From 74fc389fb382cc79f67b67591569144b3fa393bd Mon Sep 17 00:00:00 2001 From: Onur Gündüz Date: Mon, 22 Jun 2026 17:11:26 +0200 Subject: cuda: prune attention experiments --- tests/check_workspace_api.py | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-) (limited to 'tests/check_workspace_api.py') diff --git a/tests/check_workspace_api.py b/tests/check_workspace_api.py index 40bac11..05f0e8c 100644 --- a/tests/check_workspace_api.py +++ b/tests/check_workspace_api.py @@ -6,11 +6,12 @@ the synthetic fixture. Given a MODEL_DIR it loads real weights from that directory instead and exercises the workspace and batch APIs. Needs only python3 and a model directory (with vocab.json), no reference stack. -The model must be mean-pooled: the test checks that a whole-sequence span pools -to the embedding, which holds for mean pooling (pplx-embed, BERT/BGE) but not -for last-token pooling (Qwen3). Use a pplx-embed model. +For mean-pooled models, the test additionally checks that a whole-sequence span +pools to the embedding and that half-spans recombine by token-count weighting. +For last-token/CLS models, those mean-span identities are skipped because the +sentence embedding and contextual span APIs intentionally pool different rows. - tests/check_workspace_api.py --model-dir DIR # DIR = a pplx-embed model + tests/check_workspace_api.py --model-dir DIR """ import argparse -- cgit v1.3.1