Convert formatting to use ruff instead of yapf + isort (#26247)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
@@ -8,18 +8,30 @@ import torch
|
||||
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.fused_moe import FusedMoE
|
||||
from vllm.model_executor.layers.linear import (LinearBase, LinearMethodBase,
|
||||
UnquantizedLinearMethod)
|
||||
from vllm.model_executor.layers.linear import (
|
||||
LinearBase,
|
||||
LinearMethodBase,
|
||||
UnquantizedLinearMethod,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization import QuantizationMethods
|
||||
from vllm.model_executor.layers.quantization.base_config import ( # noqa: E501
|
||||
QuantizationConfig, QuantizeMethodBase)
|
||||
QuantizationConfig,
|
||||
QuantizeMethodBase,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.kv_cache import BaseKVCacheMethod
|
||||
from vllm.model_executor.layers.quantization.quark.quark_moe import ( # noqa: E501
|
||||
QuarkMoEMethod)
|
||||
QuarkMoEMethod,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.quark.schemes import (
|
||||
QuarkScheme, QuarkW4A4MXFP4, QuarkW8A8Fp8, QuarkW8A8Int8)
|
||||
QuarkScheme,
|
||||
QuarkW4A4MXFP4,
|
||||
QuarkW8A8Fp8,
|
||||
QuarkW8A8Int8,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.quark.utils import (
|
||||
deep_compare, should_ignore_layer)
|
||||
deep_compare,
|
||||
should_ignore_layer,
|
||||
)
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
__all__ = ["QuarkLinearMethod"]
|
||||
@@ -28,12 +40,13 @@ logger = init_logger(__name__)
|
||||
|
||||
|
||||
class QuarkConfig(QuantizationConfig):
|
||||
|
||||
def __init__(self,
|
||||
quant_config: dict[str, Any],
|
||||
kv_cache_group: Optional[list[str]] = None,
|
||||
kv_cache_config: Optional[dict[str, Any]] = None,
|
||||
pack_method: str = "reorder"):
|
||||
def __init__(
|
||||
self,
|
||||
quant_config: dict[str, Any],
|
||||
kv_cache_group: Optional[list[str]] = None,
|
||||
kv_cache_config: Optional[dict[str, Any]] = None,
|
||||
pack_method: str = "reorder",
|
||||
):
|
||||
super().__init__()
|
||||
if kv_cache_group is None:
|
||||
kv_cache_group = []
|
||||
@@ -55,15 +68,16 @@ class QuarkConfig(QuantizationConfig):
|
||||
def get_name(self) -> QuantizationMethods:
|
||||
return "quark"
|
||||
|
||||
def get_quant_method(self, layer: torch.nn.Module,
|
||||
prefix: str) -> Optional["QuantizeMethodBase"]:
|
||||
def get_quant_method(
|
||||
self, layer: torch.nn.Module, prefix: str
|
||||
) -> Optional["QuantizeMethodBase"]:
|
||||
from vllm.attention.layer import Attention # Avoid circular import
|
||||
|
||||
# Check if the layer is skipped for quantization.
|
||||
exclude_layers = cast(list[str], self.quant_config.get("exclude"))
|
||||
if should_ignore_layer(prefix,
|
||||
ignore=exclude_layers,
|
||||
fused_mapping=self.packed_modules_mapping):
|
||||
if should_ignore_layer(
|
||||
prefix, ignore=exclude_layers, fused_mapping=self.packed_modules_mapping
|
||||
):
|
||||
return UnquantizedLinearMethod()
|
||||
if isinstance(layer, LinearBase):
|
||||
scheme = self.get_scheme(layer=layer, layer_name=prefix)
|
||||
@@ -73,17 +87,17 @@ class QuarkConfig(QuantizationConfig):
|
||||
return QuarkKVCacheMethod(self)
|
||||
|
||||
if isinstance(layer, FusedMoE):
|
||||
return QuarkMoEMethod.get_moe_method(self,
|
||||
module=layer,
|
||||
layer_name=prefix)
|
||||
return QuarkMoEMethod.get_moe_method(self, module=layer, layer_name=prefix)
|
||||
return None
|
||||
|
||||
@classmethod
|
||||
def from_config(cls, config: dict[str, Any]) -> "QuarkConfig":
|
||||
export_config = config.get("export")
|
||||
if export_config is None:
|
||||
raise ValueError("The export key should be included in "
|
||||
"the configurations of Quark quantized model")
|
||||
raise ValueError(
|
||||
"The export key should be included in "
|
||||
"the configurations of Quark quantized model"
|
||||
)
|
||||
kv_cache_group = cast(list[str], export_config.get("kv_cache_group"))
|
||||
pack_method = cast(str, export_config.get("pack_method"))
|
||||
|
||||
@@ -96,33 +110,32 @@ class QuarkConfig(QuantizationConfig):
|
||||
kv_cache_config = None
|
||||
else:
|
||||
kv_cache_set = set(kv_cache_group)
|
||||
layer_quant_config = cast(dict[str, Any],
|
||||
config.get("layer_quant_config"))
|
||||
layer_quant_config = cast(dict[str, Any], config.get("layer_quant_config"))
|
||||
layer_quant_names = list(layer_quant_config.keys())
|
||||
layer_quant_set = set(layer_quant_names)
|
||||
|
||||
if not kv_cache_set.issubset(layer_quant_set):
|
||||
raise ValueError("The Quark quantized model has the "
|
||||
"kv_cache_group parameter setting, "
|
||||
"but no kv_cache quantization settings "
|
||||
"were found in the quantization "
|
||||
"configuration.")
|
||||
raise ValueError(
|
||||
"The Quark quantized model has the "
|
||||
"kv_cache_group parameter setting, "
|
||||
"but no kv_cache quantization settings "
|
||||
"were found in the quantization "
|
||||
"configuration."
|
||||
)
|
||||
|
||||
q_configs = [
|
||||
cast(dict[str, Any], layer_quant_config.get(name))
|
||||
for name in kv_cache_group
|
||||
]
|
||||
if not all(
|
||||
deep_compare(q_config, q_configs[0])
|
||||
for q_config in q_configs):
|
||||
if not all(deep_compare(q_config, q_configs[0]) for q_config in q_configs):
|
||||
raise ValueError(
|
||||
"The quantization method used for kv_cache should "
|
||||
"be the same, but the quantization method for the "
|
||||
"kv_cache layer in the config is different.")
|
||||
"kv_cache layer in the config is different."
|
||||
)
|
||||
kv_cache_config = q_configs[0].get("output_tensors")
|
||||
if kv_cache_config is None:
|
||||
raise ValueError(
|
||||
"The kv_cache quantization configuration is empty.")
|
||||
raise ValueError("The kv_cache quantization configuration is empty.")
|
||||
|
||||
# Since we have already set kv_cache quantization configurations,
|
||||
# we will remove the quantization configuration for the
|
||||
@@ -132,23 +145,22 @@ class QuarkConfig(QuantizationConfig):
|
||||
|
||||
# In case q_proj output is also quantized, remove the configuration
|
||||
# to keep qkv consistency.
|
||||
q_proj_q_config = cast(dict[str, Any],
|
||||
layer_quant_config.get("*q_proj"))
|
||||
q_proj_q_config = cast(dict[str, Any], layer_quant_config.get("*q_proj"))
|
||||
if q_proj_q_config is not None:
|
||||
q_proj_q_config["output_tensors"] = None
|
||||
|
||||
return cls(quant_config=config,
|
||||
kv_cache_group=kv_cache_group,
|
||||
kv_cache_config=kv_cache_config,
|
||||
pack_method=pack_method)
|
||||
return cls(
|
||||
quant_config=config,
|
||||
kv_cache_group=kv_cache_group,
|
||||
kv_cache_config=kv_cache_config,
|
||||
pack_method=pack_method,
|
||||
)
|
||||
|
||||
@classmethod
|
||||
def get_config_filenames(cls) -> list[str]:
|
||||
return []
|
||||
|
||||
def _check_scheme_supported(self,
|
||||
min_capability: int,
|
||||
error: bool = True) -> bool:
|
||||
def _check_scheme_supported(self, min_capability: int, error: bool = True) -> bool:
|
||||
capability_tuple = current_platform.get_device_capability()
|
||||
|
||||
if capability_tuple is not None:
|
||||
@@ -158,26 +170,33 @@ class QuarkConfig(QuantizationConfig):
|
||||
raise RuntimeError(
|
||||
"Quantization scheme is not supported for ",
|
||||
f"the current GPU. Min capability: {min_capability}. ",
|
||||
f"Current capability: {capability}.")
|
||||
f"Current capability: {capability}.",
|
||||
)
|
||||
return supported
|
||||
else:
|
||||
return False
|
||||
|
||||
def _is_fp8_w8a8(self, weight_quant: Optional[dict[str, Any]],
|
||||
input_quant: Optional[dict[str, Any]]) -> bool:
|
||||
def _is_fp8_w8a8(
|
||||
self,
|
||||
weight_quant: Optional[dict[str, Any]],
|
||||
input_quant: Optional[dict[str, Any]],
|
||||
) -> bool:
|
||||
# Confirm weights and input quantized.
|
||||
if weight_quant is None or input_quant is None:
|
||||
return False
|
||||
|
||||
# Confirm weight scheme is supported
|
||||
is_fp8_dtype = (weight_quant.get("dtype") == "fp8_e4m3"
|
||||
and input_quant.get("dtype") == "fp8_e4m3")
|
||||
is_fp8_dtype = (
|
||||
weight_quant.get("dtype") == "fp8_e4m3"
|
||||
and input_quant.get("dtype") == "fp8_e4m3"
|
||||
)
|
||||
is_static_weight = not weight_quant.get("is_dynamic")
|
||||
is_per_tensor_or_channel_weight = (weight_quant.get("qscheme")
|
||||
in ["per_tensor", "per_channel"])
|
||||
is_per_tensor_or_channel_weight = weight_quant.get("qscheme") in [
|
||||
"per_tensor",
|
||||
"per_channel",
|
||||
]
|
||||
|
||||
if not (is_fp8_dtype and is_static_weight
|
||||
and is_per_tensor_or_channel_weight):
|
||||
if not (is_fp8_dtype and is_static_weight and is_per_tensor_or_channel_weight):
|
||||
return False
|
||||
|
||||
# Dynamic quantization is always supported if weights supported.
|
||||
@@ -185,76 +204,86 @@ class QuarkConfig(QuantizationConfig):
|
||||
return True
|
||||
|
||||
# Confirm activation scheme is supported.
|
||||
is_per_tensor_activation = (input_quant.get("qscheme") == "per_tensor")
|
||||
is_per_tensor_activation = input_quant.get("qscheme") == "per_tensor"
|
||||
return is_per_tensor_activation
|
||||
|
||||
def _is_static_tensor_w8a8(self, weight_quant: Optional[dict[str, Any]],
|
||||
input_quant: Optional[dict[str, Any]]) -> bool:
|
||||
def _is_static_tensor_w8a8(
|
||||
self,
|
||||
weight_quant: Optional[dict[str, Any]],
|
||||
input_quant: Optional[dict[str, Any]],
|
||||
) -> bool:
|
||||
# Confirm weights and input quantized.
|
||||
if weight_quant is None or input_quant is None:
|
||||
return False
|
||||
|
||||
is_int8_dtype = (weight_quant.get("dtype") == "int8"
|
||||
and input_quant.get("dtype") == "int8")
|
||||
is_int8_dtype = (
|
||||
weight_quant.get("dtype") == "int8" and input_quant.get("dtype") == "int8"
|
||||
)
|
||||
|
||||
is_tensor = (weight_quant.get("qscheme")
|
||||
in ["per_tensor", "per_channel"]
|
||||
and input_quant.get("qscheme") == "per_tensor")
|
||||
is_tensor = (
|
||||
weight_quant.get("qscheme") in ["per_tensor", "per_channel"]
|
||||
and input_quant.get("qscheme") == "per_tensor"
|
||||
)
|
||||
|
||||
is_static = (not weight_quant.get("is_dynamic")
|
||||
and not input_quant.get("is_dynamic"))
|
||||
is_static = not weight_quant.get("is_dynamic") and not input_quant.get(
|
||||
"is_dynamic"
|
||||
)
|
||||
|
||||
is_weight_symmetric = (weight_quant.get("symmetric") is True)
|
||||
is_weight_symmetric = weight_quant.get("symmetric") is True
|
||||
|
||||
# Both symmetric and asymmetric input quantization supported.
|
||||
# Only symmetric weight quantization supported.
|
||||
return is_int8_dtype and is_tensor and is_weight_symmetric and is_static
|
||||
|
||||
def _is_mx_fp4(self, weight_quant: Optional[dict[str, Any]],
|
||||
input_quant: Optional[dict[str, Any]]) -> bool:
|
||||
def _is_mx_fp4(
|
||||
self,
|
||||
weight_quant: Optional[dict[str, Any]],
|
||||
input_quant: Optional[dict[str, Any]],
|
||||
) -> bool:
|
||||
# Confirm weights and input quantized.
|
||||
if weight_quant is None or input_quant is None:
|
||||
logger.debug("Quark model is not in MX-FP4 format: "
|
||||
"weight_quant or input_quant not set")
|
||||
logger.debug(
|
||||
"Quark model is not in MX-FP4 format: "
|
||||
"weight_quant or input_quant not set"
|
||||
)
|
||||
return False
|
||||
|
||||
# Input and weight dtype needs to be fp4.
|
||||
if weight_quant.get("dtype") != "fp4" or input_quant.get(
|
||||
"dtype") != "fp4":
|
||||
if weight_quant.get("dtype") != "fp4" or input_quant.get("dtype") != "fp4":
|
||||
logger.debug("Quark model is not in MX-FP4 format: dtype not fp4")
|
||||
return False
|
||||
|
||||
# Input and weight qscheme needs to be per group.
|
||||
if weight_quant.get("qscheme") != "per_group" or input_quant.get(
|
||||
"qscheme") != "per_group":
|
||||
if (
|
||||
weight_quant.get("qscheme") != "per_group"
|
||||
or input_quant.get("qscheme") != "per_group"
|
||||
):
|
||||
logger.debug("Quark model is not in MX-FP4 format: not per_group")
|
||||
return False
|
||||
|
||||
# Input and weight group size needs to be 32.
|
||||
if weight_quant.get("group_size") != 32 or input_quant.get(
|
||||
"group_size") != 32:
|
||||
logger.debug(
|
||||
"Quark model is not in MX-FP4 format: not group_size=32")
|
||||
if weight_quant.get("group_size") != 32 or input_quant.get("group_size") != 32:
|
||||
logger.debug("Quark model is not in MX-FP4 format: not group_size=32")
|
||||
return False
|
||||
|
||||
# Activations need to use dynamic quantization.
|
||||
if input_quant.get("is_dynamic") is False:
|
||||
logger.debug(
|
||||
"Quark model is not in MX-FP4 format: not activation dynamic")
|
||||
logger.debug("Quark model is not in MX-FP4 format: not activation dynamic")
|
||||
return False
|
||||
|
||||
# Activations and weight scales need to be in e8m0 format.
|
||||
if weight_quant.get("scale_format") != "e8m0" or input_quant.get(
|
||||
"scale_format") != "e8m0":
|
||||
logger.debug(
|
||||
"Quark model is not in MX-FP4 format: not scale_format e8m0")
|
||||
if (
|
||||
weight_quant.get("scale_format") != "e8m0"
|
||||
or input_quant.get("scale_format") != "e8m0"
|
||||
):
|
||||
logger.debug("Quark model is not in MX-FP4 format: not scale_format e8m0")
|
||||
return False
|
||||
|
||||
return True
|
||||
|
||||
def _find_matched_config(self, layer_name: str,
|
||||
module: torch.nn.Module) -> dict[str, Any]:
|
||||
|
||||
def _find_matched_config(
|
||||
self, layer_name: str, module: torch.nn.Module
|
||||
) -> dict[str, Any]:
|
||||
proj_name = layer_name.split(".")[-1]
|
||||
if proj_name in self.packed_modules_mapping:
|
||||
shard_proj_names = self.packed_modules_mapping[proj_name]
|
||||
@@ -269,59 +298,66 @@ class QuarkConfig(QuantizationConfig):
|
||||
for shard_name in shard_names
|
||||
]
|
||||
if not all(
|
||||
deep_compare(q_config, shard_configs[0])
|
||||
for q_config in shard_configs):
|
||||
deep_compare(q_config, shard_configs[0]) for q_config in shard_configs
|
||||
):
|
||||
raise ValueError(
|
||||
f"Found a different quantization configuration for "
|
||||
f"{shard_proj_names} in {layer_name}. vLLM "
|
||||
"requires all to use the same scheme.")
|
||||
"requires all to use the same scheme."
|
||||
)
|
||||
return shard_configs[0]
|
||||
else:
|
||||
layer_quant_config = cast(
|
||||
dict[str, Any], self.quant_config.get("layer_quant_config"))
|
||||
dict[str, Any], self.quant_config.get("layer_quant_config")
|
||||
)
|
||||
for name_pattern in layer_quant_config:
|
||||
if fnmatch.fnmatch(layer_name, name_pattern):
|
||||
return layer_quant_config[name_pattern]
|
||||
|
||||
layer_type = cast(str, type(module))
|
||||
layer_type_quant_config = cast(
|
||||
dict[str, Any],
|
||||
self.quant_config.get("layer_type_quant_config"))
|
||||
dict[str, Any], self.quant_config.get("layer_type_quant_config")
|
||||
)
|
||||
if layer_type in layer_type_quant_config:
|
||||
return layer_type_quant_config[layer_type]
|
||||
|
||||
global_quant_config = cast(
|
||||
dict[str, Any], self.quant_config.get("global_quant_config"))
|
||||
dict[str, Any], self.quant_config.get("global_quant_config")
|
||||
)
|
||||
return global_quant_config
|
||||
|
||||
def _get_scheme_from_config(self, config: dict[str, Any]) -> "QuarkScheme":
|
||||
if config.get("output_tensors") or config.get("bias"):
|
||||
raise NotImplementedError(
|
||||
"Currently, Quark models with output_tensors "
|
||||
"and bias quantized are not supported")
|
||||
"and bias quantized are not supported"
|
||||
)
|
||||
weight_config = cast(dict[str, Any], config.get("weight"))
|
||||
input_config = cast(dict[str, Any], config.get("input_tensors"))
|
||||
|
||||
if self._is_fp8_w8a8(weight_config, input_config):
|
||||
is_fp8_w8a8_supported = self._check_scheme_supported(
|
||||
QuarkW8A8Fp8.get_min_capability(), error=False)
|
||||
QuarkW8A8Fp8.get_min_capability(), error=False
|
||||
)
|
||||
if is_fp8_w8a8_supported:
|
||||
return QuarkW8A8Fp8(weight_config, input_config)
|
||||
elif self._is_static_tensor_w8a8(weight_config, input_config):
|
||||
weight_qscheme = cast(str, weight_config.get("qscheme"))
|
||||
return QuarkW8A8Int8(qscheme=weight_qscheme,
|
||||
is_static_input_scheme=True,
|
||||
input_symmetric=input_config.get("symmetric"))
|
||||
return QuarkW8A8Int8(
|
||||
qscheme=weight_qscheme,
|
||||
is_static_input_scheme=True,
|
||||
input_symmetric=input_config.get("symmetric"),
|
||||
)
|
||||
elif self._is_mx_fp4(weight_config, input_config):
|
||||
return QuarkW4A4MXFP4(weight_config, input_config)
|
||||
|
||||
raise NotImplementedError("No quark compatible scheme was found. "
|
||||
f"Weight config: {weight_config}, "
|
||||
f"Input config: {input_config}")
|
||||
|
||||
def get_scheme(self, layer: torch.nn.Module,
|
||||
layer_name: str) -> "QuarkScheme":
|
||||
raise NotImplementedError(
|
||||
"No quark compatible scheme was found. "
|
||||
f"Weight config: {weight_config}, "
|
||||
f"Input config: {input_config}"
|
||||
)
|
||||
|
||||
def get_scheme(self, layer: torch.nn.Module, layer_name: str) -> "QuarkScheme":
|
||||
layer_quant_config = self._find_matched_config(layer_name, layer)
|
||||
|
||||
# Find the quant_scheme
|
||||
@@ -335,7 +371,7 @@ class QuarkConfig(QuantizationConfig):
|
||||
def get_cache_scale(self, name: str) -> Optional[str]:
|
||||
"""
|
||||
Check whether the param name matches the format for k/v cache scales
|
||||
in quark. If this is the case, return its equivalent param name
|
||||
in quark. If this is the case, return its equivalent param name
|
||||
expected by vLLM
|
||||
|
||||
:param name: param name
|
||||
@@ -355,18 +391,22 @@ class QuarkConfig(QuantizationConfig):
|
||||
|
||||
|
||||
class QuarkLinearMethod(LinearMethodBase):
|
||||
|
||||
def __init__(self, quantization_config: QuarkConfig):
|
||||
self.quantization_config = quantization_config
|
||||
|
||||
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
|
||||
layer.scheme.process_weights_after_loading(layer)
|
||||
|
||||
def create_weights(self, layer: torch.nn.Module,
|
||||
input_size_per_partition: int,
|
||||
output_partition_sizes: list[int], input_size: int,
|
||||
output_size: int, params_dtype: torch.dtype,
|
||||
**extra_weight_attrs):
|
||||
def create_weights(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
input_size_per_partition: int,
|
||||
output_partition_sizes: list[int],
|
||||
input_size: int,
|
||||
output_size: int,
|
||||
params_dtype: torch.dtype,
|
||||
**extra_weight_attrs,
|
||||
):
|
||||
"""
|
||||
Use the CompressedTensorsScheme associated with each layer to create
|
||||
the necessary parameters for the layer. See LinearMethodBase for param
|
||||
@@ -380,12 +420,15 @@ class QuarkLinearMethod(LinearMethodBase):
|
||||
output_partition_sizes=output_partition_sizes,
|
||||
output_size=output_size,
|
||||
params_dtype=params_dtype,
|
||||
weight_loader=weight_loader)
|
||||
weight_loader=weight_loader,
|
||||
)
|
||||
|
||||
def apply(self,
|
||||
layer: torch.nn.Module,
|
||||
x: torch.Tensor,
|
||||
bias: Optional[torch.Tensor] = None):
|
||||
def apply(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
x: torch.Tensor,
|
||||
bias: Optional[torch.Tensor] = None,
|
||||
):
|
||||
"""
|
||||
Use the output of create_weights and the CompressedTensorsScheme
|
||||
associated with the layer to apply the forward pass with the
|
||||
@@ -422,11 +465,13 @@ class QuarkKVCacheMethod(BaseKVCacheMethod):
|
||||
if dtype != "fp8_e4m3":
|
||||
raise NotImplementedError(
|
||||
"Currently supported kv cache quantization is "
|
||||
f"dtype=fp8_e4m3, however received {dtype}")
|
||||
f"dtype=fp8_e4m3, however received {dtype}"
|
||||
)
|
||||
|
||||
qscheme = kv_cache_config.get("qscheme")
|
||||
if qscheme != "per_tensor":
|
||||
raise NotImplementedError(
|
||||
"Only support per-tensor scaling factor "
|
||||
"for quark KV cache. "
|
||||
f"Expected qscheme: per_tensor, found qscheme: {qscheme}")
|
||||
f"Expected qscheme: per_tensor, found qscheme: {qscheme}"
|
||||
)
|
||||
|
||||
@@ -8,66 +8,71 @@ import torch
|
||||
import vllm.envs as envs
|
||||
from vllm import _custom_ops as ops
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.fused_moe import (FusedMoE, FusedMoEConfig,
|
||||
FusedMoEMethodBase,
|
||||
FusedMoeWeightScaleSupported)
|
||||
from vllm.model_executor.layers.fused_moe import (
|
||||
FusedMoE,
|
||||
FusedMoEConfig,
|
||||
FusedMoEMethodBase,
|
||||
FusedMoeWeightScaleSupported,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.config import (
|
||||
FusedMoEQuantConfig, fp8_w8a8_moe_quant_config,
|
||||
mxfp4_w4a4_moe_quant_config)
|
||||
FusedMoEQuantConfig,
|
||||
fp8_w8a8_moe_quant_config,
|
||||
mxfp4_w4a4_moe_quant_config,
|
||||
)
|
||||
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
|
||||
is_rocm_aiter_moe_enabled)
|
||||
is_rocm_aiter_moe_enabled,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.utils.marlin_utils_fp8 import (
|
||||
prepare_moe_fp8_layer_for_marlin)
|
||||
from vllm.model_executor.layers.quantization.utils.mxfp4_utils import (
|
||||
OCP_MX_BLOCK_SIZE)
|
||||
from vllm.model_executor.layers.quantization.utils.quant_utils import (
|
||||
GroupShape)
|
||||
prepare_moe_fp8_layer_for_marlin,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.utils.mxfp4_utils import OCP_MX_BLOCK_SIZE
|
||||
from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
|
||||
from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
|
||||
all_close_1d, normalize_e4m3fn_to_e4m3fnuz, per_tensor_dequantize)
|
||||
all_close_1d,
|
||||
normalize_e4m3fn_to_e4m3fnuz,
|
||||
per_tensor_dequantize,
|
||||
)
|
||||
from vllm.model_executor.utils import set_weight_attrs
|
||||
from vllm.platforms import current_platform
|
||||
from vllm.scalar_type import scalar_types
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
__all__ = [
|
||||
"QuarkMoEMethod", "QuarkW8A8Fp8MoEMethod", "QuarkW4A4MXFp4MoEMethod"
|
||||
]
|
||||
__all__ = ["QuarkMoEMethod", "QuarkW8A8Fp8MoEMethod", "QuarkW4A4MXFp4MoEMethod"]
|
||||
|
||||
|
||||
class QuarkMoEMethod(FusedMoEMethodBase):
|
||||
|
||||
def __init__(self, moe: FusedMoEConfig):
|
||||
super().__init__(moe)
|
||||
|
||||
@staticmethod
|
||||
def get_moe_method(
|
||||
quant_config: "QuarkConfig", # type: ignore # noqa E501 # noqa F821
|
||||
module: torch.nn.Module,
|
||||
layer_name: str) -> "QuarkMoEMethod":
|
||||
layer_quant_config = quant_config._find_matched_config(
|
||||
layer_name, module)
|
||||
quant_config: "QuarkConfig", # type: ignore # noqa E501 # noqa F821
|
||||
module: torch.nn.Module,
|
||||
layer_name: str,
|
||||
) -> "QuarkMoEMethod":
|
||||
layer_quant_config = quant_config._find_matched_config(layer_name, module)
|
||||
|
||||
if (layer_quant_config.get("output_tensors")
|
||||
or layer_quant_config.get("bias")):
|
||||
raise NotImplementedError("Currently, Quark models with "
|
||||
"output_tensors and bias "
|
||||
"quantized are not supported")
|
||||
if layer_quant_config.get("output_tensors") or layer_quant_config.get("bias"):
|
||||
raise NotImplementedError(
|
||||
"Currently, Quark models with "
|
||||
"output_tensors and bias "
|
||||
"quantized are not supported"
|
||||
)
|
||||
weight_config = layer_quant_config.get("weight")
|
||||
input_config = layer_quant_config.get("input_tensors")
|
||||
|
||||
if quant_config._is_fp8_w8a8(weight_config, input_config):
|
||||
return QuarkW8A8Fp8MoEMethod(weight_config, input_config,
|
||||
module.moe_config)
|
||||
return QuarkW8A8Fp8MoEMethod(weight_config, input_config, module.moe_config)
|
||||
elif quant_config._is_mx_fp4(weight_config, input_config):
|
||||
return QuarkW4A4MXFp4MoEMethod(weight_config, input_config,
|
||||
module.moe_config)
|
||||
return QuarkW4A4MXFp4MoEMethod(
|
||||
weight_config, input_config, module.moe_config
|
||||
)
|
||||
else:
|
||||
raise RuntimeError("Unsupported FusedMoe scheme")
|
||||
|
||||
|
||||
class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
weight_config: dict[str, Any],
|
||||
@@ -80,38 +85,50 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
|
||||
|
||||
self.weight_qscheme = self.weight_quant.get("qscheme")
|
||||
self.input_qscheme = self.input_quant.get("qscheme")
|
||||
per_tensor = (self.weight_qscheme == "per_tensor"
|
||||
and self.input_qscheme == "per_tensor")
|
||||
per_channel = (self.weight_qscheme == "per_channel"
|
||||
and self.input_qscheme == "per_channel")
|
||||
self.act_quant_group_shape = GroupShape.PER_TOKEN \
|
||||
if per_channel else GroupShape.PER_TENSOR
|
||||
per_tensor = (
|
||||
self.weight_qscheme == "per_tensor" and self.input_qscheme == "per_tensor"
|
||||
)
|
||||
per_channel = (
|
||||
self.weight_qscheme == "per_channel" and self.input_qscheme == "per_channel"
|
||||
)
|
||||
self.act_quant_group_shape = (
|
||||
GroupShape.PER_TOKEN if per_channel else GroupShape.PER_TENSOR
|
||||
)
|
||||
if not (per_tensor or per_channel):
|
||||
raise ValueError(
|
||||
"For FP8 Fused MoE layers, only per-tensor and per-channel "
|
||||
"scales for weights and activations are supported. Found "
|
||||
f"{self.weight_qscheme}, {self.input_qscheme}") # noqa E501
|
||||
f"{self.weight_qscheme}, {self.input_qscheme}"
|
||||
) # noqa E501
|
||||
|
||||
self.static_input_scales = not self.input_quant.get("is_dynamic")
|
||||
if self.static_input_scales and per_channel:
|
||||
raise ValueError(
|
||||
"For FP8 Fused MoE layer, we require either per tensor or "
|
||||
"channelwise, dynamic per token quantization.")
|
||||
"channelwise, dynamic per token quantization."
|
||||
)
|
||||
|
||||
# For GPUs that lack FP8 hardware support, we can leverage the Marlin
|
||||
# kernel for fast weight-only FP8 quantization
|
||||
self.use_marlin = (not current_platform.has_device_capability(89)
|
||||
or envs.VLLM_TEST_FORCE_FP8_MARLIN)
|
||||
self.use_marlin = (
|
||||
not current_platform.has_device_capability(89)
|
||||
or envs.VLLM_TEST_FORCE_FP8_MARLIN
|
||||
)
|
||||
# Disable marlin for rocm
|
||||
if current_platform.is_rocm():
|
||||
self.use_marlin = False
|
||||
|
||||
self.rocm_aiter_moe_enabled = is_rocm_aiter_moe_enabled()
|
||||
|
||||
def create_weights(self, layer: torch.nn.Module, num_experts: int,
|
||||
hidden_size: int, intermediate_size_per_partition: int,
|
||||
params_dtype: torch.dtype, **extra_weight_attrs):
|
||||
|
||||
def create_weights(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
num_experts: int,
|
||||
hidden_size: int,
|
||||
intermediate_size_per_partition: int,
|
||||
params_dtype: torch.dtype,
|
||||
**extra_weight_attrs,
|
||||
):
|
||||
layer.intermediate_size_per_partition = intermediate_size_per_partition
|
||||
layer.hidden_size = hidden_size
|
||||
layer.num_experts = num_experts
|
||||
@@ -120,21 +137,27 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
|
||||
params_dtype = torch.float8_e4m3fn
|
||||
|
||||
# WEIGHTS
|
||||
w13_weight = torch.nn.Parameter(torch.empty(
|
||||
num_experts,
|
||||
2 * intermediate_size_per_partition,
|
||||
hidden_size,
|
||||
dtype=params_dtype),
|
||||
requires_grad=False)
|
||||
w13_weight = torch.nn.Parameter(
|
||||
torch.empty(
|
||||
num_experts,
|
||||
2 * intermediate_size_per_partition,
|
||||
hidden_size,
|
||||
dtype=params_dtype,
|
||||
),
|
||||
requires_grad=False,
|
||||
)
|
||||
layer.register_parameter("w13_weight", w13_weight)
|
||||
set_weight_attrs(w13_weight, extra_weight_attrs)
|
||||
|
||||
w2_weight = torch.nn.Parameter(torch.empty(
|
||||
num_experts,
|
||||
hidden_size,
|
||||
intermediate_size_per_partition,
|
||||
dtype=params_dtype),
|
||||
requires_grad=False)
|
||||
w2_weight = torch.nn.Parameter(
|
||||
torch.empty(
|
||||
num_experts,
|
||||
hidden_size,
|
||||
intermediate_size_per_partition,
|
||||
dtype=params_dtype,
|
||||
),
|
||||
requires_grad=False,
|
||||
)
|
||||
layer.register_parameter("w2_weight", w2_weight)
|
||||
set_weight_attrs(w2_weight, extra_weight_attrs)
|
||||
|
||||
@@ -142,48 +165,54 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
|
||||
if self.weight_qscheme == "per_tensor":
|
||||
# Allocate 2 scales for w1 and w3 respectively.
|
||||
# They are combined to a single scale after weight loading.
|
||||
w13_weight_scale = torch.nn.Parameter(torch.ones(
|
||||
num_experts, 2, dtype=torch.float32),
|
||||
requires_grad=False)
|
||||
w13_weight_scale = torch.nn.Parameter(
|
||||
torch.ones(num_experts, 2, dtype=torch.float32), requires_grad=False
|
||||
)
|
||||
layer.register_parameter("w13_weight_scale", w13_weight_scale)
|
||||
w2_weight_scale = torch.nn.Parameter(torch.ones(
|
||||
num_experts, dtype=torch.float32),
|
||||
requires_grad=False)
|
||||
w2_weight_scale = torch.nn.Parameter(
|
||||
torch.ones(num_experts, dtype=torch.float32), requires_grad=False
|
||||
)
|
||||
layer.register_parameter("w2_weight_scale", w2_weight_scale)
|
||||
# Add PER-TENSOR quantization for FusedMoE.weight_loader.
|
||||
extra_weight_attrs.update(
|
||||
{"quant_method": FusedMoeWeightScaleSupported.TENSOR.value})
|
||||
{"quant_method": FusedMoeWeightScaleSupported.TENSOR.value}
|
||||
)
|
||||
set_weight_attrs(w13_weight_scale, extra_weight_attrs)
|
||||
set_weight_attrs(w2_weight_scale, extra_weight_attrs)
|
||||
elif self.weight_qscheme == "per_channel":
|
||||
# quark's scale is 1 dim.
|
||||
w13_weight_scale = torch.nn.Parameter(torch.ones(
|
||||
num_experts,
|
||||
2 * intermediate_size_per_partition,
|
||||
dtype=torch.float32),
|
||||
requires_grad=False)
|
||||
w13_weight_scale = torch.nn.Parameter(
|
||||
torch.ones(
|
||||
num_experts,
|
||||
2 * intermediate_size_per_partition,
|
||||
dtype=torch.float32,
|
||||
),
|
||||
requires_grad=False,
|
||||
)
|
||||
layer.register_parameter("w13_weight_scale", w13_weight_scale)
|
||||
w2_weight_scale = torch.nn.Parameter(torch.ones(
|
||||
num_experts, hidden_size, dtype=torch.float32),
|
||||
requires_grad=False)
|
||||
w2_weight_scale = torch.nn.Parameter(
|
||||
torch.ones(num_experts, hidden_size, dtype=torch.float32),
|
||||
requires_grad=False,
|
||||
)
|
||||
layer.register_parameter("w2_weight_scale", w2_weight_scale)
|
||||
# Add PER-CHANNEL quantization for FusedMoE.weight_loader.
|
||||
extra_weight_attrs.update(
|
||||
{"quant_method": FusedMoeWeightScaleSupported.CHANNEL.value})
|
||||
{"quant_method": FusedMoeWeightScaleSupported.CHANNEL.value}
|
||||
)
|
||||
set_weight_attrs(w13_weight_scale, extra_weight_attrs)
|
||||
set_weight_attrs(w2_weight_scale, extra_weight_attrs)
|
||||
|
||||
# INPUT_SCALES
|
||||
if self.static_input_scales:
|
||||
w13_input_scale = torch.nn.Parameter(torch.ones(
|
||||
num_experts, dtype=torch.float32),
|
||||
requires_grad=False)
|
||||
w13_input_scale = torch.nn.Parameter(
|
||||
torch.ones(num_experts, dtype=torch.float32), requires_grad=False
|
||||
)
|
||||
layer.register_parameter("w13_input_scale", w13_input_scale)
|
||||
set_weight_attrs(w13_input_scale, extra_weight_attrs)
|
||||
|
||||
w2_input_scale = torch.nn.Parameter(torch.ones(
|
||||
num_experts, dtype=torch.float32),
|
||||
requires_grad=False)
|
||||
w2_input_scale = torch.nn.Parameter(
|
||||
torch.ones(num_experts, dtype=torch.float32), requires_grad=False
|
||||
)
|
||||
layer.register_parameter("w2_input_scale", w2_input_scale)
|
||||
set_weight_attrs(w2_input_scale, extra_weight_attrs)
|
||||
else:
|
||||
@@ -194,46 +223,53 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
|
||||
# Fp8 moe kernels require a single activation scale.
|
||||
# We take the max of all the scales in case they differ.
|
||||
if self.static_input_scales:
|
||||
if (layer.w13_input_scale is None or layer.w2_input_scale is None):
|
||||
if layer.w13_input_scale is None or layer.w2_input_scale is None:
|
||||
raise ValueError(
|
||||
"QuantConfig has static quantization, but found "
|
||||
"activation scales are None.")
|
||||
if (not all_close_1d(layer.w13_input_scale)
|
||||
or not all_close_1d(layer.w2_input_scale)):
|
||||
"activation scales are None."
|
||||
)
|
||||
if not all_close_1d(layer.w13_input_scale) or not all_close_1d(
|
||||
layer.w2_input_scale
|
||||
):
|
||||
logger.warning_once(
|
||||
"Found input_scales that are not equal for "
|
||||
"fp8 MoE layer. Using the maximum across experts "
|
||||
"for each layer. ")
|
||||
"for each layer. "
|
||||
)
|
||||
layer.w13_input_scale = torch.nn.Parameter(
|
||||
layer.w13_input_scale.max(), requires_grad=False)
|
||||
layer.w13_input_scale.max(), requires_grad=False
|
||||
)
|
||||
layer.w2_input_scale = torch.nn.Parameter(
|
||||
layer.w2_input_scale.max(), requires_grad=False)
|
||||
layer.w2_input_scale.max(), requires_grad=False
|
||||
)
|
||||
|
||||
if current_platform.is_fp8_fnuz():
|
||||
# Normalize the weights and scales
|
||||
w13_weight, w13_weight_scale, w13_input_scale = \
|
||||
w13_weight, w13_weight_scale, w13_input_scale = (
|
||||
normalize_e4m3fn_to_e4m3fnuz(
|
||||
layer.w13_weight, layer.w13_weight_scale,
|
||||
layer.w13_input_scale)
|
||||
w2_weight, w2_weight_scale, w2_input_scale = \
|
||||
normalize_e4m3fn_to_e4m3fnuz(
|
||||
layer.w2_weight, layer.w2_weight_scale,
|
||||
layer.w2_input_scale)
|
||||
layer.w13_weight, layer.w13_weight_scale, layer.w13_input_scale
|
||||
)
|
||||
)
|
||||
w2_weight, w2_weight_scale, w2_input_scale = normalize_e4m3fn_to_e4m3fnuz(
|
||||
layer.w2_weight, layer.w2_weight_scale, layer.w2_input_scale
|
||||
)
|
||||
# Reset the parameter
|
||||
layer.w13_weight = torch.nn.Parameter(w13_weight,
|
||||
requires_grad=False)
|
||||
layer.w13_weight_scale = torch.nn.Parameter(w13_weight_scale,
|
||||
requires_grad=False)
|
||||
layer.w13_weight = torch.nn.Parameter(w13_weight, requires_grad=False)
|
||||
layer.w13_weight_scale = torch.nn.Parameter(
|
||||
w13_weight_scale, requires_grad=False
|
||||
)
|
||||
if w13_input_scale is not None:
|
||||
layer.w13_input_scale = torch.nn.Parameter(w13_input_scale,
|
||||
requires_grad=False)
|
||||
layer.w2_weight = torch.nn.Parameter(w2_weight,
|
||||
requires_grad=False)
|
||||
layer.w2_weight_scale = torch.nn.Parameter(w2_weight_scale,
|
||||
requires_grad=False)
|
||||
layer.w13_input_scale = torch.nn.Parameter(
|
||||
w13_input_scale, requires_grad=False
|
||||
)
|
||||
layer.w2_weight = torch.nn.Parameter(w2_weight, requires_grad=False)
|
||||
layer.w2_weight_scale = torch.nn.Parameter(
|
||||
w2_weight_scale, requires_grad=False
|
||||
)
|
||||
if w2_input_scale is not None:
|
||||
layer.w2_input_scale = torch.nn.Parameter(w2_input_scale,
|
||||
requires_grad=False)
|
||||
layer.w2_input_scale = torch.nn.Parameter(
|
||||
w2_input_scale, requires_grad=False
|
||||
)
|
||||
|
||||
# For per-tensor case, Fp8 moe kernel needs single weight scale
|
||||
# for w13 per expert. Use max then dequant and requant each expert.
|
||||
@@ -245,42 +281,45 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
|
||||
start = 0
|
||||
for shard_id in range(2):
|
||||
dq_weight = per_tensor_dequantize(
|
||||
layer.w13_weight[expert_id][start:start +
|
||||
shard_size, :],
|
||||
layer.w13_weight_scale[expert_id][shard_id])
|
||||
layer.w13_weight[expert_id][
|
||||
start:start + shard_size, :], _ = ops.scaled_fp8_quant(
|
||||
dq_weight, max_w13_scales[expert_id])
|
||||
layer.w13_weight[expert_id][start : start + shard_size, :],
|
||||
layer.w13_weight_scale[expert_id][shard_id],
|
||||
)
|
||||
layer.w13_weight[expert_id][start : start + shard_size, :], _ = (
|
||||
ops.scaled_fp8_quant(dq_weight, max_w13_scales[expert_id])
|
||||
)
|
||||
start += shard_size
|
||||
|
||||
layer.w13_weight_scale = torch.nn.Parameter(max_w13_scales,
|
||||
requires_grad=False)
|
||||
layer.w13_weight_scale = torch.nn.Parameter(
|
||||
max_w13_scales, requires_grad=False
|
||||
)
|
||||
# quark's scale is 1 dim.
|
||||
elif self.weight_qscheme == "per_channel":
|
||||
if self.act_quant_group_shape == GroupShape.PER_TOKEN:
|
||||
w13_weight_scale = layer.w13_weight_scale.unsqueeze(-1)
|
||||
layer.w13_weight_scale = torch.nn.Parameter(
|
||||
w13_weight_scale, requires_grad=False)
|
||||
w13_weight_scale, requires_grad=False
|
||||
)
|
||||
w2_weight_scale = layer.w2_weight_scale.unsqueeze(-1)
|
||||
layer.w2_weight_scale = torch.nn.Parameter(w2_weight_scale,
|
||||
requires_grad=False)
|
||||
layer.w2_weight_scale = torch.nn.Parameter(
|
||||
w2_weight_scale, requires_grad=False
|
||||
)
|
||||
# Property to determine if AITER is used
|
||||
if self.rocm_aiter_moe_enabled:
|
||||
from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import ( # noqa E501
|
||||
rocm_aiter_fused_experts, shuffle_weights)
|
||||
rocm_aiter_fused_experts,
|
||||
shuffle_weights,
|
||||
)
|
||||
|
||||
# reshaping weights is required for aiter moe kernel.
|
||||
shuffled_w13, shuffled_w2 = shuffle_weights(
|
||||
layer.w13_weight.data, layer.w2_weight.data)
|
||||
layer.w13_weight.data, layer.w2_weight.data
|
||||
)
|
||||
|
||||
layer.w13_weight = torch.nn.Parameter(shuffled_w13,
|
||||
requires_grad=False)
|
||||
layer.w2_weight = torch.nn.Parameter(shuffled_w2,
|
||||
requires_grad=False)
|
||||
layer.w13_weight = torch.nn.Parameter(shuffled_w13, requires_grad=False)
|
||||
layer.w2_weight = torch.nn.Parameter(shuffled_w2, requires_grad=False)
|
||||
|
||||
self.rocm_aiter_fused_experts_func = rocm_aiter_fused_experts
|
||||
elif self.use_marlin:
|
||||
|
||||
prepare_moe_fp8_layer_for_marlin(layer, False)
|
||||
# Activations not quantized for marlin.
|
||||
del layer.w13_input_scale
|
||||
@@ -288,10 +327,12 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
|
||||
self.fused_experts_func = None
|
||||
else:
|
||||
from vllm.model_executor.layers.fused_moe import fused_experts
|
||||
|
||||
self.fused_experts_func = fused_experts
|
||||
|
||||
def get_fused_moe_quant_config(
|
||||
self, layer: torch.nn.Module) -> Optional[FusedMoEQuantConfig]:
|
||||
self, layer: torch.nn.Module
|
||||
) -> Optional[FusedMoEQuantConfig]:
|
||||
return fp8_w8a8_moe_quant_config(
|
||||
w1_scale=layer.w13_weight_scale,
|
||||
w2_scale=layer.w2_weight_scale,
|
||||
@@ -327,7 +368,8 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
|
||||
|
||||
if enable_eplb:
|
||||
raise NotImplementedError(
|
||||
"EPLB not supported for `QuarkW8A8Fp8MoEMethod` yet.")
|
||||
"EPLB not supported for `QuarkW8A8Fp8MoEMethod` yet."
|
||||
)
|
||||
|
||||
topk_weights, topk_ids, _ = FusedMoE.select_experts(
|
||||
hidden_states=x,
|
||||
@@ -341,7 +383,8 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
|
||||
scoring_func=scoring_func,
|
||||
routed_scaling_factor=routed_scaling_factor,
|
||||
e_score_correction_bias=e_score_correction_bias,
|
||||
indices_type=self.topk_indices_dtype)
|
||||
indices_type=self.topk_indices_dtype,
|
||||
)
|
||||
|
||||
if self.rocm_aiter_moe_enabled:
|
||||
return self.rocm_aiter_fused_experts_func(
|
||||
@@ -353,10 +396,10 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
|
||||
activation=activation,
|
||||
apply_router_weight_on_input=apply_router_weight_on_input,
|
||||
quant_config=self.moe_quant_config,
|
||||
expert_map=expert_map)
|
||||
expert_map=expert_map,
|
||||
)
|
||||
if self.use_marlin:
|
||||
assert activation == "silu", (
|
||||
f"{activation} not supported for Marlin MoE.")
|
||||
assert activation == "silu", f"{activation} not supported for Marlin MoE."
|
||||
return torch.ops.vllm.fused_marlin_moe(
|
||||
x,
|
||||
layer.w13_weight,
|
||||
@@ -371,7 +414,8 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
|
||||
quant_type_id=scalar_types.float8_e4m3fn.id,
|
||||
apply_router_weight_on_input=apply_router_weight_on_input,
|
||||
global_num_experts=global_num_experts,
|
||||
expert_map=expert_map)
|
||||
expert_map=expert_map,
|
||||
)
|
||||
|
||||
assert self.fused_experts_func is not None
|
||||
|
||||
@@ -386,11 +430,11 @@ class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
|
||||
apply_router_weight_on_input=apply_router_weight_on_input,
|
||||
global_num_experts=global_num_experts,
|
||||
expert_map=expert_map,
|
||||
quant_config=self.moe_quant_config)
|
||||
quant_config=self.moe_quant_config,
|
||||
)
|
||||
|
||||
|
||||
class QuarkW4A4MXFp4MoEMethod(QuarkMoEMethod):
|
||||
|
||||
def __init__(
|
||||
self,
|
||||
weight_config: dict[str, Any],
|
||||
@@ -403,19 +447,20 @@ class QuarkW4A4MXFp4MoEMethod(QuarkMoEMethod):
|
||||
|
||||
weight_qscheme = self.weight_quant.get("qscheme")
|
||||
input_qscheme = self.input_quant.get("qscheme")
|
||||
if not (weight_qscheme == "per_group"
|
||||
and input_qscheme == "per_group"):
|
||||
if not (weight_qscheme == "per_group" and input_qscheme == "per_group"):
|
||||
raise ValueError(
|
||||
"For MX(FP4) Fused MoE layers, only per-group scales "
|
||||
"for weights and activations are supported. Found "
|
||||
f"{weight_qscheme}, {input_qscheme}") # noqa E501
|
||||
f"{weight_qscheme}, {input_qscheme}"
|
||||
) # noqa E501
|
||||
|
||||
self.static_input_scales = not self.input_quant.get("is_dynamic")
|
||||
|
||||
if self.static_input_scales:
|
||||
raise NotImplementedError(
|
||||
"QuarkW4A4MXFp4MoEMethod with static input scales is currently "
|
||||
"not implemented. Please open an issue.")
|
||||
"not implemented. Please open an issue."
|
||||
)
|
||||
|
||||
if not current_platform.supports_mx():
|
||||
self.emulate = True
|
||||
@@ -423,7 +468,8 @@ class QuarkW4A4MXFp4MoEMethod(QuarkMoEMethod):
|
||||
"The current platform does not support native MXFP4 "
|
||||
"computation. Simulated weight dequantization and activation "
|
||||
"QDQ (quantize and dequantize) will be used, with the linear "
|
||||
"layers computed in high precision.")
|
||||
"layers computed in high precision."
|
||||
)
|
||||
else:
|
||||
self.emulate = True
|
||||
logger.warning_once(
|
||||
@@ -431,36 +477,49 @@ class QuarkW4A4MXFp4MoEMethod(QuarkMoEMethod):
|
||||
"computation, but kernels are not yet integrated in vLLM. "
|
||||
"Simulated weight dequantization and activation "
|
||||
"QDQ (quantize and dequantize) will be used, with the linear "
|
||||
"layers computed in high precision.")
|
||||
|
||||
def create_weights(self, layer: torch.nn.Module, num_experts: int,
|
||||
hidden_size: int, intermediate_size_per_partition: int,
|
||||
params_dtype: torch.dtype, **extra_weight_attrs):
|
||||
"layers computed in high precision."
|
||||
)
|
||||
|
||||
def create_weights(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
num_experts: int,
|
||||
hidden_size: int,
|
||||
intermediate_size_per_partition: int,
|
||||
params_dtype: torch.dtype,
|
||||
**extra_weight_attrs,
|
||||
):
|
||||
# Add the quantization method used (per tensor/grouped/channel)
|
||||
# to ensure the weight scales are loaded in properly
|
||||
extra_weight_attrs.update(
|
||||
{"quant_method": FusedMoeWeightScaleSupported.BLOCK.value})
|
||||
{"quant_method": FusedMoeWeightScaleSupported.BLOCK.value}
|
||||
)
|
||||
|
||||
params_dtype = torch.uint8
|
||||
|
||||
# WEIGHTS
|
||||
w13_weight = torch.nn.Parameter(torch.empty(
|
||||
num_experts,
|
||||
2 * intermediate_size_per_partition,
|
||||
hidden_size // 2,
|
||||
dtype=params_dtype),
|
||||
requires_grad=False)
|
||||
w13_weight = torch.nn.Parameter(
|
||||
torch.empty(
|
||||
num_experts,
|
||||
2 * intermediate_size_per_partition,
|
||||
hidden_size // 2,
|
||||
dtype=params_dtype,
|
||||
),
|
||||
requires_grad=False,
|
||||
)
|
||||
layer.register_parameter("w13_weight", w13_weight)
|
||||
|
||||
set_weight_attrs(w13_weight, extra_weight_attrs)
|
||||
|
||||
w2_weight = torch.nn.Parameter(torch.empty(
|
||||
num_experts,
|
||||
hidden_size,
|
||||
intermediate_size_per_partition // 2,
|
||||
dtype=params_dtype),
|
||||
requires_grad=False)
|
||||
w2_weight = torch.nn.Parameter(
|
||||
torch.empty(
|
||||
num_experts,
|
||||
hidden_size,
|
||||
intermediate_size_per_partition // 2,
|
||||
dtype=params_dtype,
|
||||
),
|
||||
requires_grad=False,
|
||||
)
|
||||
layer.register_parameter("w2_weight", w2_weight)
|
||||
|
||||
set_weight_attrs(w2_weight, extra_weight_attrs)
|
||||
@@ -491,7 +550,8 @@ class QuarkW4A4MXFp4MoEMethod(QuarkMoEMethod):
|
||||
layer.register_parameter("w2_weight_scale", w2_weight_scale)
|
||||
|
||||
def get_fused_moe_quant_config(
|
||||
self, layer: torch.nn.Module) -> Optional[FusedMoEQuantConfig]:
|
||||
self, layer: torch.nn.Module
|
||||
) -> Optional[FusedMoEQuantConfig]:
|
||||
return mxfp4_w4a4_moe_quant_config(
|
||||
w1_scale=layer.w13_weight_scale,
|
||||
w2_scale=layer.w2_weight_scale,
|
||||
@@ -527,7 +587,8 @@ class QuarkW4A4MXFp4MoEMethod(QuarkMoEMethod):
|
||||
|
||||
if enable_eplb:
|
||||
raise NotImplementedError(
|
||||
"EPLB not supported for `QuarkW4A4MXFp4MoEMethod` yet.")
|
||||
"EPLB not supported for `QuarkW4A4MXFp4MoEMethod` yet."
|
||||
)
|
||||
|
||||
from vllm.model_executor.layers.fused_moe import fused_experts
|
||||
|
||||
@@ -543,7 +604,8 @@ class QuarkW4A4MXFp4MoEMethod(QuarkMoEMethod):
|
||||
scoring_func=scoring_func,
|
||||
routed_scaling_factor=routed_scaling_factor,
|
||||
e_score_correction_bias=e_score_correction_bias,
|
||||
indices_type=self.topk_indices_dtype)
|
||||
indices_type=self.topk_indices_dtype,
|
||||
)
|
||||
|
||||
out = fused_experts(
|
||||
x,
|
||||
|
||||
@@ -11,7 +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.
|
||||
"""
|
||||
|
||||
@@ -26,20 +26,21 @@ class QuarkScheme(ABC):
|
||||
@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: Optional[torch.Tensor]):
|
||||
def apply_weights(
|
||||
self, layer: torch.nn.Module, x: torch.Tensor, bias: Optional[torch.Tensor]
|
||||
):
|
||||
"""
|
||||
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.
|
||||
|
||||
:param layer: torch.nn.Module with the registered weights and
|
||||
other parameters relevant to the particular scheme.
|
||||
:param layer: torch.nn.Module with the registered weights and
|
||||
other parameters relevant to the particular scheme.
|
||||
:param x: input to the layer
|
||||
:param bias: bias parameter
|
||||
|
||||
|
||||
@@ -10,17 +10,21 @@ import torch.nn.functional as F
|
||||
from vllm import envs
|
||||
from vllm.model_executor.layers.quantization.quark.schemes import QuarkScheme
|
||||
from vllm.model_executor.layers.quantization.utils.mxfp4_utils import (
|
||||
OCP_MX_BLOCK_SIZE, dequant_mxfp4, quant_dequant_mxfp4)
|
||||
from vllm.model_executor.parameter import (GroupQuantScaleParameter,
|
||||
PackedvLLMParameter)
|
||||
OCP_MX_BLOCK_SIZE,
|
||||
dequant_mxfp4,
|
||||
quant_dequant_mxfp4,
|
||||
)
|
||||
from vllm.model_executor.parameter import GroupQuantScaleParameter, PackedvLLMParameter
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
|
||||
@cache
|
||||
def is_rocm_aiter_fp4_asm_gemm_enabled() -> bool:
|
||||
return current_platform.is_rocm() \
|
||||
and envs.VLLM_ROCM_USE_AITER_FP4_ASM_GEMM \
|
||||
return (
|
||||
current_platform.is_rocm()
|
||||
and envs.VLLM_ROCM_USE_AITER_FP4_ASM_GEMM
|
||||
and envs.VLLM_ROCM_USE_AITER
|
||||
)
|
||||
|
||||
|
||||
try:
|
||||
@@ -29,6 +33,7 @@ try:
|
||||
from aiter.ops.triton.quant import dynamic_mxfp4_quant
|
||||
|
||||
from vllm.utils import direct_register_custom_op
|
||||
|
||||
if is_rocm_aiter_fp4_asm_gemm_enabled():
|
||||
from aiter import gemm_a4w4, per_1x32_f4_quant_hip
|
||||
|
||||
@@ -51,17 +56,13 @@ try:
|
||||
|
||||
# 32 alignment is enough for dim0 padding of output for
|
||||
# gemm_a4w4 kernel
|
||||
y = torch.empty((M + 31) // 32 * 32,
|
||||
weight.shape[0],
|
||||
device=x_q.device,
|
||||
dtype=out_dtype)
|
||||
y = torch.empty(
|
||||
(M + 31) // 32 * 32, weight.shape[0], device=x_q.device, dtype=out_dtype
|
||||
)
|
||||
|
||||
gemm_a4w4(x_q,
|
||||
weight,
|
||||
x_s,
|
||||
weight_scale.view(x_s.dtype),
|
||||
y,
|
||||
bpreshuffle=True)
|
||||
gemm_a4w4(
|
||||
x_q, weight, x_s, weight_scale.view(x_s.dtype), y, bpreshuffle=True
|
||||
)
|
||||
return y[:M]
|
||||
else:
|
||||
if x_scales is None:
|
||||
@@ -69,10 +70,9 @@ try:
|
||||
else:
|
||||
x_q = x
|
||||
x_s = x_scales
|
||||
y = torch.empty(x_q.shape[0],
|
||||
weight.shape[0],
|
||||
device=x_q.device,
|
||||
dtype=out_dtype)
|
||||
y = torch.empty(
|
||||
x_q.shape[0], weight.shape[0], device=x_q.device, dtype=out_dtype
|
||||
)
|
||||
|
||||
gemm_afp4wfp4(x_q, weight, x_s, weight_scale.T, out_dtype, y)
|
||||
return y
|
||||
@@ -85,9 +85,9 @@ try:
|
||||
rocm_use_aiter_fp4_asm_gemm: bool = False,
|
||||
out_dtype: Optional[torch.dtype] = torch.bfloat16,
|
||||
) -> torch.Tensor:
|
||||
return torch.empty((*x.shape[:-1], weight.shape[0]),
|
||||
dtype=out_dtype,
|
||||
device=x.device)
|
||||
return torch.empty(
|
||||
(*x.shape[:-1], weight.shape[0]), dtype=out_dtype, device=x.device
|
||||
)
|
||||
|
||||
direct_register_custom_op(
|
||||
op_name="gemm_with_dynamic_quant",
|
||||
@@ -104,46 +104,45 @@ __all__ = ["QuarkW4A4MXFP4"]
|
||||
|
||||
|
||||
class QuarkW4A4MXFP4(QuarkScheme):
|
||||
|
||||
def __init__(self, weight_quant_spec: dict[str, Any],
|
||||
input_quant_spec: dict[str, Any]):
|
||||
def __init__(
|
||||
self, weight_quant_spec: dict[str, Any], input_quant_spec: dict[str, Any]
|
||||
):
|
||||
self.out_dtype = torch.get_default_dtype()
|
||||
self.qscheme = "per_group"
|
||||
self.weight_quant_spec = weight_quant_spec
|
||||
self.input_quant_spec = input_quant_spec
|
||||
self.emulate = not current_platform.supports_mx()
|
||||
self.rocm_use_aiter_fp4_asm_gemm = is_rocm_aiter_fp4_asm_gemm_enabled()
|
||||
if not self.emulate and (dynamic_mxfp4_quant is None
|
||||
or gemm_afp4wfp4 is None):
|
||||
if not self.emulate and (dynamic_mxfp4_quant is None or gemm_afp4wfp4 is None):
|
||||
# Currently need these kernels if not emulating
|
||||
raise NotImplementedError(
|
||||
f"{self.__class__.__name__} requires AITER to be installed "
|
||||
"for non-emulation mode! Please refer to "
|
||||
"https://github.com/ROCm/aiter for installation details.")
|
||||
"https://github.com/ROCm/aiter for installation details."
|
||||
)
|
||||
|
||||
@classmethod
|
||||
def get_min_capability(cls) -> int:
|
||||
return 70
|
||||
|
||||
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
|
||||
layer.weight = torch.nn.Parameter(layer.weight.data,
|
||||
requires_grad=False)
|
||||
layer.weight = torch.nn.Parameter(layer.weight.data, requires_grad=False)
|
||||
|
||||
if self.emulate:
|
||||
layer.weight_scale = torch.nn.Parameter(layer.weight_scale.data,
|
||||
requires_grad=False)
|
||||
layer.weight_scale = torch.nn.Parameter(
|
||||
layer.weight_scale.data, requires_grad=False
|
||||
)
|
||||
try:
|
||||
from quark.torch.export.nn.modules import realquantizer
|
||||
from quark.torch.quantization.config.config import (
|
||||
QuantizationSpec)
|
||||
from quark.torch.quantization.config.config import QuantizationSpec
|
||||
except ImportError as err:
|
||||
raise ImportError(
|
||||
"The package `amd-quark` is required to use AMD Quark "
|
||||
"MX-FP4 models. Please install it with `pip install "
|
||||
"amd-quark`.") from err
|
||||
"amd-quark`."
|
||||
) from err
|
||||
|
||||
weight_quant_spec = QuantizationSpec.from_dict(
|
||||
self.weight_quant_spec)
|
||||
weight_quant_spec = QuantizationSpec.from_dict(self.weight_quant_spec)
|
||||
|
||||
weight_quantizer = realquantizer.get_real_quantizer(
|
||||
qspec=weight_quant_spec,
|
||||
@@ -170,29 +169,34 @@ class QuarkW4A4MXFP4(QuarkScheme):
|
||||
weight_scale_shuffle = layer.weight_scale.data
|
||||
sm, sn = weight_scale_shuffle.shape
|
||||
weight_scale_shuffle = weight_scale_shuffle.view(
|
||||
sm // 32, 2, 16, sn // 8, 2, 4, 1)
|
||||
sm // 32, 2, 16, sn // 8, 2, 4, 1
|
||||
)
|
||||
weight_scale_shuffle = weight_scale_shuffle.permute(
|
||||
0, 3, 5, 2, 4, 1, 6).contiguous()
|
||||
0, 3, 5, 2, 4, 1, 6
|
||||
).contiguous()
|
||||
weight_scale_shuffle = weight_scale_shuffle.view(sm, sn)
|
||||
layer.weight_scale = torch.nn.Parameter(weight_scale_shuffle,
|
||||
requires_grad=False)
|
||||
layer.weight_scale = torch.nn.Parameter(
|
||||
weight_scale_shuffle, requires_grad=False
|
||||
)
|
||||
|
||||
# shuffle weight
|
||||
weight_shuffle = layer.weight.data
|
||||
weight_shuffle = shuffle_weight(weight_shuffle,
|
||||
layout=(16, 16))
|
||||
layer.weight = torch.nn.Parameter(weight_shuffle,
|
||||
requires_grad=False)
|
||||
weight_shuffle = shuffle_weight(weight_shuffle, layout=(16, 16))
|
||||
layer.weight = torch.nn.Parameter(weight_shuffle, requires_grad=False)
|
||||
else:
|
||||
layer.weight_scale = torch.nn.Parameter(
|
||||
layer.weight_scale.data.T.contiguous(),
|
||||
requires_grad=False)
|
||||
layer.weight_scale.data.T.contiguous(), requires_grad=False
|
||||
)
|
||||
|
||||
def create_weights(self, layer: torch.nn.Module,
|
||||
output_partition_sizes: list[int],
|
||||
input_size_per_partition: int,
|
||||
params_dtype: torch.dtype, weight_loader: Callable,
|
||||
**kwargs):
|
||||
def create_weights(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
output_partition_sizes: list[int],
|
||||
input_size_per_partition: int,
|
||||
params_dtype: torch.dtype,
|
||||
weight_loader: Callable,
|
||||
**kwargs,
|
||||
):
|
||||
output_size_per_partition = sum(output_partition_sizes)
|
||||
layer.logical_widths = output_partition_sizes
|
||||
|
||||
@@ -224,16 +228,21 @@ class QuarkW4A4MXFP4(QuarkScheme):
|
||||
)
|
||||
layer.register_parameter("weight_scale", weight_scale)
|
||||
|
||||
def apply_weights(self,
|
||||
layer: torch.nn.Module,
|
||||
x: torch.Tensor,
|
||||
bias: Optional[torch.Tensor] = None) -> torch.Tensor:
|
||||
|
||||
def apply_weights(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
x: torch.Tensor,
|
||||
bias: Optional[torch.Tensor] = None,
|
||||
) -> torch.Tensor:
|
||||
if self.emulate:
|
||||
dq_w = dequant_mxfp4(layer.weight, layer.weight_scale, x.dtype)
|
||||
x = quant_dequant_mxfp4(x)
|
||||
return F.linear(x, dq_w, bias)
|
||||
else:
|
||||
return torch.ops.vllm.gemm_with_dynamic_quant(
|
||||
x, layer.weight, layer.weight_scale,
|
||||
self.rocm_use_aiter_fp4_asm_gemm, self.out_dtype)
|
||||
x,
|
||||
layer.weight,
|
||||
layer.weight_scale,
|
||||
self.rocm_use_aiter_fp4_asm_gemm,
|
||||
self.out_dtype,
|
||||
)
|
||||
|
||||
@@ -7,37 +7,43 @@ import torch
|
||||
from torch.nn import Parameter
|
||||
|
||||
from vllm.model_executor.layers.quantization.quark.schemes import QuarkScheme
|
||||
from vllm.model_executor.layers.quantization.utils.quant_utils import (
|
||||
GroupShape)
|
||||
from vllm.model_executor.layers.quantization.utils.quant_utils import GroupShape
|
||||
from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
|
||||
Fp8LinearOp, normalize_e4m3fn_to_e4m3fnuz, requantize_with_max_scale)
|
||||
from vllm.model_executor.parameter import (ChannelQuantScaleParameter,
|
||||
ModelWeightParameter,
|
||||
PerTensorScaleParameter)
|
||||
Fp8LinearOp,
|
||||
normalize_e4m3fn_to_e4m3fnuz,
|
||||
requantize_with_max_scale,
|
||||
)
|
||||
from vllm.model_executor.parameter import (
|
||||
ChannelQuantScaleParameter,
|
||||
ModelWeightParameter,
|
||||
PerTensorScaleParameter,
|
||||
)
|
||||
from vllm.platforms import current_platform
|
||||
|
||||
__all__ = ["QuarkW8A8Fp8"]
|
||||
|
||||
|
||||
class QuarkW8A8Fp8(QuarkScheme):
|
||||
|
||||
def __init__(self, weight_config: dict[str, Any],
|
||||
input_config: Optional[dict[str, Any]]):
|
||||
def __init__(
|
||||
self, weight_config: dict[str, Any], input_config: Optional[dict[str, Any]]
|
||||
):
|
||||
self.weight_qscheme = cast(str, weight_config.get("qscheme"))
|
||||
self.is_static_input_scheme: bool = False
|
||||
self.input_qscheme: Optional[str] = None
|
||||
if input_config is not None:
|
||||
self.is_static_input_scheme = not cast(
|
||||
bool, input_config.get("is_dynamic"))
|
||||
self.is_static_input_scheme = not cast(bool, input_config.get("is_dynamic"))
|
||||
self.input_qscheme = cast(str, input_config.get("qscheme"))
|
||||
|
||||
per_token = (not self.is_static_input_scheme
|
||||
and self.input_qscheme == "per_channel")
|
||||
self.act_quant_group_shape = GroupShape.PER_TOKEN \
|
||||
if per_token else GroupShape.PER_TENSOR
|
||||
per_token = (
|
||||
not self.is_static_input_scheme and self.input_qscheme == "per_channel"
|
||||
)
|
||||
self.act_quant_group_shape = (
|
||||
GroupShape.PER_TOKEN if per_token else GroupShape.PER_TENSOR
|
||||
)
|
||||
self.fp8_linear = Fp8LinearOp(
|
||||
act_quant_static=self.is_static_input_scheme,
|
||||
act_quant_group_shape=self.act_quant_group_shape)
|
||||
act_quant_group_shape=self.act_quant_group_shape,
|
||||
)
|
||||
self.out_dtype = torch.get_default_dtype()
|
||||
|
||||
@classmethod
|
||||
@@ -51,14 +57,14 @@ class QuarkW8A8Fp8(QuarkScheme):
|
||||
# requantize so we can always run per tensor
|
||||
if self.weight_qscheme == "per_tensor":
|
||||
if current_platform.is_fp8_fnuz():
|
||||
input_scale = getattr(layer, 'input_scale', None)
|
||||
input_scale = getattr(layer, "input_scale", None)
|
||||
weight, max_w_scale, input_scale = normalize_e4m3fn_to_e4m3fnuz(
|
||||
weight=layer.weight,
|
||||
weight_scale=layer.weight_scale,
|
||||
input_scale=input_scale)
|
||||
input_scale=input_scale,
|
||||
)
|
||||
if input_scale is not None:
|
||||
layer.input_scale = Parameter(input_scale,
|
||||
requires_grad=False)
|
||||
layer.input_scale = Parameter(input_scale, requires_grad=False)
|
||||
else:
|
||||
max_w_scale = layer.weight_scale
|
||||
weight = layer.weight
|
||||
@@ -77,15 +83,14 @@ class QuarkW8A8Fp8(QuarkScheme):
|
||||
weight = layer.weight
|
||||
|
||||
if current_platform.is_fp8_fnuz():
|
||||
input_scale = getattr(layer, 'input_scale', None)
|
||||
weight, weight_scale, input_scale = \
|
||||
normalize_e4m3fn_to_e4m3fnuz(
|
||||
weight=weight,
|
||||
weight_scale=layer.weight_scale,
|
||||
input_scale=input_scale)
|
||||
input_scale = getattr(layer, "input_scale", None)
|
||||
weight, weight_scale, input_scale = normalize_e4m3fn_to_e4m3fnuz(
|
||||
weight=weight,
|
||||
weight_scale=layer.weight_scale,
|
||||
input_scale=input_scale,
|
||||
)
|
||||
if input_scale is not None:
|
||||
layer.input_scale = Parameter(input_scale,
|
||||
requires_grad=False)
|
||||
layer.input_scale = Parameter(input_scale, requires_grad=False)
|
||||
else:
|
||||
weight_scale = layer.weight_scale.data
|
||||
if self.act_quant_group_shape == GroupShape.PER_TOKEN:
|
||||
@@ -95,32 +100,37 @@ class QuarkW8A8Fp8(QuarkScheme):
|
||||
layer.weight_scale = Parameter(weight_scale, requires_grad=False)
|
||||
|
||||
else:
|
||||
raise ValueError(
|
||||
f"Unknown quantization scheme {self.weight_qscheme}")
|
||||
raise ValueError(f"Unknown quantization scheme {self.weight_qscheme}")
|
||||
|
||||
# INPUT SCALE
|
||||
if self.is_static_input_scheme:
|
||||
layer.input_scale = Parameter(layer.input_scale.max(),
|
||||
requires_grad=False)
|
||||
layer.input_scale = Parameter(layer.input_scale.max(), requires_grad=False)
|
||||
else:
|
||||
layer.input_scale = None
|
||||
|
||||
def create_weights(self, layer: torch.nn.Module,
|
||||
output_partition_sizes: list[int],
|
||||
input_size_per_partition: int,
|
||||
params_dtype: torch.dtype, weight_loader: Callable,
|
||||
**kwargs):
|
||||
def create_weights(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
output_partition_sizes: list[int],
|
||||
input_size_per_partition: int,
|
||||
params_dtype: torch.dtype,
|
||||
weight_loader: Callable,
|
||||
**kwargs,
|
||||
):
|
||||
output_size_per_partition = sum(output_partition_sizes)
|
||||
layer.logical_widths = output_partition_sizes
|
||||
|
||||
# WEIGHT
|
||||
weight = ModelWeightParameter(data=torch.empty(
|
||||
output_size_per_partition,
|
||||
input_size_per_partition,
|
||||
dtype=torch.float8_e4m3fn),
|
||||
input_dim=1,
|
||||
output_dim=0,
|
||||
weight_loader=weight_loader)
|
||||
weight = ModelWeightParameter(
|
||||
data=torch.empty(
|
||||
output_size_per_partition,
|
||||
input_size_per_partition,
|
||||
dtype=torch.float8_e4m3fn,
|
||||
),
|
||||
input_dim=1,
|
||||
output_dim=0,
|
||||
weight_loader=weight_loader,
|
||||
)
|
||||
layer.register_parameter("weight", weight)
|
||||
|
||||
# WEIGHT SCALE
|
||||
@@ -128,15 +138,16 @@ class QuarkW8A8Fp8(QuarkScheme):
|
||||
# the newly added parameters
|
||||
if self.weight_qscheme == "per_channel":
|
||||
weight_scale = ChannelQuantScaleParameter(
|
||||
data=torch.empty((sum(output_partition_sizes)),
|
||||
dtype=torch.float32),
|
||||
data=torch.empty((sum(output_partition_sizes)), dtype=torch.float32),
|
||||
output_dim=0,
|
||||
weight_loader=weight_loader)
|
||||
weight_loader=weight_loader,
|
||||
)
|
||||
else:
|
||||
assert self.weight_qscheme == "per_tensor"
|
||||
weight_scale = PerTensorScaleParameter(data=torch.empty(
|
||||
len(output_partition_sizes), dtype=torch.float32),
|
||||
weight_loader=weight_loader)
|
||||
weight_scale = PerTensorScaleParameter(
|
||||
data=torch.empty(len(output_partition_sizes), dtype=torch.float32),
|
||||
weight_loader=weight_loader,
|
||||
)
|
||||
|
||||
# min requirement for fp8 kernels
|
||||
weight_scale[:] = torch.finfo(torch.float32).min
|
||||
@@ -144,20 +155,24 @@ class QuarkW8A8Fp8(QuarkScheme):
|
||||
|
||||
# INPUT SCALE
|
||||
if self.is_static_input_scheme:
|
||||
input_scale = PerTensorScaleParameter(data=torch.empty(
|
||||
len(output_partition_sizes), dtype=torch.float32),
|
||||
weight_loader=weight_loader)
|
||||
input_scale = PerTensorScaleParameter(
|
||||
data=torch.empty(len(output_partition_sizes), dtype=torch.float32),
|
||||
weight_loader=weight_loader,
|
||||
)
|
||||
input_scale[:] = torch.finfo(torch.float32).min
|
||||
layer.register_parameter("input_scale", input_scale)
|
||||
|
||||
def apply_weights(self,
|
||||
layer: torch.nn.Module,
|
||||
x: torch.Tensor,
|
||||
bias: Optional[torch.Tensor] = None) -> torch.Tensor:
|
||||
|
||||
return self.fp8_linear.apply(input=x,
|
||||
weight=layer.weight,
|
||||
weight_scale=layer.weight_scale,
|
||||
out_dtype=self.out_dtype,
|
||||
input_scale=layer.input_scale,
|
||||
bias=bias)
|
||||
def apply_weights(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
x: torch.Tensor,
|
||||
bias: Optional[torch.Tensor] = None,
|
||||
) -> torch.Tensor:
|
||||
return self.fp8_linear.apply(
|
||||
input=x,
|
||||
weight=layer.weight,
|
||||
weight_scale=layer.weight_scale,
|
||||
out_dtype=self.out_dtype,
|
||||
input_scale=layer.input_scale,
|
||||
bias=bias,
|
||||
)
|
||||
|
||||
@@ -7,12 +7,16 @@ import torch
|
||||
|
||||
from vllm.logger import init_logger
|
||||
from vllm.model_executor.layers.quantization.kernels.scaled_mm import (
|
||||
ScaledMMLinearLayerConfig, choose_scaled_mm_linear_kernel)
|
||||
ScaledMMLinearLayerConfig,
|
||||
choose_scaled_mm_linear_kernel,
|
||||
)
|
||||
from vllm.model_executor.layers.quantization.quark.schemes import QuarkScheme
|
||||
from vllm.model_executor.parameter import (BasevLLMParameter,
|
||||
ChannelQuantScaleParameter,
|
||||
ModelWeightParameter,
|
||||
PerTensorScaleParameter)
|
||||
from vllm.model_executor.parameter import (
|
||||
BasevLLMParameter,
|
||||
ChannelQuantScaleParameter,
|
||||
ModelWeightParameter,
|
||||
PerTensorScaleParameter,
|
||||
)
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
@@ -20,8 +24,12 @@ logger = init_logger(__name__)
|
||||
class QuarkW8A8Int8(QuarkScheme):
|
||||
_kernel_backends_being_used: set[str] = set()
|
||||
|
||||
def __init__(self, qscheme: str, is_static_input_scheme: Optional[bool],
|
||||
input_symmetric: Optional[bool]):
|
||||
def __init__(
|
||||
self,
|
||||
qscheme: str,
|
||||
is_static_input_scheme: Optional[bool],
|
||||
input_symmetric: Optional[bool],
|
||||
):
|
||||
self.qscheme = qscheme
|
||||
self.is_static_input_scheme = is_static_input_scheme
|
||||
self.input_symmetric = input_symmetric
|
||||
@@ -31,92 +39,101 @@ class QuarkW8A8Int8(QuarkScheme):
|
||||
# turing and up
|
||||
return 75
|
||||
|
||||
def create_weights(self, layer: torch.nn.Module,
|
||||
output_partition_sizes: list[int],
|
||||
input_size_per_partition: int,
|
||||
params_dtype: torch.dtype, weight_loader: Callable,
|
||||
**kwargs):
|
||||
def create_weights(
|
||||
self,
|
||||
layer: torch.nn.Module,
|
||||
output_partition_sizes: list[int],
|
||||
input_size_per_partition: int,
|
||||
params_dtype: torch.dtype,
|
||||
weight_loader: Callable,
|
||||
**kwargs,
|
||||
):
|
||||
layer.logical_widths = output_partition_sizes
|
||||
|
||||
scaled_mm_linear_kernel_config = ScaledMMLinearLayerConfig(
|
||||
is_channelwise=(self.qscheme == "per_channel"),
|
||||
is_static_input_scheme=(self.is_static_input_scheme is True),
|
||||
input_symmetric=(self.input_symmetric is True))
|
||||
input_symmetric=(self.input_symmetric is True),
|
||||
)
|
||||
|
||||
kernel_type = choose_scaled_mm_linear_kernel(
|
||||
scaled_mm_linear_kernel_config)
|
||||
kernel_type = choose_scaled_mm_linear_kernel(scaled_mm_linear_kernel_config)
|
||||
|
||||
if kernel_type.__name__ not in self._kernel_backends_being_used:
|
||||
logger.info("Using %s for QuarkW8A8Int8", kernel_type.__name__)
|
||||
self._kernel_backends_being_used.add(kernel_type.__name__)
|
||||
|
||||
# WEIGHT
|
||||
weight = ModelWeightParameter(data=torch.empty(
|
||||
sum(output_partition_sizes),
|
||||
input_size_per_partition,
|
||||
dtype=torch.int8),
|
||||
input_dim=1,
|
||||
output_dim=0,
|
||||
weight_loader=weight_loader)
|
||||
weight = ModelWeightParameter(
|
||||
data=torch.empty(
|
||||
sum(output_partition_sizes), input_size_per_partition, dtype=torch.int8
|
||||
),
|
||||
input_dim=1,
|
||||
output_dim=0,
|
||||
weight_loader=weight_loader,
|
||||
)
|
||||
|
||||
layer.register_parameter("weight", weight)
|
||||
|
||||
# WEIGHT SCALE
|
||||
if self.qscheme == "per_channel":
|
||||
weight_scale = ChannelQuantScaleParameter(
|
||||
data=torch.empty((sum(output_partition_sizes)),
|
||||
dtype=torch.float32),
|
||||
data=torch.empty((sum(output_partition_sizes)), dtype=torch.float32),
|
||||
output_dim=0,
|
||||
weight_loader=weight_loader)
|
||||
weight_loader=weight_loader,
|
||||
)
|
||||
ChannelQuantZPParameter = ChannelQuantScaleParameter
|
||||
weight_zero_point = ChannelQuantZPParameter(
|
||||
data=torch.empty((sum(output_partition_sizes)),
|
||||
dtype=torch.int8),
|
||||
data=torch.empty((sum(output_partition_sizes)), dtype=torch.int8),
|
||||
output_dim=0,
|
||||
weight_loader=weight_loader)
|
||||
weight_loader=weight_loader,
|
||||
)
|
||||
else:
|
||||
assert self.qscheme == "per_tensor"
|
||||
weight_scale = PerTensorScaleParameter(data=torch.empty(
|
||||
len(output_partition_sizes), dtype=torch.float32),
|
||||
weight_loader=weight_loader)
|
||||
weight_scale = PerTensorScaleParameter(
|
||||
data=torch.empty(len(output_partition_sizes), dtype=torch.float32),
|
||||
weight_loader=weight_loader,
|
||||
)
|
||||
PerTensorZPParameter = PerTensorScaleParameter
|
||||
weight_zero_point = PerTensorZPParameter(
|
||||
data=torch.empty(len(output_partition_sizes),
|
||||
dtype=torch.int8),
|
||||
weight_loader=weight_loader)
|
||||
data=torch.empty(len(output_partition_sizes), dtype=torch.int8),
|
||||
weight_loader=weight_loader,
|
||||
)
|
||||
layer.register_parameter("weight_scale", weight_scale)
|
||||
layer.register_parameter("weight_zero_point", weight_zero_point)
|
||||
|
||||
# INPUT SCALE
|
||||
if self.is_static_input_scheme:
|
||||
input_scale = BasevLLMParameter(data=torch.empty(
|
||||
1, dtype=torch.float32),
|
||||
weight_loader=weight_loader)
|
||||
input_scale = BasevLLMParameter(
|
||||
data=torch.empty(1, dtype=torch.float32), weight_loader=weight_loader
|
||||
)
|
||||
layer.register_parameter("input_scale", input_scale)
|
||||
|
||||
input_zero_point = BasevLLMParameter(data=torch.empty(
|
||||
1, dtype=torch.int8),
|
||||
weight_loader=weight_loader)
|
||||
input_zero_point = BasevLLMParameter(
|
||||
data=torch.empty(1, dtype=torch.int8), weight_loader=weight_loader
|
||||
)
|
||||
layer.register_parameter("input_zero_point", input_zero_point)
|
||||
|
||||
self.kernel = kernel_type(c=scaled_mm_linear_kernel_config,
|
||||
w_q_param_name="weight",
|
||||
w_s_param_name="weight_scale",
|
||||
i_s_param_name="input_scale",
|
||||
i_zp_param_name="input_zero_point",
|
||||
azp_adj_param_name="azp_adj")
|
||||
self.kernel = kernel_type(
|
||||
c=scaled_mm_linear_kernel_config,
|
||||
w_q_param_name="weight",
|
||||
w_s_param_name="weight_scale",
|
||||
i_s_param_name="input_scale",
|
||||
i_zp_param_name="input_zero_point",
|
||||
azp_adj_param_name="azp_adj",
|
||||
)
|
||||
|
||||
# Checkpoints are serialized in quark format, which is
|
||||
# different from the format the kernel may want. Handle repacking here.
|
||||
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
|
||||
layer.register_parameter("weight_zero_point", None)
|
||||
delattr(layer, 'weight_zero_point')
|
||||
delattr(layer, "weight_zero_point")
|
||||
if self.input_symmetric:
|
||||
layer.register_parameter("input_zero_point", None)
|
||||
delattr(layer, 'input_zero_point')
|
||||
delattr(layer, "input_zero_point")
|
||||
|
||||
self.kernel.process_weights_after_loading(layer)
|
||||
|
||||
def apply_weights(self, layer: torch.nn.Module, x: torch.Tensor,
|
||||
bias: Optional[torch.Tensor]) -> torch.Tensor:
|
||||
def apply_weights(
|
||||
self, layer: torch.nn.Module, x: torch.Tensor, bias: Optional[torch.Tensor]
|
||||
) -> torch.Tensor:
|
||||
return self.kernel.apply_weights(layer, x, bias)
|
||||
|
||||
@@ -24,7 +24,7 @@ def deep_compare(dict1: Any, dict2: Any) -> bool:
|
||||
def should_ignore_layer(
|
||||
layer_name: Optional[str],
|
||||
ignore: Iterable[str],
|
||||
fused_mapping: Mapping[str, list[str]] = MappingProxyType({})
|
||||
fused_mapping: Mapping[str, list[str]] = MappingProxyType({}),
|
||||
) -> bool:
|
||||
if layer_name is None:
|
||||
return False
|
||||
@@ -50,7 +50,8 @@ def should_ignore_layer(
|
||||
should_ignore_layer = None
|
||||
for shard_name in shard_names:
|
||||
should_ignore_shard = check_equal_or_regex_match(
|
||||
layer_name=shard_name, targets=ignore)
|
||||
layer_name=shard_name, targets=ignore
|
||||
)
|
||||
|
||||
# If shard_idx=0, set layer ignore to match shard.
|
||||
if should_ignore_layer is None:
|
||||
@@ -58,24 +59,26 @@ def should_ignore_layer(
|
||||
|
||||
# If shard_idx=1+ confirm scheme matches prior shards.
|
||||
elif should_ignore_shard != should_ignore_layer:
|
||||
raise ValueError(f"Found a different quantization schemes for "
|
||||
f"{shard_proj_names} in {layer_name}. vLLM "
|
||||
"requires all to use the same scheme.")
|
||||
raise ValueError(
|
||||
f"Found a different quantization schemes for "
|
||||
f"{shard_proj_names} in {layer_name}. vLLM "
|
||||
"requires all to use the same scheme."
|
||||
)
|
||||
|
||||
# Unfused layers like down_proj and o_proj will match
|
||||
# the safetensors checkpoint already.
|
||||
else:
|
||||
should_ignore_layer = check_equal_or_regex_match(layer_name=layer_name,
|
||||
targets=ignore)
|
||||
should_ignore_layer = check_equal_or_regex_match(
|
||||
layer_name=layer_name, targets=ignore
|
||||
)
|
||||
|
||||
assert should_ignore_layer is not None
|
||||
return should_ignore_layer
|
||||
|
||||
|
||||
def check_equal_or_regex_match(layer_name: str,
|
||||
targets: Iterable[str]) -> bool:
|
||||
def check_equal_or_regex_match(layer_name: str, targets: Iterable[str]) -> bool:
|
||||
"""
|
||||
Checks whether a layer_name is exactly equal or a regex match for
|
||||
Checks whether a layer_name is exactly equal or a regex match for
|
||||
if target starts with 're:' to any target in list.
|
||||
"""
|
||||
for target in targets:
|
||||
@@ -84,9 +87,9 @@ def check_equal_or_regex_match(layer_name: str,
|
||||
return False
|
||||
|
||||
|
||||
def _is_equal_or_regex_match(value: str,
|
||||
target: str,
|
||||
check_contains: bool = False) -> bool:
|
||||
def _is_equal_or_regex_match(
|
||||
value: str, target: str, check_contains: bool = False
|
||||
) -> bool:
|
||||
"""
|
||||
Checks whether a value is exactly equal or a regex match for target
|
||||
if target starts with 're:'. If check_contains is set to True,
|
||||
|
||||
Reference in New Issue
Block a user