Chat-v2, switch large models to GGUF, preload default
- Add GPUburnout-3B-75K-Chat-v2 (proper recipe, beats 2B on 5/6 benchmarks) - Mark old GPUburnout-3B-75K-Chat as RETIRED with explanatory description - Switch 1B, 2B, 3B-old, 3B-v2 to llama-cpp-python GGUF Q4_K_M loaders (3-5x faster CPU inference than fp32 transformers) - Preload default model at startup to eliminate first-click lag - Drop transformers + accelerate (no longer needed for large models)f6da1b3
-
Switch 1B model to s2_hf loader (use HF transformers + safetensors)
The 1B repo on HF Hub is in transformers format (model.safetensors, config.json, tokenizer.json) - same as the 2B. The old s2 loader was trying to download a non-existent pytorch_model.bin and failing with 404. Now uses _load_s2_hf which already works for the 2B. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>f4c74fc
-
Force Space rebuild - verify 2B model deployment
4245e38
-
Fix 1B model ID: gpuburnout-1b -> GPUburnout-1B-160K (correct HF repo name)
f481d97 -
Add accelerate dependency for device_map support
fb8001c -
Update Gradio SDK version to 5.50.0 for build compatibility
52dcd8b -
Trigger Space rebuild
d469d94 -
Add GPUburnout-2B model and update subtitle to 2 billion parameters
- Add 2B model to MODELS dict with HF Hub integration - Update subtitle from '1 billion' to '2 billion parameters' - Add _load_s2_hf() function to load models from HF Hub - Add generate_s2_hf() function for HF model generation - Update load_model() to handle s2_hf architecture - Add transformers library error handling - Add model examples for 2B chat model992cbd9
-
Thicker borders (2px) on all boxes
1785d90 -
Add visible borders to all boxes for better contrast on dark bg
2c71a32 -
Fix dark theme - remove invalid input_text_color variable
a0a91d1 -
Use Gradio theme system for dark mode instead of CSS overrides
Set all theme variables (body, blocks, inputs, buttons, borders) to dark colors matching blog (#08080d). Raw CSS was being ignored by Gradio's component styles.43e247e -
Force dark theme with pure CSS overrides (#08080d)
JS dark mode toggle was unreliable. Override all backgrounds, text colors, inputs, and panels directly to match blog theme.0ef17a6
-
Add weights_only=True to S1 torch.load for safety
58889f8 -
Dark theme matching blog background (#08080d)
4914171 -
Fix 1B description: 11.8B tokens trained (not 30B dataset size)
743421f -
Rename all models to GPUburnout-* branding, default to 134M for faster first run
- GPUburnout-3M, GPUburnout-134M, GPUburnout-1B - Default to 134M (loads instantly vs 3.9GB download for 1B) - Lower default max tokens from 100 to 50b087aec -
Rename Llama 1B to GPUburnout-1B to match blog branding
7ac15ba -
Use CPU-only PyTorch to speed up Space startup
Full torch with CUDA is ~2.5GB vs ~200MB for CPU-only. On a cpu-basic Space, this was causing extremely slow startup.74dec22 -
Add hf_hub_download for 1B model weights
1B weights (3.9GB) exceed Space LFS limit, so download from GPUburnout/gpuburnout-1b model repo on first load.03c0bc2 -
feat: add Tiny (3.2M) and GPT-2 Small (134M) model weights
679e4e6 -
feat: add app code, configs, and tokenizers
36bc78f -
initial commit
Verified signature514a6e1