Spaces:
Paused
Paused
Commit History
-
fix: install nvidia-cudnn-cu12 alongside CUDA torch (removes --no-deps)
0c3df82 -
fix: force-reinstall CUDA torch before gradio import; populate logs on page load
82cc86d -
fix: eliminate API leaks in training loop, set 30k steps for A100
a5c7dd0 -
Fix CPU training + 1902ms step speed bottleneck
e8360fd -
fix: push reward curve + logs every 10k steps, not just at training end
a798c01 -
perf: add encode+step benchmark to logs, validate CUDA speed
812dbcc -
perf: force ST to CUDA, disable LLM tasks, reduce to 100k steps
b71b25c -
feat: integrate trained RecurrentPPO policy into Streamlit demo UI
236d9fb -
fix: force-install CUDA torch in app.py before training starts
83cf44b -
fix: remove torch from requirements to preserve pre-installed CUDA torch
c977821 -
fix: install CUDA-enabled torch for GPU spaces
a39792d -
pre-deployment fixes for HF Spaces
aa154f2
-
Update Cell 1 & 2 with user-verified working versions
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>175b600 -
Bulletproof Cell 1: check .git dir, rm partial clone, correct REPO path
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>3bce6e3 -
Fix REPO path: clone root is /content/kuchbhi not /content/kuchbhi/spindleflow-rl
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>21923ce -
Rewrite train_colab.py: clean cells, fix audioop-lts, log file, reward curve
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>956acce -
Remove Colab notebook — using plain .py script instead
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>c8407b4 -
Fix Cell 1: skip audioop-lts on Python < 3.13 (Colab uses 3.12)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>c77f83c -
Add Colab notebook: 8 runnable cells, both secrets, log + curve
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>13dd91f -
Final Colab training script: log file, fixed reward curve, 100k steps
- Write all episode logs to /content/logs/training_log.txt (uploaded to HF) - Fix reward curve: scatter=blue, smoothed=orange, dark theme, 180 dpi - Add early baseline + final mean lines with annotated improvement - Cap at 100k timesteps (~20-25 min on T4, ~10k episodes) - Upload training_log.txt to HF Hub in Cell 8 Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>c2b373f -
Add HF Hub push cell + fix pip deps in Colab training script
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>8e0fa29 -
Remove gr.Timer: caused C-level crash with no Python output
gr.Timer in Gradio 5.50 segfaults on A100, producing zero container logs (not even the early print statements). Replace with every=10 on demo.load() which achieves the same auto-poll without native timer code. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>278b691 -
Fix logs not showing: use file-based state instead of in-memory globals
Gradio 5 SSR mode runs event handlers in separate worker processes. Those workers have their own empty copies of _logs/_status and never see updates from the training thread in the main process. Fix: _log() already writes to a file. Now _write_status() persists phase/done/error to training_status.json. _get_state() reads both files so it works regardless of which process handles the request. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>52e0424 -
Add auto-refresh timer and dark theme CSS
- gr.Timer(10) polls training state every 10s automatically - Dark theme via CSS (no theme= param to avoid deprecation noise) - Monospace log box, styled status display Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>0c2a6e7 -
Fix theme/css: keep in Blocks(), not launch()
Gradio 5.50 does not accept theme in launch(). The deprecation warnings are for Gradio 6.0 -- both params stay in Blocks() for now. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>c0c63ff -
Improve Gradio UI: fix theme, add CSS, clean up layout
Move theme to launch() per Gradio 5 deprecation notice. Add monospace log box, bolder status display, hide footer. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>c0997b4 -
Fix create_repo naming error with huggingface_hub 1.x
Newer huggingface_hub rejects 'user/repo' as repo_id in create_repo (the slash fails server-side name validation). Pass only the repo name and set the token on HfApi() instead so the namespace is inferred. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>4bad83e -
Fix gradio/huggingface_hub version conflict on Python 3.13
Gradio 5.0.0 imports HfFolder which was removed from huggingface_hub in versions >=0.26. Bump to gradio>=5.50.0 which no longer uses HfFolder and requires huggingface_hub>=0.33.5 for compatibility. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>f9a6e3e -
Add audioop-lts for Python 3.13 pydub fix (cache bust)
aada6a8 -
Add startup diagnostics
171b6d6 -
Upgrade to gradio 5 for Python 3.13 compatibility
177e110 -
Fix Python 3.13 audioop: use audioop-lts
26e4bb4 -
Fix Python 3.13 audioop compatibility
fff9a1f -
Add HF Space configuration to README
5686ade -
SpindleFlow RL — periodic push + log persistence
02ff91f