test(npu): add GSM8K accuracy test for CAMAsync token split on DP3TP2+EP2

[ShwStone]

Add GSM8K accuracy E2E coverage for CAMAsyncAFDConnector on Ascend NPU.

This PR introduces tests/e2e/accuracy/test_gsm8k_npu_async_cam.py, which mirrors the existing synchronous-connector GSM8K accuracy test (test_gsm8k_npu.py) but targets the asynchronous CAM connector.

What changed

• New test: tests/e2e/accuracy/test_gsm8k_npu_async_cam.py

  • test_gsm8k_lm_eval_async_cam_1a1f: a minimal 1-Attention + 1-FFN (2 NPU) GSM8K accuracy baseline. This runs by default when the test file is executed.
  • test_gsm8k_lm_eval_async_cam_dp3tp2_ep2: an 8-NPU non-PCP DP+TP/SP topology test parameterized over async_moe_split="request" and "token". This test is opt-in via AFD_NPU_ASYNC_CAM_RUN_DP3TP2=1 so the heavy 8-card run is not triggered by default.

• tests/e2e/conftest.py extensions

  • _launch_afd_server now accepts optional parameters for:
    • connector: override the default connector (e.g. CAMAsyncAFDConnector).
    • attention_tp_size / ffn_tp_size: per-role tensor parallelism.
    • afd_async: set additional_config['afd']['async']=true.
    • compute_gate_on_attention: required by CAMAsync.
    • afd_connector_extra_config: pass connector_extra_config such as async_moe_ubatching, async_moe_num_ubatches, and async_moe_split.

Topology under test

1A1F baseline (default)

• Attention: 1 rank, TP=1
• FFN: 1 rank, EP=1
• async_moe_split="request"

DP3TP2 + EP2 token-split validation (opt-in)

• Attention: DP=3, TP=2 → 6 ranks
• FFN: DP=2, EP=2 → 2 ranks
• async_moe_split ∈ {"request", "token"}

How to run

export AFD_NPU_E2E_MODEL=/path/to/DeepSeek-V2-Lite
export AFD_NPU_E2E_VLLM_BIN=vllm

# Default: only the 2-NPU 1A1F baseline
pytest tests/e2e/accuracy/test_gsm8k_npu_async_cam.py -v

# Opt-in: also run the 8-NPU DP3TP2+EP2 request/token comparison
export AFD_NPU_ASYNC_CAM_RUN_DP3TP2=1
pytest tests/e2e/accuracy/test_gsm8k_npu_async_cam.py -v

A staged offline GSM8K task dir is required, same as test_gsm8k_npu.py. Set AFD_NPU_GSM8K_TASK_DIR if it is not at the default sibling location.

Status

This is a draft PR. NPU E2E validation is still pending. I will mark it ready for review once the 1A1F baseline and the DP3TP2+EP2 request/token split comparison have been verified on real Ascend hardware.

[jiangkuaixue123]

Great. I will merge until test e2e pass.