Spaces:
Runtime error
Runtime error
Commit History
-
chore: untrack server/__pycache__ files
ademarteau4ed16c1 -
fix: use correct HF router URL /hf-inference/v1/ (not /models/{id}/v1/)
ademarteau9a9473a -
fix: COGS profit model - charge unit cost on sold units not on orders, preventing end-of-period bias against high-inventory strategies
ademarteau18aac4f -
fix: update HF proxy URL to router.huggingface.co (api-inference deprecated 410)
ademarteau38aa642 -
fix: serve index.html with no-cache headers to prevent CDN/browser stale cache
ademarteau2cee429 -
fix: proxy HF Inference API through FastAPI to bypass HF Spaces CSP
ademarteauc3fc8d4 -
metrics: profit first, then service level, then fill rate (React UI)
ademarteaue7f1f53 -
fix: use minimal requirements-server.txt in Docker to avoid downloading torch/ML deps
ademarteaub2065cc -
fix: add package-lock.json for npm ci in Docker build
ademarteau0209151 -
fix: switch HF Space SDK from gradio to docker
ademarteau4992459 -
feat: replace Gradio with React UI — GRPO tab, 730-day sim, 200-entry memory bank
ademarteaub413222 -
refactor: remove Unsloth, use standard transformers + PEFT
RishbhaJain +1: Claude Sonnet 4.6Replace FastLanguageModel with AutoModelForCausalLM and LoraConfig/get_peft_model across train_grpo.py and finetune_agent.py. Remove unsloth and unsloth_zoo from requirements.txt. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>355b2d5 -
fix: use nvidia/cuda devel base image so vllm can build with CUDA toolkit
Arvind SreenivasMade-with: Cursorb52921e -
fix: pin torch 2.6.0 + xformers 0.0.29, use Python 3.12 for ML compat
Arvind SreenivasMade-with: Cursord73e520 -
fix: install torch, xformers, vllm before requirements to avoid build failures
Arvind SreenivasMade-with: Cursorea9e7b6 -
fix: install torch before vllm/xformers to satisfy build deps
Arvind SreenivasMade-with: Cursor32d0699 -
feat: add vllm for Unsloth fast_inference
Arvind SreenivasMade-with: Cursora600022 -
fix: force-reinstall unsloth to fix PreTrainedConfig NameError
Arvind Sreenivasunsloth + newer transformers has a naming conflict (PreTrainedConfig vs PretrainedConfig). Force-reinstall after pip install ensures the patched version is used. Made-with: Cursor477d0a5 -
fix: align Unsloth config with recommended GRPO settings
RishbhaJain +1: Claude Sonnet 4.6- Fix import order: unsloth must be imported before trl/transformers - Add UNSLOTH_VLLM_STANDBY=1 for 30%+ memory savings during RL - Enable load_in_fp8=True for Float8 memory efficiency - Add max_lora_rank to from_pretrained for correct pre-allocation - Remove lora_dropout (0.05 → 0) to ensure deterministic rollouts - Auto-compute lora_alpha = lora_rank * 2 when not explicitly set - Expand memory bank: 15 → 100 entries, add inventory/demand fields Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>d1c6fd5 -
Remove ideas
ademarteau7519b6d -
metrics: profit first, then service level, then fill rate
ademarteau39193b5 -
Merge branch 'main' of https://github.com/ademcodesproducts/OpenEnv-Inventory-Simulations
ademarteau84565ee -
fix: pipeline-aware ordering, YoY demand signal, reward rebalancing
RishbhaJain +1: Claude Sonnet 4.6- Fix bullwhip/overshoot: order against inventory position (on-hand + pipeline) in both _simulate_rop and inventory_env.py /step endpoint. Prevents 3 overlapping orders stacking during stockout recovery. - Add demand_last_year_7d (7-day window same period last year) to observation — surfaced from pre-generated 730-day demand series. Propagated through server model, client dataclass, and train_grpo prompts. - Shorten LOOKAHEAD_DAYS 365 → 30: matches adaptive policy horizon, removes bias toward inflated constant-policy ROPs. - Raise HOLDING_RATE 0.005 → 0.02: reduces 300:1 stockout/holding asymmetry that incentivised excessive overstocking. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>c10dcd0 -
feat: integrate Unsloth into GRPO training pipeline
RishbhaJain +1: Claude Sonnet 4.6Replace standard HF Transformers + PEFT model loading with Unsloth's FastLanguageModel for ~2x faster training and ~60% less VRAM on CUDA. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>4d42a14 -
feat: full-horizon lookahead reward (365 days, <0.5ms)
Arvind Sreenivasaf5c3c7 -
feat: crash-resilient training with dataset caching and iteration resume
Arvind Sreenivas- Save rollout dataset to disk after collection (dataset_iter_N/) - On restart, reload saved dataset instead of re-collecting - Skip iterations that already have a saved adapter (adapter_config.json) - Load latest checkpoint adapter to continue training from correct weights - Fix GRPOConfig: max_new_tokens -> max_completion_length (TRL 0.13+) Made-with: Cursor9ebd26d -
feat: tee training logs to persistent volume
Arvind SreenivasMade-with: Cursor7658153 -
Remove PPO MLP agent, update README
ademarteaue21ed94 -
feat: add Northflank training Dockerfile and start.sh
Arvind SreenivasDockerfile.northflank runs uvicorn FastAPI server + GRPO training via start.sh. Original Dockerfile stays unchanged for HuggingFace Spaces. Also fixes torch_dtype deprecation and improves training logging. Made-with: Cursorc0ce96d -
feat: improve GRPO training logging and fix torch_dtype deprecation
Arvind Sreenivas7dea3a9 -
fix: let Gradio auto-select port locally, fix via env vars only
ademarteau6d9b0d9 -
fixed all environment files
ademarteau56146fe -
fixed all environment files
ademarteauff33a83 -
fix: use 127.0.0.1 locally, 0.0.0.0 in Docker via env var
ademarteaub7d695d -
fix: Dockerfile runs Gradio app, not FastAPI server
ademarteaub5c037f -
Merge branch 'main' of https://huggingface.co/spaces/ademarteau/RL-Inventory-Simulations
ademarteau7f56785 -
fix: fall back to HF_TOKEN env var for Inference API auth
ademarteauf8a94b0 -
Upload ppo_inventory.zip with huggingface_hub
Verified signatureff80aea -
Merge: accept remote ppo model and gitattributes
ademarteauca85d91 -
Upload ppo_inventory.zip with huggingface_hub
c865376 -
Track zip files with LFS
ademarteaucc5ff9b -
Merge teammate changes, unify reward via reward.py, add PPO model
ademarteau043e4e9 -
Added PPO model and reward.py
ademarteau7ed1454 -
feat: improve training logging with tqdm, timings, GPU memory, ETA
Arvind Sreenivas766dc8c -
Merge branch 'main' of https://github.com/ademcodesproducts/OpenEnv-Inventory-Simulations
ademarteau920573d -
Added trained PPO model + app.py UI changes for HF Spaces
ademarteau3cad082 -
fix: add missing ML and simulation packages to requirements.txt
Arvind Sreenivas2344156 -
fix: remove pywin32 Windows-only packages, use Python 3.13
Arvind Sreenivas1091939 -
fix: bump to Python 3.13 to match requirements.txt (audioop-lts)
Arvind Sreenivas5482efa -
fix: remove scripts/ COPY, scripts dir not in repo
Arvind Sreenivasdd1dcec