soup-cli 0.71.28__tar.gz → 0.71.30__tar.gz

{soup_cli-0.71.28 → soup_cli-0.71.30}/CHANGELOG.md

@@ -12,6 +12,73 @@ reproducing 70+ versions of notes.
 
 ## [Unreleased]
 
+## [0.71.30] - 2026-07-05
+
+### Added
+- **PRM-guided GRPO** — use a trained Process Reward Model as the *per-step*
+  reward inside GRPO (the o1-era process-supervision signal). Set
+  `training.prm_reward: <PRM dir|id>` (a model produced by `soup train`
+  `task=prm`) and `training.prm_aggregate: min|prod|last`; the PRM splits each
+  generated completion into reasoning steps, scores every step with its reward
+  head, and folds the per-step scores into one scalar reward that GRPO
+  optimises. The PRM reward *replaces* `reward_fn` and rides the existing
+  reward-shaping + reward-hack-mitigation seam, so the v0.71.26 controller
+  observes it. Cross-validators gate `task='grpo'` + `backend='transformers'` +
+  `modality='text'`. Default aggregation is `min` (weakest-link); `prod`
+  assumes calibrated `[0,1]` step scores.
+- **Bundled rollout environments** — three pure-Python toy environments
+  (`soup_cli.envs.calculator` / `retrieval_qa` / `guess_number`) exposing a
+  `rollout(prompts)` entry point so the live `openenv` GRPO rollout path runs
+  out-of-the-box: `training.rollout_backend=openenv` +
+  `training.rollout_func=soup_cli.envs.calculator:rollout`. Three ready-made
+  recipes added (`grpo-env-calculator` / `grpo-env-retrieval-qa` /
+  `grpo-env-guess-number`); catalog 134 → 137.
+
+### Fixed
+- **`soup train task=prm` producer conformance** (surfaced by the v0.71.30 live
+  smoke): the PRM trainer now casts its reward head to the base-model dtype
+  (bf16 CUDA runs previously crashed on the first `compute_loss`), saves the
+  tokenizer alongside the model (so a PRM checkpoint is loadable standalone),
+  and returns the standard trainer-result shape (previously the CLI crashed with
+  a `KeyError: 'initial_loss'` right after saving).
+
+### Notes
+- Proof-of-mechanism only: validated on a tiny model (SmolLM2-135M) with a tiny
+  synthetic PRM and synthetic reward — not a production reward-model claim
+  (scale ask tracked in #286). The bundled environments are deterministic
+  single-shot *seeders*, not interactive multi-turn model-in-the-loop episodes
+  (the live `openenv` contract passes only prompts). Step split is a newline
+  heuristic; PRM completions are scored one forward pass each.
+
+## [0.71.29] - 2026-07-05
+
+### Added
+- **`soup shrink`** — depth-prune a model + optional distill-heal
+  ("The Unreasonable Ineffectiveness of the Deeper Layers", arXiv:2403.17887).
+  Ranks decoder layers by the angular distance of the residual stream across a
+  contiguous block over a calibration set, drops the least-important block
+  (first and last layer always protected), optionally *heals* by distilling the
+  original model into the pruned student, and emits a single dense smaller model
+  plus a one-screen **SHIP / DON'T SHIP** perplexity verdict.
+  `soup shrink --model <id|path> [--drop-ratio 0.25 | --drop-layers N] --calib
+  <calib.jsonl> [--heal <heal.jsonl> --heal-steps N] [--tolerance 0.10]
+  [-o <dir>] [--device cpu] [--attach-to-registry <id>] [--plan-only]`.
+  Exit codes: 0 = SHIP, 2 = DON'T SHIP, 1 = error. The heal runs as an isolated
+  `soup train` subprocess (LoRA-student logit distillation) and the adapter is
+  fused back so the shipped artifact stays a single dense model. Arch allowlist
+  v1: Llama / Qwen / SmolLM. Validated live on SmolLM2-135M (drop 25 %: 30 -> 22
+  layers, ppl x2.98 unhealed; drop 4 + heal: ppl x1.35, recovered).
+
+### Security
+- `soup shrink` contains `--calib` / `--heal` / `--output-dir` (and every
+  derived write path: `<out>/model`, `<out>/heal_adapter`, the fuse staging dir)
+  under cwd with `os.path.realpath` + `commonpath` + O_NOFOLLOW + symlink
+  rejection, re-validating derived paths right before each write (TOCTOU). The
+  heal subprocess uses an argv list (no shell) with a timeout; its config is
+  schema-validated before spawn; subprocess output is C0/ESC-stripped before it
+  reaches the terminal. `--model` defaults `trust_remote_code=False` with a
+  probe + warn.
+
 ## [0.71.28] - 2026-07-04
 
 ### Added

{soup_cli-0.71.28 → soup_cli-0.71.30}/CONTRIBUTING.md

@@ -110,17 +110,17 @@ src/soup_cli/
   experiment/         - SQLite experiment tracking
   eval/               - Eval platform (custom tasks, LLM judge, human eval, leaderboard)
   migrate/            - Config migration (LLaMA-Factory, Axolotl, Unsloth)
-  recipes/            - Ready-made configs for popular models (134 recipes)
+  recipes/            - Ready-made configs for popular models (137 recipes)
   autopilot/          - Zero-config decision engine (v0.25.0)
   registry/           - Model Registry (hashing, store, diff, attach) (v0.26.0 + v0.33.0)
   cans/               - Shareable .can artifact format + run/publish orchestrator (v0.26.0 + v0.33.0)
   data/traces/        - Trace-to-Preference harvester (v0.26.0)
   data/collators.py   - CrossDocCollator for sample packing (v0.33.0)
-  utils/              - GPU, errors, MoE, GaLore, QAT, Unsloth, vLLM, SGLang, Liger, FlashAttn, FSDP, Ring Attention, long-context, quality, curriculum, freeze, dataset-registry, mlx, peft_builder, paths, topology, launcher, mii, pipeline, cut_ce, fp8, gradient_ckpt, kernel_picker, cross_doc_attn, activation_offload, hf, spec_pairing, structured_output, metrics, tracing, auto_quant, lr_finder, grad_accum, mixed_precision, warmup, spike_recovery, convergence, v028_features, multipack_sampler, multipack, neat_packing, jinja_analyzer, quant_menu, relora, peft_patches, peft_wiring, dpo_variants, optimizer_zoo, lr_groups, loftq_init, block_expansion, tts, classifier, distill, bitnet, ebft_gdpo, moe_quant, reasoning_effort, gguf_quant, kv_cache, advanced_precision, save_formats, deploy_measure, advise, advise_history, adapter_diff, adapter_merge, blame, adapter_branch, unlearning, unlearning_eval, knowledge_edit, edit_governor, edit_diff, ra_dit, steering, citation_faithful, grace_codebook, ingest_sources, prune_prompt, active_sampler, ab_test, drift_alarm, tunability, terraform_plan, env_lock, hardware_fit, completions, license_advisor, behavior_battery, capability_suite, checklist_dsl, irt, sae_diff, sleeper_probe, interference, probe_pack, cmaes_merge, vector_bank, mole_routing, adapter_pr, soup_lock, adapter_bisect, prompt_compile, prompt_distill, compile_tools, apple_adapter, local_rl, build_dag, expectations, magpie, persona_hub, brain_rot, reward_hacking, uld, minillm, rl_checkpoint, iterative_dpo, echo_trap, mod, local_rl_scheduler, spectrum_scan, ship_verdict, reward_hack_control, data_doctor, data_lint
+  utils/              - GPU, errors, MoE, GaLore, QAT, Unsloth, vLLM, SGLang, Liger, FlashAttn, FSDP, Ring Attention, long-context, quality, curriculum, freeze, dataset-registry, mlx, peft_builder, paths, topology, launcher, mii, pipeline, cut_ce, fp8, gradient_ckpt, kernel_picker, cross_doc_attn, activation_offload, hf, spec_pairing, structured_output, metrics, tracing, auto_quant, lr_finder, grad_accum, mixed_precision, warmup, spike_recovery, convergence, v028_features, multipack_sampler, multipack, neat_packing, jinja_analyzer, quant_menu, relora, peft_patches, peft_wiring, dpo_variants, optimizer_zoo, lr_groups, loftq_init, block_expansion, tts, classifier, distill, bitnet, ebft_gdpo, moe_quant, reasoning_effort, gguf_quant, kv_cache, advanced_precision, save_formats, deploy_measure, advise, advise_history, adapter_diff, adapter_merge, blame, adapter_branch, unlearning, unlearning_eval, knowledge_edit, edit_governor, edit_diff, ra_dit, steering, citation_faithful, grace_codebook, ingest_sources, prune_prompt, active_sampler, ab_test, drift_alarm, tunability, terraform_plan, env_lock, hardware_fit, completions, license_advisor, behavior_battery, capability_suite, checklist_dsl, irt, sae_diff, sleeper_probe, interference, probe_pack, cmaes_merge, vector_bank, mole_routing, adapter_pr, soup_lock, adapter_bisect, prompt_compile, prompt_distill, compile_tools, apple_adapter, local_rl, build_dag, expectations, magpie, persona_hub, brain_rot, reward_hacking, uld, minillm, rl_checkpoint, iterative_dpo, echo_trap, mod, local_rl_scheduler, spectrum_scan, ship_verdict, reward_hack_control, data_doctor, data_lint, shrink
   templates/          - 17 built-in soup.yaml templates (YAML + manifest.json) with load_template loader (v0.39.0, +bco v0.40.0)
   ui/                 - Web UI (FastAPI + HTML/JS SPA)
 
-tests/                - Test suite (305 files, 15140 tests)
+tests/                - Test suite (307 files, 15334 tests)
 examples/             - Real-world config examples and datasets
 ```
 
@@ -150,7 +150,7 @@ pytest tests/test_data.py::test_detect_alpaca_format -v
 pytest tests/ --cov=soup_cli --cov-report=html
 ```
 
-### Test Files (305 files)
+### Test Files (306 files)
 
 > A representative sample of the suite below. The full table lives in
 > [`.claude/CLAUDE.md`](.claude/CLAUDE.md); run `pytest tests/ -v` for the complete list.

{soup_cli-0.71.28 → soup_cli-0.71.30}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: soup-cli
-Version: 0.71.28
+Version: 0.71.30
 Summary: Fine-tune and post-train LLMs in one command. No SSH, no config hell.
 Project-URL: Homepage, https://github.com/MakazhanAlpamys/Soup
 Project-URL: Repository, https://github.com/MakazhanAlpamys/Soup
@@ -193,25 +193,27 @@ infrastructure instead of improving models. Soup fixes that.
 
 ## What's New
 
-**v0.71.28 — Drive Soup from your coding agent: `soup mcp serve`.** A Model Context Protocol server, so Claude Code / Cursor / Cline / Continue can run Soup conversationally. No other fine-tuning CLI ships one.
-
-- **14 read-only tools, JSON out.** `advise`, `data_inspect` / `validate` / `score` / `doctor`,
-  `recipes_search` / `show`, `runs_list` / `show`, `registry_list` / `show`, `profile`,
-  `diagnose_evidence`, `ship_evidence` — your agent can inspect data, search recipes, read runs,
-  and get a ship verdict without you leaving the chat.
-- **Plan-only by default, safe by design.** stdio transport only (no network listener); the two
-  mutating tools (`train_start`, `export`) are gated behind `--allow-mutating` and only ever
-  render the exact command that would run. Every path argument re-enters cwd containment +
-  symlink rejection; output is control-char sanitized; errors are path-free.
-- **Light install.** The official `mcp` SDK lives behind a new `[mcp]` extra
-  (`pip install 'soup-cli[mcp]'`), lazy-imported so the core CLI stays torch-free and fast.
-- **Live-validated on Windows.** A real stdio MCP client drove `initialize → list_tools →
-  call` end-to-end on Windows + RTX 3050 — read tools return real JSON, the mutating tools
-  refuse without the flag, and a path-traversal argument is rejected.
-
-```jsonc
-// .mcp.json (Claude Code) or claude_desktop_config.json
-{ "mcpServers": { "soup": { "command": "soup", "args": ["mcp", "serve"] } } }
+**v0.71.30 — Process-supervised RL: PRM-guided GRPO.** Use a trained Process Reward Model to score each *reasoning step* of a GRPO completion — the o1-era training signal — plus bundled toy environments so the openenv rollout path runs out-of-the-box.
+
+- **PRM as the GRPO reward.** Point `training.prm_reward` at a PRM you trained with
+  `soup train task=prm`; it splits each completion into steps, scores every step with the PRM's
+  reward head, and folds them (`min` / `prod` / `last`) into one reward that GRPO optimises —
+  replacing `reward_fn`. It rides the reward-shaping + reward-hack-mitigation seam, so the
+  v0.71.26 controller still watches it.
+- **Bundled rollout environments.** `soup_cli.envs.calculator` / `retrieval_qa` / `guess_number`
+  each expose a `rollout(prompts)` entry point; wire one with `rollout_backend=openenv` +
+  `rollout_func=soup_cli.envs.calculator:rollout` (three ready-made `grpo-env-*` recipes ship it).
+- **Producer fixes too.** `soup train task=prm` now saves its tokenizer and no longer crashes on
+  its own summary — a PRM checkpoint is loadable standalone.
+- **Proof-of-mechanism, honestly.** Live-validated on SmolLM2-135M (CPU): the PRM reward scores
+  good completions above bad and drives GRPO's advantages (`rewards/prm_reward` logged). Tiny
+  model + synthetic PRM — not a production reward-model claim (scale help wanted: #286).
+
+```yaml
+task: grpo
+training:
+  prm_reward: ./my-prm            # a `soup train task=prm` checkpoint dir
+  prm_aggregate: min              # weakest-link (default) | prod | last
 ```
 
 Full history: [CHANGELOG.md](CHANGELOG.md) · [GitHub Releases](https://github.com/MakazhanAlpamys/Soup/releases).

{soup_cli-0.71.28 → soup_cli-0.71.30}/README.md

@@ -49,25 +49,27 @@ infrastructure instead of improving models. Soup fixes that.
 
 ## What's New
 
-**v0.71.28 — Drive Soup from your coding agent: `soup mcp serve`.** A Model Context Protocol server, so Claude Code / Cursor / Cline / Continue can run Soup conversationally. No other fine-tuning CLI ships one.
-
-- **14 read-only tools, JSON out.** `advise`, `data_inspect` / `validate` / `score` / `doctor`,
-  `recipes_search` / `show`, `runs_list` / `show`, `registry_list` / `show`, `profile`,
-  `diagnose_evidence`, `ship_evidence` — your agent can inspect data, search recipes, read runs,
-  and get a ship verdict without you leaving the chat.
-- **Plan-only by default, safe by design.** stdio transport only (no network listener); the two
-  mutating tools (`train_start`, `export`) are gated behind `--allow-mutating` and only ever
-  render the exact command that would run. Every path argument re-enters cwd containment +
-  symlink rejection; output is control-char sanitized; errors are path-free.
-- **Light install.** The official `mcp` SDK lives behind a new `[mcp]` extra
-  (`pip install 'soup-cli[mcp]'`), lazy-imported so the core CLI stays torch-free and fast.
-- **Live-validated on Windows.** A real stdio MCP client drove `initialize → list_tools →
-  call` end-to-end on Windows + RTX 3050 — read tools return real JSON, the mutating tools
-  refuse without the flag, and a path-traversal argument is rejected.
-
-```jsonc
-// .mcp.json (Claude Code) or claude_desktop_config.json
-{ "mcpServers": { "soup": { "command": "soup", "args": ["mcp", "serve"] } } }
+**v0.71.30 — Process-supervised RL: PRM-guided GRPO.** Use a trained Process Reward Model to score each *reasoning step* of a GRPO completion — the o1-era training signal — plus bundled toy environments so the openenv rollout path runs out-of-the-box.
+
+- **PRM as the GRPO reward.** Point `training.prm_reward` at a PRM you trained with
+  `soup train task=prm`; it splits each completion into steps, scores every step with the PRM's
+  reward head, and folds them (`min` / `prod` / `last`) into one reward that GRPO optimises —
+  replacing `reward_fn`. It rides the reward-shaping + reward-hack-mitigation seam, so the
+  v0.71.26 controller still watches it.
+- **Bundled rollout environments.** `soup_cli.envs.calculator` / `retrieval_qa` / `guess_number`
+  each expose a `rollout(prompts)` entry point; wire one with `rollout_backend=openenv` +
+  `rollout_func=soup_cli.envs.calculator:rollout` (three ready-made `grpo-env-*` recipes ship it).
+- **Producer fixes too.** `soup train task=prm` now saves its tokenizer and no longer crashes on
+  its own summary — a PRM checkpoint is loadable standalone.
+- **Proof-of-mechanism, honestly.** Live-validated on SmolLM2-135M (CPU): the PRM reward scores
+  good completions above bad and drives GRPO's advantages (`rewards/prm_reward` logged). Tiny
+  model + synthetic PRM — not a production reward-model claim (scale help wanted: #286).
+
+```yaml
+task: grpo
+training:
+  prm_reward: ./my-prm            # a `soup train task=prm` checkpoint dir
+  prm_aggregate: min              # weakest-link (default) | prod | last
 ```
 
 Full history: [CHANGELOG.md](CHANGELOG.md) · [GitHub Releases](https://github.com/MakazhanAlpamys/Soup/releases).

{soup_cli-0.71.28 → soup_cli-0.71.30}/docs/README.md

@@ -8,7 +8,7 @@ feature reference — every `soup` capability, grouped by area.
 | Guide | Covers |
 |---|---|
 | [Training tasks & methods](training.md) | SFT, DPO/GRPO/PPO/KTO/ORPO/SimPO/IPO/BCO, tool-calling, PRM, pre-training, distillation, classification, vision/audio/TTS, unlearning, RAFT/RA-DIT, loop-hardening detectors |
-| [PEFT, long context & efficiency](peft-and-efficiency.md) | DoRA, LoRA+, rsLoRA, VeRA, OLoRA, NEFTune, PiSSA, ReLoRA, optimizer & PEFT zoo, LLaMA Pro, GaLore, YaRN/LongLoRA, packing, curriculum, auto-tuning |
+| [PEFT, long context & efficiency](peft-and-efficiency.md) | DoRA, LoRA+, rsLoRA, VeRA, OLoRA, NEFTune, PiSSA, ReLoRA, optimizer & PEFT zoo, LLaMA Pro, GaLore, YaRN/LongLoRA, packing, curriculum, auto-tuning, depth pruning + distill-heal (`soup shrink`) |
 | [Performance & quantization](performance-and-quantization.md) | QAT, FP8, Quant Menu (I + II), KV-cache, NVFP4, save formats, Cut Cross-Entropy, gradient checkpointing, kernels, activation offloading, multi-GPU / DeepSpeed / FSDP |
 | [Data engineering](data.md) | Formats, the Axolotl/LF-parity pipeline, data tools, synthetic generation & forge, quality scorecards, trace tooling, remote datasets, mixing, recipe DAGs |
 | [Evaluation & probes](evaluation.md) | Eval design/gate, eval-gated training, benchmarks, NLG metrics, calibration, Elo arena, diagnose, `soup ship` verdict, post-train X-ray probes, A/B, drift, tunability, `soup advise` |

{soup_cli-0.71.28 → soup_cli-0.71.30}/docs/commands.md

@@ -133,7 +133,7 @@ soup migrate --from llamafactory config.yaml  Import config from LLaMA-Factory
 soup migrate --from axolotl config.yml        Import config from Axolotl
 soup migrate --from unsloth notebook.ipynb    Import config from Unsloth notebook
 soup migrate --from llamafactory c.yaml --dry-run  Preview without writing
-soup recipes list                             List all 134 ready-made recipes
+soup recipes list                             List all 137 ready-made recipes
 soup recipes show llama3.1-8b-sft            Print recipe YAML
 soup recipes use llama3.1-8b-sft             Copy recipe to soup.yaml
 soup recipes search "reasoning"              Search by keyword/task/size
@@ -158,6 +158,9 @@ soup ship ... --task-mode judge_score --judge-model ollama://llama3.1  Leg-1 via
 soup ship ... --general-suite mmlu,gsm8k --baseline base.json  lm-eval leg-2 + recorded base scores
 soup mcp serve                                MCP server over stdio (drive Soup from Claude Code / Cursor / Cline; requires [mcp] extra) (v0.71.28)
 soup mcp serve --allow-mutating               Also expose plan-only train_start / export tools (never execute) (v0.71.28)
+soup shrink --model <id|path> --drop-ratio 0.25 --calib c.jsonl -o shrunk  Depth-prune least-important layer block + SHIP/DON'T-SHIP ppl verdict (exit 0/2/1) (v0.71.29)
+soup shrink ... --drop-layers N --heal h.jsonl --heal-steps 200 --device cpu  Drop N layers + distill-heal (fuse LoRA back to one dense model)
+soup shrink ... --tolerance 0.10 --plan-only [--attach-to-registry <id>]  Ppl-regression tolerance / print importance table only / registry attach
 soup tui                                      Full-screen Textual dashboard (requires [tui] extra)
 soup train --config soup.yaml --profile       Record torch.profiler trace to <output>/profiles/
 soup --log-level quiet|normal|verbose|debug   Global logging tier (Rich-formatted)

{soup_cli-0.71.28 → soup_cli-0.71.30}/docs/peft-and-efficiency.md

@@ -24,6 +24,57 @@
 - [Training Stability & Auto-Tuning](#training-stability--auto-tuning)
 - [Training Intelligence (Forgetting + Checkpoint Quality)](#training-intelligence-forgetting--checkpoint-quality)
 - [GaLore (Memory-Efficient Full-Parameter Training)](#galore-memory-efficient-full-parameter-training)
+- [Depth Pruning + Distill-Heal (`soup shrink`)](#depth-pruning--distill-heal-soup-shrink)
+
+---
+
+## Depth Pruning + Distill-Heal (`soup shrink`)
+
+`soup shrink` makes a model smaller by dropping its least-important **contiguous
+block of decoder layers**, then optionally *healing* the loss with knowledge
+distillation. It implements "The Unreasonable Ineffectiveness of the Deeper
+Layers" (Gromov et al., arXiv:2403.17887), with a Minitron-style distillation
+heal instead of the paper's plain LoRA fine-tune.
+
+**How it ranks layers.** For each candidate block `[L, L+n)`, one forward pass
+per calibration prompt (`output_hidden_states=True`) measures the **angular
+distance** of the residual stream entering (`hidden_states[L]`) vs leaving
+(`hidden_states[L+n]`) the block, averaged over every non-pad token across the
+calibration set. The block with the *lowest* distance transforms the residual
+stream least, so it is the safest to drop. The first and last decoder layers are
+always protected (they carry the most transformation).
+
+```bash
+# See the importance table + chosen block without writing anything:
+soup shrink --model HuggingFaceTB/SmolLM2-135M-Instruct --drop-ratio 0.25 \
+    --calib calib.jsonl --device cpu --plan-only
+
+# Prune 25% + heal (distill the original into the pruned student, fuse to one
+# dense model) + get a SHIP/DON'T-SHIP perplexity verdict:
+soup shrink --model HuggingFaceTB/SmolLM2-135M-Instruct --drop-ratio 0.25 \
+    --calib calib.jsonl --heal heal.jsonl --heal-steps 200 \
+    --tolerance 0.10 -o shrunk --device cpu --attach-to-registry <id>
+```
+
+- **`--drop-ratio F` / `--drop-layers N`** (exactly one) — how many contiguous
+  layers to drop; the *position* is chosen automatically by the importance scan.
+- **`--calib <jsonl>`** — calibration prompts (`{"text": ...}` / `{"prompt":
+  ...}` / chat `messages`). Must stay under cwd.
+- **`--heal <jsonl> --heal-steps N`** — distill the full-depth original into the
+  pruned student (LoRA logit-KD) as an isolated `soup train` subprocess, then
+  fuse the adapter back into the pruned base so the output is a single dense
+  model. Heal keeps the teacher resident (~2× model memory); with `--device
+  cpu` the heal runs on CPU (validated ≤ 3 B on a 4 GB card).
+- **`--tolerance F`** — ship if the perplexity regression stays within `F`
+  (default `0.10` = 10 %). Exit code `0 = SHIP`, `2 = DON'T SHIP`, `1 = error`.
+- **`-o <dir>`** — the shrunk model lands in `<dir>/model`; the verdict in
+  `<dir>/shrink_report.json`.
+
+**Arch support (v1):** Llama / Qwen / SmolLM. Others are a friendly reject.
+The importance pass loads the model, so live-validated on ≤ 3 B; larger models
+work but are unvalidated on the reference hardware. Perplexity is an unweighted
+mean of per-example perplexities — valid for the before/after *ratio* the
+verdict uses, not directly comparable to `soup eval` absolute numbers.
 
 ---
 

soup_cli-0.71.30/docs/superpowers/plans/2026-07-05-prm-guided-grpo.md

@@ -0,0 +1,133 @@
+# PRM-guided GRPO + bundled rollout envs — Implementation Plan
+
+> **For agentic workers:** REQUIRED SUB-SKILL: superpowers:test-driven-development, task-by-task. Steps use checkbox (`- [ ]`) syntax.
+
+**Goal:** Let a trained Soup PRM score each reasoning step of a GRPO completion as the reward signal, and bundle 3 pure-python toy environments that seed the live `openenv` rollout path out-of-the-box.
+
+**Architecture:** A pure-kernel + torch-lazy `PRMScorer` in `utils/prm_reward.py` becomes the GRPO `reward_fn` when `training.prm_reward` is set (replacing the configured reward, riding the existing shaping+buffer seam). Three `soup_cli/envs/*.py` modules expose `rollout(prompts)` entry points wired via `training.rollout_func`. Schema fields + cross-validators gate both.
+
+**Tech Stack:** Python 3.10+, Pydantic v2, torch/transformers/safetensors (lazy), TRL GRPO, pytest.
+
+## Global Constraints
+
+- Line length 100 (ruff E,F,I,N,W). No bare `print()` — use `rich.console.Console`.
+- Lazy imports for torch/transformers/peft/safetensors inside functions, never module-level.
+- Path containment: `os.path.realpath` + `os.path.commonpath`, NOT `Path.resolve()+relative_to()`.
+- Pydantic v2 models in `config/schema.py`; assert specific error keywords in tests.
+- Test file: `tests/test_v07130.py`. Use `assert result.exit_code == 0, (result.output, repr(result.exception))`.
+- Version target: `0.71.29 → 0.71.30` (bump in `pyproject.toml` + `src/soup_cli/__init__.py`).
+- Proof-of-mechanism honesty: tiny model / synthetic data only; scale ask extends #286.
+
+---
+
+### Task 1: PRM reward pure kernels
+
+**Files:**
+- Create: `src/soup_cli/utils/prm_reward.py`
+- Test: `tests/test_v07130.py`
+
+**Interfaces:**
+- Produces: `split_steps(text: str) -> list[str]`, `aggregate_step_scores(scores: list[float], mode: str) -> float`, module constants `_MAX_STEPS`, `_MAX_STEP_CHARS`, `AGGREGATE_MODES: tuple[str,...]`.
+
+- [ ] Step 1: Write failing tests `TestSplitSteps` (newline split; drop empty/whitespace; count cap `_MAX_STEPS`; per-step char cap; non-str→`[]`) and `TestAggregate` (min/prod/last correctness; empty→0.0; single; non-finite→treated safe; bad mode `ValueError` names modes; bool reject).
+- [ ] Step 2: Run — expect ImportError / fail.
+- [ ] Step 3: Implement kernels (no top-level torch). `split_steps`: `str.splitlines()`, strip, drop empties, truncate each to `_MAX_STEP_CHARS`, cap list to `_MAX_STEPS`. `aggregate_step_scores`: validate mode in `AGGREGATE_MODES`; empty→0.0; `min`→min, `last`→scores[-1], `prod`→math.prod; coerce non-finite to 0.0.
+- [ ] Step 4: Run tests → PASS.
+- [ ] Step 5: `ruff check` + commit `feat(prm): PRM reward pure kernels (split_steps/aggregate) (v0.71.30)`.
+
+---
+
+### Task 2: Schema fields + cross-validators
+
+**Files:**
+- Modify: `src/soup_cli/config/schema.py` (TrainingConfig fields near `reward_fn` ~L1043; field_validator; SoupConfig cross-validator near `_validate_rollout_backend` ~L4183)
+- Test: `tests/test_v07130.py`
+
+**Interfaces:**
+- Produces: `TrainingConfig.prm_reward: Optional[str]`, `TrainingConfig.prm_aggregate: Literal["min","prod","last"]`; a `SoupConfig` `model_validator(mode="after")` `_validate_prm_reward`.
+
+- [ ] Step 1: Write failing `TestPrmSchema` — happy config (task=grpo, transformers, text, prm_reward set, prm_aggregate=prod) parses via `load_config_from_string`; rejects: task≠grpo ("requires task='grpo'"), backend=mlx/unsloth ("transformers"), modality≠text ("modality='text'"), `prm_aggregate` non-default while `prm_reward None` (footgun keyword), null-byte prm_reward, oversize prm_reward.
+- [ ] Step 2: Run → fail (fields absent).
+- [ ] Step 3: Add fields + `field_validator("prm_reward", mode="before")` (null-byte/non-str/len≤512 reject) + SoupConfig `_validate_prm_reward` cross-validator (gates task/backend/modality; footgun on aggregate). Mirror `_validate_rollout_backend`.
+- [ ] Step 4: Run → PASS.
+- [ ] Step 5: ruff + commit `feat(prm): schema prm_reward/prm_aggregate + cross-validators (v0.71.30)`.
+
+---
+
+### Task 3: PRMScorer + build_prm_reward_fn (torch-lazy load + scoring)
+
+**Files:**
+- Modify: `src/soup_cli/utils/prm_reward.py`
+- Test: `tests/test_v07130.py`
+
+**Interfaces:**
+- Consumes: `split_steps`, `aggregate_step_scores` (Task 1); `TrainingConfig` (Task 2).
+- Produces: `PRMScorer(prm_path, aggregate, device="cpu", trust_remote_code=False)` with `__call__(completions, **kwargs) -> list[float]` and `__name__="prm_reward"`; `build_prm_reward_fn(tcfg, device, trust_remote_code) -> PRMScorer`; `load_reward_head_weights(prm_path) -> dict` helper.
+
+- [ ] Step 1: Write failing `TestPRMScorer` — fixture: build a tiny `AutoModelForCausalLM` from a 2-layer config, attach `reward_head=nn.Linear(hidden,1)`, `save_pretrained(tmp)`; `PRMScorer(tmp,"min","cpu")` scores `[[{"role":"assistant","content":"step1\nstep2"}]]` → list len 1 of finite float. "not a PRM" fixture (base with no reward_head) → friendly `ValueError`/`RuntimeError` naming "reward_head". `build_prm_reward_fn` returns callable with `__name__=="prm_reward"`; outside-cwd path → containment reject.
+- [ ] Step 2: Run → fail.
+- [ ] Step 3: Implement. `load_reward_head_weights`: iterate `*.safetensors` in dir via `safetensors.safe_open`, collect `reward_head.*` tensors; empty→raise. `PRMScorer._ensure_loaded`: lazy tokenizer + `AutoModelForCausalLM.from_pretrained`, `reward_head=nn.Linear(hidden,1)`, `load_state_dict` the head, `eval()`+`requires_grad_(False)`+`.to(device)`. `__call__`: per completion → render prompt (from `kwargs["prompts"]`) + steps, tokenize, `no_grad` `output_hidden_states` forward, gather boundary hidden, `reward_head`→scalars, `aggregate_step_scores`. `build_prm_reward_fn`: realpath+commonpath containment for local paths, trust_remote_code probe+warn.
+- [ ] Step 4: Run → PASS.
+- [ ] Step 5: ruff + commit `feat(prm): PRMScorer + build_prm_reward_fn (safetensors head load) (v0.71.30)`.
+
+---
+
+### Task 4: GRPO integration
+
+**Files:**
+- Modify: `src/soup_cli/trainer/grpo.py` (`setup()`, the `load_reward_fn` block ~L224-229)
+- Test: `tests/test_v07130.py`
+
+**Interfaces:**
+- Consumes: `build_prm_reward_fn` (Task 3).
+
+- [ ] Step 1: Write failing `TestGrpoPrmWiring` — construct a SoupConfig with `prm_reward` set, monkeypatch `build_prm_reward_fn` to a sentinel, call the reward-selection path (extract a small helper `_select_reward_fn(tcfg, device, trust)` if cleaner, or assert via the setup path with mocked model load), assert the PRM scorer is chosen (not `load_reward_fn`).
+- [ ] Step 2: Run → fail.
+- [ ] Step 3: In `setup()`: `if tcfg.prm_reward is not None: reward_fn = build_prm_reward_fn(tcfg, self.device, self._trust_remote_code) else: reward_fn = load_reward_fn(...)`. Keep the shaping+buffer wrapping downstream unchanged.
+- [ ] Step 4: Run → PASS.
+- [ ] Step 5: ruff + commit `feat(prm): wire PRM reward into GRPO setup (v0.71.30)`.
+
+---
+
+### Task 5: Bundled rollout envs
+
+**Files:**
+- Create: `src/soup_cli/envs/__init__.py`, `calculator.py`, `retrieval_qa.py`, `guess_number.py`
+- Test: `tests/test_v07130.py`
+
+**Interfaces:**
+- Produces: `soup_cli.envs.calculator:rollout`, `...retrieval_qa:rollout`, `...guess_number:rollout` — each `rollout(prompts) -> list[{"prompt": str, "answer": str}]`, deterministic (seeded `random.Random`), no I/O.
+
+- [ ] Step 1: Write failing `TestEnvs` — each `rollout([])` returns non-empty rows; every row passes `agent_rollout._normalise_rollout_rows(rows, "openenv")`; determinism (two calls equal); calculator answers verified (parse "a OP b" from prompt, compute, assert == answer).
+- [ ] Step 2: Run → fail.
+- [ ] Step 3: Implement 3 modules. Each builds a fixed count of {prompt, answer} rows via `random.Random(seed)` (fixed seed → deterministic; no `Math.random`/wall-clock). calculator: arithmetic; retrieval_qa: doc+question, answer=span; guess_number: constraint deduction puzzles with a unique integer answer.
+- [ ] Step 4: Run → PASS.
+- [ ] Step 5: ruff + commit `feat(envs): bundled calculator/retrieval-qa/guess-number rollout envs (v0.71.30)`.
+
+---
+
+### Task 6: Recipes + no-top-level-torch AST test
+
+**Files:**
+- Modify: `src/soup_cli/recipes/catalog.py` (+3 GRPO env recipes)
+- Test: `tests/test_v07130.py`
+
+- [ ] Step 1: Write failing `TestRecipes` (3 new recipe names resolve + their YAML parses via `load_config_from_string`) + `TestNoTopLevelTorch` (AST-parse `utils/prm_reward.py`, assert no top-level `import torch/transformers/peft`).
+- [ ] Step 2: Run → fail.
+- [ ] Step 3: Add `grpo-env-calculator` / `grpo-env-retrieval-qa` / `grpo-env-guess-number` RecipeMeta entries (SmolLM2-135M, task=grpo, rollout_backend=openenv, rollout_func=`soup_cli.envs.*:rollout`, reward_fn=math/accuracy). Confirm the no-top-level-torch already holds.
+- [ ] Step 4: Run → PASS.
+- [ ] Step 5: ruff + commit `feat(recipes): 3 GRPO env recipes + AST purity test (v0.71.30)`.
+
+---
+
+### Task 7: Release docs (Release Checklist steps 7-13)
+
+Version bump (`pyproject.toml` + `__init__.py` → 0.71.30); CHANGELOG; CLAUDE.md (arch line for `utils/prm_reward.py` + `envs/`, schema field list, CLI/train notes, recipe count 134→137, test-count refs, history-roll to `.claude/history/release-notes.md` + `tests-index.md`); README What's New; `docs/training.md` + `docs/commands.md`; CONTRIBUTING counts; plan.md flip. Handled after Tasks 1-6 pass the 5-review round + Step-6 smoke.
+
+---
+
+## Self-review
+
+- **Spec coverage:** Part A schema (T2) + kernels (T1) + scorer (T3) + wiring (T4); Part B envs (T5) + recipes (T6); docs (T7). ✅
+- **Type consistency:** `split_steps`/`aggregate_step_scores` (T1) consumed by `PRMScorer` (T3); `build_prm_reward_fn` (T3) consumed by grpo (T4); `rollout` signature uniform (T5). ✅
+- **Placeholders:** none — each task has concrete test targets + implementation sketch (full code written during TDD RED/GREEN).