zhangj1an/audiox_random

AudioX random / test fixture

A tiny random-init bundle of vLLM-Omni's AudioXPipeline. Used by the L1/L2 core_model CI tests (tests/e2e/offline_inference/test_audiox_model.py, tests/e2e/online_serving/test_audiox_online.py) so they can verify the full pipeline (load → forward → trim → return numpy WAV) end-to-end without paying the cost of the real ~11 GB checkpoint.

It follows the same config.json schema as zhangj1an/AudioX, but with much smaller transformer dimensions:

• embed_dim: 1536 → 384
• depth: 24 → 4
• num_heads: 24 → 6
• gate_type_config.num_experts_per_modality: 64 → 16
• gate_type_config.num_fusion_layers: 8 → 2
• sample_size: 485100 → 483328 (still gives latent_len = sample_size // 2048 = 236, matching the transformer's RoPE precompute)

All weights are random, fp16, generated by running the AudioXPipeline.__init__ with the small config and dumping its state_dict() with the bundle's legacy naming convention. Do not use for actual generation — outputs are noise.