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
@@ -382,7 +382,8 @@ def test_inc_get_quant_method_unquantized_layer_with_model_prefix(
|
||||
layer_factory,
|
||||
) -> None:
|
||||
"""``model.``-prefixed extra_config entries trigger early exit for both
|
||||
LinearBase and ParallelLMHead layers."""
|
||||
LinearBase and ParallelLMHead layers.
|
||||
"""
|
||||
config = make_config(extra_config={"model.layer": {"bits": 16}})
|
||||
layer = layer_factory()
|
||||
|
||||
@@ -395,7 +396,8 @@ def test_inc_get_quant_method_unquantized_routed_experts_with_model_prefix(
|
||||
monkeypatch,
|
||||
) -> None:
|
||||
"""``model.``-prefixed extra_config entries trigger early exit for
|
||||
RoutedExperts layers, returning UnquantizedFusedMoEMethod."""
|
||||
RoutedExperts layers, returning UnquantizedFusedMoEMethod.
|
||||
"""
|
||||
|
||||
class DummyUnquantizedFusedMoEMethod:
|
||||
def __init__(self, moe_config) -> None:
|
||||
@@ -418,7 +420,8 @@ def test_inc_get_quant_method_unquantized_routed_experts_with_model_prefix(
|
||||
|
||||
def test_inc_get_quant_method_unknown_layer_with_model_prefix_returns_none() -> None:
|
||||
"""``model.``-prefixed extra_config entries return None for unhandled
|
||||
layer types (non-Linear, non-MoE)."""
|
||||
layer types (non-Linear, non-MoE).
|
||||
"""
|
||||
config = make_config(extra_config={"model.unknown": {"bits": 16}})
|
||||
|
||||
method = config.get_quant_method(DummyLayer(), "unknown")
|
||||
@@ -733,7 +736,8 @@ def test_wna16_cuda_low_bit_moe_routes_to_humming(monkeypatch, bits) -> None:
|
||||
@pytest.mark.parametrize("bits", [4, 8])
|
||||
def test_wna16_cuda_high_bit_skips_humming(monkeypatch, bits) -> None:
|
||||
"""4/8-bit int stays on the Marlin/GPTQ/AWQ path even on CUDA so a single
|
||||
model can mix high-bit Marlin and low-bit humming layers."""
|
||||
model can mix high-bit Marlin and low-bit humming layers.
|
||||
"""
|
||||
called = {"humming": False}
|
||||
|
||||
monkeypatch.setattr(current_platform, "is_cuda", lambda: True)
|
||||
@@ -1664,7 +1668,8 @@ def test_inc_get_quant_method_unquantized_moe_returns_unquantized(
|
||||
monkeypatch,
|
||||
) -> None:
|
||||
"""Early-exit returns UnquantizedFusedMoEMethod for FusedMoE layers
|
||||
when extra_config has bits >= 16."""
|
||||
when extra_config has bits >= 16.
|
||||
"""
|
||||
config = make_config(extra_config={"layer": {"bits": 16}})
|
||||
layer = object.__new__(RoutedExperts)
|
||||
layer.moe_config = None # UnquantizedFusedMoEMethod accepts moe_config
|
||||
|
||||
@@ -574,8 +574,7 @@ def test_compressed_tensors_fp8_block_enabled(
|
||||
reason="This test is not for non-CUDA platforms",
|
||||
)
|
||||
def test_compressed_tensors_moe_ignore_with_model(vllm_runner):
|
||||
"""
|
||||
Integration test for MoE layer ignore functionality with a real model.
|
||||
"""Integration test for MoE layer ignore functionality with a real model.
|
||||
|
||||
This test would verify that when loading a compressed-tensors quantized
|
||||
MoE model where some MoE layers are in the ignore list, those layers
|
||||
@@ -587,7 +586,6 @@ def test_compressed_tensors_moe_ignore_with_model(vllm_runner):
|
||||
- Config with ignore list containing specific MoE layers
|
||||
- Multiple MoE layers where some are quantized and some are not
|
||||
"""
|
||||
|
||||
# model_path = "nm-testing/tinysmokeqwen3moe-W4A16-first-only" # CT 12.3
|
||||
model_path = "nm-testing/tinysmokeqwen3moe-W4A16-first-only-CTstable" # CT 12.2
|
||||
|
||||
|
||||
@@ -213,7 +213,8 @@ def test_deepseek_v41_vl_mapper_routes_linear_scales(
|
||||
):
|
||||
"""The VL wrapper must map ``.scale`` keys to the parameter the linear
|
||||
quant method registers, as the text model does. A hardcoded
|
||||
``weight_scale_inv`` raised KeyError for native MXFP8 checkpoints."""
|
||||
``weight_scale_inv`` raised KeyError for native MXFP8 checkpoints.
|
||||
"""
|
||||
from vllm.models.deepseek_v41.nvidia import model as model_module
|
||||
from vllm.models.deepseek_v41.nvidia import vl_model as vl_module
|
||||
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
Tests for MXFP4 MoE oracle backend selection on mi355x (GFX950).
|
||||
"""Tests for MXFP4 MoE oracle backend selection on mi355x (GFX950).
|
||||
|
||||
These tests run on real hardware — no mocks. Skipped on non-GFX950 platforms.
|
||||
"""
|
||||
@@ -161,7 +160,8 @@ def test_qwen38_tep8_auto_fallback_respects_explicit_backend(
|
||||
@pytest.fixture
|
||||
def mxfp4_oracle_config():
|
||||
"""Stub the config the oracle reads (``model_config.quantization_config``)
|
||||
so backend dispatch resolves without a real model / user override."""
|
||||
so backend dispatch resolves without a real model / user override.
|
||||
"""
|
||||
from unittest.mock import patch
|
||||
|
||||
with patch(
|
||||
|
||||
@@ -85,7 +85,8 @@ def test_block_fp8_activation_quant_config_is_block_quantized():
|
||||
(FP8 dtype, [1, 128] block, not per-act-token). This is what makes the
|
||||
DeepEP prepare/finalize step quantize activations to block FP8 *before* the
|
||||
all-to-all dispatch instead of deferring to Humming -- and what makes
|
||||
HummingExpertsBase.expects_unquantized_inputs return False."""
|
||||
HummingExpertsBase.expects_unquantized_inputs return False.
|
||||
"""
|
||||
from vllm.model_executor.layers.quantization.utils.humming_utils import (
|
||||
make_humming_moe_quant_config,
|
||||
)
|
||||
@@ -107,7 +108,8 @@ def test_block_fp8_activation_quant_config_is_block_quantized():
|
||||
def test_default_activation_quant_config_defers_to_humming():
|
||||
"""Without a block activation shape the config stays per-token (the deferred
|
||||
path): Humming quantizes internally, preserving pre-existing behavior for
|
||||
every non-block-FP8 scheme."""
|
||||
every non-block-FP8 scheme.
|
||||
"""
|
||||
from vllm.model_executor.layers.quantization.utils.humming_utils import (
|
||||
make_humming_moe_quant_config,
|
||||
)
|
||||
|
||||
@@ -1,7 +1,6 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""
|
||||
Tests for INT8 (W8A8) fused-MoE oracle backend selection.
|
||||
"""Tests for INT8 (W8A8) fused-MoE oracle backend selection.
|
||||
|
||||
These exercise ``select_int8_moe_backend`` only (no kernels are launched), so
|
||||
they run on any platform where the Triton INT8 MoE kernel is available — CUDA
|
||||
@@ -71,7 +70,8 @@ def _make_int8_moe_config(moe_backend: str = "auto") -> FusedMoEConfig:
|
||||
)
|
||||
def test_int8_dynamic_schemes_dispatch_to_triton(weight_key, activation_key):
|
||||
"""Both dynamic-activation INT8 MoE schemes (per-channel + per-tensor
|
||||
weights) select the Triton backend."""
|
||||
weights) select the Triton backend.
|
||||
"""
|
||||
config = _make_int8_moe_config()
|
||||
backend, experts_cls = select_int8_moe_backend(
|
||||
config, weight_key=weight_key, activation_key=activation_key
|
||||
|
||||
@@ -587,7 +587,8 @@ def test_modelopt_fp8_pb_wo_checkpoint_setup(monkeypatch, dist_init, workspace_i
|
||||
def test_modelopt_nvfp4_config_dispatches_w4a4_method():
|
||||
"""``quant_method="NVFP4"`` (W4A4) resolves to a
|
||||
``(kNvfp4Static, kNvfp4Dynamic)`` QuantSpec under the generic
|
||||
``ModelOptLinearMethod``."""
|
||||
``ModelOptLinearMethod``.
|
||||
"""
|
||||
from vllm.model_executor.layers.linear import LinearBase
|
||||
|
||||
config = ModelOptNvFp4Config(
|
||||
@@ -636,7 +637,8 @@ def test_modelopt_linear_method_builder_registry_override(monkeypatch):
|
||||
"""The bespoke-method escape hatch: a format registered in
|
||||
``LINEAR_METHOD_BUILDERS`` routes that algo to its own method instead of the
|
||||
generic ``ModelOptLinearMethod``. This is how a format that cannot be a
|
||||
``(weight, activation)`` key pair plugs into dispatch."""
|
||||
``(weight, activation)`` key pair plugs into dispatch.
|
||||
"""
|
||||
from vllm.model_executor.layers.linear import LinearBase
|
||||
from vllm.model_executor.layers.quantization import modelopt as m
|
||||
|
||||
@@ -667,7 +669,8 @@ def test_modelopt_linear_method_builder_registry_override(monkeypatch):
|
||||
def test_modelopt_w4a16_respects_linear_backend(linear_backend, kernel_cls):
|
||||
"""W4A16 (`activation=None`) kernel selection honors ``--linear-backend``:
|
||||
``use_a16=True`` defaults to Marlin, but an explicit backend wins. The
|
||||
generic method routes this through ``select_linear_kernel``."""
|
||||
generic method routes this through ``select_linear_kernel``.
|
||||
"""
|
||||
from vllm.config.quantization import QuantSpec
|
||||
from vllm.model_executor.layers.quantization.modelopt import (
|
||||
RuntimeDtypes,
|
||||
@@ -938,7 +941,8 @@ def test_modelopt_fp8_pb_wo_hides_output_padding(monkeypatch):
|
||||
def test_modelopt_fp8_pb_wo_rejects_non_128_input():
|
||||
"""Input width must still be a multiple of 128 (same as #53132, which only
|
||||
pads the output). A partial input block is refused loudly rather than
|
||||
silently loading wrong scales."""
|
||||
silently loading wrong scales.
|
||||
"""
|
||||
from vllm.model_executor.layers.quantization import modelopt as mo
|
||||
from vllm.model_executor.layers.quantization.utils.quant_utils import (
|
||||
kFp8Static128BlockSym,
|
||||
|
||||
@@ -449,7 +449,6 @@ def test_online_shorthand_selects_checkpoint_or_online_config(
|
||||
tmp_path, quantization, checkpoint_quantization_config, is_config_file
|
||||
) -> None:
|
||||
"""Online shorthands defer to checkpoint metadata when available."""
|
||||
|
||||
if is_config_file:
|
||||
(tmp_path / "hf_quant_config.json").write_text(
|
||||
'{"quant_method": "mxfp8", "ignored_layers": []}'
|
||||
@@ -609,14 +608,12 @@ def test_online_quantization(
|
||||
dist_init,
|
||||
workspace_init,
|
||||
) -> None:
|
||||
"""
|
||||
Tests that online quantization frontend configuration works -
|
||||
"""Tests that online quantization frontend configuration works -
|
||||
selecting quant schemes, overriding quant schemes by type, ignoring
|
||||
layers.
|
||||
|
||||
Does not test performance, peak memory usage, etc.
|
||||
"""
|
||||
|
||||
# TODO: Relax this condition once there is a native MXFP4_MXFP4
|
||||
# linear/moe backend supported on cuda.
|
||||
if quant_scheme == "mxfp4" and not (on_gfx950() or on_gfx942()):
|
||||
|
||||
@@ -135,8 +135,7 @@ def _skip_reason_if_unavailable(backend: str, dtype: torch.dtype) -> str | None:
|
||||
@pytest.mark.parametrize("dtype", [torch.bfloat16, torch.float16])
|
||||
@pytest.mark.parametrize("backend", ["triton", "aiter", "xpu", "quark"])
|
||||
def test_mxfp4_quantization_correctness(backend: str, dtype: torch.dtype):
|
||||
"""
|
||||
Tests that the different implementations of mxfp4_quantize
|
||||
"""Tests that the different implementations of mxfp4_quantize
|
||||
in mxfp4_utils.py all match.
|
||||
"""
|
||||
skip_reason = _skip_reason_if_unavailable(backend, dtype)
|
||||
@@ -290,8 +289,7 @@ def test_online_mxfp4_moe_matches_quark(
|
||||
dist_init,
|
||||
monkeypatch,
|
||||
):
|
||||
"""
|
||||
Ensures `Mxfp4OnlineMoEMethod` (online quantization)
|
||||
"""Ensures `Mxfp4OnlineMoEMethod` (online quantization)
|
||||
and `QuarkOCP_MX_MoEMethod` (AMD Quark checkpoints) produce the same weights,
|
||||
with same MOE backend used.
|
||||
|
||||
@@ -413,7 +411,8 @@ def test_online_mxfp4_moe_matches_quark(
|
||||
) -> tuple[torch.Tensor, torch.Tensor]:
|
||||
"""Place the source weights into buffers of the padded size, the
|
||||
way the weight loader writes only the unpadded slice of a larger
|
||||
allocation."""
|
||||
allocation.
|
||||
"""
|
||||
w13 = torch.full(
|
||||
(num_experts, 2 * intermediate_size, hidden_size),
|
||||
padding_value,
|
||||
@@ -495,8 +494,7 @@ def test_online_mxfp4_moe_matches_quark(
|
||||
def test_online_mxfp4_dense_matches_quark(
|
||||
linear_backend: str, default_vllm_config, dist_init, monkeypatch
|
||||
):
|
||||
"""
|
||||
Ensures `Mxfp4OnlineLinearMethod` (online quantization)
|
||||
"""Ensures `Mxfp4OnlineLinearMethod` (online quantization)
|
||||
and `QuarkOCP_MX` (AMD Quark checkpoints) produce the same weights,
|
||||
with same linear backend used.
|
||||
"""
|
||||
|
||||
@@ -490,7 +490,8 @@ def test_per_token_head_negative_slot_skipped(qcfg: QuantConfig):
|
||||
)
|
||||
def test_process_weights_sets_placeholder_scales(kv_cache_dtype: str):
|
||||
"""Per-token-head should set _k_scale=1.0, _v_scale=1.0
|
||||
and delete checkpoint attrs."""
|
||||
and delete checkpoint attrs.
|
||||
"""
|
||||
from vllm.model_executor.layers.quantization.kv_cache import (
|
||||
BaseKVCacheMethod,
|
||||
)
|
||||
|
||||
@@ -48,7 +48,6 @@ class FakeQuantLinearMethod(UnquantizedLinearMethod):
|
||||
bias: torch.Tensor | None = None,
|
||||
) -> torch.Tensor:
|
||||
"""Perform fake quantization before the linear layer."""
|
||||
|
||||
# Calculate the scales dynamically
|
||||
max_val = torch.amax(x, dim=(0, -1), keepdims=True)
|
||||
min_val = torch.amin(x, dim=(0, -1), keepdims=True)
|
||||
@@ -108,7 +107,6 @@ class CustomQuantConfig(QuantizationConfig):
|
||||
|
||||
def test_register_quantization_config(caplog_vllm):
|
||||
"""Test register custom quantization config."""
|
||||
|
||||
# The quantization method `custom_quant` should be registered.
|
||||
assert get_quantization_config("custom_quant") == CustomQuantConfig
|
||||
|
||||
|
||||
Reference in New Issue
Block a user