vllm-cpu 0.8.5.post2__cp310-cp310-manylinux_2_17_x86_64.whl

vllm/model_executor/layers/quantization/quark/quark_moe.py

@@ -0,0 +1,236 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Any, Callable, Dict, Optional
+
+import torch
+
+import vllm.model_executor.layers.fused_moe  # noqa
+from vllm import _custom_ops as ops
+from vllm.logger import init_logger
+from vllm.model_executor.layers.fused_moe import (FusedMoE, FusedMoEMethodBase,
+                                                  FusedMoeWeightScaleSupported)
+from vllm.model_executor.layers.quantization.utils.w8a8_utils import (
+    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
+
+logger = init_logger(__name__)
+
+__all__ = ["QuarkMoEMethod", "QuarkW8A8Fp8MoEMethod"]
+
+
+class QuarkMoEMethod(FusedMoEMethodBase):
+
+    @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)
+
+        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)
+        else:
+            raise RuntimeError("Unsupported FusedMoe scheme")
+
+
+class QuarkW8A8Fp8MoEMethod(QuarkMoEMethod):
+
+    def __init__(self, weight_config: Dict[str, Any], input_config: Dict[str,
+                                                                         Any]):
+        self.weight_quant = weight_config
+        self.input_quant = input_config
+
+        weight_qscheme = self.weight_quant.get("qscheme")
+        input_qscheme = self.input_quant.get("qscheme")
+        if not (weight_qscheme == "per_tensor"
+                and input_qscheme == "per_tensor"):
+            raise ValueError(
+                "For FP8 Fused MoE layers, only per-tensor scales "
+                "for weights and activations are supported. Found "
+                f"{weight_qscheme}, {input_qscheme}")  # noqa E501
+
+        self.static_input_scales = not self.input_quant.get("is_dynamic")
+
+    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):
+
+        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)
+        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)
+        layer.register_parameter("w2_weight", w2_weight)
+        set_weight_attrs(w2_weight, extra_weight_attrs)
+
+        # WEIGHT_SCALES
+        # Allocate 2 scales for w1 and w3 respectively.
+        # They will be 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)
+        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)
+        layer.register_parameter("w2_weight_scale", w2_weight_scale)
+        # 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.TENSOR.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)
+            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)
+            layer.register_parameter("w2_input_scale", w2_input_scale)
+            set_weight_attrs(w2_input_scale, extra_weight_attrs)
+        else:
+            layer.w13_input_scale = None
+            layer.w2_input_scale = None
+
+    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
+        # 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):
+                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)):
+                logger.warning_once(
+                    "Found input_scales that are not equal for "
+                    "fp8 MoE layer. Using the maximum across experts "
+                    "for each layer. ")
+            layer.w13_input_scale = torch.nn.Parameter(
+                layer.w13_input_scale.max(), requires_grad=False)
+            layer.w2_input_scale = torch.nn.Parameter(
+                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 = \
+                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)
+            # 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)
+            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)
+            if w2_input_scale is not None:
+                layer.w2_input_scale = torch.nn.Parameter(w2_input_scale,
+                                                          requires_grad=False)
+
+        # Fp8 moe kernel needs single weight scale for w13 per expert.
+        # We take the max then dequant and requant each expert.
+        assert layer.w13_weight_scale is not None
+        shard_size = layer.intermediate_size_per_partition
+        max_w13_scales = layer.w13_weight_scale.max(dim=1).values
+        for expert_id in range(layer.local_num_experts):
+            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])
+                start += shard_size
+
+        layer.w13_weight_scale = torch.nn.Parameter(max_w13_scales,
+                                                    requires_grad=False)
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        router_logits: torch.Tensor,
+        top_k: int,
+        renormalize: bool,
+        use_grouped_topk: bool = False,
+        topk_group: Optional[int] = None,
+        num_expert_group: Optional[int] = None,
+        global_num_experts: int = -1,
+        expert_map: Optional[torch.Tensor] = None,
+        custom_routing_function: Optional[Callable] = None,
+        scoring_func: str = "softmax",
+        e_score_correction_bias: Optional[torch.Tensor] = None,
+        apply_router_weight_on_input: bool = False,
+        activation: str = "silu",
+    ) -> torch.Tensor:
+        from vllm.model_executor.layers.fused_moe import fused_experts
+
+        topk_weights, topk_ids = FusedMoE.select_experts(
+            hidden_states=x,
+            router_logits=router_logits,
+            use_grouped_topk=use_grouped_topk,
+            top_k=top_k,
+            renormalize=renormalize,
+            topk_group=topk_group,
+            num_expert_group=num_expert_group,
+            custom_routing_function=custom_routing_function,
+            scoring_func=scoring_func,
+            e_score_correction_bias=e_score_correction_bias)
+
+        return fused_experts(
+            x,
+            layer.w13_weight,
+            layer.w2_weight,
+            topk_weights=topk_weights,
+            topk_ids=topk_ids,
+            inplace=True,
+            use_fp8_w8a8=True,
+            global_num_experts=global_num_experts,
+            apply_router_weight_on_input=apply_router_weight_on_input,
+            expert_map=expert_map,
+            w1_scale=layer.w13_weight_scale,
+            w2_scale=layer.w2_weight_scale,
+            a1_scale=layer.w13_input_scale,
+            a2_scale=layer.w2_input_scale)

vllm/model_executor/layers/quantization/quark/schemes/__init__.py

@@ -0,0 +1,7 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from .quark_scheme import QuarkScheme
+from .quark_w8a8_fp8 import QuarkW8A8Fp8
+from .quark_w8a8_int8 import QuarkW8A8Int8
+
+__all__ = ["QuarkScheme", "QuarkW8A8Fp8", "QuarkW8A8Int8"]

vllm/model_executor/layers/quantization/quark/schemes/quark_scheme.py

@@ -0,0 +1,54 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from abc import ABC, abstractmethod
+from typing import Optional
+
+import torch
+
+__all__ = ["QuarkScheme"]
+
+
+class QuarkScheme(ABC):
+    """
+    Abstract class used to describe the weight creation and forward pass 
+    of different quantization schemes supported by Quark.
+    """
+
+    @classmethod
+    @abstractmethod
+    def get_min_capability(cls) -> int:
+        """
+        Get minimum device capability.
+        """
+        raise NotImplementedError
+
+    @abstractmethod
+    def create_weights(self, *args, **kwargs):
+        """
+        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]):
+        """
+        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 x: input to the layer
+        :param 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
+        needs to occur.
+        """
+        raise NotImplementedError

vllm/model_executor/layers/quantization/quark/schemes/quark_w8a8_fp8.py

@@ -0,0 +1,142 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Callable, List, Optional
+
+import torch
+from torch.nn import Parameter
+
+from vllm.model_executor.layers.quantization.quark.schemes import QuarkScheme
+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)
+from vllm.platforms import current_platform
+
+__all__ = ["QuarkW8A8Fp8"]
+
+
+class QuarkW8A8Fp8(QuarkScheme):
+
+    def __init__(self, qscheme: str, is_static_input_scheme: Optional[bool]):
+        self.qscheme = qscheme
+        self.is_static_input_scheme = is_static_input_scheme
+        self.fp8_linear = Fp8LinearOp(use_per_token_if_dynamic=False)
+        self.out_dtype = torch.get_default_dtype()
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        # lovelace and up
+        return 89
+
+    def process_weights_after_loading(self, layer) -> None:
+        # If per tensor, when we have a fused module (e.g. QKV) with per
+        # tensor scales (thus N scales being passed to the kernel),
+        # requantize so we can always run per tensor
+        if self.qscheme == "per_tensor":
+            max_w_scale, weight = requantize_with_max_scale(
+                weight=layer.weight,
+                weight_scale=layer.weight_scale,
+                logical_widths=layer.logical_widths,
+            )
+
+            if current_platform.is_fp8_fnuz():
+                input_scale = getattr(layer, 'input_scale', None)
+                weight, max_w_scale, input_scale = normalize_e4m3fn_to_e4m3fnuz(
+                    weight=weight,
+                    weight_scale=max_w_scale,
+                    input_scale=input_scale)
+                if input_scale is not None:
+                    layer.input_scale = Parameter(input_scale,
+                                                  requires_grad=False)
+
+            layer.weight = Parameter(weight.t(), requires_grad=False)
+            layer.weight_scale = Parameter(max_w_scale, requires_grad=False)
+
+        # If channelwise, scales are already lined up, so just transpose.
+        elif self.qscheme == "per_channel":
+            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)
+                if input_scale is not None:
+                    layer.input_scale = Parameter(input_scale,
+                                                  requires_grad=False)
+            else:
+                weight_scale = layer.weight_scale.data
+
+            layer.weight = Parameter(weight.t(), requires_grad=False)
+            # required by torch.compile to be torch.nn.Parameter
+            layer.weight_scale = Parameter(weight_scale, requires_grad=False)
+
+        else:
+            raise ValueError(f"Unknown quantization scheme {self.qscheme}")
+
+        # INPUT SCALE
+        if self.is_static_input_scheme:
+            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):
+        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)
+        layer.register_parameter("weight", weight)
+
+        # WEIGHT SCALE
+        # TODO: update create_xxx_parameter functions to return
+        # the newly added parameters
+        if self.qscheme == "per_channel":
+            weight_scale = ChannelQuantScaleParameter(
+                data=torch.empty((sum(output_partition_sizes)),
+                                 dtype=torch.float32),
+                output_dim=0,
+                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)
+
+        # min requirement for fp8 kernels
+        weight_scale[:] = torch.finfo(torch.float32).min
+        layer.register_parameter("weight_scale", weight_scale)
+
+        # 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[:] = 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)

vllm/model_executor/layers/quantization/quark/schemes/quark_w8a8_int8.py

@@ -0,0 +1,121 @@
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Callable, List, Optional, Set
+
+import torch
+
+from vllm.logger import init_logger
+from vllm.model_executor.layers.quantization.kernels.scaled_mm import (
+    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)
+
+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]):
+        self.qscheme = qscheme
+        self.is_static_input_scheme = is_static_input_scheme
+        self.input_symmetric = input_symmetric
+
+    @classmethod
+    def get_min_capability(cls) -> int:
+        # 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):
+        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))
+
+        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)
+
+        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),
+                output_dim=0,
+                weight_loader=weight_loader)
+            ChannelQuantZPParameter = ChannelQuantScaleParameter
+            weight_zero_point = ChannelQuantZPParameter(
+                data=torch.empty((sum(output_partition_sizes)),
+                                 dtype=torch.int8),
+                output_dim=0,
+                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)
+            PerTensorZPParameter = PerTensorScaleParameter
+            weight_zero_point = PerTensorZPParameter(
+                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)
+            layer.register_parameter("input_scale", input_scale)
+
+            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")
+
+    # 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')
+        if self.input_symmetric:
+            layer.register_parameter("input_zero_point", None)
+            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:
+        return self.kernel.apply_weights(layer, x, bias)

vllm/model_executor/layers/quantization/quark/utils.py

@@ -0,0 +1,102 @@
+# SPDX-License-Identifier: Apache-2.0
+
+import re
+from types import MappingProxyType
+from typing import Any, Iterable, List, Mapping, Optional
+
+
+def deep_compare(dict1: Any, dict2: Any) -> bool:
+    if type(dict1) is not type(dict2):
+        return False
+    if isinstance(dict1, dict):
+        if dict1.keys() != dict2.keys():
+            return False
+        return all(deep_compare(dict1[k], dict2[k]) for k in dict1)
+    elif isinstance(dict1, list):
+        return set(dict1) == set(dict2)
+    else:
+        return dict1 == dict2
+
+
+def should_ignore_layer(
+    layer_name: Optional[str],
+    ignore: Iterable[str],
+    fused_mapping: Mapping[str, List[str]] = MappingProxyType({})
+) -> bool:
+    if layer_name is None:
+        return False
+
+    # layer_name = model.layers.0.self_attn.qkv_proj
+    # proj_name = qkv_proj
+    proj_name = layer_name.split(".")[-1]
+
+    # Fused layers like gate_up_proj or qkv_proj will not be fused
+    # in the safetensors checkpoint. So, we convert the name
+    # from the fused version to unfused + check to make sure that
+    # each shard of the fused layer has the same scheme.
+    if proj_name in fused_mapping:
+        shard_proj_names = fused_mapping[proj_name]
+
+        # Convert fused_name --> [shard_names]
+        shard_names = [
+            layer_name.replace(proj_name, shard_proj_name)
+            for shard_proj_name in shard_proj_names
+        ]
+
+        # Layer should be ignored if shards are ignored.
+        should_ignore_layer = None
+        for shard_name in shard_names:
+            should_ignore_shard = check_equal_or_regex_match(
+                layer_name=shard_name, targets=ignore)
+
+            # If shard_idx=0, set layer ignore to match shard.
+            if should_ignore_layer is None:
+                should_ignore_layer = should_ignore_shard
+
+            # 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.")
+
+    # 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)
+
+    assert should_ignore_layer is not None
+    return should_ignore_layer
+
+
+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 
+    if target starts with 're:' to any target in list.
+    """
+    for target in targets:
+        if _is_equal_or_regex_match(layer_name, target):
+            return True
+    return False
+
+
+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,
+    additionally checks if the target string is contained within the value.
+    """
+
+    if target.startswith("re:"):
+        pattern = target[3:]
+        if re.match(pattern, value):
+            return True
+    elif check_contains:
+        if target.lower() in value.lower():
+            return True
+    elif target == value:
+        return True
+    return False