src/tests/test_multimodal.py

"""
Frox AI Morph 1.1 — Multimodal Fusion Tests
Run with: pytest tests/test_multimodal.py -v
"""
from __future__ import annotations

import pytest
import torch

from multimodal.fusion.morph_multimodal import MorphMultimodalModel


class TestImageTokenInjection:
    def test_single_image_expands_sequence_correctly(self, tiny_config, device):
        model = MorphMultimodalModel(tiny_config).to(device).eval()

        num_image_tokens = (tiny_config.vision.image_size // tiny_config.vision.patch_size) ** 2
        text_len = 6

        input_ids = torch.randint(
            0, tiny_config.text.vocab_size, (1, text_len), device=device
        )
        input_ids[0, 3] = model.image_token_id

        pixel_values = torch.randn(
            1, 3, tiny_config.vision.image_size, tiny_config.vision.image_size, device=device
        )

        with torch.no_grad():
            inputs_embeds, mask = model.prepare_multimodal_inputs(input_ids, pixel_values)

        expected_len = (text_len - 1) + num_image_tokens
        assert inputs_embeds.shape[1] == expected_len
        assert inputs_embeds.shape[2] == tiny_config.text.hidden_size

    def test_no_image_falls_back_to_text_only(self, tiny_config, device):
        model = MorphMultimodalModel(tiny_config).to(device).eval()
        input_ids = torch.randint(0, tiny_config.text.total_vocab_size, (1, 8), device=device)

        with torch.no_grad():
            embeds, mask = model.prepare_multimodal_inputs(input_ids, pixel_values=None)

        assert embeds.shape[1] == 8
        assert mask is None

    def test_forward_pass_with_image_produces_finite_logits(self, tiny_config, device):
        model = MorphMultimodalModel(tiny_config).to(device).eval()

        input_ids = torch.randint(0, tiny_config.text.vocab_size, (1, 6), device=device)
        input_ids[0, 2] = model.image_token_id
        pixel_values = torch.randn(
            1, 3, tiny_config.vision.image_size, tiny_config.vision.image_size, device=device
        )

        with torch.no_grad():
            out = model(input_ids=input_ids, pixel_values=pixel_values)

        assert not torch.isnan(out.logits).any()
        assert not torch.isinf(out.logits).any()


class TestParamCount:
    def test_param_count_breakdown_sums_correctly(self, tiny_config, device):
        model = MorphMultimodalModel(tiny_config).to(device)
        counts = model.param_count()
        # lm + vision should roughly account for the total (small rounding tolerance
        # from the shared image_sep_embed parameter, which isn't double counted)
        assert counts["total_billions"] >= counts["lm_billions"] + counts["vision_billions"] - 1e-6