[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
@@ -46,24 +46,21 @@ class QuantizeMethodBase(ABC):
|
||||
):
|
||||
"""Create weights for a layer.
|
||||
|
||||
The weights will be set as attributes of the layer.
|
||||
"""
|
||||
The weights will be set as attributes of the layer."""
|
||||
raise NotImplementedError
|
||||
|
||||
@abstractmethod
|
||||
def apply(self, layer: torch.nn.Module, *args, **kwargs) -> torch.Tensor:
|
||||
"""Apply the weights in layer to the input tensor.
|
||||
|
||||
Expects create_weights to have been called before on the layer.
|
||||
"""
|
||||
Expects create_weights to have been called before on the layer."""
|
||||
raise NotImplementedError
|
||||
|
||||
# Not required functions
|
||||
def embedding(self, layer: torch.nn.Module, *args, **kwargs) -> torch.Tensor:
|
||||
"""Gather embeddings in the layer based on indices in the input tensor.
|
||||
|
||||
Expects create_weights to have been called before on the layer.
|
||||
"""
|
||||
Expects create_weights to have been called before on the layer."""
|
||||
raise NotImplementedError
|
||||
|
||||
# Not required functions
|
||||
@@ -76,8 +73,7 @@ class QuantizeMethodBase(ABC):
|
||||
Quantization methods that need special weight handling (e.g. repacked
|
||||
weights) override this.
|
||||
|
||||
Expects create_weights to have been called before on the layer.
|
||||
"""
|
||||
Expects create_weights to have been called before on the layer."""
|
||||
layer.weight = embed_tokens.weight
|
||||
return layer
|
||||
|
||||
|
||||
@@ -665,8 +665,7 @@ class CompressedTensorsConfig(QuantizationConfig):
|
||||
) -> bool:
|
||||
"""Weight N-bit INT (pack-quantized for sub-byte, int-quantized for 8-bit)
|
||||
with static per-tensor INT8 input/output activation quant, applied as a float
|
||||
fake-quant around a weight-only matmul.
|
||||
"""
|
||||
fake-quant around a weight-only matmul."""
|
||||
is_int_pack_format = format in (
|
||||
CompressionFormat.pack_quantized.value,
|
||||
CompressionFormat.int_quantized.value,
|
||||
@@ -702,8 +701,7 @@ class CompressedTensorsConfig(QuantizationConfig):
|
||||
format: str | None,
|
||||
) -> bool:
|
||||
"""Weight N-bit INT with symmetric dynamic INT activation quant
|
||||
via Humming kernel.
|
||||
"""
|
||||
via Humming kernel."""
|
||||
if input_quant is None:
|
||||
return False
|
||||
is_pack_format = format == CompressionFormat.pack_quantized.value
|
||||
|
||||
+1
-2
@@ -37,8 +37,7 @@ def _dequant_gather_kernel(
|
||||
BLOCK: tl.constexpr,
|
||||
):
|
||||
"""Gather embedding rows by token id, unpack int32-packed INT weights, and
|
||||
dequantize to ``out`` dtype in one pass (no int8 intermediate).
|
||||
"""
|
||||
dequantize to ``out`` dtype in one pass (no int8 intermediate)."""
|
||||
row = tl.program_id(0)
|
||||
col = tl.program_id(1) * BLOCK + tl.arange(0, BLOCK)
|
||||
col_mask = col < hidden
|
||||
|
||||
+1
-2
@@ -211,8 +211,7 @@ class CompressedTensorsWNA8O8Int(CompressedTensorsScheme):
|
||||
|
||||
def _pack_int_quantized_weight(self, layer: torch.nn.Module) -> None:
|
||||
"""Normalize an int-quantized (plain int8) weight to the canonical
|
||||
``weight_packed`` int32 + ``weight_shape`` layout the MP kernels expect.
|
||||
"""
|
||||
``weight_packed`` int32 + ``weight_shape`` layout the MP kernels expect."""
|
||||
weight = layer.weight
|
||||
out_features, in_features = weight.shape
|
||||
packed = pack_to_int32(weight.data.contiguous(), self.num_bits)
|
||||
|
||||
@@ -86,8 +86,7 @@ class INCConfigParser:
|
||||
|
||||
def is_explicitly_configured(name: str) -> bool:
|
||||
"""Return True if *name* has an explicit entry in extra_config,
|
||||
either via exact key match or via a regex pattern key.
|
||||
"""
|
||||
either via exact key match or via a regex pattern key."""
|
||||
if not self._config.extra_config:
|
||||
return False
|
||||
if name in self._config.extra_config:
|
||||
|
||||
@@ -48,8 +48,7 @@ class INCScheme(ABC):
|
||||
layer_config: "INCLayerConfig",
|
||||
) -> "FusedMoEMethodBase | None":
|
||||
"""Optional. Override if this scheme supports MoE.
|
||||
Default raises NotImplementedError.
|
||||
"""
|
||||
Default raises NotImplementedError."""
|
||||
raise NotImplementedError(
|
||||
f"{type(self).__name__} does not support MoE layers. "
|
||||
f"Layer config: {layer_config}"
|
||||
|
||||
@@ -1988,8 +1988,7 @@ class KNvfp4Static(QuantKeyScheme):
|
||||
|
||||
class KNvfp4Dynamic(QuantKeyScheme):
|
||||
"""NVFP4 activation scheme (W4A4). Has a static global input scale on disk;
|
||||
the per-group scale is computed at runtime inside the kernel.
|
||||
"""
|
||||
the per-group scale is computed at runtime inside the kernel."""
|
||||
|
||||
key = kNvfp4Dynamic
|
||||
|
||||
@@ -2026,8 +2025,7 @@ class KNvfp4Dynamic(QuantKeyScheme):
|
||||
|
||||
class KFp8StaticTensor(QuantKeyScheme):
|
||||
"""Plain per-tensor static FP8 — bivalent: serves BOTH the weight slot and
|
||||
the activation slot (W8A8). One key in both QuantSpec slots.
|
||||
"""
|
||||
the activation slot (W8A8). One key in both QuantSpec slots."""
|
||||
|
||||
key = kFp8StaticTensorSym
|
||||
|
||||
@@ -2097,8 +2095,7 @@ class KFp8StaticTensor(QuantKeyScheme):
|
||||
|
||||
class KFp8StaticChannel(QuantKeyScheme):
|
||||
"""Per-channel static FP8 weight (the 'PcPt' weight). Weight-role only —
|
||||
there is no static per-channel *activation* today.
|
||||
"""
|
||||
there is no static per-channel *activation* today."""
|
||||
|
||||
key = kFp8StaticTokenSym
|
||||
|
||||
@@ -2139,8 +2136,7 @@ class KFp8StaticChannel(QuantKeyScheme):
|
||||
class KFp8Block128(QuantKeyScheme):
|
||||
"""128x128 block-static FP8 weight ('PbWo'). Weight-role only. ModelOpt
|
||||
exports the scale 4-D [out_blk,1,in_blk,1]; process squeezes to 2-D.
|
||||
No transpose (block kernel keeps [out,in]).
|
||||
"""
|
||||
No transpose (block kernel keeps [out,in])."""
|
||||
|
||||
key = kFp8Static128BlockSym
|
||||
|
||||
@@ -2196,8 +2192,7 @@ class KFp8Block128(QuantKeyScheme):
|
||||
|
||||
class KMxfp8Static(QuantKeyScheme):
|
||||
"""MXFP8 weight: fp8-e4m3 values + per-32-block e8m0 (uint8) scale.
|
||||
Weight-role only. process is validate-only plus an idempotency guard.
|
||||
"""
|
||||
Weight-role only. process is validate-only plus an idempotency guard."""
|
||||
|
||||
key = kMxfp8Static
|
||||
|
||||
@@ -2272,8 +2267,7 @@ class KDynamicNoParam(QuantKeyScheme):
|
||||
"""Dynamic activation with no stored scale (W8A8): quantized at runtime in
|
||||
the kernel. NOT the same as activation=None (weight-only) — init_fp8 needs a
|
||||
non-None activation key. Activation-role only. Serves the fp8 per-token, fp8
|
||||
per-block, and mxfp8 dynamic activation keys.
|
||||
"""
|
||||
per-block, and mxfp8 dynamic activation keys."""
|
||||
|
||||
def create_weights(self, layer, role, ctx, shapes, wl) -> None:
|
||||
if role is not ACT:
|
||||
|
||||
@@ -79,8 +79,7 @@ class Mxfp4Config(QuantizationConfig):
|
||||
|
||||
def _make_moe_method(self, moe: FusedMoEConfig) -> FusedMoEMethodBase:
|
||||
"""MoE method for RoutedExperts. Subclasses override to pick a
|
||||
checkpoint-specific kernel family.
|
||||
"""
|
||||
checkpoint-specific kernel family."""
|
||||
return Mxfp4MoEMethod(moe)
|
||||
|
||||
def get_quant_method(
|
||||
|
||||
@@ -139,8 +139,7 @@ def _find_matching_targets(
|
||||
|
||||
class OnlineQuantizationConfig(QuantizationConfig):
|
||||
"""Model-level config for online quantization (quantize fp16/bf16 weights
|
||||
during model loading, without requiring a pre-quantized checkpoint).
|
||||
"""
|
||||
during model loading, without requiring a pre-quantized checkpoint)."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
|
||||
@@ -118,8 +118,7 @@ def _is_tp_sharded(layer: Module, *, reduces_output_dim: bool = True) -> bool:
|
||||
|
||||
class OnlineLinearBase(LinearMethodBase):
|
||||
"""Shared base for online FP8 linear methods. Loads fp16/bf16 checkpoint
|
||||
weights onto meta device and materializes them just-in-time.
|
||||
"""
|
||||
weights onto meta device and materializes them just-in-time."""
|
||||
|
||||
uses_meta_device: bool = True
|
||||
|
||||
@@ -163,8 +162,7 @@ class OnlineLinearBase(LinearMethodBase):
|
||||
|
||||
class Fp8PerTensorOnlineLinearMethod(OnlineLinearBase):
|
||||
"""Online tensorwise FP8 linear quantization.
|
||||
Loads fp16/bf16 weights and quantizes them per-tensor during loading.
|
||||
"""
|
||||
Loads fp16/bf16 weights and quantizes them per-tensor during loading."""
|
||||
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
@@ -265,8 +263,7 @@ class Fp8PerTensorOnlineLinearMethod(OnlineLinearBase):
|
||||
|
||||
class Fp8PerBlockOnlineLinearMethod(OnlineLinearBase):
|
||||
"""Online blockwise FP8 linear quantization.
|
||||
Loads fp16/bf16 weights and quantizes them per-block during loading.
|
||||
"""
|
||||
Loads fp16/bf16 weights and quantizes them per-block during loading."""
|
||||
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
@@ -440,8 +437,7 @@ class Fp8PtpcOnlineLinearMethod(OnlineLinearBase):
|
||||
|
||||
class _Fp8OnlineMoEBase(OnlineMoEMethodBase):
|
||||
"""Shared base for online FP8 MoE methods. Loads fp16/bf16 checkpoint
|
||||
weights onto meta device and materializes them just-in-time.
|
||||
"""
|
||||
weights onto meta device and materializes them just-in-time."""
|
||||
|
||||
# Declared here for mypy; actual values are set in __init__.
|
||||
fp8_backend: "Fp8MoeBackend"
|
||||
@@ -562,8 +558,7 @@ class _Fp8OnlineMoEBase(OnlineMoEMethodBase):
|
||||
|
||||
class Fp8PerTensorOnlineMoEMethod(_Fp8OnlineMoEBase):
|
||||
"""Online tensorwise FP8 MoE quantization.
|
||||
Loads fp16/bf16 weights and quantizes them per-tensor during loading.
|
||||
"""
|
||||
Loads fp16/bf16 weights and quantizes them per-tensor during loading."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
@@ -620,8 +615,7 @@ class Fp8PerTensorOnlineMoEMethod(_Fp8OnlineMoEBase):
|
||||
|
||||
class Fp8PerBlockOnlineMoEMethod(_Fp8OnlineMoEBase):
|
||||
"""Online blockwise FP8 MoE quantization.
|
||||
Loads fp16/bf16 weights and quantizes them per-block during loading.
|
||||
"""
|
||||
Loads fp16/bf16 weights and quantizes them per-block during loading."""
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
|
||||
@@ -69,8 +69,7 @@ def input_to_float8(
|
||||
x: torch.Tensor, dtype: torch.dtype | None = None
|
||||
) -> tuple[torch.Tensor, torch.Tensor]:
|
||||
"""This function quantizes input values to float8 values "
|
||||
"with tensor-wise quantization.
|
||||
"""
|
||||
"with tensor-wise quantization."""
|
||||
dtype = current_platform.fp8_dtype() if dtype is None else dtype
|
||||
finfo = torch.finfo(dtype)
|
||||
min_val, max_val = x.aminmax()
|
||||
@@ -1195,8 +1194,7 @@ def prepare_fp8_moe_layer_for_deepgemm(
|
||||
|
||||
def _maybe_pad_fp8_weight(weight: torch.Tensor) -> torch.Tensor:
|
||||
"""Pad the weight tensor. This is an optimization on ROCm platform, which
|
||||
can benefit from tensors located far enough from one another in memory
|
||||
"""
|
||||
can benefit from tensors located far enough from one another in memory"""
|
||||
if (
|
||||
envs.VLLM_ROCM_FP8_PADDING
|
||||
and current_platform.is_rocm()
|
||||
|
||||
@@ -297,8 +297,7 @@ def _humming_input_schema_to_quant_key(
|
||||
schema: "HummingInputSchema",
|
||||
) -> QuantKey | None:
|
||||
"""Convert a HummingInputSchema to a QuantKey. Returns None if
|
||||
the schema represents unquantized (bf16/fp16) inputs.
|
||||
"""
|
||||
the schema represents unquantized (bf16/fp16) inputs."""
|
||||
if schema.a_dtype is None or schema.a_dtype.num_bits >= 16:
|
||||
return None
|
||||
|
||||
@@ -427,8 +426,7 @@ def convert_linear_layer_to_humming_standard(
|
||||
):
|
||||
"""Rename/reshape a linear layer's quantized params (the canonical MPLinear
|
||||
layout: ``weight_packed`` int32 + ``weight_scale``) into the parameter names
|
||||
and layout humming's weight schema expects (``weight`` / ``weight_scale``).
|
||||
"""
|
||||
and layout humming's weight schema expects (``weight`` / ``weight_scale``)."""
|
||||
for name, checkpoint_name in name_map.items():
|
||||
tensor = getattr(layer, checkpoint_name)
|
||||
delattr(layer, checkpoint_name)
|
||||
|
||||
@@ -245,8 +245,7 @@ def marlin_padded_nk(size_n: int, size_k: int, group_size: int = -1) -> tuple[in
|
||||
|
||||
def marlin_repacked_nk(qweight: torch.Tensor, num_bits: int) -> tuple[int, int]:
|
||||
"""Recover the (size_n, size_k) a Marlin weight was repacked with
|
||||
(including any tile padding) from its packed shape.
|
||||
"""
|
||||
(including any tile padding) from its packed shape."""
|
||||
pack_factor = 32 // num_bits
|
||||
size_k = qweight.size(0) * GPTQ_MARLIN_TILE
|
||||
size_n = qweight.size(1) * pack_factor // GPTQ_MARLIN_TILE
|
||||
@@ -257,8 +256,7 @@ def marlin_pad_qweight(
|
||||
qweight: torch.Tensor, size_n: int, size_k: int, padded_n: int, padded_k: int
|
||||
) -> torch.Tensor:
|
||||
"""Zero-pad a GPTQ-layout packed weight (size_k / pack, size_n) for
|
||||
gptq_marlin_repack.
|
||||
"""
|
||||
gptq_marlin_repack."""
|
||||
if (padded_n, padded_k) == (size_n, size_k):
|
||||
return qweight
|
||||
pack_factor = size_k // qweight.size(0)
|
||||
@@ -276,8 +274,7 @@ def marlin_pad_scales(
|
||||
group_size: int,
|
||||
) -> torch.Tensor:
|
||||
"""Zero-pad weight scales (num_groups, size_n); call before
|
||||
marlin_permute_scales and pass the padded extents to it.
|
||||
"""
|
||||
marlin_permute_scales and pass the padded extents to it."""
|
||||
if (padded_n, padded_k) == (size_n, size_k):
|
||||
return scales
|
||||
pad_rows = padded_k // group_size - scales.size(0) if group_size > 0 else 0
|
||||
|
||||
@@ -41,8 +41,7 @@ def _nvfp4_compute_scale_factor(
|
||||
) -> float:
|
||||
"""Compute the power-of-2 scale_factor needed so that all non-zero
|
||||
values in marlin_scales * 2^7 are >= 2 after rescaling.
|
||||
Returns a Python float (power of 2, >= 1.0).
|
||||
"""
|
||||
Returns a Python float (power of 2, >= 1.0)."""
|
||||
# Since half has a smaller dynamic range compared to bfloat16,
|
||||
# no rescaling is applied here if active dtype is half.
|
||||
if a_dtype is not None and a_dtype == torch.half:
|
||||
@@ -376,8 +375,7 @@ def prepare_nvfp4_moe_layer_for_marlin(
|
||||
|
||||
def pad_w13(x: torch.Tensor) -> torch.Tensor:
|
||||
"""Zero-pad each gate/up shard of a (E, num_shards * N, cols)
|
||||
tensor to padded_N rows.
|
||||
"""
|
||||
tensor to padded_N rows."""
|
||||
if padded_N == N:
|
||||
return x
|
||||
x = x.view(E, num_shards, N, x.size(-1))
|
||||
@@ -386,8 +384,7 @@ def prepare_nvfp4_moe_layer_for_marlin(
|
||||
|
||||
def pad_w2(x: torch.Tensor, packing: int) -> torch.Tensor:
|
||||
"""Zero-pad the packed N (last) dim of a (E, K, N / packing)
|
||||
tensor.
|
||||
"""
|
||||
tensor."""
|
||||
if padded_N == N:
|
||||
return x
|
||||
return torch.nn.functional.pad(x, (0, (padded_N - N) // packing))
|
||||
|
||||
@@ -217,8 +217,7 @@ def prepare_fp8_layer_for_marlin(
|
||||
|
||||
def _moe_pad_shard_rows(x: torch.Tensor, n: int, padded_n: int) -> torch.Tensor:
|
||||
"""Zero-pad each gate/up shard of a ``(E, 2 * n, ...)`` tensor to padded_n
|
||||
rows. FP8 zero decodes to 0.0, so the padded rows contribute nothing.
|
||||
"""
|
||||
rows. FP8 zero decodes to 0.0, so the padded rows contribute nothing."""
|
||||
if padded_n == n:
|
||||
return x
|
||||
e = x.size(0)
|
||||
|
||||
Reference in New Issue
Block a user