iHore/vLLM
8
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Disable mypy arg-type and assignment checks in tests (#59428)

Browse Source 
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
This commit is contained in:
Harry Mellor
2026-09-30 15:18:50 +00:00
committed by GitHub
co-authored by Claude Opus 5.5
parent 17e9295dd5
commit fc2c801ad9
131 changed files with 402 additions and 409 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tests/quantization/test_auto_gptq.py
+2 -2
View File
@@ -123,8 +123,8 @@ def test_auto_gptq_moe_creates_zero_initialized_expert_biases():
method = object.__new__(AutoGPTQMoEMethod)
method.quant_config = AutoGPTQConfig(4, 128, False, True, False, {}, {})
method.input_dtype = None
method.experts_cls = None # type: ignore[assignment] # Weight creation does not select experts.
method.moe = SimpleNamespace(w13_num_shards=2) # type: ignore[assignment]
method.experts_cls = None
method.moe = SimpleNamespace(w13_num_shards=2)
layer = torch.nn.Module()
method.create_weights(
tests/quantization/test_auto_round.py
+6 -6
View File
@@ -423,7 +423,7 @@ def test_inc_get_quant_method_unquantized_routed_experts_with_model_prefix(
config = make_config(extra_config={"model.moe_layer": {"bits": 16}})
layer = object.__new__(RoutedExperts)
layer.moe_config = None # type: ignore[assignment] # No MoE config used by this stub.
layer.moe_config = None
method = config.get_quant_method(layer, "moe_layer")
@@ -730,7 +730,7 @@ def test_wna16_cuda_low_bit_moe_routes_to_humming(monkeypatch, bits) -> None:
)
layer = object.__new__(RoutedExperts)
layer.moe_config = DummyMoeConfig() # type: ignore[assignment]
layer.moe_config = DummyMoeConfig()
layer_config = make_layer_config(bits=bits)
method = INCWna16Scheme().get_moe_method(
make_config(), layer, "model.layers.0.mlp", layer_config
@@ -861,7 +861,7 @@ def test_qwen3_30b_a3b_w4a16_autoround_routes_to_gptq_moe(
config = make_qwen3_autoround_config("qwen3_30b_a3b_w4a16")
layer = object.__new__(RoutedExperts)
layer.moe_config = DummyMoeConfig() # type: ignore[assignment]
layer.moe_config = DummyMoeConfig()
method = config.get_quant_method(layer, "model.layers.0.mlp")
@@ -890,7 +890,7 @@ def test_qwen3_30b_a3b_mxfp4_autoround_routes_to_mxfp4_moe(
config = make_qwen3_autoround_config("qwen3_30b_a3b_mxfp4")
layer = object.__new__(RoutedExperts)
layer.moe_config = DummyMoeConfig() # type: ignore[assignment]
layer.moe_config = DummyMoeConfig()
ignored_method = config.get_quant_method(
object.__new__(LinearBase), "model.layers.0.self_attn.q_proj"
@@ -1186,7 +1186,7 @@ def test_wna16_xpu_moe_routes_to_gptq_moe(monkeypatch) -> None:
)
layer = object.__new__(RoutedExperts)
layer.moe_config = DummyMoeConfig() # type: ignore[assignment]
layer.moe_config = DummyMoeConfig()
method = INCWna16Scheme().get_moe_method(
make_config(),
layer,
@@ -1669,7 +1669,7 @@ def test_inc_get_quant_method_unquantized_moe_returns_unquantized(
when extra_config has bits >= 16."""
config = make_config(extra_config={"layer": {"bits": 16}})
layer = object.__new__(RoutedExperts)
layer.moe_config = None # type: ignore[assignment] # No MoE config used by this stub.
layer.moe_config = None
class DummyUnquantizedFusedMoEMethod:
def __init__(self, moe_config) -> None:
tests/quantization/test_fp8.py
+2 -2
View File
@@ -220,7 +220,7 @@ def test_deepseek_v41_vl_mapper_routes_linear_scales(
quant_config=SimpleNamespace(weight_block_size=weight_block_size)
)
resolved = model_module._linear_scale_param_name(
vllm_config, # type: ignore[arg-type]
vllm_config,
expert_dtype,
)
assert resolved == scale_name
@@ -798,7 +798,7 @@ def test_kv_cache_scale_sync_to_host_copies():
set_default_quant_scales(layer, register_buffer=True)
layer.kv_cache_dtype = "fp8"
method = BaseKVCacheMethod(quant_config=None) # type: ignore[arg-type] # Scale synchronization does not use quant_config.
method = BaseKVCacheMethod(quant_config=None)
method.create_weights(layer)
# 0.3 stays != 1.0 even after the fp8_fnuz x2 rescale.
checkpoint_scale = torch.tensor(0.3, dtype=torch.float32)
tests/quantization/test_moe_wna16.py
+4 -4
View File
@@ -218,9 +218,9 @@ def test_gptq_nfirst_weights_pass_through_for_triton():
def test_moe_wna16_setup_forwards_selected_backend(monkeypatch):
method = object.__new__(MoeWNA16Method)
method.experts_cls = object # type: ignore[assignment] # Opaque mocked-expert sentinel.
method.experts_cls = object
method.wna16_backend = WNA16MoEBackend.HUMMING
method.moe = object() # type: ignore[assignment] # Opaque mocked-config sentinel.
method.moe = object()
quant_config = object()
monkeypatch.setattr(
method, "get_fused_moe_quant_config", lambda layer: quant_config
@@ -235,7 +235,7 @@ def test_moe_wna16_setup_forwards_selected_backend(monkeypatch):
monkeypatch.setattr(moe_wna16, "make_wna16_moe_kernel", fake_make_wna16_moe_kernel)
method._setup_kernel(layer) # type: ignore[arg-type] # Routing-table stub.
method._setup_kernel(layer)
assert method.moe_kernel is kernel
assert captured["backend"] == WNA16MoEBackend.HUMMING
@@ -280,7 +280,7 @@ def test_moe_wna16_uses_humming_quant_config(monkeypatch):
),
)
assert method.get_fused_moe_quant_config(layer) is quant_config # type: ignore[arg-type] # Identity sentinel.
assert method.get_fused_moe_quant_config(layer) is quant_config
@pytest.mark.skipif(
tests/quantization/test_online.py
+4 -4
View File
@@ -210,9 +210,9 @@ def test_online_nvfp4_reuses_kernel_when_weights_are_reprocessed(
monkeypatch,
) -> None:
method = object.__new__(Nvfp4OnlineMoEMethod)
method.moe = SimpleNamespace(is_act_and_mul=True) # type: ignore[assignment]
method.nvfp4_backend = object() # type: ignore[assignment] # Opaque mocked-backend sentinel.
method.experts_cls = object # type: ignore[assignment] # Opaque mocked-expert sentinel.
method.moe = SimpleNamespace(is_act_and_mul=True)
method.nvfp4_backend = object()
method.experts_cls = object
method.moe_quant_config = None
method.moe_kernel = None
@@ -1432,7 +1432,7 @@ def test_online_int8_moe_w2_scale_matches_unsharded(monkeypatch) -> None:
layer.w2_weight = torch.nn.Parameter(w2_in, requires_grad=False)
layer.num_experts = layer.local_num_experts = w13.shape[0]
method = SimpleNamespace(moe=SimpleNamespace(tp_size=moe_tp_size))
Int8OnlineMoEMethod._quantize_weights(method, layer) # type: ignore[arg-type]
Int8OnlineMoEMethod._quantize_weights(method, layer)
return layer.w2_weight, layer.w2_scale
full_weight, full_scale = quantize(w2, 1)
tests/quantization/test_quark.py
+2 -2
View File
@@ -2314,7 +2314,7 @@ class TestQuarkInt4Format:
None,
quant_config.quant_config["global_quant_config"]["weight"],
quant_config.pack_method,
moe_config, # type: ignore[arg-type]
moe_config,
)
layer = _FakeLayer(moe_config)
@@ -2326,7 +2326,7 @@ class TestQuarkInt4Format:
requires_grad=False,
)
loader = method.get_weight_loader(
layer, # type: ignore[arg-type]
layer,
weight_loader=None,
)
loader(
tests/quantization/test_trtllm_nvfp4_hidden_dim_padding.py
+1 -1
View File
@@ -34,7 +34,7 @@ def test_shared_nvfp4_input_scales_have_writable_storage(monkeypatch):
outputs = prepare_nvfp4_moe_layer_for_fi_or_cutlass(
backend=NvFp4MoeBackend.FLASHINFER_CUTLASS,
layer=layer, # type: ignore[arg-type]
layer=layer,
w13=w13,
w13_scale=w13_scale,
w13_scale_2=weight_scale,
tests/quantization/test_turboquant.py
+1 -1
View File
@@ -300,7 +300,7 @@ class TestTurboQuantKVCacheSpec:
# The layer builds an unpacked spec; the worker's spec-collection
# loop applies TQ slot packing via the backend's customize_spec hook.
spec = Attention.get_kv_cache_spec(layer, vllm_config) # type: ignore[arg-type]
spec = Attention.get_kv_cache_spec(layer, vllm_config)
assert isinstance(spec, FullAttentionSpec)
assert spec.kv_quant_mode.is_turboquant
assert spec.state_content_bytes is None