Add pydocstyle to the ruff rules (#52136)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com> Signed-off-by: Harry Mellor <hej.mellor@gmail.com> Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
co-authored by
Claude Opus 5
parent
804e5377aa
commit
8c1557a79c
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
EAGLE3 Acceptance Length Regression Tests.
|
||||
"""EAGLE3 Acceptance Length Regression Tests.
|
||||
|
||||
These tests verify that acceptance lengths for EAGLE3 speculative decoding
|
||||
do not regress across vLLM commits. Each test runs inference on the MT-Bench
|
||||
|
||||
@@ -131,7 +131,8 @@ class TestBackupTokenAsyncSpec:
|
||||
|
||||
def test_buggy_code_was_always_off_by_one(self):
|
||||
"""The original code used seq_len as index, which is always one past
|
||||
the end of output_token_ids even without async inflation."""
|
||||
the end of output_token_ids even without async inflation.
|
||||
"""
|
||||
req_ids = ["r0"]
|
||||
requests = {"r0": _FakeRequest([0, 1, 2], [1000, 1001])}
|
||||
batch = _FakeInputBatch(req_ids, [5])
|
||||
|
||||
@@ -39,7 +39,6 @@ def _create_vllm_config_for_dsd(
|
||||
a schedule covering every query length in ``[1, max_decode_query_len]`` is
|
||||
generated.
|
||||
"""
|
||||
|
||||
max_decode_query_len = max_spec_tokens + 1
|
||||
max_capture_tokens = max_num_seqs * max_decode_query_len
|
||||
|
||||
@@ -91,7 +90,6 @@ def test_dynamic_sd_full_cudagraph_covers_all_uniform_decode_shapes(monkeypatch)
|
||||
shapes have been built, dispatch() should pick a FULL graph for every
|
||||
uniform decode batch shape produced by DSD up to max_num_seqs.
|
||||
"""
|
||||
|
||||
max_num_seqs = 512
|
||||
max_spec_tokens = 7
|
||||
max_decode_query_len = max_spec_tokens + 1
|
||||
@@ -190,7 +188,6 @@ def test_dynamic_sd_non_uniform_batch_falls_back_to_piecewise(monkeypatch):
|
||||
FULL candidates should be skipped in favor of the mixed-batch PIECEWISE
|
||||
graph under FULL_AND_PIECEWISE mode.
|
||||
"""
|
||||
|
||||
max_spec_tokens = 4
|
||||
|
||||
monkeypatch.setattr(
|
||||
@@ -236,7 +233,6 @@ def test_prompt_chunks_shaped_like_spec_decode_miss_the_full_graph(monkeypatch):
|
||||
by coincidence; replaying the FULL graph over them corrupts every request
|
||||
in the batch. See https://github.com/vllm-project/vllm/issues/49918.
|
||||
"""
|
||||
|
||||
max_spec_tokens = 7
|
||||
decode_query_len = max_spec_tokens + 1
|
||||
|
||||
@@ -307,7 +303,6 @@ def test_basic_sd_does_not_capture_shorter_full_decode_shapes(monkeypatch):
|
||||
Uniform batches with smaller query lengths should therefore miss the FULL
|
||||
path entirely when using FULL_AND_PIECEWISE.
|
||||
"""
|
||||
|
||||
max_num_seqs = 512
|
||||
max_spec_tokens = 7
|
||||
max_decode_query_len = max_spec_tokens + 1
|
||||
@@ -368,7 +363,6 @@ def test_dynamic_sd_only_captures_scheduled_query_lengths(monkeypatch):
|
||||
FULL graphs, while every other query length (e.g. the lower values 1, 2, 3)
|
||||
must fall back to the mixed-batch PIECEWISE graph.
|
||||
"""
|
||||
|
||||
max_num_seqs = 128
|
||||
max_spec_tokens = 7
|
||||
max_decode_query_len = max_spec_tokens + 1
|
||||
|
||||
@@ -116,8 +116,7 @@ def _create_proposer(
|
||||
|
||||
|
||||
def test_prepare_next_token_ids():
|
||||
"""
|
||||
Test for prepare_next_token_ids_cpu and prepare_next_token_ids_padded.
|
||||
"""Test for prepare_next_token_ids_cpu and prepare_next_token_ids_padded.
|
||||
Each will produce a device tensor of next_token_ids, taking as input
|
||||
either the GPU tensor of sampled_token_ids with -1 for rejected tokens,
|
||||
or the CPU python list[list[int]] with the rejected tokens removed.
|
||||
@@ -196,8 +195,7 @@ def test_prepare_next_token_ids():
|
||||
|
||||
|
||||
def test_prepare_inputs():
|
||||
"""
|
||||
cu_target_query_lens: [0, a, a + b, a + b + c]
|
||||
"""cu_target_query_lens: [0, a, a + b, a + b + c]
|
||||
num_rejected_tokens: [n1, n2, n3]
|
||||
num_tokens_per_req: [a - n1, b - n2, c - n3]
|
||||
cu_num_tokens: [0, a - n1, a + b - n1 - n2, a + b + c - n1 - n2 - n3]
|
||||
@@ -286,8 +284,7 @@ def test_prepare_inputs():
|
||||
|
||||
|
||||
def test_prepare_inputs_padded():
|
||||
"""
|
||||
Input scenario is 3 requests with num_speculative_tokens == 2 and:
|
||||
"""Input scenario is 3 requests with num_speculative_tokens == 2 and:
|
||||
- Request 1: query_len = 3, rejected = 1
|
||||
- Request 2: query_len = 3, rejected = 0
|
||||
- Request 3: query_len = 3, rejected = 2
|
||||
@@ -297,7 +294,6 @@ def test_prepare_inputs_padded():
|
||||
Reason: After accounting for rejections, these are the valid token positions
|
||||
from the original indices to sample from.
|
||||
"""
|
||||
|
||||
device = torch.device(DEVICE_TYPE)
|
||||
|
||||
expected_token_indices_to_sample = torch.tensor(
|
||||
@@ -350,8 +346,7 @@ def test_prepare_inputs_padded():
|
||||
|
||||
|
||||
def test_set_inputs_first_pass_default_eagle():
|
||||
"""
|
||||
Test for set_inputs_first_pass without extra input slots (default EAGLE).
|
||||
"""Test for set_inputs_first_pass without extra input slots (default EAGLE).
|
||||
|
||||
This tests the path where needs_extra_input_slots=False, which is the
|
||||
default EAGLE pathway. In this case:
|
||||
@@ -436,8 +431,7 @@ def test_set_inputs_first_pass_default_eagle():
|
||||
|
||||
|
||||
def test_set_inputs_first_pass_draft_model():
|
||||
"""
|
||||
Test for set_inputs_first_pass with a draft model (extra input slots,
|
||||
"""Test for set_inputs_first_pass with a draft model (extra input slots,
|
||||
no shift).
|
||||
|
||||
This tests the path where needs_extra_input_slots=True and
|
||||
@@ -578,8 +572,7 @@ def test_set_inputs_first_pass_draft_model():
|
||||
|
||||
|
||||
def test_set_inputs_first_pass_parallel_drafting():
|
||||
"""
|
||||
Test for set_inputs_first_pass with parallel drafting (extra input slots,
|
||||
"""Test for set_inputs_first_pass with parallel drafting (extra input slots,
|
||||
with shift).
|
||||
|
||||
This tests the path where needs_extra_input_slots=True and
|
||||
@@ -1107,8 +1100,7 @@ def test_propose_stores_probabilistic_draft_probs(attn_backend, monkeypatch):
|
||||
|
||||
|
||||
def test_set_inputs_first_pass_dflash():
|
||||
"""
|
||||
Test for DFlash set_inputs_first_pass.
|
||||
"""Test for DFlash set_inputs_first_pass.
|
||||
|
||||
DFlash uses cross-attention: context tokens become K/V and only
|
||||
query tokens (bonus + mask) are Q. This tests the DFlash-specific
|
||||
|
||||
@@ -105,7 +105,8 @@ def test_build_draft_attn_metadata_sets_seq_lens_cpu_upper_bound():
|
||||
"""The fix: every per-step ``CommonAttentionMetadata`` carries a non-None
|
||||
``seq_lens_cpu_upper_bound`` derived from the target-side upper bound plus
|
||||
the current draft-step offset. Padded entries are zeroed (matching the
|
||||
main model runner's convention)."""
|
||||
main model runner's convention).
|
||||
"""
|
||||
fake = _make_fake_speculator()
|
||||
base = torch.tensor([100, 200, 300, 0], dtype=torch.int32)
|
||||
|
||||
@@ -126,7 +127,8 @@ def test_build_draft_attn_metadata_sets_seq_lens_cpu_upper_bound():
|
||||
|
||||
def test_build_draft_attn_metadata_zeroes_padded_upper_bound_tail():
|
||||
"""The padded tail of the upper-bound tensor is zeroed, so it stays a
|
||||
valid tensor of length ``num_reqs_padded`` regardless of padding."""
|
||||
valid tensor of length ``num_reqs_padded`` regardless of padding.
|
||||
"""
|
||||
fake = _make_fake_speculator()
|
||||
base = torch.zeros(2, dtype=torch.int32)
|
||||
|
||||
@@ -141,7 +143,8 @@ def test_build_draft_attn_metadata_zeroes_padded_upper_bound_tail():
|
||||
|
||||
def test_build_draft_attn_metadata_clamps_to_max_model_len():
|
||||
"""The per-request upper bound (target bound + step) is clamped to the
|
||||
model length so it never exceeds the allocated KV range."""
|
||||
model length so it never exceeds the allocated KV range.
|
||||
"""
|
||||
fake = _make_fake_speculator(max_model_len=1024)
|
||||
base = torch.tensor([1023, 500], dtype=torch.int32)
|
||||
|
||||
|
||||
@@ -132,7 +132,8 @@ def test_eagle_step_slot_mapping_kernel_exceeds_max():
|
||||
|
||||
def test_eagle_step_slot_mapping_kernel_cudagraph_padding():
|
||||
"""Test that padding threads write PADDING_SLOT_ID when
|
||||
input_batch_size > batch_size (cudagraph padding)."""
|
||||
input_batch_size > batch_size (cudagraph padding).
|
||||
"""
|
||||
device = torch.device(DEVICE_TYPE)
|
||||
batch_size = 4
|
||||
input_batch_size = 8
|
||||
|
||||
@@ -130,8 +130,7 @@ def test_proposer_initialization_missing_layer_ids():
|
||||
|
||||
|
||||
def test_prepare_next_token_ids_padded():
|
||||
"""
|
||||
Test for prepare_next_token_ids_padded with extract_hidden_states.
|
||||
"""Test for prepare_next_token_ids_padded with extract_hidden_states.
|
||||
|
||||
Since num_speculative_tokens == 1, sampled_token_ids has shape (batch_size, 1).
|
||||
For each request we either use the sampled token (if valid and not discarded)
|
||||
@@ -196,8 +195,7 @@ def test_prepare_next_token_ids_padded():
|
||||
|
||||
|
||||
def test_propose():
|
||||
"""
|
||||
Test the propose() method of ExtractHiddenStatesProposer.
|
||||
"""Test the propose() method of ExtractHiddenStatesProposer.
|
||||
|
||||
This should:
|
||||
1. Accept target hidden states and sampled token IDs
|
||||
|
||||
@@ -69,7 +69,8 @@ def _make_kv_cache_config(layer_names: set[str]) -> SimpleNamespace:
|
||||
|
||||
def test_block_size_uses_kernel_block_size(monkeypatch: pytest.MonkeyPatch):
|
||||
"""The proposer's slot-mapping math runs against the kernel-granularity
|
||||
block table, so block_size must come from kernel_block_sizes."""
|
||||
block table, so block_size must come from kernel_block_sizes.
|
||||
"""
|
||||
layer_names = {"draft.0.self_attn.attn"}
|
||||
proposer = _make_proposer(monkeypatch, layer_names)
|
||||
|
||||
@@ -97,7 +98,8 @@ def test_block_size_falls_back_to_kv_cache_spec(monkeypatch: pytest.MonkeyPatch)
|
||||
|
||||
def test_draft_layer_iteration_is_deterministic(monkeypatch: pytest.MonkeyPatch):
|
||||
"""_draft_attn_layer_names is a set; the attention groups built from it
|
||||
must not depend on its (process-random) iteration order."""
|
||||
must not depend on its (process-random) iteration order.
|
||||
"""
|
||||
layer_names = {"draft.c.attn", "draft.a.attn", "draft.b.attn"}
|
||||
expected_order = sorted(layer_names)
|
||||
|
||||
|
||||
@@ -93,7 +93,8 @@ def test_eagle_max_len(
|
||||
@pytest.mark.parametrize("spec_max_model_len", [80, 150])
|
||||
def test_mtp_speculative_config_max_model_len(spec_max_model_len: int):
|
||||
"""Regression test for #41456: max_model_len in speculative config
|
||||
should be respected for the draft model."""
|
||||
should be respected for the draft model.
|
||||
"""
|
||||
model_config = ModelConfig(
|
||||
model="XiaomiMiMo/MiMo-7B-Base",
|
||||
runner="generate",
|
||||
|
||||
@@ -66,7 +66,6 @@ def _create_mtp_proposer(num_speculative_tokens: int) -> EagleProposer:
|
||||
@mock.patch("vllm.v1.spec_decode.llm_base_proposer.get_model")
|
||||
def test_mtp_load_model_unified(mock_get_model, mock_get_layers, mock_get_pp_group):
|
||||
"""Test MTP-specific model loading with unified model approach."""
|
||||
|
||||
# Setup mocks
|
||||
mock_model = mock.MagicMock()
|
||||
mock_model.model.embed_tokens.weight.shape = (131072, 4096)
|
||||
@@ -117,8 +116,7 @@ def test_mtp_load_model_unified(mock_get_model, mock_get_layers, mock_get_pp_gro
|
||||
|
||||
@pytest.mark.parametrize("num_speculative_tokens", [1])
|
||||
def test_mtp_propose(num_speculative_tokens, monkeypatch):
|
||||
"""Test that MTP's forward method returns hidden states directly"""
|
||||
|
||||
"""Test that MTP's forward method returns hidden states directly."""
|
||||
device = torch.device(DEVICE_TYPE)
|
||||
batch_size = 2
|
||||
seq_lens = [5, 3]
|
||||
|
||||
@@ -93,8 +93,7 @@ def _assert_distribution_match(
|
||||
label: str = "",
|
||||
min_expected: float = 5.0,
|
||||
):
|
||||
"""
|
||||
Assert sampled tokens match the target distribution via a
|
||||
"""Assert sampled tokens match the target distribution via a
|
||||
chi-squared goodness-of-fit test. This is done by computing
|
||||
observed vs expected token counts (target_probs * num_samples),
|
||||
then checking that the chi-squared statistic is below a conservative
|
||||
@@ -242,8 +241,7 @@ def test_watermarked_recovery_supports_smaller_draft_vocabulary():
|
||||
def test_stochastic_rejection_sample(
|
||||
num_speculative_steps: int, temperature: float, draft_logits_dtype: torch.dtype
|
||||
):
|
||||
"""
|
||||
Verify that rejection sampling produces the target distribution.
|
||||
"""Verify that rejection sampling produces the target distribution.
|
||||
This is done by simulating many independent trials of speculative
|
||||
decoding (from a fixed target and draft distribution). We then
|
||||
run rejection sample on all of the trials (requests), and verify
|
||||
@@ -253,7 +251,6 @@ def test_stochastic_rejection_sample(
|
||||
Parametrized over the draft-logits dtype: storing them in the draft head's
|
||||
dtype must not bias the output distribution.
|
||||
"""
|
||||
|
||||
torch.manual_seed(42)
|
||||
device = "cuda"
|
||||
num_trials = 10 * VOCAB_SIZE
|
||||
@@ -383,11 +380,9 @@ def test_gumbel_drafted_rejection_sample_is_unbiased(num_speculative_steps: int)
|
||||
|
||||
@pytest.mark.parametrize("num_speculative_steps", [1, 3])
|
||||
def test_greedy_rejection_sample(num_speculative_steps: int):
|
||||
"""
|
||||
Verify that greedy (temperature=0) always outputs the target argmax
|
||||
"""Verify that greedy (temperature=0) always outputs the target argmax
|
||||
at every accepted position.
|
||||
"""
|
||||
|
||||
torch.manual_seed(42)
|
||||
device = "cuda"
|
||||
num_trials = 10 * VOCAB_SIZE
|
||||
@@ -435,8 +430,7 @@ def test_synthetic_rejection_sample(
|
||||
temperature: float,
|
||||
unconditional_rates: list[float],
|
||||
):
|
||||
"""
|
||||
Verify that synthetic rejection sampling produces the expected
|
||||
"""Verify that synthetic rejection sampling produces the expected
|
||||
per-position acceptance rates. The unconditional rate at position i
|
||||
is P(all draft steps 0..i accepted) = product(conditional_rates[0:i+1]).
|
||||
This is approximately mean(num accepted >= i + 1) over many trials.
|
||||
@@ -689,14 +683,12 @@ def test_placeholder_blocks_later_draft_tokens(use_block_verification: bool):
|
||||
def test_block_verification_rejection_sample(
|
||||
num_speculative_steps: int, temperature: float, has_draft_logits: bool
|
||||
):
|
||||
"""
|
||||
Verify that block verification (Sun et al.) preserves the target
|
||||
"""Verify that block verification (Sun et al.) preserves the target
|
||||
distribution at every accepted position, for both the full draft-logits
|
||||
case and the one-hot (no draft logits) case. Block verification changes
|
||||
*which* prefix is accepted, but the marginal of every output position must
|
||||
still match the target distribution p(x).
|
||||
"""
|
||||
|
||||
torch.manual_seed(42)
|
||||
device = "cuda"
|
||||
num_trials = 10 * VOCAB_SIZE
|
||||
@@ -733,12 +725,10 @@ def test_block_verification_rejection_sample(
|
||||
|
||||
@pytest.mark.parametrize("num_speculative_steps", [3, 5])
|
||||
def test_block_verification_accepts_at_least_as_many(num_speculative_steps: int):
|
||||
"""
|
||||
Block verification is designed to accept at least as long a prefix as
|
||||
"""Block verification is designed to accept at least as long a prefix as
|
||||
token verification in expectation. Verify the mean accepted length is no
|
||||
worse than the standard method on the same inputs.
|
||||
"""
|
||||
|
||||
torch.manual_seed(0)
|
||||
device = "cuda"
|
||||
num_trials = 20 * VOCAB_SIZE
|
||||
|
||||
@@ -134,8 +134,7 @@ def print_spec_decode_stats(stats: dict) -> None:
|
||||
|
||||
@pytest.mark.parametrize("config", SPECULATOR_CONFIGS)
|
||||
def test_speculators_model(vllm_runner, example_prompts, monkeypatch, config):
|
||||
"""
|
||||
Test speculators model properly initializes speculative decoding.
|
||||
"""Test speculators model properly initializes speculative decoding.
|
||||
|
||||
Verifies:
|
||||
1. Speculative config is automatically initialized from speculators config
|
||||
@@ -185,8 +184,7 @@ def test_speculators_model(vllm_runner, example_prompts, monkeypatch, config):
|
||||
@large_gpu_mark(min_gb=40)
|
||||
@pytest.mark.parametrize("config", SPECULATOR_CONFIGS)
|
||||
def test_speculators_correctness(monkeypatch, config):
|
||||
"""
|
||||
E2E correctness test via the speculators auto-detect path.
|
||||
"""E2E correctness test via the speculators auto-detect path.
|
||||
|
||||
Evaluates GSM8k accuracy to ensure the speculators-format model produces
|
||||
correct outputs, and checks that acceptance length does not collapse under
|
||||
|
||||
@@ -32,8 +32,7 @@ from vllm.platforms import current_platform
|
||||
def test_eagle3_speculators_model(
|
||||
vllm_runner, example_prompts, model_path, monkeypatch
|
||||
):
|
||||
"""
|
||||
Test Eagle3 speculators models properly initialize speculative decoding.
|
||||
"""Test Eagle3 speculators models properly initialize speculative decoding.
|
||||
|
||||
This test verifies:
|
||||
1. Eagle3 support is detected for the model
|
||||
|
||||
Reference in New Issue
Block a user