Skip to content

Local UI to run and train LLMs and diffusion models. Supports GGUF, MLX, Qwen3.8, DeepSeek-V4.1, MiniMax-H3, Gemma 4, FLUX and more.

PP env GA test: compare the trained adapter at 5e-5 (#16457)

Daniel Hand4dcba7dd506
diff --git a/tests/test_pp_env_gloo.py b/tests/test_pp_env_gloo.py
index d19149dc3ea..a40136e89e5 100644
--- a/tests/test_pp_env_gloo.py
+++ b/tests/test_pp_env_gloo.py
@@ -884,7 +884,10 @@ def test_pp_env_gradient_accumulation_and_microbatches(ga, bs, mb, schedule, arc
_spawn("env", 2, env, bs = bs, ga = ga, evaluate = False, post = False, arch = arch)
):
_close(res, ref, what = rank)
- _same_tensors(res["adapter"] or ref["adapter"], ref["adapter"])
+ # _close already pins every loss and grad norm. The trained adapter after several steps carries
+ # more of the reduction-order noise: GitHub runners put gemma3_ga1_bs4's lora_A 1.6e-5 from the
+ # reference on two PRs (1 of 64 elements), against at most 3.7e-6 here.
+ _same_tensors(res["adapter"] or ref["adapter"], ref["adapter"], atol = 5e-5)
@needs_pipelining