Upload folder using huggingface_hub

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+tokenizer.json filter=lfs diff=lfs merge=lfs -text

.mv

@@ -0,0 +1 @@
+Revision:master,CreatedAt:1753851721

README.md

@@ -0,0 +1,110 @@
+---
+library_name: transformers
+pipeline_tag: text-generation
+tags:
+- glm4_moe
+- AWQ
+- FP16Mix
+- 量化修复
+- vLLM
+base_model:
+  - ZhipuAI/GLM-4.5-Air
+base_model_relation: quantized
+---
+# GLM-4.5-Air-AWQ-FP16Mix
+基础型 [ZhipuAI/GLM-4.5-Air](https://www.modelscope.cn/models/ZhipuAI/GLM-4.5-Air)
+
+
+### 【Vllm 单机8卡启动命令】
+<i>注: 启动该模型一定要跟`--enable-expert-parallel` ，否则其专家张量TP整除除不尽；即使是2卡也需要。 </i>
+```
+$CONTEXT_LENGTH=32768
+
+vllm serve \
+    tclf90/GLM-4.5-Air-AWQ-FP16Mix \
+    --served-model-name GLM-4.5-Air-AWQ-FP16Mix \
+    --enable-expert-parallel \
+    --swap-space 16 \
+    --max-num-seqs 512 \
+    --max-model-len $CONTEXT_LENGTH \
+    --max-seq-len-to-capture $CONTEXT_LENGTH \
+    --gpu-memory-utilization 0.9 \
+    --tensor-parallel-size 8 \
+    --trust-remote-code \
+    --disable-log-requests \
+    --host 0.0.0.0 \
+    --port 8000
+```
+
+### 【依赖】
+
+```
+vllm==0.10.0
+```
+
+### 【❗❗vllm==0.10.0 临时补丁❗❗】
+`vllm`内`awq_marlin` 在加载 awq moe模型的时候，遗漏检查 `modules_to_not_convert` 参数，导致moe的混合量化不生效/报错 [[Issue #21888]](https://github.com/vllm-project/vllm/pull/21888)。
+
+PR merge之前，先临时将 `awq_marlin.py` 替换置 `vllm/model_executor/layers/quantization/awq_marlin.py`
+
+### 【模型更新日期】
+``` 
+2025-07-30
+1. 首次commit
+```
+
+### 【模型列表】
+
+| 文件大小   | 最近更新时间       |
+|--------|--------------|
+| `69GB` | `2025-07-30` |
+
+
+
+### 【模型下载】
+
+```python
+from modelscope import snapshot_download
+snapshot_download('tclf90/GLM-4.5-Air-AWQ-FP16Mix', cache_dir="本地路径")
+```
+
+
+### 【介绍】
+
+# GLM-4.5
+
+<div align="center">
+<img src=https://raw.githubusercontent.com/zai-org/GLM-4.5/refs/heads/main/resources/logo.svg width="15%"/>
+</div>
+<p align="center">
+    👋 加入我们的<a href="https://github.com/zai-org/GLM-4.5/blob/main/resources/WECHAT.md" target="_blank"> 微信群 </a>。
+    <br>
+    📖 查看GLM-4.5<a href="https://z.ai/blog/glm-4.5" target="_blank"> 技术博客 </a>。
+    <br>
+    📍 在<a href="https://docs.bigmodel.cn/cn/guide/models/text/glm-4.5"> 智谱AI开放平台 </a>上使用GLM-4.5 API服务。
+    <br>
+    👉 一键体验 <a href="https://chat.z.ai" >GLM-4.5 </a>。
+</p>
+
+## 模型介绍
+
+**GLM-4.5** 系列模型是专为智能体设计的基础模型。GLM-4.5拥有 **3550** 亿总参数量，其中 **320** 亿活跃参数；GLM-4.5-Air采用更紧凑的设计，拥有
+ **1060** 亿总参数量，其中 **120** 亿活跃参数。GLM-4.5模型统一了推理、编码和智能体能力，以满足智能体应用的复杂需求。
+
+GLM-4.5 和 GLM-4.5-Air 都是混合推理模型，提供两种模式：用于复杂推理和工具使用的思考模式，以及用于即时响应的非思考模式。
+
+我们已开源了 GLM-4.5 和 GLM-4.5-Air 的基础模型、混合推理模型以及混合推理模型的FP8版本。它们采用MIT开源许可证发布，可用于商业用途和二次开发。
+
+在我们对12项行业标准基准的全面评估中，GLM-4.5表现卓越，得分 **63.2**，在所有专有和开源模型中排名**第3**
+。值得注意的是，GLM-4.5-Air在保持优异效率的同时，仍取得了 **59.8** 的竞争性成绩。
+
+![bench](https://raw.githubusercontent.com/zai-org/GLM-4.5/refs/heads/main/resources/bench.png)
+
+如需了解更多评估结果、展示案例和技术细节，请访问我们的 [技术博客](https://z.ai/blog/glm-4.5)。技术报告将很快发布。
+
+模型代码、工具解析器和推理解析器可在 [transformers](https://github.com/huggingface/transformers/tree/main/src/transformers/models/glm4_moe)、 [vLLM](https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/glm4_moe_mtp.py)
+和 [SGLang](https://github.com/sgl-project/sglang/blob/main/python/sglang/srt/models/glm4_moe.py) 的实现中找到。
+
+## 快速开始
+
+请参考我们的[github](https://github.com/zai-org/GLM-4.5)项目。

awq_marlin.py

@@ -0,0 +1,537 @@
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
+from copy import deepcopy
+from typing import Any, Callable, Optional
+
+import torch
+from torch.nn import Parameter
+
+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.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 import QuantizationMethods
+from vllm.model_executor.layers.quantization.awq import (AWQConfig,
+                                                         is_layer_skipped_awq)
+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_scales,
+    moe_awq_to_marlin_zero_points, verify_marlin_supported,
+    verify_marlin_supports_shape)
+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
+
+logger = init_logger(__name__)
+
+
+def get_moe_quant_method(
+    config: QuantizationConfig,
+    layer: torch.nn.Module,
+    prefix: str,
+    moe_method_cls: type,
+):
+    if isinstance(layer, FusedMoE) and is_layer_skipped_awq(prefix, getattr(config, "modules_to_not_convert", [])):
+        return UnquantizedFusedMoEMethod(layer.moe_config)
+    return moe_method_cls(config)
+
+
+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: Optional[list[str]],
+                 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_awq(prefix, self.modules_to_not_convert):
+                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 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 get_moe_quant_method(self, layer, prefix,
+                                        AWQMoEMethod)
+        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)
+
+
+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:
+        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)
+
+    def apply(
+        self,
+        layer: torch.nn.Module,
+        x: torch.Tensor,
+        bias: Optional[torch.Tensor] = 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)
+
+
+class AWQMoEMethod(FusedMoEMethodBase):
+
+    def __init__(self, quant_config: AWQMarlinConfig):
+        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
+
+    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:
+        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)
+
+    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",
+        enable_eplb: bool = False,
+        expert_load_view: Optional[torch.Tensor] = None,
+        logical_to_physical_map: Optional[torch.Tensor] = None,
+        logical_replica_count: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        if enable_eplb:
+            raise NotImplementedError(
+                "EPLB not supported for `AWQMoEMethod` yet.")
+
+        assert activation == "silu", "Only SiLU activation is supported."
+
+        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 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,
+            quant_type_id=self.quant_type.id,
+            apply_router_weight_on_input=apply_router_weight_on_input,
+            global_num_experts=global_num_experts,
+            expert_map=expert_map,
+            w1_zeros=layer.w13_qzeros,
+            w2_zeros=layer.w2_qzeros,
+            workspace=layer.workspace)

chat_template.jinja

@@ -0,0 +1,103 @@
+[gMASK]<sop>
+{%- if tools -%}
+<|system|>
+# Tools
+
+You may call one or more functions to assist with the user query.
+
+You are provided with function signatures within <tools></tools> XML tags:
+<tools>
+{% for tool in tools %}
+{{ tool | tojson(ensure_ascii=False) }}
+{% endfor %}
+</tools>
+
+For each function call, output the function name and arguments within the following XML format:
+<tool_call>{function-name}
+<arg_key>{arg-key-1}</arg_key>
+<arg_value>{arg-value-1}</arg_value>
+<arg_key>{arg-key-2}</arg_key>
+<arg_value>{arg-value-2}</arg_value>
+...
+</tool_call>{%- endif -%}
+{%- macro visible_text(content) -%}
+    {%- if content is string -%}
+        {{- content }}
+    {%- elif content is iterable and content is not mapping -%}
+        {%- for item in content -%}
+            {%- if item is mapping and item.type == 'text' -%}
+                {{- item.text }}
+            {%- elif item is string -%}
+                {{- item }}
+            {%- endif -%}
+        {%- endfor -%}
+    {%- else -%}
+        {{- content }}
+    {%- endif -%}
+{%- endmacro -%}
+{%- set ns = namespace(last_user_index=-1) %}
+{%- for m in messages %}
+    {%- if m.role == 'user' %}
+        {% set ns.last_user_index = loop.index0 -%}
+    {%- endif %}
+{%- endfor %}
+{% for m in messages %}
+{%- if m.role == 'user' -%}<|user|>
+{{ visible_text(m.content) }}
+{{- '/nothink' if (enable_thinking is defined and not enable_thinking and not visible_text(m.content).endswith("/nothink")) else '' -}}
+{%- elif m.role == 'assistant' -%}
+<|assistant|>
+{%- set reasoning_content = '' %}
+{%- set content = visible_text(m.content) %}
+{%- if m.reasoning_content is string %}
+    {%- set reasoning_content = m.reasoning_content %}
+{%- else %}
+    {%- if '</think>' in content %}
+        {%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
+        {%- set content = content.split('</think>')[-1].lstrip('\n') %}
+    {%- endif %}
+{%- endif %}
+{%- if loop.index0 > ns.last_user_index and reasoning_content -%}
+{{ '\n<think>' + reasoning_content.strip() +  '</think>'}}
+{%- else -%}
+{{ '\n<think></think>' }}
+{%- endif -%}
+{%- if content.strip() -%}
+{{ '\n' + content.strip() }}
+{%- endif -%}
+{% if m.tool_calls %}
+{% for tc in m.tool_calls %}
+{%- if tc.function %}
+    {%- set tc = tc.function %}
+{%- endif %}
+{{ '\n<tool_call>' + tc.name }}
+{% set _args = tc.arguments %}
+{% for k, v in _args.items() %}
+<arg_key>{{ k }}</arg_key>
+<arg_value>{{ v | tojson(ensure_ascii=False) if v is not string else v }}</arg_value>
+{% endfor %}
+</tool_call>{% endfor %}
+{% endif %}
+{%- elif m.role == 'tool' -%}
+{%- if m.content is string -%}
+{%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
+    {{- '<|observation|>' }}
+{%- endif %}
+{{- '\n<tool_response>\n' }}
+{{- m.content }}
+{{- '\n</tool_response>' }}
+{%- else -%}
+<|observation|>{% for tr in m.content %}
+
+<tool_response>
+{{ tr.output if tr.output is defined else tr }}
+</tool_response>{% endfor -%}
+{% endif -%}
+{%- elif m.role == 'system' -%}
+<|system|>
+{{ visible_text(m.content) }}
+{%- endif -%}
+{%- endfor -%}
+{%- if add_generation_prompt -%}
+    <|assistant|>{{- '\n<think></think>' if (enable_thinking is defined and not enable_thinking) else '' -}}
+{%- endif -%}

config.json

@@ -0,0 +1,52 @@
+{
+  "name_or_path": "tclf90/GLM-4.5-Air-AWQ-FP16Mix",
+  "architectures": [
+    "Glm4MoeForCausalLM"
+  ],
+  "attention_bias": true,
+  "attention_dropout": 0.0,
+  "pad_token_id": 151329,
+  "eos_token_id": [
+    151329,
+    151336,
+    151338
+  ],
+  "head_dim": 128,
+  "hidden_act": "silu",
+  "hidden_size": 4096,
+  "partial_rotary_factor": 0.5,
+  "initializer_range": 0.02,
+  "intermediate_size": 11264,
+  "max_position_embeddings": 131072,
+  "model_type": "glm4_moe",
+  "moe_intermediate_size": 1408,
+  "norm_topk_prob": true,
+  "num_attention_heads": 96,
+  "n_group": 1,
+  "topk_group": 1,
+  "n_routed_experts": 128,
+  "n_shared_experts": 1,
+  "routed_scaling_factor": 1.0,
+  "num_experts_per_tok": 8,
+  "first_k_dense_replace": 1,
+  "num_hidden_layers": 46,
+  "num_key_value_heads": 8,
+  "rms_norm_eps": 1e-05,
+  "rope_scaling": null,
+  "rope_theta": 1000000,
+  "num_nextn_predict_layers": 1,
+  "tie_word_embeddings": false,
+  "torch_dtype": "float16",
+  "transformers_version": "4.54.0",
+  "use_cache": true,
+  "use_qk_norm": false,
+  "vocab_size": 151552,
+  "quantization_config": {
+    "quant_method": "awq_marlin",
+    "bits": 4,
+    "group_size": 128,
+    "version": "gemm",
+    "zero_point": true,
+    "modules_to_not_convert": ["model.embed_tokens", "model.layers.0.", "model.layers.1.", "model.layers.45.", "model.layers.46.", "model.layers.1.mlp.shared_experts.", "model.layers.2.mlp.shared_experts.", "model.layers.3.mlp.shared_experts.", "model.layers.4.mlp.shared_experts.", "model.layers.5.mlp.shared_experts.", "model.layers.6.mlp.shared_experts.", "model.layers.7.mlp.shared_experts.", "model.layers.8.mlp.shared_experts.", "model.layers.9.mlp.shared_experts.", "model.layers.10.mlp.shared_experts.", "model.layers.11.mlp.shared_experts.", "model.layers.12.mlp.shared_experts.", "model.layers.13.mlp.shared_experts.", "model.layers.14.mlp.shared_experts.", "model.layers.15.mlp.shared_experts.", "model.layers.16.mlp.shared_experts.", "model.layers.17.mlp.shared_experts.", "model.layers.18.mlp.shared_experts.", "model.layers.19.mlp.shared_experts.", "model.layers.20.mlp.shared_experts.", "model.layers.21.mlp.shared_experts.", "model.layers.22.mlp.shared_experts.", "model.layers.23.mlp.shared_experts.", "model.layers.24.mlp.shared_experts.", "model.layers.25.mlp.shared_experts.", "model.layers.26.mlp.shared_experts.", "model.layers.27.mlp.shared_experts.", "model.layers.28.mlp.shared_experts.", "model.layers.29.mlp.shared_experts.", "model.layers.30.mlp.shared_experts.", "model.layers.31.mlp.shared_experts.", "model.layers.32.mlp.shared_experts.", "model.layers.33.mlp.shared_experts.", "model.layers.34.mlp.shared_experts.", "model.layers.35.mlp.shared_experts.", "model.layers.36.mlp.shared_experts.", "model.layers.37.mlp.shared_experts.", "model.layers.38.mlp.shared_experts.", "model.layers.39.mlp.shared_experts.", "model.layers.40.mlp.shared_experts.", "model.layers.41.mlp.shared_experts.", "model.layers.42.mlp.shared_experts.", "model.layers.43.mlp.shared_experts.", "model.layers.44.mlp.shared_experts.", "model.layers.45.mlp.shared_experts.", "model.layers.46.mlp.shared_experts.", "model.layers.46.shared_head.", "lm_head"]
+  }
+}

configuration.json

@@ -0,0 +1 @@
+{"framework":"Pytorch","task":"text-generation"}

model-00001-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1fc26a957fd5241a3646ad8d4af0ecfb6c4afb3d5a331e2a5fccba165292049a
+size 4996541272

model-00002-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8636bec0729cc1d553c1033cfe05bdbf9b31321c24e834f9deb7bd44caef4f96
+size 4998307520

model-00003-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8a8c5c1b6b142cff5cf2c854ddba8dd5caeaea231a047aaee3bd073629a4713e
+size 4999151480

model-00004-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8e2b4c25d8591c75d2d5066bf68b3ef8c032d96f5c228a50c745d3c8dbe2d3ba
+size 4999153696

model-00005-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c224ab8dffd72a759ac18067b76fb26081356b2011c533ebe0125a87be261483
+size 4990763840

model-00006-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:14d52bddbf5812b8262e12e40eb921ccc53b36e40a8ad591b6f12046e1a41f63
+size 4997445712

model-00007-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2a1b1955fc16166ad1bbd33e027b91f0cbbf1667bba4180e230283f5143aa32
+size 4998360800

model-00008-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8f152246c1b439ecebf28559a2ac5508ebe6834a2e56549692e0ac21bb8939de
+size 5000525936

model-00009-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e37674f17722dee14b814ae6d5b3672c6e3f25c6dedede1a5707fe63fd122824
+size 4999156080

model-00010-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c8f5364deb67daceb52a427deccfb399d0fe8080c84a115ff1b9f02688010f48
+size 4999156120

model-00011-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:32b4de7acea1825f8963e287a589b5814e8510283aa38bc49865a76636d64d08
+size 4999156120

model-00012-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8458af9bd76d476cc7c53b456064677ad9697590d447108339ef84cb80eda365
+size 4999156120

model-00013-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9ac2c934733de485a23aea34609888eeb13d3ef4fe620254425114c5b40d337c
+size 4995004840

model-00014-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be46467218653568d2ebbb08a5ffb78c4c94b841426f151df6ee8675f0395392
+size 4999719624

model-00015-of-00015.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:77c8f54a9f3b2d23793cb870fb465bcee910af595db1d9d71adcd854ab793154
+size 3056676280

tokenizer.json

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9340665016419c825c4bdabbcc9acc43b7ca2c68ce142724afa829abb1be5efd
+size 19970699

tokenizer_config.json

@@ -0,0 +1,325 @@
+{
+  "added_tokens_decoder": {
+    "151329": {
+      "content": "<|endoftext|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151330": {
+      "content": "[MASK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151331": {
+      "content": "[gMASK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151332": {
+      "content": "[sMASK]",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151333": {
+      "content": "<sop>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151334": {
+      "content": "<eop>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151335": {
+      "content": "<|system|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151336": {
+      "content": "<|user|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151337": {
+      "content": "<|assistant|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151338": {
+      "content": "<|observation|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151339": {
+      "content": "<|begin_of_image|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151340": {
+      "content": "<|end_of_image|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151341": {
+      "content": "<|begin_of_video|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151342": {
+      "content": "<|end_of_video|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151343": {
+      "content": "<|begin_of_audio|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151344": {
+      "content": "<|end_of_audio|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151345": {
+      "content": "<|begin_of_transcription|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151346": {
+      "content": "<|end_of_transcription|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151347": {
+      "content": "<|code_prefix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151348": {
+      "content": "<|code_middle|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151349": {
+      "content": "<|code_suffix|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151350": {
+      "content": "<think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151351": {
+      "content": "</think>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151352": {
+      "content": "<tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151353": {
+      "content": "</tool_call>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151354": {
+      "content": "<tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151355": {
+      "content": "</tool_response>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151356": {
+      "content": "<arg_key>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151357": {
+      "content": "</arg_key>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151358": {
+      "content": "<arg_value>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151359": {
+      "content": "</arg_value>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151360": {
+      "content": "/nothink",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": true
+    },
+    "151361": {
+      "content": "<|begin_of_box|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151362": {
+      "content": "<|end_of_box|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151363": {
+      "content": "<|image|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    },
+    "151364": {
+      "content": "<|video|>",
+      "lstrip": false,
+      "normalized": false,
+      "rstrip": false,
+      "single_word": false,
+      "special": false
+    }
+  },
+  "additional_special_tokens": [
+    "<|endoftext|>",
+    "[MASK]",
+    "[gMASK]",
+    "[sMASK]",
+    "<sop>",
+    "<eop>",
+    "<|system|>",
+    "<|user|>",
+    "<|assistant|>",
+    "<|observation|>",
+    "<|begin_of_image|>",
+    "<|end_of_image|>",
+    "<|begin_of_video|>",
+    "<|end_of_video|>",
+    "<|begin_of_audio|>",
+    "<|end_of_audio|>",
+    "<|begin_of_transcription|>",
+    "<|end_of_transcription|>",
+    "<|code_prefix|>",
+    "<|code_middle|>",
+    "<|code_suffix|>",
+    "/nothink"
+  ],
+  "clean_up_tokenization_spaces": false,
+  "do_lower_case": false,
+  "eos_token": "<|endoftext|>",
+  "extra_special_tokens": {},
+  "model_max_length": 128000,
+  "pad_token": "<|endoftext|>",
+  "padding_side": "left",
+  "remove_space": false,
+  "tokenizer_class": "PreTrainedTokenizer"
+}