[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:
co-authored by
Claude Opus 5
parent
1085b64425
commit
651a88c09a
@@ -36,8 +36,7 @@ def is_aiter_mxfp8_moe_available() -> bool:
|
||||
nightly lacking #3811 would wrongly select FlyDSL instead of falling back to
|
||||
the native Triton dot_scaled path. #3811 added no probe-able public symbol,
|
||||
so detect the ``minimax_m3_mxfp8`` tuned config it shipped. Every check fails
|
||||
closed (returns False -> triton dot_scaled), which is always safe.
|
||||
"""
|
||||
closed (returns False -> triton dot_scaled), which is always safe."""
|
||||
if not (current_platform.is_rocm() and current_platform.supports_mx()):
|
||||
return False
|
||||
try:
|
||||
|
||||
@@ -192,8 +192,7 @@ class CPUUnquantizedExperts(mk.FusedMoEExpertsModular):
|
||||
|
||||
def _pad_moe_intermediate(self, layer: torch.nn.Module) -> None:
|
||||
"""Zero-pad the per-partition MoE intermediate dim of both weights and
|
||||
the expert bias, see `_padded_intermediate_size`.
|
||||
"""
|
||||
the expert bias, see `_padded_intermediate_size`."""
|
||||
intermediate_size = self.moe_config.intermediate_size_per_partition
|
||||
padded_size = self._padded_intermediate_size(self.moe_config)
|
||||
if padded_size == intermediate_size:
|
||||
@@ -953,8 +952,7 @@ class CPUExpertsInt4(mk.FusedMoEExpertsModular):
|
||||
|
||||
class CPUExpertsInt8(mk.FusedMoEExpertsModular):
|
||||
"""CPU INT8 W8A8 per-channel weight / dynamic per-token activation
|
||||
modular MoE experts.
|
||||
"""
|
||||
modular MoE experts."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
@@ -1287,8 +1285,7 @@ class ArmCPUExpertsInt8(mk.FusedMoEExpertsModular):
|
||||
|
||||
class ZenCPUExpertsInt8(mk.FusedMoEExpertsModular):
|
||||
"""AMD Zen INT8 MoE with per-token activation and channelwise weight
|
||||
quantization, dispatched through zentorch.
|
||||
"""
|
||||
quantization, dispatched through zentorch."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
|
||||
@@ -34,8 +34,7 @@ from vllm.triton_utils import tl, triton
|
||||
|
||||
def _select_cfg(M, N, K, block_m):
|
||||
"""Pick the launch config from host constants only (M=num_valid_tokens, N, K,
|
||||
block_m) — graph-capture safe (no GPU-scalar branch).
|
||||
"""
|
||||
block_m) — graph-capture safe (no GPU-scalar branch)."""
|
||||
# Per-regime winners (measured, isolated cuda-event A/B on gfx950, GPU 3):
|
||||
# BLOCK_K=256 (fewer K-iters + bigger MX scale-load coalesced with the dot),
|
||||
# num_stages=2 (software-pipeline overlaps the E8M0 scale-load with the scaled
|
||||
|
||||
Reference in New Issue
Block a user