Add disagg_smoke: minimal-Qwen Modal test for pool claim/reset + cookbook consolidation

[jvmncs]

Summary

A small Modal app (cookbook/disagg_smoke/) that validates the pool-control work end to end on a minimal Qwen (Qwen/Qwen2.5-0.5B-Instruct) — no Kimi/large config, no Megatron trainer. It exercises both PR #5 (explicit pool claim/reset, fresh-run-id ownership, monotonic pointer, 1:1 trainer↔pool) and PR #6 (the consolidated cookbook.sidecar/serving/trainer_helpers/rollout_control spine + thin adapters + whole-cookbook image mount). Stacked on PR #6's branch.

Instead of training, it synthesizes real slime disk-delta versions (xor + zstd + xxh3-128) from the base checkpoint and publishes them through the real bulletin board, so the actual host-side delta-apply + reconcile path runs.

delta.py — synthetic delta encoder

Writes byte-for-byte valid weight_v{N:06d}/ dirs that slime's real apply_deltas accepts. DeltaPublisher tracks cumulative new bytes per tensor, so expected_tensor(name) is exactly what the decoder must produce in the local checkpoint:

delta_file[name] = zstd(old ^ new)            # safetensors-format payload
header["__metadata__"][name] = xxh3_128(new)  # what the decoder re-checksums post-apply
index.json.metadata = {version, base_version, xor, zstd, xxh3-128}

Perturbs only the low (mantissa) byte of a few elements, so bf16 weights stay finite and the perturbed model still generates.

app.py::control_plane_test (GPU-free, primary)

Drives the real WeightSyncManager/build_manager against a fake SGLang upstream (an in-process http.server answering flush_cache/pause/continue/update_weights_from_disk). Asserts everything itself; raises on failure:

• claim(run_a) → empty pointer <run_a>/weight_v000000, reset=True; startup converges to base.
• publish v1+v2 → reconcile → local checkpoint bytes == pub.expected_tensor(...) (real slime decoder), engine reloaded at weight_version="2".
• advance(run_a, 1) and re-claim(run_a) both raise PointerRewind (fresh-run-id enforcement).
• fresh run_b re-claims → engine resets to base → replays its own chain (1:1 epoch).
• imports every consolidated module + thin adapter off the mount — with include_source=False, the PR Consolidate cookbook trainers onto a shared sidecar/serving/launch spine #6 whole-cookbook mount regression would ImportError here.

app.py::serving_smoke (1× GPU)

Runs the real consolidated sidecar (python3 -m cookbook.slime_disagg.sidecar) in front of a real SGLang server on tiny Qwen, claims the run, publishes one delta, triggers a reconcile, and asserts the engine reloaded to v1 (real update_weights_from_disk) and serves a completion pinned to min_required_version: 1. Uses a vanilla single-GPU SGLang image (not the Blackwell fa4 fork build, which targets B200s) so it runs on a common 1×GPU; the whole-cookbook mount + --no-deps decoder layers are identical.

Run

modal run cookbook/disagg_smoke/app.py::control_plane_test   # GPU-free
modal run cookbook/disagg_smoke/app.py::serving_smoke        # 1× GPU

Pin SLIME_SMOKE_REPO/SLIME_SMOKE_REF to the same slime ref the trainer encodes with (encoder == decoder).

Validation

Static only (per request — I don't have Modal creds; you run modal run):

• py_compile clean on all four files.
• Full local suite: 116 passed, 1 skipped. delta_test.py (encoder structural + xor/checksum self-consistency) passes (3 passed) once numpy/zstandard/xxhash are present; it importorskips otherwise so the bare-env suite stays green.

The byte-for-byte round-trip against slime's real decoder is asserted by control_plane_test on Modal.

Link to Devin session: https://modal.devinenterprise.com/sessions/b0d6110c71924c24943bcb5cea1eb5f4
Requested by: @jvmncs

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR that start with 'DevinAI' or '@devin'.
• Look at CI failures and help fix them

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring