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
@@ -126,7 +126,6 @@ class QuarkConfig(QuantizationConfig):
|
||||
revision: str | None = None,
|
||||
):
|
||||
"""Enable dynamic MXFP4 only for DeepSeek-V3-family fp4 checkpoints."""
|
||||
|
||||
if hf_config is None:
|
||||
return
|
||||
|
||||
@@ -162,13 +161,13 @@ class QuarkConfig(QuantizationConfig):
|
||||
def apply_vllm_mapper( # noqa: B027
|
||||
self, hf_to_vllm_mapper: "WeightsMapper"
|
||||
):
|
||||
"""
|
||||
Interface for models to update module names referenced in
|
||||
"""Interface for models to update module names referenced in
|
||||
quantization configs in order to reflect the vllm model structure
|
||||
|
||||
Args:
|
||||
hf_to_vllm_mapper: maps from hf model structure (the assumed
|
||||
structure of the qconfig) to vllm model structure
|
||||
|
||||
"""
|
||||
quant_config_with_hf_to_vllm_mapper: dict[str, Any] = {}
|
||||
|
||||
@@ -657,8 +656,7 @@ class QuarkConfig(QuantizationConfig):
|
||||
input_quant: dict[str, Any] | None,
|
||||
allow_static_fp8: bool = False,
|
||||
) -> QuantKeyMatch:
|
||||
"""
|
||||
This check returns True only if it is an OCP-MX weight quantization.
|
||||
"""This check returns True only if it is an OCP-MX weight quantization.
|
||||
The activation can be FP16/BF16, OCP MXFP4, MXFP8, FP8.
|
||||
The rationale for checking only the weight type is that
|
||||
the model loading concept and process primarily concerns the weights themselves.
|
||||
@@ -984,8 +982,7 @@ class QuarkLinearMethod(LinearMethodBase):
|
||||
params_dtype: torch.dtype,
|
||||
**extra_weight_attrs,
|
||||
):
|
||||
"""
|
||||
Use the CompressedTensorsScheme associated with each layer to create
|
||||
"""Use the CompressedTensorsScheme associated with each layer to create
|
||||
the necessary parameters for the layer. See LinearMethodBase for param
|
||||
details
|
||||
"""
|
||||
@@ -1006,8 +1003,7 @@ class QuarkLinearMethod(LinearMethodBase):
|
||||
x: torch.Tensor,
|
||||
bias: torch.Tensor | None = None,
|
||||
):
|
||||
"""
|
||||
Use the output of create_weights and the CompressedTensorsScheme
|
||||
"""Use the output of create_weights and the CompressedTensorsScheme
|
||||
associated with the layer to apply the forward pass with the
|
||||
layer input. See LinearMethodBase for param details
|
||||
|
||||
@@ -1020,9 +1016,7 @@ class QuarkLinearMethod(LinearMethodBase):
|
||||
|
||||
|
||||
class QuarkKVCacheMethod(BaseKVCacheMethod):
|
||||
"""
|
||||
Supports loading kv-cache scaling factors from quark checkpoints.
|
||||
"""
|
||||
"""Supports loading kv-cache scaling factors from quark checkpoints."""
|
||||
|
||||
def __init__(self, quant_config: QuarkConfig):
|
||||
self.validate_kv_cache_config(quant_config.kv_cache_config)
|
||||
@@ -1030,12 +1024,12 @@ class QuarkKVCacheMethod(BaseKVCacheMethod):
|
||||
|
||||
@staticmethod
|
||||
def validate_kv_cache_config(kv_cache_config: dict[str, Any] | None):
|
||||
"""
|
||||
Validator for the kv cache configuration. Useful for controlling the
|
||||
"""Validator for the kv cache configuration. Useful for controlling the
|
||||
kv cache quantization schemes, that are being supported in vLLM
|
||||
|
||||
Args:
|
||||
kv_cache_config: the quark kv cache scheme
|
||||
|
||||
"""
|
||||
if kv_cache_config is None:
|
||||
return
|
||||
|
||||
@@ -1748,10 +1748,7 @@ class QuarkNvfp4MoEMethod(QuarkMoEMethod):
|
||||
set_weight_attrs(w2_input_scale_2, extra_weight_attrs)
|
||||
|
||||
def process_weights_after_loading(self, layer: RoutedExperts) -> None:
|
||||
"""
|
||||
Convert NVFP4 MoE weights into kernel format and setup the kernel.
|
||||
"""
|
||||
|
||||
"""Convert NVFP4 MoE weights into kernel format and setup the kernel."""
|
||||
# Match existing NVFP4 MoE paths: fused w13 uses the w1 global scale.
|
||||
if self.moe.is_act_and_mul and not torch.allclose(
|
||||
layer.w13_weight_scale_2[:, 0], layer.w13_weight_scale_2[:, 1]
|
||||
|
||||
@@ -31,8 +31,7 @@ logger = init_logger(__name__)
|
||||
|
||||
|
||||
class QuarkNVFP4(QuarkScheme):
|
||||
"""
|
||||
Quark NVFP4 quantization scheme.
|
||||
"""Quark NVFP4 quantization scheme.
|
||||
|
||||
Supports loading NVFP4 checkpoints with the following structure:
|
||||
- weight: uint8, shape [out_features, in_features // 2] (packed FP4)
|
||||
|
||||
@@ -11,8 +11,7 @@ __all__ = ["QuarkScheme"]
|
||||
|
||||
|
||||
class QuarkScheme(ABC):
|
||||
"""
|
||||
Abstract class used to describe the weight creation and forward pass
|
||||
"""Abstract class used to describe the weight creation and forward pass
|
||||
of different quantization schemes supported by Quark.
|
||||
"""
|
||||
|
||||
@@ -36,25 +35,19 @@ class QuarkScheme(ABC):
|
||||
@classmethod
|
||||
@abstractmethod
|
||||
def get_min_capability(cls) -> int:
|
||||
"""
|
||||
Get minimum device capability.
|
||||
"""
|
||||
"""Get minimum device capability."""
|
||||
raise NotImplementedError
|
||||
|
||||
@abstractmethod
|
||||
def create_weights(self, *args, **kwargs):
|
||||
"""
|
||||
Weight creation for the particular scheme. Inputs to this function
|
||||
|
||||
"""
|
||||
"""Weight creation for the particular scheme. Inputs to this function."""
|
||||
raise NotImplementedError
|
||||
|
||||
@abstractmethod
|
||||
def apply_weights(
|
||||
self, layer: torch.nn.Module, x: torch.Tensor, bias: torch.Tensor | None
|
||||
):
|
||||
"""
|
||||
Run the forward pass for the particular scheme. This is where
|
||||
"""Run the forward pass for the particular scheme. This is where
|
||||
scheme-specific dequant/quant steps/kernels should be applied.
|
||||
|
||||
Args:
|
||||
@@ -62,13 +55,13 @@ class QuarkScheme(ABC):
|
||||
other parameters relevant to the particular scheme.
|
||||
x: input to the layer
|
||||
bias: bias parameter
|
||||
|
||||
"""
|
||||
raise NotImplementedError
|
||||
|
||||
@abstractmethod
|
||||
def process_weights_after_loading(self, layer: torch.nn.Module):
|
||||
"""
|
||||
Called after weight loading is complete for any cleanup that
|
||||
"""Called after weight loading is complete for any cleanup that
|
||||
needs to occur.
|
||||
"""
|
||||
raise NotImplementedError
|
||||
|
||||
@@ -33,8 +33,7 @@ OCP_MX_BLOCK_SIZE = 32
|
||||
|
||||
|
||||
class QuarkW4A8_MXFP4_FP8(QuarkScheme):
|
||||
"""
|
||||
- Weights: MXFP4 with E8M0 scales per block of 32
|
||||
"""- Weights: MXFP4 with E8M0 scales per block of 32
|
||||
- Activations: FP8 E4M3 (static per-tensor quantization)
|
||||
|
||||
Uses the AITER Triton kernel and falls back to emulation if AITER not available.
|
||||
|
||||
Reference in New Issue
Block a user