[amarquic]

Title

Add weight-free MoE expert-parallel and decode-BMM coverage

Summary

This PR adds weight-free QAIC tests for MoE prefill and decode configurations, makes QAIC test core selection device-aware, and configures Jenkins to store prepared weight-free checkpoints in the test-owned cache.

Changes

• Cover MoE prefill with expert parallelism and decode with decode_bmm.

• Add expert-parallel decode coverage for selected VLM and continuous-batching tests.

• Check the configured MoE flavor and expert-parallel chunk size through model hash parameters.

• Compare QAIC-generated tokens with Hugging Face output for the additional expert-parallel decode paths.

• Replace hardcoded QAIC core counts with DEFAULT_AIC_NUM_CORES.

• Add the use_onnx_subfunctions guard to the Qwen3-MoE parity test.

• Keep generated weight-free checkpoints under QEFF_HOME for test cleanup.

• Set QEFF_WF_HOME in Jenkins to:

  /efficient-transformers/weight_free_qaic/weight_free_checkpoints
  

• Preserve the Gemma4 vision overflow fix by folding standardization scales before multiplying by the root hidden size.

Test coverage and configurations

Test (test_file_path::test_name)	Model / parameter	Configuration covered
tests/weight_free/test_disaggregated.py::test_weight_free_disaggregated_prefill_and_decode[qwen3-moe]	tiny-random/qwen3-moe	Disaggregated expert-parallel prefill; decode_bmm decode; mxint8_kv_cache=False
tests/weight_free/test_disaggregated.py::test_weight_free_disaggregated_prefill_and_decode[gpt-oss]	tiny-random/gpt-oss	Disaggregated expert-parallel prefill; decode_bmm decode; mxint8_kv_cache=False
tests/weight_free/test_disaggregated.py::test_weight_free_vlm_disaggregated_prefill_and_decode[qwen3-vl-moe]	tiny-random/qwen3-vl-moe	Vision and expert-parallel prefill; separate decode_bmm and expert-parallel decode paths; HF token parity; mxint8_kv_cache=False
tests/weight_free/test_disaggregated.py::test_weight_free_vlm_disaggregated_prefill_and_decode[gemma-4-moe]	tiny-random/gemma-4-moe	Vision and expert-parallel prefill; separate decode_bmm and expert-parallel decode paths; HF token parity; mxint8_kv_cache=False
tests/weight_free/test_disaggregated.py::test_weight_free_qwen3_moe_disagg_cb_kv_handoff_and_hf_parity[local_qwen3_moe_prefill2_decode1_stages2]	Local tiny Qwen3-MoE FP32 checkpoint	Continuous batching, disaggregated KV handoff, expert-parallel decode, and HF token parity; mxint8_kv_cache=False
tests/weight_free/test_disaggregated.py::test_weight_free_gpt_oss_disagg_cb_kv_handoff_and_hf_parity[tiny_model_prefill2_decode1_stages2]	tiny-random/gpt-oss	Continuous batching, disaggregated KV handoff, expert-parallel decode, and HF token parity; mxint8_kv_cache=False
tests/weight_free/test_on_qaic_vlm.py::test_weight_free_vlm_multi_specialization_hf_qaic_parity[qwen3.5-moe]	tiny-random/qwen3.5-moe	Two vision resolutions; expert-parallel prefill; separate decode_bmm and expert-parallel decode paths; HF token parity. MXINT8 KV cache is omitted and remains disabled by default.

[Amit Raj (quic-amitraj)]

amarquic lets fix the DCO issue and correct the PR title. Other changes LGTM.

[Ann Kuruvilla (quic-akuruvil)]

Test coverage table is descriptive and good. It would be helpful to include the test file path also in the table for easy lookup. test_file_path::test_name

[amarquic]

Test coverage table is descriptive and good. It would be helpful to include the test file path also in the table for easy lookup. test_file_path::test_name

Test coverage table is descriptive and good. It would be helpful to include the test file path also in the table for easy lookup. test_file_path::test_name

Updated