Spaces:
Sleeping
Sleeping
Commit History
-
added wandb links
45d21c6 -
added notebook link blog
b545786 -
updated broking links in readme and blog
c1aa48c -
google collab links in readme and blog
d18f9e7 -
updated blog
07438f2 -
Merge branch 'master' into hf_branch
9e3d4af -
updated training eval
a70fbb9 -
updated the blog
3d661ae -
sorted imports
8f2eab9 -
blog: expand GRPO mechanics, reward shaping narrative, and lessons learned
- Replace thin GRPO bullet list with full algorithm walkthrough: G rollouts, group normalization math, zero-std problem, num_iterations (μ) semantics - Expand two-layer reward section with credit-assignment framing and clean shaping table (training-only, invisible at inference) - Replace sparse lessons-learned bullets with full specific anecdotes: high random baseline, cold-start failure, epochs/max_steps confusion, base LLM intuition, what didn't work (num_gen=2, short completions, LoRA string path, eval batch size), what mattered most Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>2fd537a -
fix: sync README with current codebase
- Add /metadata and /health endpoints to server table - Add scripts/training_utils.py to project structure - Add scenario split tests to test coverage list Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>832ac5b -
Merge branch 'master' into hf_branch
7516665 -
updated notebok name and links in readme and blog
bc11b8d -
Merge branch 'master' into hf_branch
0dfb14e -
fixed formatting of blog
4133e36 -
added hf configurations in readme
c056ac8 -
added telemetric details to blog
6664bbd -
added more detailed content to the blog
2dcc78c -
updated blog
6d655de -
updated readme
0c7e75e -
final runs and update blog
8e2a258 -
seprated blog from readme and added new readme
fb7b148 -
docs: replace README with clean developer-facing doc; blog post in BLOG.md
Standard README covers: overview, install, server endpoints, inference usage, training pipeline, project structure, test suite, action/reward reference, deployment. Full technical write-up moved to BLOG.md. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>85118cd -
updated readme
112e90e -
ignore few files and update readme
77e6bd1 -
fix: add training plots and clean up duplicate README headings
- Extract pre_training_reward_distribution.png, post_training_eval.png, training_results.png from notebook outputs into artifacts/plots/ - Remove two stale H1 headings (Neural Tuner Env Environment, NeuralTuner: Hardware-Aware RL Environment for LLMs) — only the full descriptive title remains Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>5905982 -
Merge branch 'master' into hf_branch
28fdd67 -
updated the readme
e194352 -
docs: rewrite README as technical writeup for HF Spaces submission
Full storytelling README covering: - Problem statement: manual Snapdragon HTP optimization bottleneck - HTP hardware context (INT4/INT8 acceleration, sparse weights, DVFS) - RL environment design: partial observability, action space, simulator - Key term glossary: sensitivity, oracle ceiling, random baseline, GRPO, SFT warm-up - Reward function design and anti-hacking properties - 19 scenarios table with difficulty scaling - Training pipeline diagram (SFT → GRPO → curriculum) - Future work: live DLBC/SWC/sysmon/farf/power-config hardware integration Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>93702ff -
added more tasks
3b7f446
-
updated open env yaml config
4f52a38 -
updated build failure in uv.lock
c407759 -
Merge branch 'master' into hf_branch
9f3f578 -
chore: refresh training artifacts and rename consume_reward_components to private
- Rename consume_reward_components → _consume_reward_components to signal it's an internal helper - Update train/eval/comparison metrics from latest training run: avg_tool_calls_per_episode 2.9 → 4.125, frac_zero_std_steps 0.2 → 0.0, heldout_size 20 → 5, train_loss_final corrected from 0.0 → 0.284 - Regenerate training plots and train_prompts dataset7909885 -
Merge branch 'master' into hf_branch
6998881 -
feat: add episode trace, refresh training dataset, and update eval metrics
- Add rollout_eval.py --trace flag to generate side-by-side random vs heuristic episode trace (artifacts/eval/episode_trace.md) - Replace training prompts with expanded curriculum-focused scenarios (easy-heavy, edge cases, fast-path patterns) and remove hard/medium scenarios that may have been causing instability - Correct oracle ceiling from 0.786 → 0.6428 (heuristic agent, not theoretical max) - Add train/heldout split metrics to eval_metrics.json for cleaner generalization tracking - Update training plots and metrics with latest 403s training run - Expand README with model zoo table, reward component breakdown, episode trace comparison, updated submission checklist, and HF Space link - Add training_utils.pya422c8d -
refactor: move training code to scripts/, add train/eval split, tune GRPO hyperparams
- Relocate NeuralTunerOpenEnv and training utilities from server/ to scripts/ - Add 80/20 train/eval scenario split (TRAIN_FRACTION, SPLIT_SEED) for proper held-out evaluation - Refactor post_training_logs to evaluate held-out set against random/baseline/heuristic baselines - Switch oracle from run_baseline_episode to run_heuristic_episode (stronger ceiling) - Tune GRPO: lr 1e-6→2e-6, max_completion_length 384→256, NUM_STEPS 1→20, TRAIN_EPISODES 100→80 - Add prune_layer to ALLOWED_ACTIONS in inference.py - Replace weave with standard W&B; add WANDB_PROJECT env var and conditional report_to - Add device smoke check cell for MPS/CPU validation before trainingfad16c9 -
feat: add OpenEnv TRL wrapper, expand dataset, and add W&B eval tracking
- Add server/neural_tuner.py: OpenEnv-compatible wrapper for TRL environment_factory with scenario scheduling and step dispatch - Expand training dataset from 16 to 100 scenarios/prompts covering more model/difficulty combinations - Re-run training: non-zero loss achieved (-0.094), runtime down to 141s from 425s - Add wandb + weave dependencies for experiment tracking - Add eval artifacts: comparison_metrics.json, eval_metrics.json, and training plots - Update .gitignore to exclude wandb/ run directories6fa4fbd -
feat: add structured pruning action and random baseline policy
Extends NeuralTunerAction with a new prune_layer op (LOW/MEDIUM/HIGH sparsity) alongside quantization, so the LLM agent can now compress models via both axes simultaneously. Pruning and quantization effects stack multiplicatively for latency/memory and additively for accuracy penalty in the simulator, matching real Snapdragon HTP behaviour. Key changes: - models.py: add prune_layer action_type and sparsity field (LOW=25%, MEDIUM=50%, HIGH=75%) - server/simulator.py: PRUNE_CONFIGS table, simulate() and get_benchmark_report() now accept a prune_map; profile report includes per-layer pruning advice based on sensitivity - server/neural_tuner_env_environment.py: _prune_map state, _handle_prune() handler, revert_layer resets both dtype and sparsity, benchmark passes prune_map through to simulator - client.py: forward sparsity field in action payload - rollout_eval.py: add run_random_episode() (zero-knowledge baseline, expected reward 0.20–0.45), run_policy_sweep() to run all thr…d064b19 -
refactor: harden imports, add training extras, and rewrite README
- Wrap all openenv and relative imports in try/except so the package works both as an installed module and when run from the repo root without openenv present (models.py, client.py, server/app.py, server/neural_tuner_env_environment.py, __init__.py) - Add optional `training` dependency group to pyproject.toml (trl, transformers, datasets, accelerate, jupyter, matplotlib, pandas) - Wire main.py CLI entrypoint to rollout_eval harness - Rewrite README to describe the actual NeuralTuner RL environment (loop, reward design, quick-start, test commands) instead of the generic echo-env template - Extend .gitignore with .env.local, .hf_cache, .pytest_cache5dd60b9 -
fix: resolve merge conflict markers in openenv.yaml and uv.lock
c524b25 -
updated readme for hf deployment
3e57db4 -
Merge branch 'master' into hf_branch
0575199 -
implement NeuralTuner RL environment for Snapdragon quantization
Replace the echo stub with a full quantization optimization environment where an LLM agent profiles neural network layers and applies FP16/INT8/INT4 dtype decisions to meet Snapdragon HTP latency, memory, and accuracy budgets. - Add five agent actions: profile_layer, quantize_layer, revert_layer, benchmark, submit — dispatched via a handler table in NeuralTunerEnvironment - Introduce NeuralTunerState (extends State) and redesign Action/Observation models to carry structured fields (action_type, layer_id, dtype, output, success, error) instead of the old echo message contract - Integrate model_zoo, scenarios, and HardwareSimulator for realistic episode generation with difficulty tiers - Wire multi-component reward: latency improvement (40%), memory constraint (30%), accuracy retention (20%), efficiency bonus (10%) - Add black and isort as dev dependencies782222a -
removed hf content from readme
af8e43a -
added deployment configuration for hugging face
77bfc70 -
merge hf space
c2f781b -
first commit
15c3238