[bong-water-water-bong]

third_party/laya: ca5045a3f1fa -> d74e8be46ef5 (NandhaKishorM/laya main). config/laya.json revision: 7b928d828b7b -> 7b928d828b7b (convaiinnovations/laya).

Before merging: scripts/fetch-laya.sh

(verifies every file) and the Laya tests.

[github-actions]

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

⏱️ Estimated effort to review: 1 🔵⚪⚪⚪⚪

🧪 No relevant tests

🔒 No security concerns identified

⚡ No major issues detected

[bong-water-water-bong]

Laya pin check on ryzen (CPU only): ca5045a (0.4.9, current pin) vs d74e8be (0.4.11, this PR).

Method: both commits checked out as detached worktrees of NandhaKishorM/laya, every tests/test_*.py run from inside its own tree with one shared venv (Python 3.14.4 + torch 2.14.1, transformers 5.20.1, mcp, onnx/onnxruntime/onnxscript, fastapi/uvicorn). pytest files (those that import pytest) ran under pytest, the assert-script files ran directly. scripts/fetch-laya.sh /tmp/laya-pinned against the pin in this PR's config/laya.json: 21 files verified (sha256/git-blob), plus the 1 pinned gguf.

Result: 101 test files exist at both commits, 104 at the new one (added test_agno.py, test_onnx_export_batch.py, test_tools.py). Same 12 files fail at both pins, no file that passes at ca5045a fails at d74e8be, and no file that fails at the old pin passes at the new one:

test file	old ca5045a	new d74e8be
test_audit_regressions	fail	fail
test_context_manager	fail	fail
test_criteria_normalization	fail	fail
test_fast	fail (no CUDA/tilelang)	fail (no CUDA/tilelang)
test_fast_cpu	fail	fail
test_imports	fail	fail
test_langchain_real_mode	fail	fail
test_local_e2e	fail	fail
test_mcp_local_e2e	fail	fail
test_onnx_quantize	fail	fail
test_shortlist_cosine	fail	fail
test_truncation_direction	fail	fail
test_onnx_export_batch	-	fail (host gap: onnx absent in the first run)
all other 89 files	pass	pass

Every failure is an environment gap or a test that needs the full three-checkpoint model root, not a code difference:

• test_local_e2e needs <root>/laya, laya-multilingual, laya-typed-decisions as three separate dirs; our ~/models/laya-st keeps only typed-decisions. Pointed at the fetched pin (three symlinks to one tree) both commits give the identical 37-line PASS/FAIL list (35 passed, 2 failed): the two failures are noul labels/multilingual/positive/plain -- P(true)=0.0000 and hindi answer is billing -- got technical, identical at both pins.
• test_mcp_local_e2e passes at both pins when run with PYTHONPATH set to the tree (133 passed) - it downloads its checkpoints itself.
• test_serve is a pytest file and passes 301/301 at both pins; run as a plain script it has no __main__, so a bare python tests/test_serve.py exits 0 without testing anything.
• test_audit_regressions, test_context_manager, test_criteria_normalization, test_imports, test_shortlist_cosine, test_truncation_direction, test_onnx_quantize exit 1 only because they import pytest and are driven through python -m pytest; with PYTHONPATH set and pytest driving them they all pass at both pins.
• test_langchain_real_mode needs laya[langchain]; test_fast/test_fast_cpu need CUDA + tilelang; test_onnx_export_batch is the same missing-onnx gap (99/99 pass once onnx/onnxruntime/onnxscript are installed) and is new at this pin, so it has no old-pin counterpart.

Nothing new fails. The engine-side dependency is clean too: the PR touches only third_party/laya; config/laya.json is byte-identical to main (same 7b928d828b7b model revision), so only the source submodule moves and no registry regen is needed. The fetched model tree (/tmp/laya-pinned) is the same shape as the existing ~/models/laya-st.

PR-READY per the pin-bump rule: no new test failure against the old pin.