Skip to content

vllm.model_executor.layers.quantization.awq_marlin

logger module-attribute 

logger = init_logger(__name__)

AWQMarlinConfig

Bases: QuantizationConfig

Config class for AWQ Marlin

Source code in vllm/model_executor/layers/quantization/awq_marlin.py 
 39
 40
 41
 42
 43
 44
 45
 46
 47
 48
 49
 50
 51
 52
 53
 54
 55
 56
 57
 58
 59
 60
 61
 62
 63
 64
 65
 66
 67
 68
 69
 70
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
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 AWQMoEMethod(self)
        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)

TYPE_MAP class-attribute instance-attribute 

TYPE_MAP = {4: uint4, 8: uint8}

full_config instance-attribute 

full_config = full_config

group_size instance-attribute 

group_size = group_size

lm_head_quantized instance-attribute 

lm_head_quantized = lm_head_quantized

modules_to_not_convert instance-attribute 

modules_to_not_convert = modules_to_not_convert or []

pack_factor instance-attribute 

pack_factor = 32 // weight_bits

quant_type instance-attribute 

quant_type = TYPE_MAP[weight_bits]

weight_bits instance-attribute 

weight_bits = weight_bits

zero_point instance-attribute 

zero_point = zero_point

__init__ 

__init__(
    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
Source code in vllm/model_executor/layers/quantization/awq_marlin.py 
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
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)

__repr__ 

__repr__() -> str
Source code in vllm/model_executor/layers/quantization/awq_marlin.py 
71
72
73
74
75
76
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})")

from_config classmethod 

from_config(config: dict[str, Any]) -> AWQMarlinConfig
Source code in vllm/model_executor/layers/quantization/awq_marlin.py 
 94
 95
 96
 97
 98
 99
100
101
102
103
104
@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)

get_config_filenames classmethod 

get_config_filenames() -> list[str]
Source code in vllm/model_executor/layers/quantization/awq_marlin.py 
90
91
92
@classmethod
def get_config_filenames(cls) -> list[str]:
    return ["quantize_config.json"]

get_min_capability classmethod 

get_min_capability() -> int
Source code in vllm/model_executor/layers/quantization/awq_marlin.py 
86
87
88
@classmethod
def get_min_capability(cls) -> int:
    return 80

get_name classmethod 

get_name() -> QuantizationMethods
Source code in vllm/model_executor/layers/quantization/awq_marlin.py 
78
79
80
@classmethod
def get_name(cls) -> QuantizationMethods:
    return "awq_marlin"

get_quant_method 

get_quant_method(
    layer: Module, prefix: str
) -> Optional[QuantizeMethodBase]
Source code in vllm/model_executor/layers/quantization/awq_marlin.py 
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
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 AWQMoEMethod(self)
    return None

get_supported_act_dtypes classmethod 

get_supported_act_dtypes() -> list[dtype]
Source code in vllm/model_executor/layers/quantization/awq_marlin.py 
82
83
84
@classmethod
def get_supported_act_dtypes(cls) -> list[torch.dtype]:
    return [torch.half, torch.bfloat16]

is_awq_marlin_compatible classmethod 

is_awq_marlin_compatible(quant_config: dict[str, Any])
Source code in vllm/model_executor/layers/quantization/awq_marlin.py 
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
@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)

override_quantization_method classmethod 

override_quantization_method(
    hf_quant_cfg, user_quant
) -> Optional[QuantizationMethods]
Source code in vllm/model_executor/layers/quantization/awq_marlin.py 
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
@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

AWQMarlinLinearMethod

Bases: LinearMethodBase

Linear method for AWQ Marlin.

Parameters:

Name Type Description Default
quant_config AWQMarlinConfig

The AWQ Marlin quantization config.

 required
Source code in vllm/model_executor/layers/quantization/awq_marlin.py 
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
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)

quant_config instance-attribute 

quant_config = quant_config

__init__ 

__init__(quant_config: AWQMarlinConfig) -> None
Source code in vllm/model_executor/layers/quantization/awq_marlin.py 
186
187
def __init__(self, quant_config: AWQMarlinConfig) -> None:
    self.quant_config = quant_config

apply 

apply(
    layer: Module, x: Tensor, bias: Optional[Tensor] = None
) -> Tensor
Source code in vllm/model_executor/layers/quantization/awq_marlin.py 
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
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)

create_weights 

create_weights(
    layer: Module,
    input_size_per_partition: int,
    output_partition_sizes: list[int],
    input_size: int,
    output_size: int,
    params_dtype: dtype,
    **extra_weight_attrs,
) -> None
Source code in vllm/model_executor/layers/quantization/awq_marlin.py 
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
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

process_weights_after_loading 

process_weights_after_loading(layer: Module) -> None
Source code in vllm/model_executor/layers/quantization/awq_marlin.py 
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
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)

AWQMoEMethod

Bases: FusedMoEMethodBase

Source code in vllm/model_executor/layers/quantization/awq_marlin.py 
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
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)

quant_config instance-attribute 

quant_config = quant_config

quant_type instance-attribute 

quant_type = uint4

__init__ 

__init__(quant_config: AWQMarlinConfig)
Source code in vllm/model_executor/layers/quantization/awq_marlin.py 
324
325
326
327
328
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

apply 

apply(
    layer: Module,
    x: Tensor,
    router_logits: 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[Tensor] = None,
    custom_routing_function: Optional[Callable] = None,
    scoring_func: str = "softmax",
    e_score_correction_bias: Optional[Tensor] = None,
    apply_router_weight_on_input: bool = False,
    activation: str = "silu",
    enable_eplb: bool = False,
    expert_load_view: Optional[Tensor] = None,
    logical_to_physical_map: Optional[Tensor] = None,
    logical_replica_count: Optional[Tensor] = None,
) -> Tensor
Source code in vllm/model_executor/layers/quantization/awq_marlin.py 
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
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)

create_weights 

create_weights(
    layer: Module,
    num_experts: int,
    hidden_size: int,
    intermediate_size_per_partition: int,
    params_dtype: dtype,
    **extra_weight_attrs,
)
Source code in vllm/model_executor/layers/quantization/awq_marlin.py 
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
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)

process_weights_after_loading 

process_weights_after_loading(layer: Module) -> None
Source code in vllm/model_executor/layers/quantization/awq_marlin.py 
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
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)