Disable mypy arg-type and assignment checks in tests (#59428)
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:
co-authored by
Claude Opus 5.5
parent
17e9295dd5
commit
fc2c801ad9
@@ -27,7 +27,7 @@ def make_manager(
|
||||
manager.num_speculative_steps = num_steps
|
||||
manager._stale_confidences = [SimpleNamespace(np=confidences)] # type: ignore[list-item] # CPU-only confidence buffer.
|
||||
manager._stale_idx = 0
|
||||
manager.req_states = SimpleNamespace( # type: ignore[assignment] # CPU-only request fixture.
|
||||
manager.req_states = SimpleNamespace(
|
||||
req_id_to_index={"low": 0, "high": 1},
|
||||
num_computed_tokens_np=np.ones(num_reqs, dtype=np.int32),
|
||||
prefill_len=SimpleNamespace(np=np.ones(num_reqs, dtype=np.int32)),
|
||||
@@ -153,11 +153,11 @@ def test_manager_checks_target_varlen_cudagraph_bound(
|
||||
manager = maybe_create_adaptive_verification_manager(
|
||||
enable_adaptive_verification=True,
|
||||
attn_groups=groups,
|
||||
req_states=SimpleNamespace(num_speculative_steps=7), # type: ignore[arg-type]
|
||||
req_states=SimpleNamespace(num_speculative_steps=7),
|
||||
query_start_loc=object(),
|
||||
num_bonus_tokens=1,
|
||||
max_total_logits=1,
|
||||
vllm_config=None, # type: ignore[arg-type]
|
||||
vllm_config=None,
|
||||
target_layer_names={"target"},
|
||||
additional_attn_cg_support=additional,
|
||||
)
|
||||
@@ -186,9 +186,7 @@ def test_budget_stops_where_marginal_drafts_stop_paying_for_themselves():
|
||||
|
||||
def test_profiled_batches_seed_cost_curves_via_consumer(monkeypatch):
|
||||
manager = AdaptiveVerificationManager.__new__(AdaptiveVerificationManager)
|
||||
manager.req_states = SimpleNamespace( # type: ignore[assignment] # CPU-only request fixture.
|
||||
max_num_batched_tokens=4096, max_num_reqs=64
|
||||
)
|
||||
manager.req_states = SimpleNamespace(max_num_batched_tokens=4096, max_num_reqs=64)
|
||||
manager.num_speculative_steps = 7
|
||||
manager.num_bonus_tokens = 1
|
||||
curves: dict[str, list[tuple[int, float]]] = {}
|
||||
|
||||
@@ -98,7 +98,7 @@ def test_draft_quant_exclusions_include_global_layer_indices():
|
||||
]
|
||||
)
|
||||
|
||||
_add_global_draft_layer_exclusions(quant_config, 88, 5) # type: ignore[arg-type]
|
||||
_add_global_draft_layer_exclusions(quant_config, 88, 5)
|
||||
|
||||
assert "layers.88.mlp_conv*" in quant_config.exclude_modules
|
||||
assert "*layers.92.self_attn.q_proj" in quant_config.exclude_modules
|
||||
@@ -171,7 +171,7 @@ def test_selector_leaves_greedy_drafting_without_proposal_logits(monkeypatch):
|
||||
allocating one here would claim a proposal the walk never sampled from.
|
||||
"""
|
||||
_stub_base(monkeypatch, None)
|
||||
speculator = DFlash2Speculator(None, torch.device("cpu")) # type: ignore[arg-type] # Base initialization is stubbed.
|
||||
speculator = DFlash2Speculator(None, torch.device("cpu"))
|
||||
|
||||
assert speculator.draft_logits is None
|
||||
|
||||
@@ -183,7 +183,7 @@ def test_selector_asks_for_fp32_proposal_logits():
|
||||
candidate row often enough that the walk and the rejection sampler checking it
|
||||
would no longer read the same distribution.
|
||||
"""
|
||||
dtype, fill = DFlash2Speculator.draft_logits_spec(None, None) # type: ignore[arg-type] # Constant spec ignores self/config.
|
||||
dtype, fill = DFlash2Speculator.draft_logits_spec(None, None)
|
||||
|
||||
assert dtype is torch.float32
|
||||
assert fill == float("-inf")
|
||||
@@ -309,9 +309,9 @@ def test_candidate_model_decoder_layer_cls(monkeypatch, variant):
|
||||
vllm_config.compilation_config = mock_current_vllm_config.compilation_config
|
||||
|
||||
# 3. Instantiate the model under meta device to avoid parameter allocation issues
|
||||
with set_current_vllm_config(mock_current_vllm_config), torch.device("meta"): # type: ignore[arg-type]
|
||||
with set_current_vllm_config(mock_current_vllm_config), torch.device("meta"):
|
||||
model_cls = DFlash2Qwen3Model if variant == "dflash2" else LiLiCorr
|
||||
model = model_cls(vllm_config=vllm_config) # type: ignore[arg-type]
|
||||
model = model_cls(vllm_config=vllm_config)
|
||||
|
||||
# 4. Assert that the layers are DFlash2Qwen3DecoderLayer (the subclass)
|
||||
assert len(model.layers) == 2
|
||||
@@ -350,7 +350,7 @@ def test_conv_projections_use_draft_quant_config(monkeypatch):
|
||||
quant_method="W4A16_NVFP4", is_checkpoint_nvfp4_serialized=True
|
||||
)
|
||||
layer = DFlash2Qwen3DecoderLayer(
|
||||
SimpleNamespace( # type: ignore[arg-type]
|
||||
SimpleNamespace(
|
||||
speculative_config=SimpleNamespace(num_speculative_tokens=7),
|
||||
model_config=SimpleNamespace(dtype=torch.bfloat16),
|
||||
),
|
||||
@@ -434,7 +434,7 @@ def test_context_kv_uses_quantized_projection_fallback(monkeypatch):
|
||||
for projection in projections
|
||||
]
|
||||
qwen3_dflash.DFlashQwen3Model._build_context_kv_buffers(
|
||||
model, # type: ignore[arg-type]
|
||||
model,
|
||||
layers_attn,
|
||||
has_bias=False,
|
||||
)
|
||||
@@ -442,7 +442,7 @@ def test_context_kv_uses_quantized_projection_fallback(monkeypatch):
|
||||
assert all(not hasattr(projection, "weight") for projection in projections)
|
||||
|
||||
actual_k, actual_v = qwen3_dflash.DFlashQwen3Model._project_context_kv(
|
||||
model, # type: ignore[arg-type]
|
||||
model,
|
||||
context_states,
|
||||
num_ctx=2,
|
||||
num_layers=2,
|
||||
|
||||
@@ -141,9 +141,9 @@ def test_dflash_drafter_window_reserves_bonus_token():
|
||||
speculative_config=_dflash_speculative_config(NUM_SPECULATIVE_TOKENS),
|
||||
)
|
||||
# window = 4, so 96 fits (96 + 4 == 100) but 97 does not (97 + 4 == 101)
|
||||
assert input_fits_in_drafter(dflash_runner, SimpleNamespace(max_seq_len=96)) # type: ignore[arg-type]
|
||||
assert not input_fits_in_drafter(dflash_runner, SimpleNamespace(max_seq_len=97)) # type: ignore[arg-type]
|
||||
assert not input_fits_in_drafter(dflash_runner, None) # type: ignore[arg-type] # No metadata.
|
||||
assert input_fits_in_drafter(dflash_runner, SimpleNamespace(max_seq_len=96))
|
||||
assert not input_fits_in_drafter(dflash_runner, SimpleNamespace(max_seq_len=97))
|
||||
assert not input_fits_in_drafter(dflash_runner, None)
|
||||
|
||||
# Other drafters don't reserve the bonus token, so 97 fits (97 + 3 == 100).
|
||||
plain_runner = SimpleNamespace(
|
||||
@@ -151,4 +151,4 @@ def test_dflash_drafter_window_reserves_bonus_token():
|
||||
effective_drafter_max_model_len=100,
|
||||
speculative_config=SimpleNamespace(use_dflash=lambda: False),
|
||||
)
|
||||
assert input_fits_in_drafter(plain_runner, SimpleNamespace(max_seq_len=97)) # type: ignore[arg-type]
|
||||
assert input_fits_in_drafter(plain_runner, SimpleNamespace(max_seq_len=97))
|
||||
|
||||
@@ -158,9 +158,7 @@ def test_wrapper_forwards_spec_step_idx():
|
||||
|
||||
hidden_states = torch.empty(0)
|
||||
wrapper = object.__new__(DeepSeekV4MTP)
|
||||
wrapper.model = SimpleNamespace( # type: ignore[assignment]
|
||||
get_top_tokens=lambda *args: args
|
||||
)
|
||||
wrapper.model = SimpleNamespace(get_top_tokens=lambda *args: args)
|
||||
|
||||
assert DeepSeekV4MTP.get_top_tokens(wrapper, hidden_states, 2) == (hidden_states, 2)
|
||||
assert DeepSeekV4MTP.get_top_tokens(wrapper, hidden_states) == (hidden_states, 0)
|
||||
|
||||
@@ -51,7 +51,7 @@ def _create_proposer(
|
||||
target_parallel_config=ParallelConfig(),
|
||||
method="extract_hidden_states",
|
||||
num_speculative_tokens=num_speculative_tokens,
|
||||
draft_model_config={ # type: ignore[arg-type]
|
||||
draft_model_config={
|
||||
"hf_config": {
|
||||
"eagle_aux_hidden_state_layer_ids": layer_ids,
|
||||
}
|
||||
@@ -103,7 +103,7 @@ def test_proposer_initialization_missing_layer_ids():
|
||||
target_parallel_config=ParallelConfig(),
|
||||
method="extract_hidden_states",
|
||||
num_speculative_tokens=1,
|
||||
draft_model_config={ # type: ignore[arg-type]
|
||||
draft_model_config={
|
||||
"hf_config": {} # Missing eagle_aux_hidden_state_layer_ids
|
||||
},
|
||||
)
|
||||
@@ -363,7 +363,7 @@ def test_extract_hidden_states_text_only_config_regression():
|
||||
target_parallel_config=ParallelConfig(),
|
||||
method="extract_hidden_states",
|
||||
num_speculative_tokens=1,
|
||||
draft_model_config={ # type: ignore[arg-type]
|
||||
draft_model_config={
|
||||
"hf_config": {
|
||||
"eagle_aux_hidden_state_layer_ids": [1, 2, 3, 4],
|
||||
}
|
||||
@@ -450,7 +450,7 @@ def test_extract_hidden_states_speculative_config_vlm():
|
||||
target_parallel_config=ParallelConfig(),
|
||||
method="extract_hidden_states",
|
||||
num_speculative_tokens=1,
|
||||
draft_model_config={ # type: ignore[arg-type]
|
||||
draft_model_config={
|
||||
"hf_config": {
|
||||
"eagle_aux_hidden_state_layer_ids": [1, 2],
|
||||
}
|
||||
|
||||
@@ -190,17 +190,17 @@ def test_runtime_length_keeps_trained_checkpoint_geometry(monkeypatch):
|
||||
for slots in (1, 3):
|
||||
spec.num_speculative_tokens = slots
|
||||
head = LiLiCorr(
|
||||
vllm_config=SimpleNamespace(speculative_config=spec) # type: ignore[arg-type]
|
||||
vllm_config=SimpleNamespace(speculative_config=spec)
|
||||
).lilicorr
|
||||
assert head.slot_embedding.shape == (1, 1, 3, 1, 8)
|
||||
assert head.relative_slot_bias.shape == (2, 7)
|
||||
spec.draft_model_config.hf_config.dflash_config["lilicorr_enabled"] = False
|
||||
with pytest.raises(ValueError, match="lilicorr_enabled"):
|
||||
LiLiCorr(vllm_config=SimpleNamespace(speculative_config=spec)) # type: ignore[arg-type]
|
||||
LiLiCorr(vllm_config=SimpleNamespace(speculative_config=spec))
|
||||
spec.draft_model_config.hf_config.dflash_config["lilicorr_enabled"] = True
|
||||
spec.num_speculative_tokens = 4
|
||||
with pytest.raises(ValueError, match="trained block_size"):
|
||||
LiLiCorr(vllm_config=SimpleNamespace(speculative_config=spec)) # type: ignore[arg-type]
|
||||
LiLiCorr(vllm_config=SimpleNamespace(speculative_config=spec))
|
||||
|
||||
|
||||
@pytest.mark.parametrize("tp_size", [1, 2])
|
||||
@@ -269,8 +269,8 @@ def test_candidates_use_global_partition_and_exclude_padding(monkeypatch, tp_siz
|
||||
)
|
||||
monkeypatch.setattr(logits_processor, "_topk", lambda x, k: x.topk(k, dim=-1))
|
||||
ids, values = logits_processor.LogitsProcessor.get_top_k_tokens(
|
||||
fake, # type: ignore[arg-type]
|
||||
lm_head, # type: ignore[arg-type]
|
||||
fake,
|
||||
lm_head,
|
||||
torch.empty(2, 1),
|
||||
2,
|
||||
return_log_probs=True,
|
||||
@@ -291,13 +291,13 @@ def test_context_anchor_is_last_committed_normalized_feature():
|
||||
anchor_valid=torch.ones(4, dtype=torch.bool),
|
||||
)
|
||||
batch = SimpleNamespace(num_reqs=2, query_start_loc=torch.tensor([0, 2, 6]))
|
||||
LiLiCorrSpeculator.prepare_context_anchor(spec, batch, torch.tensor([0, 2])) # type: ignore[arg-type]
|
||||
LiLiCorrSpeculator.prepare_context_anchor(spec, batch, torch.tensor([0, 2]))
|
||||
torch.testing.assert_close(spec.anchor_hidden[:2], norm(hidden[[1, 3]]))
|
||||
assert spec.anchor_valid.tolist() == [True, True, False, False]
|
||||
assert not spec.anchor_hidden[2:].any()
|
||||
# A smaller reordered batch must not retain an old request's anchor.
|
||||
batch = SimpleNamespace(num_reqs=1, query_start_loc=torch.tensor([0, 3]))
|
||||
LiLiCorrSpeculator.prepare_context_anchor(spec, batch, torch.tensor([1])) # type: ignore[arg-type]
|
||||
LiLiCorrSpeculator.prepare_context_anchor(spec, batch, torch.tensor([1]))
|
||||
torch.testing.assert_close(spec.anchor_hidden[0], norm(hidden[1]))
|
||||
assert not spec.anchor_hidden[1:].any()
|
||||
|
||||
@@ -321,7 +321,7 @@ def test_candidate_generation_routes_scores_and_adaptive_inputs(monkeypatch, lil
|
||||
|
||||
monkeypatch.setattr(DFlashSpeculator, "__init__", init_base)
|
||||
cls = LiLiCorrSpeculator if lilicorr else DFlash2Speculator
|
||||
spec = cls(None, torch.device("cpu")) # type: ignore[arg-type]
|
||||
spec = cls(None, torch.device("cpu"))
|
||||
hidden = torch.arange(36).float().view(9, 4)
|
||||
spec._run_model = lambda *args: hidden
|
||||
spec.sample_indices = torch.tensor([1, 2, 4, 5, 7, 8])
|
||||
@@ -334,7 +334,7 @@ def test_candidate_generation_routes_scores_and_adaptive_inputs(monkeypatch, lil
|
||||
spec.draft_logits = None
|
||||
spec.use_fp64_gumbel = False
|
||||
spec.enable_adaptive_verification = True
|
||||
spec.input_buffers = SimpleNamespace(input_ids=torch.arange(9)) # type: ignore[assignment]
|
||||
spec.input_buffers = SimpleNamespace(input_ids=torch.arange(9))
|
||||
candidates = torch.arange(24).view(6, 4)
|
||||
log_probs = candidates.float().log_softmax(-1)
|
||||
scores = torch.randn(3, 2, 4, 4)
|
||||
@@ -636,7 +636,7 @@ def owned_head_model(monkeypatch):
|
||||
compilation_config=CompilationConfig(mode=CompilationMode.NONE),
|
||||
)
|
||||
):
|
||||
model = LiLiCorrForCausalLM(vllm_config=vllm_config) # type: ignore[arg-type]
|
||||
model = LiLiCorrForCausalLM(vllm_config=vllm_config)
|
||||
return model, vllm_config
|
||||
|
||||
return build
|
||||
@@ -746,7 +746,7 @@ def test_quantized_head_calls_methods_without_reading_packed_weights(monkeypatch
|
||||
block_size=4,
|
||||
rms_norm_eps=1e-6,
|
||||
config=_config(hidden_size=8),
|
||||
quant_config=quant_config, # type: ignore[arg-type]
|
||||
quant_config=quant_config,
|
||||
prefix="model.lilicorr",
|
||||
)
|
||||
with torch.no_grad():
|
||||
|
||||
@@ -53,7 +53,7 @@ def _make_proposer(
|
||||
monkeypatch.setattr(llm_base_proposer, "AttentionGroup", _FakeAttentionGroup)
|
||||
|
||||
proposer = EagleProposer.__new__(EagleProposer)
|
||||
proposer.vllm_config = None # type: ignore[assignment] # Patched backend discovery does not read config.
|
||||
proposer.vllm_config = None
|
||||
proposer.device = None
|
||||
proposer._draft_attn_layer_names = set(layer_names)
|
||||
proposer.kv_cache_gid = -1
|
||||
|
||||
@@ -85,7 +85,7 @@ def test_compute_probs_and_sample_next_token_uses_fp64_exponential_race():
|
||||
def test_mtp_model_returns_tuple(architecture: str, expected: bool):
|
||||
proposer = object.__new__(SpecDecodeBaseProposer)
|
||||
proposer.method = "mtp"
|
||||
proposer.draft_model_config = SimpleNamespace( # type: ignore[assignment] # Architecture-only config.
|
||||
proposer.draft_model_config = SimpleNamespace(
|
||||
hf_config=SimpleNamespace(architectures=[architecture])
|
||||
)
|
||||
|
||||
|
||||
@@ -700,7 +700,7 @@ def test_verify_rejects_unproposed_drafts():
|
||||
first_pos = inputs["pos"].view(num_trials, K + 1)[:, 0]
|
||||
prefill_len = first_pos + 1
|
||||
rejection_sampler = object.__new__(RejectionSampler)
|
||||
rejection_sampler.sampler = SimpleNamespace( # type: ignore[assignment]
|
||||
rejection_sampler.sampler = SimpleNamespace(
|
||||
apply_sampling_params=lambda logits, *args: logits,
|
||||
sampling_states=SimpleNamespace(
|
||||
temperature=SimpleNamespace(gpu=inputs["temperature"]),
|
||||
|
||||
Reference in New Issue
Block a user