Run E2e

Run AFD E2E Tests

Scope

Run one of the model suites:

• tests/e2e/models/deepseek_v2_lite/test_deepseek_v2_lite.py on GPU or NPU
• tests/e2e/models/qwen3_moe/test_qwen3_moe.py on GPU
• tests/e2e/models/qwen3_6/test_qwen3_6.py on CUDA (text-only Qwen3.6 evidence for the Qwen3.5/3.6 adapter family)

Each suite contains four gate scenarios:

• baseline-graph
• afd-eager-2a2f (DeepSeek-V2-Lite gate only; Qwen3 MoE/Qwen3.6 use afd-eager-2a1f)
• afd-graph-2a2f (DeepSeek-V2-Lite gate only; Qwen3 MoE/Qwen3.6 use afd-graph-2a1f)
• afd-graph-dbo-2a2f (DeepSeek-V2-Lite gate only; Qwen3 MoE/Qwen3.6 use afd-graph-dbo-2a1f)

Each gate scenario evaluates the first 7 GSM8K samples. The 2A2F AFD gate scenarios (DeepSeek-V2-Lite only) use 2 Attention ranks and 2 FFN ranks. baseline-graph uses native DP4/TP1/EP4. The 2A1F cases (afd-eager-2a1f, afd-graph-2a1f, afd-graph-dbo-2a1f) use 2 Attention ranks and 1 FFN rank. Note that for DeepSeek-V2-Lite 2A1F are local-only scenarios, while for Qwen3 MoE and Qwen3.6 MoE they are the suite's gate scenarios.

Workflow

1. Select the backend

Honor an explicit backend. Otherwise inspect nvidia-smi -L and npu-smi info. If both are available, ask which to use. If neither is available, stop.

2. Validate prerequisites

Before starting pytest, confirm:

• AFD_E2E_DEVICES contains the device IDs required by test cases.
• The backend model variable is set to a local path, or the environment can download the selected suite's checkpoint via huggingface_hub.
• The selected vllm command runs.
• pytest, afd_plugin, lm_eval, datasets, and huggingface_hub are importable.
• HF_HOME points to the Hugging Face cache used for GSM8K and model weights.
• HF_ENDPOINT is reachable: gsm8k.py defaults the lm-eval child to https://hf-mirror.com, so an unreachable mirror stalls or fails at GSM8K dataset resolution only after the servers are already up and devices reserved. Confirm the default mirror is reachable or override it.
• GPU: the selected devices are visible to CUDA.
• NPU: torch_npu and the Ascend runtime work.

Install missing lm_eval only in the runner environment, never in pyproject.toml or uv.lock.

Fail before pytest when a prerequisite is missing; never turn it into a skip.

Set HF_HOME before every run. The pytest entrypoint downloads/caches GSM8K and the model when the backend model env var is unset.

3. Configure the run

For GPU:

export HF_HOME=/path/to/huggingface
export AFD_E2E_BACKEND=gpu
export AFD_E2E_DEVICES=0,1,2,3
# Optional if the model is already local:
# export AFD_GPU_E2E_MODEL=/path/to/model
# Optional: export AFD_GPU_E2E_VLLM_BIN=/path/to/vllm

For NPU:

export HF_HOME=/path/to/huggingface
export AFD_E2E_BACKEND=npu
export AFD_E2E_DEVICES=0,1,2,3
# Optional if the model is already local:
# export AFD_NPU_E2E_MODEL=/path/to/DeepSeek-V2-Lite
# Optional: export AFD_NPU_E2E_VLLM_BIN=/path/to/vllm

Device order defines roles: the 2A2F AFD scenarios use the first two devices for Attention DP2/TP1 and the last two for FFN DP2/TP1/EP2. baseline-graph uses the first four for native DP4/TP1/EP4. The 2A1F scenarios use the first two for Attention and the third for FFN (used local-only for DeepSeek-V2-Lite; and gate topology for Qwen3 MoE and Qwen3.6 MoE).

4. Run

From the repository root, stream output in the foreground:

python -m pytest -q -s \
  "tests/e2e/models/deepseek_v2_lite/test_deepseek_v2_lite.py::test_deepseek_v2_lite[baseline-graph]" \
  "tests/e2e/models/deepseek_v2_lite/test_deepseek_v2_lite.py::test_deepseek_v2_lite[afd-eager-2a2f]" \
  "tests/e2e/models/deepseek_v2_lite/test_deepseek_v2_lite.py::test_deepseek_v2_lite[afd-graph-2a2f]" \
  "tests/e2e/models/deepseek_v2_lite/test_deepseek_v2_lite.py::test_deepseek_v2_lite[afd-graph-dbo-2a2f]"

For Qwen3 MoE on GPU, run:

python -m pytest -q -s \
  tests/e2e/models/qwen3_moe/test_qwen3_moe.py

For Qwen3.6 MoE on CUDA, run:

export AFD_E2E_BACKEND=gpu
export AFD_E2E_DEVICES=0,1,2,3
export AFD_GPU_E2E_MODEL=/path/to/Qwen3.6-35B-A3B
python -m pytest -q -s \
  tests/e2e/models/qwen3_6/test_qwen3_6.py

The Qwen3.6 suite uses Qwen/Qwen3.6-35B-A3B, vLLM V1, and passes --language-model-only to every runner invocation. It is CUDA-only and text-only: baseline-graph is native DP4/TP1/EP4, while the AFD scenarios are synchronous 2A1F with Attention on the first two devices and FFN on the third. Multimodal, NPU, compute_gate_on_attention=true, asynchronous, pipeline-parallel, and multi-node execution are not covered; quantization is unverified.

Do not add backend markers or run scenarios in parallel; they share devices.

For the local DeepSeek-V2-Lite 2A1F cases, run the same pytest entrypoint with [afd-eager-2a1f], [afd-graph-2a1f], or [afd-graph-dbo-2a1f].

On cancellation, forward SIGTERM and allow over 90 seconds for cleanup.

5. Report

Success means the selected suite reports 4 passed and 0 skipped. Report the failed scenario, first actionable error, and cleanup status. Any skip is a gate failure.

Environment reference