inductive-mlxrl 0.2.2__tar.gz → 0.2.3__tar.gz

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (103) hide show
  1. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/CHANGELOG.md +23 -0
  2. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/CONTRIBUTING.md +5 -0
  3. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/PKG-INFO +2 -2
  4. inductive_mlxrl-0.2.3/examples/personal_taste_grpo/README.md +763 -0
  5. inductive_mlxrl-0.2.3/examples/personal_taste_grpo/assets/terminal-results.jpg +0 -0
  6. inductive_mlxrl-0.2.3/examples/personal_taste_grpo/assets/terminal-training.jpg +0 -0
  7. inductive_mlxrl-0.2.3/examples/personal_taste_grpo/assets/ui-ranking.jpg +0 -0
  8. inductive_mlxrl-0.2.3/examples/personal_taste_grpo/assets/ui-scorer-eval.jpg +0 -0
  9. inductive_mlxrl-0.2.3/examples/personal_taste_grpo/compare.py +164 -0
  10. inductive_mlxrl-0.2.3/examples/personal_taste_grpo/config.toml +48 -0
  11. inductive_mlxrl-0.2.3/examples/personal_taste_grpo/data/briefs.json +120 -0
  12. inductive_mlxrl-0.2.3/examples/personal_taste_grpo/data/scorer.json +1584 -0
  13. inductive_mlxrl-0.2.3/examples/personal_taste_grpo/taste_reward.py +535 -0
  14. inductive_mlxrl-0.2.3/examples/personal_taste_grpo/test_reward.py +54 -0
  15. inductive_mlxrl-0.2.3/examples/personal_taste_grpo/train.py +282 -0
  16. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/__init__.py +1 -1
  17. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/cli.py +13 -13
  18. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/policy/logprobs.py +50 -7
  19. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/rollout/agentic.py +3 -3
  20. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/rollout/naive.py +2 -2
  21. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/rollout/optimized.py +20 -9
  22. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/train/grpo.py +6 -6
  23. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/train/trajectory.py +9 -9
  24. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/pyproject.toml +5 -1
  25. inductive_mlxrl-0.2.3/tests/test_cache_compatibility.py +67 -0
  26. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/tests/test_echo.py +7 -7
  27. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/tests/test_gigpo.py +2 -2
  28. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/tests/test_grpo.py +16 -16
  29. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/tests/test_optimized_rollout.py +52 -3
  30. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/tests/test_policy_logprobs.py +195 -8
  31. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/tests/test_train_grpo.py +2 -2
  32. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/tests/test_trajectory_logprobs.py +1 -1
  33. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/.gitignore +0 -0
  34. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/DESIGN.md +0 -0
  35. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/LICENSE +0 -0
  36. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/README.md +0 -0
  37. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/THIRD_PARTY_LICENSES.md +0 -0
  38. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/__init__.py +0 -0
  39. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/audit_phase4_config.py +0 -0
  40. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/external_baseline_worker.py +0 -0
  41. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate1_mlxrl_after.jsonl +0 -0
  42. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate1_mlxrl_after.md +0 -0
  43. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate1_mlxrl_after_final.jsonl +0 -0
  44. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate1_mlxrl_after_final.md +0 -0
  45. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate1_mlxrl_after_noprefixsync.jsonl +0 -0
  46. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate1_mlxrl_after_noprefixsync.md +0 -0
  47. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate1_mlxrl_before.jsonl +0 -0
  48. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate1_mlxrl_before.md +0 -0
  49. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate2_unified_smoke.jsonl +0 -0
  50. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate2_unified_smoke.md +0 -0
  51. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate3_mlx_tune_sync_256.jsonl +0 -0
  52. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate3_mlx_tune_sync_256.md +0 -0
  53. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate3_sync_blocked.jsonl +0 -0
  54. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate3_sync_blocked.md +0 -0
  55. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate3_sync_smoke.jsonl +0 -0
  56. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate3_sync_smoke.md +0 -0
  57. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate4_config_audit.json +0 -0
  58. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate4_config_audit.md +0 -0
  59. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate4_sampler_smoke.jsonl +0 -0
  60. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate4_sampler_smoke.md +0 -0
  61. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate5_full_reconciled.jsonl +0 -0
  62. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/gate5_full_reconciled.md +0 -0
  63. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/phase4_external_smoke.jsonl +0 -0
  64. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/phase4_external_smoke.md +0 -0
  65. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/phase4_full_all_targets.jsonl +0 -0
  66. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/phase4_full_all_targets.md +0 -0
  67. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/phase4_local_mlxrl_vs_mlx_lm.jsonl +0 -0
  68. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/phase4_local_mlxrl_vs_mlx_lm.md +0 -0
  69. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/phase4_mlx_lm_g4_smoke.jsonl +0 -0
  70. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/results/phase4_mlx_lm_g4_smoke.md +0 -0
  71. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/benchmarks/run_phase4.py +0 -0
  72. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/examples/qwen35_9b_g2_checkpoint.toml +0 -0
  73. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/examples/qwen3_0_6b_gigpo_recurring.toml +0 -0
  74. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/examples/qwen3_0_6b_grpo.toml +0 -0
  75. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/algo/__init__.py +0 -0
  76. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/algo/gigpo.py +0 -0
  77. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/algo/grpo.py +0 -0
  78. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/algorithm.py +0 -0
  79. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/config.py +0 -0
  80. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/data/__init__.py +0 -0
  81. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/data/gsm8k.py +0 -0
  82. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/data/rewards.py +0 -0
  83. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/echo.py +0 -0
  84. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/env/__init__.py +0 -0
  85. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/env/base.py +0 -0
  86. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/env/reference.py +0 -0
  87. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/policy/__init__.py +0 -0
  88. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/policy/model.py +0 -0
  89. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/policy/trajectory_logprobs.py +0 -0
  90. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/py.typed +0 -0
  91. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/rollout/__init__.py +0 -0
  92. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/train/__init__.py +0 -0
  93. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/train/reference.py +0 -0
  94. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/mlxrl/trajectory.py +0 -0
  95. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/tests/test_agentic_env.py +0 -0
  96. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/tests/test_agentic_rollout.py +0 -0
  97. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/tests/test_benchmark_metrics.py +0 -0
  98. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/tests/test_cli.py +0 -0
  99. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/tests/test_config.py +0 -0
  100. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/tests/test_import_direction.py +0 -0
  101. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/tests/test_policy_model.py +0 -0
  102. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/tests/test_rewards.py +0 -0
  103. {inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/tests/test_train_trajectory.py +0 -0
{inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/CHANGELOG.md RENAMED
@@ -5,6 +5,29 @@ All notable changes to this project are documented here.
5
5
  The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
6
6
  and this project uses semantic versioning once releases are cut.
7
7
 
8
+ ## [0.2.3] - 2026-10-21
9
+
10
+ ### Changed
11
+
12
+ - Project only completion positions through the output head for stock Qwen2,
13
+ Qwen3, and Llama models, reducing training logit memory while preserving
14
+ logprobs and gradients. Custom subclasses and wrappers retain the full
15
+ forward path. Thanks to @TarangKhanna for PR #6.
16
+ - Refresh runtime, development, and benchmark dependencies while preserving
17
+ the seven-day package-age cutoff in the lockfile.
18
+ - Include the Pillow, aiohttp, AnyIO, and urllib3 updates from Dependabot
19
+ PRs #1 through #4.
20
+
21
+ ### Fixed
22
+
23
+ - Support MLX-LM 0.32 prompt-cache cloning while retaining older cache support.
24
+ Thanks to @TarangKhanna for PR #5.
25
+ - Use valid prefix lengths when building compiled decode caches from MLX-LM
26
+ 0.32 cache buffers, and restore random state through MLX 0.32's mutable key.
27
+ - Keep scalar metric conversions compatible with the current MLX type stubs.
28
+ - Override the benchmark-only mlx-tune dependency's old datasets constraint
29
+ to use patched datasets 5.x (CVE-2026-66007).
30
+
8
31
  ## [0.2.2] - 2026-06-25
9
32
 
10
33
  ### Added
{inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/CONTRIBUTING.md RENAMED
@@ -72,6 +72,11 @@ runtime dependencies marked for macOS arm64 and Python `<3.15` so universal
72
72
  resolvers do not try to solve MLX wheels for unsupported platform or Python
73
73
  splits.
74
74
 
75
+ The development environment overrides mlx-tune 0.6's `datasets<4` constraint
76
+ with `datasets>=5.0.1,<6` to avoid CVE-2026-66007. This only affects the external
77
+ benchmark tools; the published package depends on MLX and MLX-LM. Remove the
78
+ override when mlx-tune accepts a patched datasets release.
79
+
75
80
  ## Release Cycle
76
81
 
77
82
  Releases are tag-driven and publish to PyPI through Trusted Publishing. Do not
{inductive_mlxrl-0.2.2 → inductive_mlxrl-0.2.3}/PKG-INFO RENAMED
@@ -1,6 +1,6 @@
1
- Metadata-Version: 2.4
1
+ Metadata-Version: 2.5
2
2
  Name: inductive-mlxrl
3
- Version: 0.2.2
3
+ Version: 0.2.3
4
4
  Summary: Small single-process RL post-training for LLMs on Apple Silicon with MLX.
5
5
  Project-URL: Homepage, https://github.com/inductiveML/mlxrl
6
6
  Project-URL: Repository, https://github.com/inductiveML/mlxrl