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
@@ -11,7 +11,8 @@ from ...utils import compute_acceptance_len, get_test_prompts
|
||||
@single_gpu_only
|
||||
def test_synthetic_acceptance_rate(vllm_runner):
|
||||
"""Verify that synthetic rejection sampling produces an acceptance
|
||||
length close to the requested mean acceptance length."""
|
||||
length close to the requested mean acceptance length.
|
||||
"""
|
||||
num_spec_tokens = 3
|
||||
expected_acceptance_len = 1.875
|
||||
tolerance = 0.15
|
||||
|
||||
@@ -43,8 +43,7 @@ def run_acceptance_length_eval(
|
||||
chat_template_kwargs: dict[str, Any],
|
||||
use_mrv2: bool,
|
||||
):
|
||||
"""
|
||||
E2E acceptance-rate validation for speculative decoding.
|
||||
"""E2E acceptance-rate validation for speculative decoding.
|
||||
|
||||
Drives one or more datasets (keyed in ``expected_acceptance_lengths``)
|
||||
through the spec decode engine and asserts the mean acceptance length
|
||||
|
||||
@@ -167,7 +167,8 @@ def test_draft_model_tensor_parallelism(vllm_runner):
|
||||
@multi_gpu_only(num_gpus=2)
|
||||
def test_draft_model_engine_args_tensor_parallelism():
|
||||
"""Ensure the vllm_config for the draft model is created correctly,
|
||||
and independently of the target model (quantization, TP, etc.)"""
|
||||
and independently of the target model (quantization, TP, etc.)
|
||||
"""
|
||||
engine_args = EngineArgs(
|
||||
model="Qwen/Qwen3-1.7B-FP8", # <<< tgt quantized
|
||||
tensor_parallel_size=2,
|
||||
@@ -198,7 +199,8 @@ def test_draft_model_engine_args_tensor_parallelism():
|
||||
|
||||
def _apply_draft_moe_backend(vllm_config: VllmConfig) -> VllmConfig:
|
||||
"""Replicate SpecDecodeBaseProposer._create_draft_vllm_config logic
|
||||
so we can test it without instantiating a full proposer."""
|
||||
so we can test it without instantiating a full proposer.
|
||||
"""
|
||||
spec_cfg = vllm_config.speculative_config
|
||||
if spec_cfg.moe_backend is not None:
|
||||
return replace(
|
||||
@@ -223,7 +225,8 @@ def _platform_moe_backend(monkeypatch: pytest.MonkeyPatch) -> MoEBackend:
|
||||
|
||||
def test_draft_model_moe_backend_override(monkeypatch: pytest.MonkeyPatch):
|
||||
"""When moe_backend is set in speculative_config, the draft VllmConfig
|
||||
should use it while the target keeps its own setting."""
|
||||
should use it while the target keeps its own setting.
|
||||
"""
|
||||
target_moe_backend = _platform_moe_backend(monkeypatch)
|
||||
engine_args = EngineArgs(
|
||||
model="Qwen/Qwen3-1.7B",
|
||||
@@ -248,7 +251,8 @@ def test_draft_model_moe_backend_override(monkeypatch: pytest.MonkeyPatch):
|
||||
|
||||
def test_draft_model_moe_backend_inherits_target(monkeypatch: pytest.MonkeyPatch):
|
||||
"""When moe_backend is not set in speculative_config, the draft should
|
||||
inherit the target's moe_backend."""
|
||||
inherit the target's moe_backend.
|
||||
"""
|
||||
target_moe_backend = _platform_moe_backend(monkeypatch)
|
||||
engine_args = EngineArgs(
|
||||
model="Qwen/Qwen3-1.7B",
|
||||
@@ -271,7 +275,8 @@ def test_draft_model_moe_backend_inherits_target(monkeypatch: pytest.MonkeyPatch
|
||||
|
||||
def test_draft_model_moe_backend_default_auto():
|
||||
"""When neither target nor draft set moe_backend explicitly, both should
|
||||
default to 'auto'."""
|
||||
default to 'auto'.
|
||||
"""
|
||||
engine_args = EngineArgs(
|
||||
model="Qwen/Qwen3-1.7B",
|
||||
tensor_parallel_size=1,
|
||||
@@ -292,8 +297,8 @@ def test_draft_model_moe_backend_default_auto():
|
||||
|
||||
def test_draft_model_engine_args_rejects_invalid_tp_argname():
|
||||
"""The user should pass "draft_tensor_parallel_size" rather than
|
||||
"tensor_parallel_size". We enforce this with validation."""
|
||||
|
||||
"tensor_parallel_size". We enforce this with validation.
|
||||
"""
|
||||
engine_args = EngineArgs(
|
||||
model="Qwen/Qwen3-1.7B",
|
||||
tensor_parallel_size=1,
|
||||
@@ -310,7 +315,8 @@ def test_draft_model_engine_args_rejects_invalid_tp_argname():
|
||||
|
||||
def assert_draft_model_correctness(args: ArgsTest, vllm_runner):
|
||||
"""Compare the outputs using and not using speculative decoding.
|
||||
In the greedy decoding case, the outputs must match EXACTLY."""
|
||||
In the greedy decoding case, the outputs must match EXACTLY.
|
||||
"""
|
||||
test_prompts: list[Messages] = get_messages(
|
||||
dataset=args.dataset, n=args.num_prompts
|
||||
)
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
This script contains:
|
||||
"""This script contains:
|
||||
1. test lora with speculative decoding for batch inference
|
||||
"""
|
||||
|
||||
@@ -59,8 +58,7 @@ def test_batch_inference_correctness(
|
||||
model_setup: tuple[str, str, str, str, int],
|
||||
vllm_runner,
|
||||
):
|
||||
"""
|
||||
Compare the outputs of a LLM with only Lora and a LLM with both SD and Lora.
|
||||
"""Compare the outputs of a LLM with only Lora and a LLM with both SD and Lora.
|
||||
Should be the same and no failure when doing batch inference.
|
||||
model_setup: (method, model_name, spec_model_name, lora_path, tp_size)
|
||||
"""
|
||||
|
||||
@@ -28,8 +28,7 @@ def _run_eagle_correctness(
|
||||
attn_backend: str,
|
||||
vllm_runner,
|
||||
):
|
||||
"""
|
||||
Compare the outputs of an original LLM and a speculative LLM
|
||||
"""Compare the outputs of an original LLM and a speculative LLM
|
||||
which should be the same when using eagle speculative decoding.
|
||||
"""
|
||||
method, model_name, spec_model_name, tp_size = model_setup
|
||||
|
||||
@@ -157,8 +157,7 @@ def test_mtp_correctness(
|
||||
expected_accuracy_threshold: float,
|
||||
vllm_runner,
|
||||
):
|
||||
"""
|
||||
Compare the outputs of a original LLM and a speculative LLM
|
||||
"""Compare the outputs of a original LLM and a speculative LLM
|
||||
which should be the same when using MTP speculative decoding. Due to some variance
|
||||
in the engine, it is possible for some outputs to differ, so we expect that at least
|
||||
6/10 output tokens match exactly, and that the GSM8k accuracy is above a precomputed
|
||||
|
||||
@@ -64,8 +64,7 @@ def test_ngram_gpu_default_with_async_scheduling(
|
||||
async_scheduling: bool,
|
||||
vllm_runner,
|
||||
):
|
||||
"""
|
||||
Test ngram_gpu speculative decoding (k=3) correctness with and without
|
||||
"""Test ngram_gpu speculative decoding (k=3) correctness with and without
|
||||
async scheduling, validated via GSM8K accuracy.
|
||||
Uses Qwen/Qwen3-8B (ref GSM8K accuracy: 87%-92%).
|
||||
"""
|
||||
@@ -99,8 +98,7 @@ def test_suffix_decoding_acceptance(
|
||||
model_name: str,
|
||||
vllm_runner,
|
||||
):
|
||||
"""
|
||||
Check that suffix decoding caching takes effect and improves acceptance
|
||||
"""Check that suffix decoding caching takes effect and improves acceptance
|
||||
lengths and acceptance rates over multiple runs of the same prompts.
|
||||
"""
|
||||
test_prompts = get_test_prompts(mm_enabled=False)
|
||||
|
||||
@@ -32,8 +32,7 @@ def test_speculators_model_integration(
|
||||
expected_accuracy_threshold: float,
|
||||
vllm_runner,
|
||||
):
|
||||
"""
|
||||
Test that speculators models work with the simplified integration.
|
||||
"""Test that speculators models work with the simplified integration.
|
||||
|
||||
This verifies the `vllm serve <speculator-model>` use case where
|
||||
speculative config is automatically detected from the model config
|
||||
|
||||
@@ -28,7 +28,8 @@ def test_mtp_sharded_sampling_equivalence(monkeypatch: pytest.MonkeyPatch):
|
||||
sampling under MTP spec decoding: the collectives move the same logits
|
||||
bytes, Gumbel keys derive from (request slot, position, seed), and slot
|
||||
assignment is rank-deterministic. Both runs here are spec decode with
|
||||
identical math, so outputs must match exactly."""
|
||||
identical math, so outputs must match exactly.
|
||||
"""
|
||||
tp_size = 2
|
||||
_skip_if_insufficient_gpus_for_tp(tp_size)
|
||||
model_name = "Qwen/Qwen3.5-0.8B-Base"
|
||||
|
||||
Reference in New Issue
Block a user