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
@@ -31,8 +31,7 @@ def test_v1_generation_is_deterministic_across_batch_sizes_with_needle(
|
||||
backend,
|
||||
rms_norm_impl,
|
||||
):
|
||||
"""
|
||||
Ensures that the same request (the 'needle' prompt) yields identical output
|
||||
"""Ensures that the same request (the 'needle' prompt) yields identical output
|
||||
whether run alone (bs=1) or mixed into a larger batch (e.g., bs=64),
|
||||
using the high-level v1 LLM() API only (no manual batching).
|
||||
|
||||
@@ -51,6 +50,7 @@ def test_v1_generation_is_deterministic_across_batch_sizes_with_needle(
|
||||
to produce a more random-sounding phrase, yet remain deterministic by
|
||||
seed.
|
||||
- Keep max_tokens and max_model_len bounded for speed and memory use.
|
||||
|
||||
"""
|
||||
seed = int(os.getenv("VLLM_TEST_SEED", "12345"))
|
||||
random.seed(seed)
|
||||
@@ -395,8 +395,7 @@ def test_logprobs_bitwise_batch_invariance_bs1_vs_bsN(
|
||||
BACKENDS,
|
||||
)
|
||||
def test_simple_generation(backend):
|
||||
"""
|
||||
Simple test that runs the model with a basic prompt and prints the output.
|
||||
"""Simple test that runs the model with a basic prompt and prints the output.
|
||||
Useful for quick smoke testing and debugging.
|
||||
"""
|
||||
model = TEST_MODEL
|
||||
@@ -447,8 +446,7 @@ def test_simple_generation(backend):
|
||||
def test_logprobs_without_batch_invariance_should_fail(
|
||||
backend, monkeypatch: pytest.MonkeyPatch
|
||||
):
|
||||
"""
|
||||
This test is the inverse of test_logprobs_bitwise_batch_invariance_bs1_vs_bsN.
|
||||
"""This test is the inverse of test_logprobs_bitwise_batch_invariance_bs1_vs_bsN.
|
||||
It DISABLES batch invariance mode and expects to see non-deterministic behavior
|
||||
between BS=1 and BS=N runs. This demonstrates that batch invariance is actually
|
||||
doing something useful.
|
||||
@@ -660,8 +658,7 @@ def test_logprobs_without_batch_invariance_should_fail(
|
||||
def test_decode_logprobs_match_prefill_logprobs(
|
||||
backend,
|
||||
):
|
||||
"""
|
||||
Test that verifies decode logprobs match prefill logprobs.
|
||||
"""Test that verifies decode logprobs match prefill logprobs.
|
||||
|
||||
For each decoded token at position i:
|
||||
1. Run decode to generate N tokens and collect their logprobs
|
||||
@@ -927,8 +924,7 @@ def LLM_with_max_seqs(
|
||||
attention_config: dict | None = None,
|
||||
kernel_config: dict | None = None,
|
||||
) -> LLM:
|
||||
"""
|
||||
Helper to construct an LLM with a specific max_num_seqs (batch-size limit)
|
||||
"""Helper to construct an LLM with a specific max_num_seqs (batch-size limit)
|
||||
using the high-level v1 LLM API, while constraining memory usage.
|
||||
"""
|
||||
extra_kwargs: dict = {}
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
Test batch-invariant matmul against torch.matmul for various shape combinations.
|
||||
"""Test batch-invariant matmul against torch.matmul for various shape combinations.
|
||||
|
||||
Tests correctness (matches torch.matmul) and batch invariance (result for one
|
||||
item doesn't change based on other items in the batch).
|
||||
@@ -53,9 +52,7 @@ DEVICE_TYPE = current_platform.device_type
|
||||
)
|
||||
@pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16])
|
||||
def test_matmul_correctness(a_shape, b_shape, dtype):
|
||||
"""
|
||||
Compare matmul_batch_invariant against torch.matmul for various shapes.
|
||||
"""
|
||||
"""Compare matmul_batch_invariant against torch.matmul for various shapes."""
|
||||
device = torch.device(DEVICE_TYPE)
|
||||
|
||||
torch.manual_seed(42)
|
||||
@@ -87,11 +84,9 @@ def test_matmul_correctness(a_shape, b_shape, dtype):
|
||||
@skip_unsupported
|
||||
@pytest.mark.parametrize("dtype", [torch.float16, torch.bfloat16])
|
||||
def test_matmul_batch_invariance(dtype):
|
||||
"""
|
||||
Verify that the result for one item is bitwise identical regardless
|
||||
"""Verify that the result for one item is bitwise identical regardless
|
||||
of what other items are in the batch.
|
||||
"""
|
||||
|
||||
device = torch.device(DEVICE_TYPE)
|
||||
|
||||
torch.manual_seed(42)
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
HTTP-based batch invariance test: send requests to a running
|
||||
"""HTTP-based batch invariance test: send requests to a running
|
||||
vLLM server and compare BS=1 vs BS=N results (tokens and per-step logprobs).
|
||||
|
||||
Environment variables:
|
||||
|
||||
@@ -41,8 +41,7 @@ def test_rms_norm_batch_invariant_vs_reference(
|
||||
eps: float,
|
||||
seed: int,
|
||||
):
|
||||
"""
|
||||
Compare batch-invariant Triton RMS norm against a PyTorch reference.
|
||||
"""Compare batch-invariant Triton RMS norm against a PyTorch reference.
|
||||
|
||||
Tests that the Triton-based batch-invariant RMS norm produces numerically
|
||||
equivalent results to an independent implementation across configurations.
|
||||
@@ -90,9 +89,7 @@ def test_fused_add_rms_norm_batch_invariant_residual_path(
|
||||
n_extra: int,
|
||||
seed: int,
|
||||
):
|
||||
"""
|
||||
Test the batch-invariant fused residual-add + RMSNorm helper directly.
|
||||
"""
|
||||
"""Test the batch-invariant fused residual-add + RMSNorm helper directly."""
|
||||
device = torch.device(DEVICE_TYPE)
|
||||
|
||||
torch.manual_seed(seed)
|
||||
@@ -278,7 +275,8 @@ def test_rms_norm_per_block_quant_batch_invariant(
|
||||
hidden_size: int, dtype: torch.dtype, seed: int
|
||||
):
|
||||
"""C++ ``rms_norm_per_block_quant`` must be batch invariant across the
|
||||
block threshold (compiled fp8 block-quant path; block pinned to 512)."""
|
||||
block threshold (compiled fp8 block-quant path; block pinned to 512).
|
||||
"""
|
||||
import vllm._custom_ops as ops
|
||||
|
||||
device = torch.device(DEVICE_TYPE)
|
||||
@@ -313,8 +311,7 @@ def test_rms_norm_per_block_quant_batch_invariant(
|
||||
def test_rms_norm_3d_input(
|
||||
default_vllm_config, batch_size: int, seq_len: int, hidden_size: int
|
||||
):
|
||||
"""
|
||||
Test RMS norm with 3D input tensors (batch, seq_len, hidden_size).
|
||||
"""Test RMS norm with 3D input tensors (batch, seq_len, hidden_size).
|
||||
|
||||
Ensures that the batch-invariant RMS norm correctly handles multi-dimensional
|
||||
inputs that are common in transformer models.
|
||||
@@ -349,8 +346,7 @@ def test_rms_norm_3d_input(
|
||||
|
||||
@skip_if_not_cuda
|
||||
def test_rms_norm_numerical_stability(default_vllm_config):
|
||||
"""
|
||||
Test RMS norm numerical stability with extreme values.
|
||||
"""Test RMS norm numerical stability with extreme values.
|
||||
|
||||
Ensures that both implementations handle edge cases like very small or large
|
||||
values without producing NaN or Inf.
|
||||
@@ -406,8 +402,7 @@ def test_rms_norm_numerical_stability(default_vllm_config):
|
||||
|
||||
@skip_unsupported
|
||||
def test_rms_norm_formula(default_vllm_config):
|
||||
"""
|
||||
Test that RMS norm follows the correct mathematical formula.
|
||||
"""Test that RMS norm follows the correct mathematical formula.
|
||||
|
||||
Verifies: output = input / sqrt(mean(input^2) + eps) * weight
|
||||
"""
|
||||
@@ -440,8 +435,7 @@ def test_rms_norm_formula(default_vllm_config):
|
||||
@skip_if_not_cuda
|
||||
@pytest.mark.parametrize("hidden_size", [128, 1024, 4096, 16384])
|
||||
def test_rms_norm_different_hidden_sizes(default_vllm_config, hidden_size: int):
|
||||
"""
|
||||
Test RMS norm with various hidden sizes to ensure block size handling.
|
||||
"""Test RMS norm with various hidden sizes to ensure block size handling.
|
||||
|
||||
The Triton kernel uses a fixed BLOCK_SIZE=1024, so this tests that it
|
||||
correctly handles hidden sizes both smaller and larger than the block size.
|
||||
@@ -474,8 +468,7 @@ def test_rms_norm_different_hidden_sizes(default_vllm_config, hidden_size: int):
|
||||
|
||||
@skip_unsupported
|
||||
def test_rms_norm_determinism(default_vllm_config):
|
||||
"""
|
||||
Test that batch-invariant RMS norm produces deterministic results.
|
||||
"""Test that batch-invariant RMS norm produces deterministic results.
|
||||
|
||||
Runs the same input through the kernel multiple times and verifies
|
||||
identical outputs.
|
||||
|
||||
Reference in New Issue
Block a user