From 4ca004eac66faec24591b59f594bbbd4c08c5acc Mon Sep 17 00:00:00 2001 From: Joao Gante Date: Thu, 31 Oct 2024 15:42:13 +0000 Subject: [PATCH] Qwen2VL: skip base `input_ids`-`inputs_embeds` equivalence check (#34535) it has complex inputs_embeds computation --- tests/generation/test_utils.py | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/tests/generation/test_utils.py b/tests/generation/test_utils.py index 545b696d67..cbcb0665eb 100644 --- a/tests/generation/test_utils.py +++ b/tests/generation/test_utils.py @@ -1610,7 +1610,7 @@ class GenerationTesterMixin: inputs_dict.pop("pixel_values_images", None) # 2.C - No easy fix, let's skip the check that compares the outputs from `input_ids` and `inputs_embeds` has_complex_embeds_computation = any( - model_name in model_class.__name__.lower() for model_name in ["moshi"] + model_name in model_class.__name__.lower() for model_name in ["moshi", "qwen2vl"] ) # 3 - `inputs_dict` doesn't contain `attention_mask`. When `attention_mask` is not passed to generate, # we infer it from `input_ids`. The last test case will fail if there is a pad token in the original input.