mstar.model.command_a_plus.tests.test_vision

mstar.model.command_a_plus.tests.test_vision#

The image encoder’s weight loading, and its output against Hugging Face’s.

MSTAR_TEST_HF=1 enables the comparison with transformers’ SigLIP tower and Cohere2Vision projector.

Functions

checkpoint_name(name)

tiny_encoder()

Classes

HuggingFaceVisionParityTests([methodName])

VisionEncoderTests([methodName])

class mstar.model.command_a_plus.tests.test_vision.HuggingFaceVisionParityTests(methodName='runTest')[source]#

Bases: TestCase

test_matches_siglip_and_the_cohere2_vision_projector()[source]#
class mstar.model.command_a_plus.tests.test_vision.VisionEncoderTests(methodName='runTest')[source]#

Bases: TestCase

test_loads_checkpoint_names_strictly()[source]#
mstar.model.command_a_plus.tests.test_vision.checkpoint_name(name)[source]#
Parameters:

name (str)

Return type:

str

mstar.model.command_a_plus.tests.test_vision.tiny_encoder()[source]#
Return type:

CommandAPlusVisionEncoder