[CI] Revert flaky test_quark_int8_w8a8_moe (#54991)

Signed-off-by: Felix Marty <Felix.Marty@amd.com>
This commit is contained in:
fxmarty-amd
2026-09-02 13:58:14 -04:00
committed by GitHub
parent 605c3ddcba
commit 488e6fd53c
+21 -21
View File
@@ -1020,32 +1020,32 @@ def test_quark_int8_w_per_tensor_a_per_tensor(monkeypatch, dist_init, workspace_
assert torch.isfinite(logits).all()
def test_quark_int8_w8a8_moe(monkeypatch, dist_init, workspace_init):
@pytest.mark.parametrize("tp", [1])
def test_quark_int8_w8a8_moe(vllm_runner, tp):
"""Test W8A8 INT8 MoE quantization with a tiny Qwen3 MoE model."""
model_path = "amd/tiny-qwen3-moe-w8a8-int8"
model, vllm_config = load_model_without_vllm_runner(
with vllm_runner(
model_path,
model_config_kwargs={"hf_overrides": {"num_hidden_layers": 3}},
)
enforce_eager=True,
tensor_parallel_size=tp,
gpu_memory_utilization=0.1,
) as llm:
layer = model.model.layers[0]
moe = layer.mlp.experts
assert isinstance(moe._quant_method, QuarkW8A8Int8MoEMethod), (
f"Expected QuarkW8A8Int8MoEMethod, got {type(moe._quant_method)}"
)
qkv_proj = layer.self_attn.qkv_proj
assert isinstance(qkv_proj.scheme, QuarkW8A8Int8)
def check_model(model):
layer = model.model.layers[0]
# MoE experts should use QuarkW8A8Int8MoEMethod
moe = layer.mlp.experts
assert isinstance(moe._quant_method, QuarkW8A8Int8MoEMethod), (
f"Expected QuarkW8A8Int8MoEMethod, got {type(moe._quant_method)}"
)
# Non-MoE linear layers should use QuarkW8A8Int8
qkv_proj = layer.self_attn.qkv_proj
assert isinstance(qkv_proj.scheme, QuarkW8A8Int8)
monkeypatch.setattr(Attention, "forward", lambda _, q, k, v: q.contiguous())
input_ids = torch.tensor([1, 2, 3, 4], device=DEVICE_TYPE)
positions = torch.arange(input_ids.numel(), device=DEVICE_TYPE)
with (
set_current_vllm_config(vllm_config),
set_forward_context(None, vllm_config, num_tokens=input_ids.numel()),
):
hidden_states = model(input_ids, positions, None)
logits = model.compute_logits(hidden_states)
assert torch.isfinite(logits).all()
llm.apply_model(check_model)
output = llm.generate_greedy("Hello", max_tokens=4)
assert output
@pytest.mark.skipif(