[Fix] Load Laya without the throwaway random weight init

[cacheline999]

Why

--model laya spends about 35 s loading a 0.4B model, and every CLI call pays it again because the model loads in process. Most of that time is spent initializing weights that get overwritten right away.

How

laya.load builds the ModernBERT encoder with AutoModel.from_config, so transformers draws every weight at random, and then laya loads the checkpoint over all of them with strict=True. Profiling the load on CPU puts ~33 s of it in _init_weights / trunc_normal_.

LayaModel.from_env now runs laya.load under transformers' no_init_weights(). The helper moved from transformers.modeling_utils (4.x) to transformers.initialization (5.x); without_weight_init() in s1a/decision_models/laya.py tries both and falls back to a plain load if neither is there. Because laya loads with strict=True, a checkpoint that misses a weight still fails loudly instead of leaving it uninitialized.

The from_env tests stub the helper. Without the stub the real one imports torch inside patch.dict(sys.modules), which drops torch again on exit, and the next torch import in the same process segfaults. This only happens with the laya extra installed, not in CI.

What

--model laya (agents, decide, probe, MCP decide) loads in about 3 s. No flags, outputs or records change.

Verification

Default checkpoint, laya 0.3.5, M1 Pro 16 GB. Each load in a fresh process:

transformers	device	load before	load after	state dict + buffers	answers
5.17.0 (lock)	cpu	35.3 s	2.5 s	identical	identical
5.17.0 (lock)	mps	35.0 s	3.7 s	identical	identical
4.57.1	cpu	31.4 s	1.0 s	identical	identical
4.48.0 (laya's minimum)	cpu	33.2 s	1.5 s	identical	identical

s1a decide --model laya end to end: 37.6 s → 9.9 s on CPU, 39.7 s → 8.5 s on MPS, same output.

Warm latency is unchanged. On CPU the first request after load is ~0.2–0.3 s slower, since the random init no longer warms up torch's CPU path during load.

• uv run ruff format --check . && uv run ruff check . && uv run ty check
• uv run pytest -q and scripts/smoke.sh: all pass on the core install (no torch, as in CI). With the laya extra installed, test_an_agent_without_system_one_is_a_config_error_naming_the_method fails on main too, because it expects laya's version to be unknown.
• CHANGELOG.md

[Yunaik]

great find! thanks!

[Yunaik]

Follow-ups:

• [Deps] Require laya 0.3.9 and drop the local weight-init skip #28: laya 0.3.9 ships the same no_init_weights skip inside laya.load. Bump the extra and remove without_weight_init() and its test stubs.
• [Docs] Time a cold model load when adding an in-process backend #29: add a cold-load timing step to "Adding a backend" in docs/decision-models.md, starting with Julia-1 ([Feat] Integrate and evaluate Julia-1 #25).

[Yunaik]

this issue seems to be fixed with laya 0.3.9 btw