inductive-mlxrl 0.2.1__tar.gz → 0.2.3__tar.gz

{inductive_mlxrl-0.2.1 → inductive_mlxrl-0.2.3}/CHANGELOG.md

@@ -5,6 +5,40 @@ All notable changes to this project are documented here.
 The format follows [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 and this project uses semantic versioning once releases are cut.
 
+## [0.2.3] - 2026-10-21
+
+### Changed
+
+- Project only completion positions through the output head for stock Qwen2,
+  Qwen3, and Llama models, reducing training logit memory while preserving
+  logprobs and gradients. Custom subclasses and wrappers retain the full
+  forward path. Thanks to @TarangKhanna for PR #6.
+- Refresh runtime, development, and benchmark dependencies while preserving
+  the seven-day package-age cutoff in the lockfile.
+- Include the Pillow, aiohttp, AnyIO, and urllib3 updates from Dependabot
+  PRs #1 through #4.
+
+### Fixed
+
+- Support MLX-LM 0.32 prompt-cache cloning while retaining older cache support.
+  Thanks to @TarangKhanna for PR #5.
+- Use valid prefix lengths when building compiled decode caches from MLX-LM
+  0.32 cache buffers, and restore random state through MLX 0.32's mutable key.
+- Keep scalar metric conversions compatible with the current MLX type stubs.
+- Override the benchmark-only mlx-tune dependency's old datasets constraint
+  to use patched datasets 5.x (CVE-2026-66007).
+
+## [0.2.2] - 2026-06-25
+
+### Added
+
+- ECHO world-modeling support for tagged multi-turn trajectories, composing an
+  independently normalized ECHO SFT loss with the existing GiGPO action loss.
+- `echo_alpha` scheduling with constant and linear-taper-to-zero modes, plus
+  separate `loss_echo` and `echo_accuracy` diagnostics.
+- Real-Qwen smoke coverage for tagged ECHO batches on Qwen3-0.6B and
+  Qwen3.5-9B MLX 4-bit models.
+
 ## [0.2.1] - 2026-06-25
 
 ### Fixed

{inductive_mlxrl-0.2.1 → inductive_mlxrl-0.2.3}/CONTRIBUTING.md

@@ -72,6 +72,11 @@ runtime dependencies marked for macOS arm64 and Python `<3.15` so universal
 resolvers do not try to solve MLX wheels for unsupported platform or Python
 splits.
 
+The development environment overrides mlx-tune 0.6's `datasets<4` constraint
+with `datasets>=5.0.1,<6` to avoid CVE-2026-66007. This only affects the external
+benchmark tools; the published package depends on MLX and MLX-LM. Remove the
+override when mlx-tune accepts a patched datasets release.
+
 ## Release Cycle
 
 Releases are tag-driven and publish to PyPI through Trusted Publishing. Do not

{inductive_mlxrl-0.2.1 → inductive_mlxrl-0.2.3}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.4
+Metadata-Version: 2.5
 Name: inductive-mlxrl
-Version: 0.2.1
+Version: 0.2.3
 Summary: Small single-process RL post-training for LLMs on Apple Silicon with MLX.
 Project-URL: Homepage, https://github.com/inductiveML/mlxrl
 Project-URL: Repository, https://github.com/inductiveML/mlxrl
@@ -300,6 +300,27 @@ currently exact for token-mean policy losses: base GRPO, DAPO, GSPO token mode,
 RLOO, and Dr. GRPO with `loss_reduction="token_mean"`. Sequence-reduced losses
 should keep `micro_batch_size=0`.
 
+## ECHO World Modeling
+
+ECHO is an additive GiGPO trainer feature for tagged multi-turn trajectories.
+Producer code may attach `token_roles` to each `Trajectory`, aligned to
+`full_token_ids`: `0 = MASKED`, `1 = ACTION`, `2 = ECHO`. ACTION tokens keep
+the normal GiGPO policy-gradient loss. ECHO tokens add an independently
+normalized SFT term, `alpha * NLL`, in the same forward/backward pass.
+
+`echo_alpha` defaults to `0.0`, which leaves current GiGPO behavior untouched.
+For early world-modeling, start small, roughly `0.005` to `0.05`, and consider
+`echo_schedule = "linear_taper_to_zero"` with `echo_taper_steps = N` so training
+returns to pure RL. `alpha` is model-size-sensitive; too-high values can
+overfit tool responses or collapse RL.
+
+ECHO is content-agnostic by design. `mlxrl` never inspects token text to decide
+what is an environment or tool response. The rollout producer owns the tags:
+MASKED tokens have zero gradient, ACTION tokens receive the RL advantage, and
+ECHO tokens receive only the SFT loss with no KL penalty or importance ratio.
+The trainer logs `loss_echo` and token-level `echo_acc` separately from the
+ACTION loss.
+
 ## Policy Semantics
 
 - The base model is frozen before LoRA injection.

{inductive_mlxrl-0.2.1 → inductive_mlxrl-0.2.3}/README.md

@@ -275,6 +275,27 @@ currently exact for token-mean policy losses: base GRPO, DAPO, GSPO token mode,
 RLOO, and Dr. GRPO with `loss_reduction="token_mean"`. Sequence-reduced losses
 should keep `micro_batch_size=0`.
 
+## ECHO World Modeling
+
+ECHO is an additive GiGPO trainer feature for tagged multi-turn trajectories.
+Producer code may attach `token_roles` to each `Trajectory`, aligned to
+`full_token_ids`: `0 = MASKED`, `1 = ACTION`, `2 = ECHO`. ACTION tokens keep
+the normal GiGPO policy-gradient loss. ECHO tokens add an independently
+normalized SFT term, `alpha * NLL`, in the same forward/backward pass.
+
+`echo_alpha` defaults to `0.0`, which leaves current GiGPO behavior untouched.
+For early world-modeling, start small, roughly `0.005` to `0.05`, and consider
+`echo_schedule = "linear_taper_to_zero"` with `echo_taper_steps = N` so training
+returns to pure RL. `alpha` is model-size-sensitive; too-high values can
+overfit tool responses or collapse RL.
+
+ECHO is content-agnostic by design. `mlxrl` never inspects token text to decide
+what is an environment or tool response. The rollout producer owns the tags:
+MASKED tokens have zero gradient, ACTION tokens receive the RL advantage, and
+ECHO tokens receive only the SFT loss with no KL penalty or importance ratio.
+The trainer logs `loss_echo` and token-level `echo_acc` separately from the
+ACTION loss.
+
 ## Policy Semantics
 
 - The base model is frozen before LoRA injection.