# SPDX-License-Identifier: Apache-2.0 # SPDX-FileCopyrightText: Copyright contributors to the vLLM project from collections.abc import Callable from typing import TYPE_CHECKING, Any, Optional import torch from safetensors.torch import _TYPES as _SAFETENSORS_TO_TORCH_DTYPE from torch.nn import Parameter import vllm.model_executor.layers.fused_moe # noqa from vllm import _custom_ops as ops from vllm.forward_context import ForwardContext, get_forward_context from vllm.logger import init_logger from vllm.model_executor.layers.fused_moe.config import ( FusedMoEConfig, FusedMoEQuantConfig, ) from vllm.model_executor.layers.fused_moe.fused_marlin_moe import fused_marlin_moe from vllm.model_executor.layers.fused_moe.layer import ( FusedMoE, FusedMoEMethodBase, FusedMoeWeightScaleSupported, UnquantizedFusedMoEMethod, ) from vllm.model_executor.layers.linear import ( LinearBase, LinearMethodBase, UnquantizedLinearMethod, set_weight_attrs, ) from vllm.model_executor.layers.quantization.awq import AWQConfig from vllm.model_executor.layers.quantization.base_config import ( QuantizationConfig, QuantizeMethodBase, ) from vllm.model_executor.layers.quantization.utils import replace_parameter from vllm.model_executor.layers.quantization.utils.marlin_utils import ( apply_awq_marlin_linear, awq_to_marlin_zero_points, check_marlin_supported, check_marlin_supports_layer, check_moe_marlin_supports_layer, marlin_make_empty_g_idx, marlin_make_workspace_new, marlin_moe_permute_scales, marlin_permute_bias, marlin_permute_scales, moe_awq_to_marlin_zero_points, verify_marlin_supported, verify_marlin_supports_shape, ) from vllm.model_executor.layers.quantization.utils.quant_utils import is_layer_skipped from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead from vllm.model_executor.parameter import GroupQuantScaleParameter, PackedvLLMParameter from vllm.platforms import current_platform from vllm.scalar_type import scalar_types from vllm.transformers_utils.config import get_safetensors_params_metadata if TYPE_CHECKING: from vllm.model_executor.layers.quantization import QuantizationMethods from vllm.model_executor.models.utils import WeightsMapper import ixformer.inference.functions as ixfops logger = init_logger(__name__) class AWQMarlinConfig(QuantizationConfig): """Config class for AWQ Marlin""" # num_bits -> type TYPE_MAP = { 4: scalar_types.uint4, 8: scalar_types.uint8, } def __init__( self, weight_bits: int, group_size: int, zero_point: bool, lm_head_quantized: bool, modules_to_not_convert: list[str] | None, full_config: dict[str, Any], ) -> None: super().__init__() self.pack_factor = 32 // weight_bits # packed into int32 self.group_size = group_size self.zero_point = zero_point self.lm_head_quantized = lm_head_quantized self.weight_bits = weight_bits self.modules_to_not_convert = modules_to_not_convert or [] self.full_config = full_config if self.weight_bits not in self.TYPE_MAP: raise ValueError( f"Unsupported num_bits = {self.weight_bits}. " f"Supported num_bits = {self.TYPE_MAP.keys()}" ) self.quant_type = self.TYPE_MAP[self.weight_bits] verify_marlin_supported( self.quant_type, group_size=self.group_size, has_zp=self.zero_point ) def __repr__(self) -> str: return ( f"AWQMarlinConfig(quant_type={self.quant_type}, " f"group_size={self.group_size}, " f"zero_point={self.zero_point}, " f"lm_head_quantized={self.lm_head_quantized}, " f"modules_to_not_convert={self.modules_to_not_convert})" ) @classmethod def get_name(cls) -> "QuantizationMethods": return "awq_marlin" @classmethod def get_supported_act_dtypes(cls) -> list[torch.dtype]: return [torch.half, torch.bfloat16] @classmethod def get_min_capability(cls) -> int: return 80 @classmethod def get_config_filenames(cls) -> list[str]: return ["quantize_config.json"] @classmethod def from_config(cls, config: dict[str, Any]) -> "AWQMarlinConfig": weight_bits = cls.get_from_keys(config, ["bits"]) group_size = cls.get_from_keys(config, ["group_size"]) zero_point = cls.get_from_keys(config, ["zero_point"]) lm_head_quantized = cls.get_from_keys_or(config, ["lm_head"], default=False) modules_to_not_convert = cls.get_from_keys_or( config, ["modules_to_not_convert"], None ) return cls( weight_bits, group_size, zero_point, lm_head_quantized, modules_to_not_convert, config, ) @classmethod def override_quantization_method( cls, hf_quant_cfg, user_quant ) -> Optional["QuantizationMethods"]: can_convert = cls.is_awq_marlin_compatible(hf_quant_cfg) is_valid_user_quant = ( user_quant is None or user_quant == "marlin" or user_quant == "awq_marlin" ) if can_convert and is_valid_user_quant: msg = ( "The model is convertible to {} during runtime." " Using {} kernel.".format(cls.get_name(), cls.get_name()) ) logger.info(msg) return cls.get_name() if can_convert and user_quant == "awq": logger.info( "Detected that the model can run with awq_marlin" ", however you specified quantization=awq explicitly," " so forcing awq. Use quantization=awq_marlin for" " faster inference" ) return None def get_quant_method( self, layer: torch.nn.Module, prefix: str ) -> Optional["QuantizeMethodBase"]: if isinstance(layer, LinearBase) or ( isinstance(layer, ParallelLMHead) and self.lm_head_quantized ): if is_layer_skipped( prefix, self.modules_to_not_convert, self.packed_modules_mapping, skip_with_substr=True, ): return UnquantizedLinearMethod() # Check if the layer is supported by AWQMarlin. if not check_marlin_supports_layer(layer, self.group_size): logger.warning_once( "Layer '%s' is not supported by AWQMarlin. Falling back to unoptimized AWQ kernels.", # noqa: E501 prefix, ) return AWQConfig.from_config(self.full_config).get_quant_method( layer, prefix ) return AWQMarlinLinearMethod(self) elif isinstance(layer, FusedMoE): # from vllm.model_executor.layers.quantization.moe_wna16 import ( # MoeWNA16Config) # if is_layer_skipped_awq( # prefix, getattr(self, "modules_to_not_convert", [])): # return UnquantizedFusedMoEMethod(layer.moe_config) # if not check_moe_marlin_supports_layer(layer, self.group_size): # logger.warning_once( # f"Layer '{prefix}' is not supported by AWQMoeMarlin. " # "Falling back to Moe WNA16 kernels.") # return MoeWNA16Config.from_config( # self.full_config).get_quant_method(layer, prefix) # return AWQMoEMethod(self, layer.moe_config) return AWQMoEMethod(self, layer.moe_config) return None @classmethod def is_awq_marlin_compatible(cls, quant_config: dict[str, Any]): # Extract data from quant config. quant_method = quant_config.get("quant_method", "").lower() num_bits = quant_config.get("bits") group_size = quant_config.get("group_size") zero_point = quant_config.get("zero_point") if not current_platform.is_cuda(): return False if quant_method != "awq": return False # If we cannot find the info needed in the config, cannot convert. if num_bits is None or group_size is None or zero_point is None: return False if num_bits not in cls.TYPE_MAP: return False return check_marlin_supported( quant_type=cls.TYPE_MAP[num_bits], group_size=group_size, has_zp=zero_point ) def apply_vllm_mapper(self, hf_to_vllm_mapper: "WeightsMapper"): if self.modules_to_not_convert: self.modules_to_not_convert = hf_to_vllm_mapper.apply_list( self.modules_to_not_convert ) def maybe_update_config(self, model_name: str, revision: str | None = None): if self.modules_to_not_convert: return unquant_dtypes = [torch.float16, torch.bfloat16, torch.float32] metadata = get_safetensors_params_metadata(model_name, revision=revision) layers = {param_name.rsplit(".", 1)[0] for param_name in metadata} quant_layers: set[str] = { param_name.rsplit(".", 1)[0] for param_name, info in metadata.items() if (dtype := info.get("dtype", None)) and _SAFETENSORS_TO_TORCH_DTYPE[dtype] not in unquant_dtypes } self.modules_to_not_convert = list(layers - quant_layers) class AWQMarlinLinearMethod(LinearMethodBase): """Linear method for AWQ Marlin. Args: quant_config: The AWQ Marlin quantization config. """ def __init__(self, quant_config: AWQMarlinConfig) -> None: self.quant_config = quant_config 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, ) -> None: del output_size output_size_per_partition = sum(output_partition_sizes) weight_loader = extra_weight_attrs.get("weight_loader") # Normalize group_size if self.quant_config.group_size != -1: group_size = self.quant_config.group_size else: group_size = input_size verify_marlin_supports_shape( output_size_per_partition=output_size_per_partition, input_size_per_partition=input_size_per_partition, input_size=input_size, group_size=group_size, ) qweight = PackedvLLMParameter( data=torch.empty( input_size_per_partition, output_size_per_partition // self.quant_config.pack_factor, dtype=torch.int32, ), input_dim=0, output_dim=1, packed_dim=1, packed_factor=self.quant_config.pack_factor, weight_loader=weight_loader, ) num_groups = input_size_per_partition // group_size qzeros = PackedvLLMParameter( data=torch.empty( num_groups, output_size_per_partition // self.quant_config.pack_factor, dtype=torch.int32, ), input_dim=0, output_dim=1, packed_dim=1, packed_factor=self.quant_config.pack_factor, weight_loader=weight_loader, ) scales = GroupQuantScaleParameter( data=torch.empty( num_groups, output_size_per_partition, dtype=params_dtype, ), input_dim=0, output_dim=1, weight_loader=weight_loader, ) layer.register_parameter("qweight", qweight) layer.register_parameter("qzeros", qzeros) layer.register_parameter("scales", scales) layer.input_size_per_partition = input_size_per_partition layer.output_size_per_partition = output_size_per_partition layer.num_groups = num_groups # TODO: Update this docs # Checkpoints are serialized in AutoAWQ format, which is different from the # marlin format. This function is called after the weights are loaded. # Here, we handle the repacking def process_weights_after_loading(self, layer: torch.nn.Module) -> None: layer.output_size_per_partition = layer.qweight.data.shape[1] * self.quant_config.pack_factor align_bits = 64 * 8 align_size = align_bits // self.quant_config.weight_bits if layer.output_size_per_partition % align_size != 0: padding_output_size_per_partition = (layer.output_size_per_partition + align_size - 1) // align_size * align_size layer.output_padding_size = padding_output_size_per_partition - layer.output_size_per_partition device = layer.qweight.device pad_qweight = torch.zeros( layer.input_size_per_partition, padding_output_size_per_partition // self.quant_config.pack_factor, dtype=torch.int32, device=device, ) pad_qzeros = torch.zeros( layer.num_groups, padding_output_size_per_partition // self.quant_config.pack_factor, dtype=torch.int32, device=device, ) pad_scales = torch.zeros( layer.num_groups, padding_output_size_per_partition, dtype=layer.scales.data.dtype, device=device, ) pad_qweight[..., :layer.output_size_per_partition // self.quant_config.pack_factor] = layer.qweight.data pad_qzeros[..., :layer.output_size_per_partition // self.quant_config.pack_factor] = layer.qzeros.data pad_scales[..., :layer.output_size_per_partition] = layer.scales.data replace_parameter(layer, "qweight", pad_qweight) replace_parameter(layer, "qzeros", pad_qzeros) replace_parameter(layer, "scales", pad_scales) return # TODO(gyf) Marlin format is not support for now.. device = layer.qweight.device layer.qweight = torch.nn.Parameter(layer.qweight.data, requires_grad=False) layer.qzeros = torch.nn.Parameter(layer.qzeros.data, requires_grad=False) layer.scales = torch.nn.Parameter(layer.scales.data, requires_grad=False) # Allocate marlin workspace layer.workspace = marlin_make_workspace_new(device) # Repack weights from AWQ format to marlin format. marlin_qweight = ops.awq_marlin_repack( layer.qweight, size_k=layer.input_size_per_partition, size_n=layer.output_size_per_partition, num_bits=self.quant_config.quant_type.size_bits, ) replace_parameter(layer, "qweight", marlin_qweight) # Permute scales from AWQ format to marlin format. marlin_scales = marlin_permute_scales( layer.scales, size_k=layer.input_size_per_partition, size_n=layer.output_size_per_partition, group_size=self.quant_config.group_size, ) replace_parameter(layer, "scales", marlin_scales) # Permute zero-points from AWQ format to marlin format. marlin_zp = awq_to_marlin_zero_points( layer.qzeros, size_k=layer.num_groups, size_n=layer.output_size_per_partition, num_bits=self.quant_config.quant_type.size_bits, ) replace_parameter(layer, "qzeros", marlin_zp) # Not-used layer.g_idx = marlin_make_empty_g_idx(device) layer.g_idx_sort_indices = marlin_make_empty_g_idx(device) if hasattr(layer, "bias") and layer.bias is not None: layer.bias.data = marlin_permute_bias(layer.bias) def apply( self, layer: torch.nn.Module, x: torch.Tensor, bias: torch.Tensor | None = None, ) -> torch.Tensor: # return apply_awq_marlin_linear( # input=x, # weight=layer.qweight, # weight_scale=layer.scales, # weight_zp=layer.qzeros, # g_idx=layer.g_idx, # g_idx_sort_indices=layer.g_idx_sort_indices, # workspace=layer.workspace, # quant_type=self.quant_config.quant_type, # output_size_per_partition=layer.output_size_per_partition, # input_size_per_partition=layer.input_size_per_partition, # bias=bias, # ) # TODO use awq kernel temporarily.. qweight = layer.qweight scales = layer.scales qzeros = layer.qzeros pack_factor = self.quant_config.pack_factor out_shape = (x.shape[:-1] + (qweight.shape[-1] * pack_factor, )) reshaped_x = x.reshape(-1, x.shape[-1]) out = ops.awq_gemm(reshaped_x, qweight, scales, qzeros, pack_factor, group_size=self.quant_config.group_size) if bias is not None: out.add_(bias) return out.reshape(out_shape) class AWQMoEMethod(FusedMoEMethodBase): def __init__( self, quant_config: AWQMarlinConfig, moe: FusedMoEConfig, ): super().__init__(moe) self.quant_config = quant_config if self.quant_config.weight_bits != 4: raise ValueError("AWQMoEMethod only supports 4bit now.") self.quant_type = scalar_types.uint4 self.use_marlin = True 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, ): extra_weight_attrs.update( { "is_transposed": True, "quant_method": FusedMoeWeightScaleSupported.GROUP.value, } ) w13_qweight = Parameter( torch.empty( num_experts, hidden_size, 2 * intermediate_size_per_partition // self.quant_config.pack_factor, dtype=torch.int32, ), requires_grad=False, ) layer.register_parameter("w13_qweight", w13_qweight) set_weight_attrs(w13_qweight, extra_weight_attrs) w2_qweight = Parameter( torch.empty( num_experts, intermediate_size_per_partition, hidden_size // self.quant_config.pack_factor, dtype=torch.int32, ), requires_grad=False, ) layer.register_parameter("w2_qweight", w2_qweight) set_weight_attrs(w2_qweight, extra_weight_attrs) num_groups_w13 = hidden_size // self.quant_config.group_size num_groups_w2 = intermediate_size_per_partition // self.quant_config.group_size # WEIGHT_SCALES # Allocate 2 scales for w1 and w3 respectively. w13_scales = Parameter( torch.empty( num_experts, num_groups_w13, intermediate_size_per_partition * 2, dtype=params_dtype, ), requires_grad=False, ) layer.register_parameter("w13_scales", w13_scales) set_weight_attrs(w13_scales, extra_weight_attrs) w2_scales = Parameter( torch.empty(num_experts, num_groups_w2, hidden_size, dtype=params_dtype), requires_grad=False, ) layer.register_parameter("w2_scales", w2_scales) set_weight_attrs(w2_scales, extra_weight_attrs) # WEIGHT_ZERO_POINT # Allocate 2 zero points for w1 and w3 respectively. w13_qzeros = Parameter( torch.empty( num_experts, num_groups_w13, 2 * intermediate_size_per_partition // self.quant_config.pack_factor, dtype=torch.int32, ), requires_grad=False, ) layer.register_parameter("w13_qzeros", w13_qzeros) set_weight_attrs(w13_qzeros, extra_weight_attrs) w2_qzeros = Parameter( torch.empty( num_experts, num_groups_w2, hidden_size // self.quant_config.pack_factor, dtype=torch.int32, ), requires_grad=False, ) layer.register_parameter("w2_qzeros", w2_qzeros) set_weight_attrs(w2_qzeros, extra_weight_attrs) device = layer.w13_qweight.device layer.workspace = marlin_make_workspace_new(device, 4) def process_weights_after_loading(self, layer: torch.nn.Module) -> None: return # TODO(gyf) Marlin format is not support for now.. num_experts = layer.w13_qweight.shape[0] device = layer.w13_qweight.device layer.w13_g_idx_sort_indices = torch.nn.Parameter( torch.empty((num_experts, 0), dtype=torch.int32, device=device), requires_grad=False, ) layer.w2_g_idx_sort_indices = torch.nn.Parameter( torch.empty((num_experts, 0), dtype=torch.int32, device=device), requires_grad=False, ) marlin_w13_qweight = ops.awq_marlin_moe_repack( layer.w13_qweight, layer.w13_g_idx_sort_indices, size_k=layer.w13_qweight.shape[1], size_n=layer.w13_qweight.shape[2] * self.quant_config.pack_factor, num_bits=self.quant_config.weight_bits, ) replace_parameter(layer, "w13_qweight", marlin_w13_qweight) marlin_w2_qweight = ops.awq_marlin_moe_repack( layer.w2_qweight, layer.w2_g_idx_sort_indices, size_k=layer.w2_qweight.shape[1], size_n=layer.w2_qweight.shape[2] * self.quant_config.pack_factor, num_bits=self.quant_config.weight_bits, ) replace_parameter(layer, "w2_qweight", marlin_w2_qweight) # Why does this take the intermediate size for size_k? marlin_w13_scales = marlin_moe_permute_scales( s=layer.w13_scales, size_k=layer.intermediate_size_per_partition, size_n=layer.w13_scales.shape[2], group_size=self.quant_config.group_size, ) replace_parameter(layer, "w13_scales", marlin_w13_scales) marlin_w2_scales = marlin_moe_permute_scales( s=layer.w2_scales, size_k=layer.intermediate_size_per_partition, size_n=layer.w2_scales.shape[2], group_size=self.quant_config.group_size, ) replace_parameter(layer, "w2_scales", marlin_w2_scales) marlin_w13_zp = moe_awq_to_marlin_zero_points( layer.w13_qzeros, size_k=layer.w13_qzeros.shape[1], size_n=layer.w13_qzeros.shape[2] * self.quant_config.pack_factor, num_bits=self.quant_config.weight_bits, ) replace_parameter(layer, "w13_qzeros", marlin_w13_zp) marlin_w2_zp = moe_awq_to_marlin_zero_points( layer.w2_qzeros, size_k=layer.w2_qzeros.shape[1], size_n=layer.w2_qzeros.shape[2] * self.quant_config.pack_factor, num_bits=self.quant_config.weight_bits, ) replace_parameter(layer, "w2_qzeros", marlin_w2_zp) if hasattr(layer, "w13_bias") and layer.w13_bias is not None: layer.w13_bias.data = marlin_permute_bias(layer.w13_bias) if hasattr(layer, "w2_bias") and layer.w2_bias is not None: layer.w2_bias.data = marlin_permute_bias(layer.w2_bias) def get_fused_moe_quant_config( self, layer: torch.nn.Module ) -> FusedMoEQuantConfig | None: return None 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: int | None = None, num_expert_group: int | None = None, global_num_experts: int = -1, expert_map: torch.Tensor | None = None, custom_routing_function: Callable | None = None, scoring_func: str = "softmax", routed_scaling_factor: float = 1.0, e_score_correction_bias: torch.Tensor | None = None, apply_router_weight_on_input: bool = False, activation: str = "silu", enable_eplb: bool = False, expert_load_view: torch.Tensor | None = None, logical_to_physical_map: torch.Tensor | None = None, logical_replica_count: torch.Tensor | None = None, ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]: if enable_eplb: raise NotImplementedError("EPLB not supported for `AWQMoEMethod` yet.") assert activation == "silu", "Only SiLU activation is supported." use_ep = expert_map is not None attn_metadata = get_forward_context().attn_metadata if attn_metadata: if isinstance(attn_metadata, dict): only_decode = (use_ep == False and all(t.num_decodes > 0 and t.num_prefills ==0 for t in list(attn_metadata.values()))) else: only_decode = use_ep == False and attn_metadata.num_decodes > 0 and attn_metadata.num_prefills == 0 else: only_decode = False if use_ep: start_eid = layer.ep_rank * layer.local_num_experts end_eid = min((layer.ep_rank + 1) * layer.local_num_experts, global_num_experts) if apply_router_weight_on_input: raise NotImplementedError( "Apply router weight on input is not supported for" "fused Marlin MoE method.") 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, routed_scaling_factor=routed_scaling_factor, e_score_correction_bias=e_score_correction_bias, indices_type=self.topk_indices_dtype) num_tokens, num_experts = router_logits.shape if use_ep: hidden_size = x.shape[1] ( src_to_dst, sorted_token_ids, expert_sizes_gpu, expert_sizes_cpu, expand_tokens, ) = ixfops.moe_compute_token_index_ep( topk_ids=topk_ids, num_experts=num_experts, start_expert_id=start_eid, end_expert_id=end_eid, ) if expert_sizes_cpu.sum() == 0: return torch.zeros( (num_tokens, hidden_size), device=x.device, dtype=x.dtype, ) else: expand_tokens = num_tokens * top_k ( src_to_dst, sorted_token_ids, expert_sizes_gpu, expert_sizes_cpu, ) = ixfops.moe_compute_token_index( topk_ids=topk_ids, num_experts=num_experts, ) # expand + reorder # TODO use kernel expand_hidden_states = ixfops.moe_expand_input( hidden_states=x, dst_to_src=sorted_token_ids, dst_tokens=expand_tokens, topk=top_k, src_to_dst=src_to_dst, ) # w4a16 group gemm 1 # pt_output_1: (expand_tokens, 2n) dtype if only_decode: pt_output_1 = ixfops.moe_w4a16_group_gemv( input=expand_hidden_states, weight=layer.w13_qweight, w_scales=layer.w13_scales, quant_type="awq", w_zeros=layer.w13_qzeros, group_size=self.quant_config.group_size, dst_to_src=None, format="NN", tokens_per_experts_gpu=expert_sizes_gpu, ) # act pt_output_2 = ixfops.silu_and_mul(pt_output_1) pt_output_3 = ixfops.moe_w4a16_group_gemv( input=pt_output_2, weight=layer.w2_qweight, w_scales=layer.w2_scales, quant_type="awq", w_zeros=layer.w2_qzeros, group_size=self.quant_config.group_size, dst_to_src=sorted_token_ids, format="NN", tokens_per_experts_gpu=expert_sizes_gpu, ) # mul + reduce_sum # final_hidden_states: (num_tokens, k) final_hidden_states = ixfops.moe_output_reduce_sum( input=pt_output_3.view(num_tokens, top_k, -1), topk_weight=topk_weights, scaling_factor=routed_scaling_factor ) else: expert_sizes_cpu = expert_sizes_gpu.cpu() pt_output_1 = ixfops.moe_w4a16_group_gemm( input=expand_hidden_states, weight=layer.w13_qweight, w_scales=layer.w13_scales, quant_type="awq", tokens_per_experts=expert_sizes_cpu, w_zeros=layer.w13_qzeros, group_size=self.quant_config.group_size, dst_to_src=None, format="NN", tokens_per_experts_gpu=expert_sizes_gpu, ) # act pt_output_2 = ixfops.silu_and_mul(pt_output_1) # w4a16 group gemm 2 + reorder # pt_output_3: (expand_tokens, k) dtype if use_ep: pt_output_3 = torch.empty( (num_tokens * top_k, hidden_size), device=x.device, dtype=x.dtype, ) ixfops.moe_w4a16_group_gemm( input=pt_output_2, weight=layer.w2_qweight, w_scales=layer.w2_scales, quant_type="awq", tokens_per_experts=expert_sizes_cpu, w_zeros=layer.w2_qzeros, group_size=self.quant_config.group_size, dst_to_src=sorted_token_ids, format="NN", output=pt_output_3, tokens_per_experts_gpu=expert_sizes_gpu, ) reduce_mask = src_to_dst == -1 final_hidden_states = ixfops.moe_output_reduce_sum( input=pt_output_3.view(num_tokens, top_k, -1), topk_weight=topk_weights, scaling_factor=routed_scaling_factor, mask=reduce_mask, ) else: pt_output_3 = ixfops.moe_w4a16_group_gemm( input=pt_output_2, weight=layer.w2_qweight, w_scales=layer.w2_scales, quant_type="awq", tokens_per_experts=expert_sizes_cpu, w_zeros=layer.w2_qzeros, group_size=self.quant_config.group_size, dst_to_src=sorted_token_ids, format="NN", tokens_per_experts_gpu=expert_sizes_gpu, ) # mul + reduce_sum # final_hidden_states: (num_tokens, k) final_hidden_states = ixfops.moe_output_reduce_sum( input=pt_output_3.view(num_tokens, top_k, -1), topk_weight=topk_weights, scaling_factor=routed_scaling_factor ) return final_hidden_states # return torch.ops.vllm.fused_marlin_moe( # x, # layer.w13_qweight, # layer.w2_qweight, # layer.w13_scales, # layer.w2_scales, # router_logits, # topk_weights, # topk_ids, # w1_zeros=layer.w13_qzeros, # w2_zeros=layer.w2_qzeros, # num_bits=self.quant_config.weight_bits, # )