SUMIN/HuggingFace_transformer
1
0
Fork 0
Code Issues Pull Requests Actions 7 Packages Projects Releases Wiki Activity

Qwen2VL: skip base input_ids-inputs_embeds equivalence check (#34535)

Browse Source 
it has complex inputs_embeds computation
This commit is contained in:
Joao Gante
2024-10-31 15:42:13 +00:00
committed by GitHub
parent ab98f0b0a1
commit 4ca004eac6
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
tests/generation/test_utils.py
View File

@@ -1610,7 +1610,7 @@ class GenerationTesterMixin:
inputs_dict.pop("pixel_values_images", None) 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` # 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( 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, # 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. # we infer it from `input_ids`. The last test case will fail if there is a pad token in the original input.