[CI] Ignore ruff D209, rejoin the docstrings it split, and silence incompatible-rule warnings (#57212)

Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
This commit is contained in:
Harry Mellor
2026-09-16 10:57:24 -07:00
committed by GitHub
co-authored by Claude Opus 5
parent 1085b64425
commit 651a88c09a
693 changed files with 2088 additions and 4171 deletions
+2 -4
View File
@@ -41,8 +41,7 @@ def _rmsnorm(x: torch.Tensor, w: torch.Tensor, eps: float) -> torch.Tensor:
def test_fused_embed_norm_matches_reference():
"""Main-model fusion: the residual is the exact gather and the second output
is a correct RMSNorm. The norm matches a full-precision reference to ~2 bf16
ulp (rtol 1e-2) -- that gap is bf16 rounding, not the kernel.
"""
ulp (rtol 1e-2) -- that gap is bf16 rounding, not the kernel."""
set_random_seed(13)
table = torch.randn(VOCAB, HIDDEN, dtype=DTYPE, device="cuda")
ids = torch.randint(0, VOCAB, (NUM_TOKENS,), dtype=torch.int32, device="cuda")
@@ -61,8 +60,7 @@ def test_fused_embed_norm_matches_reference():
@torch.inference_mode()
def test_fused_embed_eh_norm_matches_reference():
"""MTP fusion (folded gather) is bit-exact vs gathering the embeds and
feeding the untouched model-local ``fused_eh_norm``.
"""
feeding the untouched model-local ``fused_eh_norm``."""
set_random_seed(13)
table = torch.randn(VOCAB, HIDDEN, dtype=DTYPE, device="cuda")
ids = torch.randint(0, VOCAB, (NUM_TOKENS,), dtype=torch.int32, device="cuda")
@@ -62,8 +62,7 @@ def test_fused_q_kv_rmsnorm_outputs_are_packed(num_tokens: int):
a [1, q_size] slice of a [1, q_size + kv_size] buffer used to produce a
qr_out with row stride q_size + kv_size. Downstream dispatchers that
require packed row-major inputs then reject the tensor and silently fall
back to a slower GEMM path on every decode step.
"""
back to a slower GEMM path on every decode step."""
device = "cuda"
dtype = torch.bfloat16
q_size, kv_size = 192, 576
@@ -89,8 +88,7 @@ def test_fused_q_kv_rmsnorm_outputs_are_packed(num_tokens: int):
def test_fused_q_kv_rmsnorm_launches_past_grid_y_cap(num_tokens: int):
"""Regression guard: grid used to be (2, num_tokens), hitting CUDA's
65535 grid-y cap at num_tokens >= 65536. The new grid (num_tokens, 2)
lifts that bound to 2**31-1.
"""
lifts that bound to 2**31-1."""
device = "cuda"
dtype = torch.bfloat16
q_size, kv_size = 192, 576
@@ -2,8 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Tests that FusedRMSNormGated decomposes correctly under torch.compile,
matching the eager triton kernel output.
"""
matching the eager triton kernel output."""
import pytest
import torch
+1 -2
View File
@@ -579,8 +579,7 @@ def test_rocm_aiter_mla_dual_rmsnorm_group_quant_vs_sequential(m):
def test_rocm_aiter_mla_dual_rmsnorm_group_quant_strided_split_view():
"""Split-view inputs, as produced by the DeepSeek-V4 q/kv split, match
contiguous inputs.
"""
contiguous inputs."""
require_aiter()
require_fp8()
from vllm._aiter_ops import rocm_aiter_ops
@@ -87,8 +87,7 @@ def _shard(full_table, si, rows_per_rank, dtype):
@pytest.mark.parametrize("id_dtype", [torch.int32, torch.int64])
def test_matches_masked_gather(tp_size, vocab_size, org_vocab_size, hidden, id_dtype):
"""Every rank's partial output is bit-exact with the eager path, including
the added-vocab (LoRA) rows and the ids owned by another rank.
"""
the added-vocab (LoRA) rows and the ids owned by another rank."""
set_random_seed(7)
dtype = torch.bfloat16
rows = pad_vocab_size(vocab_size, 64) // tp_size