Skip to content

vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors_moe

__all__ module-attribute 

__all__ = [
    "CompressedTensorsMoEMethod",
    "CompressedTensorsW8A8Fp8MoEMethod",
    "CompressedTensorsW8A8Fp8MoECutlassMethod",
    "CompressedTensorsW8A8Int8MoEMethod",
    "CompressedTensorsWNA16MarlinMoEMethod",
    "CompressedTensorsWNA16MoEMethod",
    "CompressedTensorsW4A4MoeMethod",
]

logger module-attribute 

logger = init_logger(__name__)

CompressedTensorsMoEMethod

Bases: FusedMoEMethodBase

Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
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
class CompressedTensorsMoEMethod(FusedMoEMethodBase):

    @staticmethod
    def get_moe_method(
        quant_config: "CompressedTensorsConfig",  # type: ignore # noqa E501
        layer: torch.nn.Module,
    ) -> "CompressedTensorsMoEMethod":
        # TODO: @dsikka: refactor this to use schemes as other kernels
        # are supported + check if the layer is being ignored.
        weight_quant = quant_config.target_scheme_map["Linear"].get("weights")
        input_quant = quant_config.target_scheme_map["Linear"].get(
            "input_activations")

        if quant_config._is_wNa16_group_channel(weight_quant, input_quant):
            # group_size=None means channelwise
            group_size = weight_quant.group_size or -1
            # Prefer to use the MarlinMoE kernel when it is supported.
            if not check_moe_marlin_supports_layer(layer, group_size):
                if (weight_quant.strategy in QuantizationStrategy.GROUP and
                        weight_quant.actorder in (ActivationOrdering.GROUP,
                                                  ActivationOrdering.DYNAMIC)):
                    raise ValueError(
                        "WNA16MoE is not supported with actorder=group/dynamic."
                    )
                logger.info_once("Using CompressedTensorsWNA16MoEMethod")
                return CompressedTensorsWNA16MoEMethod(quant_config)
            else:
                logger.info_once("Using CompressedTensorsWNA16MarlinMoEMethod")
                return CompressedTensorsWNA16MarlinMoEMethod(quant_config)
        elif quant_config._is_fp4a4_nvfp4(weight_quant, input_quant):
            return CompressedTensorsW4A4MoeMethod()
        elif quant_config._is_fp8_w8a8_sm90(weight_quant, input_quant):
            return CompressedTensorsW8A8Fp8MoECutlassMethod(quant_config)
        elif quant_config._is_fp8_w8a8(weight_quant, input_quant):
            return CompressedTensorsW8A8Fp8MoEMethod(quant_config)
        elif quant_config._is_dynamic_token_w8a8(weight_quant, input_quant):
            return CompressedTensorsW8A8Int8MoEMethod(quant_config)
        else:
            raise RuntimeError(
                f"Unsupported FusedMoe scheme: {weight_quant}, {input_quant}")

get_moe_method staticmethod 

get_moe_method(
    quant_config: CompressedTensorsConfig, layer: Module
) -> CompressedTensorsMoEMethod
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
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
@staticmethod
def get_moe_method(
    quant_config: "CompressedTensorsConfig",  # type: ignore # noqa E501
    layer: torch.nn.Module,
) -> "CompressedTensorsMoEMethod":
    # TODO: @dsikka: refactor this to use schemes as other kernels
    # are supported + check if the layer is being ignored.
    weight_quant = quant_config.target_scheme_map["Linear"].get("weights")
    input_quant = quant_config.target_scheme_map["Linear"].get(
        "input_activations")

    if quant_config._is_wNa16_group_channel(weight_quant, input_quant):
        # group_size=None means channelwise
        group_size = weight_quant.group_size or -1
        # Prefer to use the MarlinMoE kernel when it is supported.
        if not check_moe_marlin_supports_layer(layer, group_size):
            if (weight_quant.strategy in QuantizationStrategy.GROUP and
                    weight_quant.actorder in (ActivationOrdering.GROUP,
                                              ActivationOrdering.DYNAMIC)):
                raise ValueError(
                    "WNA16MoE is not supported with actorder=group/dynamic."
                )
            logger.info_once("Using CompressedTensorsWNA16MoEMethod")
            return CompressedTensorsWNA16MoEMethod(quant_config)
        else:
            logger.info_once("Using CompressedTensorsWNA16MarlinMoEMethod")
            return CompressedTensorsWNA16MarlinMoEMethod(quant_config)
    elif quant_config._is_fp4a4_nvfp4(weight_quant, input_quant):
        return CompressedTensorsW4A4MoeMethod()
    elif quant_config._is_fp8_w8a8_sm90(weight_quant, input_quant):
        return CompressedTensorsW8A8Fp8MoECutlassMethod(quant_config)
    elif quant_config._is_fp8_w8a8(weight_quant, input_quant):
        return CompressedTensorsW8A8Fp8MoEMethod(quant_config)
    elif quant_config._is_dynamic_token_w8a8(weight_quant, input_quant):
        return CompressedTensorsW8A8Int8MoEMethod(quant_config)
    else:
        raise RuntimeError(
            f"Unsupported FusedMoe scheme: {weight_quant}, {input_quant}")

CompressedTensorsW4A4MoeMethod

Bases: CompressedTensorsMoEMethod

Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
 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
177
178
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
320
321
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
class CompressedTensorsW4A4MoeMethod(CompressedTensorsMoEMethod):

    def __init__(self):
        self.use_marlin = not cutlass_fp4_supported()
        self.group_size = 16

    def create_weights(self, layer: torch.nn.Module, num_experts: int,
                       hidden_size: int, intermediate_size_per_partition: int,
                       params_dtype: torch.dtype, **extra_weight_attrs):

        layer.num_experts = num_experts
        layer.params_dtype = params_dtype

        w13_weight = torch.nn.Parameter(
            torch.empty(
                num_experts,
                2 * intermediate_size_per_partition,
                # 2 fp4 items are packed in the input dimension
                hidden_size // 2,
                requires_grad=False,
                dtype=torch.uint8),
            requires_grad=False)
        layer.register_parameter("w13_weight_packed", w13_weight)
        set_weight_attrs(w13_weight, extra_weight_attrs)

        w2_weight = torch.nn.Parameter(
            torch.empty(
                num_experts,
                hidden_size,
                # 2 fp4 items are packed in the input dimension
                intermediate_size_per_partition // 2,
                dtype=torch.uint8),
            requires_grad=False)
        layer.register_parameter("w2_weight_packed", w2_weight)
        set_weight_attrs(w2_weight, extra_weight_attrs)

        # Weight Scales
        w13_weight_scale = torch.nn.Parameter(
            torch.empty(
                num_experts,
                2 * intermediate_size_per_partition,
                # 2 fp4 items are packed in the input dimension
                hidden_size // self.group_size,
                dtype=torch.float8_e4m3fn),
            requires_grad=False)
        layer.register_parameter("w13_weight_scale", w13_weight_scale)
        extra_weight_attrs.update(
            {"quant_method": FusedMoeWeightScaleSupported.GROUP.value})
        set_weight_attrs(w13_weight_scale, extra_weight_attrs)

        w2_weight_scale = torch.nn.Parameter(
            torch.empty(
                num_experts,
                hidden_size,
                # 2 fp4 items are packed in the input dimension
                intermediate_size_per_partition // self.group_size,
                dtype=torch.float8_e4m3fn),
            requires_grad=False)
        layer.register_parameter("w2_weight_scale", w2_weight_scale)
        extra_weight_attrs.update(
            {"quant_method": FusedMoeWeightScaleSupported.GROUP.value})
        set_weight_attrs(w2_weight_scale, extra_weight_attrs)

        # Weight Global Scales
        w13_weight_scale_2 = torch.nn.Parameter(torch.empty(
            num_experts, 2, dtype=torch.float32),
                                                requires_grad=False)
        layer.register_parameter("w13_weight_global_scale", w13_weight_scale_2)
        extra_weight_attrs.update(
            {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value})
        set_weight_attrs(w13_weight_scale_2, extra_weight_attrs)

        w2_weight_scale_2 = torch.nn.Parameter(torch.empty(
            num_experts, dtype=torch.float32),
                                               requires_grad=False)
        layer.register_parameter("w2_weight_global_scale", w2_weight_scale_2)
        extra_weight_attrs.update(
            {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value})
        set_weight_attrs(w2_weight_scale_2, extra_weight_attrs)

        # Input Global Scales
        w13_input_scale = torch.nn.Parameter(torch.empty(num_experts,
                                                         2,
                                                         dtype=torch.float32),
                                             requires_grad=False)
        layer.register_parameter("w13_input_global_scale", w13_input_scale)
        extra_weight_attrs.update(
            {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value})
        set_weight_attrs(w13_input_scale, extra_weight_attrs)

        w2_input_scale = torch.nn.Parameter(torch.empty(num_experts,
                                                        dtype=torch.float32),
                                            requires_grad=False)
        layer.register_parameter("w2_input_global_scale", w2_input_scale)
        extra_weight_attrs.update(
            {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value})
        set_weight_attrs(w2_input_scale, extra_weight_attrs)

    def swizzle_blockscale(self, scale: torch.tensor):
        assert (scale.dtype == torch.float8_e4m3fn)
        # Pad and blockwise interleave weight_scale
        scale_ndim = scale.ndim
        if scale.ndim == 2:
            scale = scale.unsqueeze(0)
        assert scale.ndim == 3
        B, M, K = scale.shape
        round_up_multiple = lambda x, m: (x + m - 1) // m * m
        M_padded = round_up_multiple(M, 128)
        K_padded = round_up_multiple(K, 4)
        padded_scale = torch.zeros((B, M_padded, K_padded), dtype=scale.dtype)
        padded_scale[:B, :M, :K] = scale
        batches, rows, cols = padded_scale.shape
        assert rows % 128 == 0
        assert cols % 4 == 0
        padded_scale = padded_scale.reshape(batches, rows // 128, 4, 32,
                                            cols // 4, 4)
        swizzled_scale = padded_scale.permute((0, 1, 4, 3, 2, 5))
        swizzled_scale = swizzled_scale.contiguous().cuda()
        return (swizzled_scale.reshape(M, K)
                if scale_ndim == 2 else swizzled_scale.reshape(B, M, K))

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:

        # From packed to weight
        layer.w13_weight = torch.nn.Parameter(layer.w13_weight_packed.data,
                                              requires_grad=False)

        layer.w2_weight = torch.nn.Parameter(layer.w2_weight_packed.data,
                                             requires_grad=False)

        if not torch.allclose(layer.w13_weight_global_scale[:, 0],
                              layer.w13_weight_global_scale[:, 1]):
            logger.warning_once(
                "w1_weight_global_scale must match w3_weight_global_scale. "
                "Accuracy may be affected.")

        # Take inverse of global scale saved to disk
        layer.w13_weight_scale_2 = torch.nn.Parameter(
            1 / layer.w13_weight_global_scale[:, 0], requires_grad=False)

        layer.w2_weight_scale_2 = torch.nn.Parameter(
            1 / layer.w2_weight_global_scale.data, requires_grad=False)

        if self.use_marlin:
            prepare_moe_fp4_layer_for_marlin(layer)
            return

        # swizzle weight scales
        layer.w13_blockscale_swizzled = torch.nn.Parameter(
            self.swizzle_blockscale(layer.w13_weight_scale),
            requires_grad=False)

        layer.w2_blockscale_swizzled = torch.nn.Parameter(
            self.swizzle_blockscale(layer.w2_weight_scale),
            requires_grad=False)

        # w13
        w13_input_global_scale = layer.w13_input_global_scale.max(
            dim=1).values.to(torch.float32)

        layer.g1_alphas = torch.nn.Parameter(
            ((1 / w13_input_global_scale) * layer.w13_weight_scale_2),
            requires_grad=False)

        layer.w13_input_scale_quant = torch.nn.Parameter(
            (w13_input_global_scale), requires_grad=False)

        # w2
        layer.g2_alphas = torch.nn.Parameter(
            ((1 / layer.w2_input_global_scale) * layer.w2_weight_scale_2).to(
                torch.float32),
            requires_grad=False)

        layer.w2_input_scale_quant = torch.nn.Parameter(
            (layer.w2_input_global_scale), requires_grad=False)

    def apply(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        router_logits: torch.Tensor,
        top_k: int,
        renormalize: bool,
        use_grouped_topk: bool = False,
        topk_group: Optional[int] = None,
        num_expert_group: Optional[int] = None,
        global_num_experts: int = -1,
        expert_map: Optional[torch.Tensor] = None,
        custom_routing_function: Optional[Callable] = None,
        scoring_func: str = "softmax",
        e_score_correction_bias: Optional[torch.Tensor] = None,
        apply_router_weight_on_input: bool = False,
        activation: str = "silu",
        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 "
                                      "`CompressedTensorsW4A4MoeMethod` yet.")

        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,
        )

        if self.use_marlin:
            return torch.ops.vllm.fused_marlin_moe(
                x,
                layer.w13_weight,
                layer.w2_weight,
                layer.w13_weight_scale,
                layer.w2_weight_scale,
                router_logits,
                topk_weights,
                topk_ids,
                global_scale1=layer.w13_weight_scale_2,
                global_scale2=layer.w2_weight_scale_2,
                quant_type_id=scalar_types.float4_e2m1f.id,
                apply_router_weight_on_input=apply_router_weight_on_input,
                global_num_experts=global_num_experts,
                expert_map=expert_map)

        assert activation == "silu", "Only SiLU activation is supported."
        assert not apply_router_weight_on_input, (
            "Router weight on input is not "
            "supported for CompressedTensorsW4A4MoeMethod.")
        assert expert_map is None, ("Expert Parallelism / expert_map "
                                    "is currently not supported for "
                                    "CompressedTensorsW4A4MoeMethod.")

        from vllm.model_executor.layers.fused_moe.cutlass_moe import (
            cutlass_moe_fp4)

        # Cutlass moe takes in activations in BF16/Half precision
        # and fp4 quantized weights loaded from the checkpoint
        return cutlass_moe_fp4(a=x,
                               w1_fp4=layer.w13_weight,
                               w1_blockscale=layer.w13_blockscale_swizzled,
                               w1_alphas=layer.g1_alphas,
                               w2_fp4=layer.w2_weight,
                               w2_blockscale=layer.w2_blockscale_swizzled,
                               w2_alphas=layer.g2_alphas,
                               topk_weights=topk_weights,
                               topk_ids=topk_ids,
                               m=x.shape[0],
                               n=layer.w2_weight.shape[2] * 2,
                               k=x.shape[1],
                               e=layer.w13_weight.shape[0],
                               a1_gscale=layer.w13_input_scale_quant,
                               a2_gscale=layer.w2_input_scale_quant,
                               device=x.device).to(x.dtype)

group_size instance-attribute 

group_size = 16

use_marlin instance-attribute 

use_marlin = not cutlass_fp4_supported()

__init__ 

__init__()
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
 99
100
101
def __init__(self):
    self.use_marlin = not cutlass_fp4_supported()
    self.group_size = 16

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/compressed_tensors/compressed_tensors_moe.py 
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
320
321
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
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 "
                                  "`CompressedTensorsW4A4MoeMethod` yet.")

    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,
    )

    if self.use_marlin:
        return torch.ops.vllm.fused_marlin_moe(
            x,
            layer.w13_weight,
            layer.w2_weight,
            layer.w13_weight_scale,
            layer.w2_weight_scale,
            router_logits,
            topk_weights,
            topk_ids,
            global_scale1=layer.w13_weight_scale_2,
            global_scale2=layer.w2_weight_scale_2,
            quant_type_id=scalar_types.float4_e2m1f.id,
            apply_router_weight_on_input=apply_router_weight_on_input,
            global_num_experts=global_num_experts,
            expert_map=expert_map)

    assert activation == "silu", "Only SiLU activation is supported."
    assert not apply_router_weight_on_input, (
        "Router weight on input is not "
        "supported for CompressedTensorsW4A4MoeMethod.")
    assert expert_map is None, ("Expert Parallelism / expert_map "
                                "is currently not supported for "
                                "CompressedTensorsW4A4MoeMethod.")

    from vllm.model_executor.layers.fused_moe.cutlass_moe import (
        cutlass_moe_fp4)

    # Cutlass moe takes in activations in BF16/Half precision
    # and fp4 quantized weights loaded from the checkpoint
    return cutlass_moe_fp4(a=x,
                           w1_fp4=layer.w13_weight,
                           w1_blockscale=layer.w13_blockscale_swizzled,
                           w1_alphas=layer.g1_alphas,
                           w2_fp4=layer.w2_weight,
                           w2_blockscale=layer.w2_blockscale_swizzled,
                           w2_alphas=layer.g2_alphas,
                           topk_weights=topk_weights,
                           topk_ids=topk_ids,
                           m=x.shape[0],
                           n=layer.w2_weight.shape[2] * 2,
                           k=x.shape[1],
                           e=layer.w13_weight.shape[0],
                           a1_gscale=layer.w13_input_scale_quant,
                           a2_gscale=layer.w2_input_scale_quant,
                           device=x.device).to(x.dtype)

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/compressed_tensors/compressed_tensors_moe.py 
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
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
def create_weights(self, layer: torch.nn.Module, num_experts: int,
                   hidden_size: int, intermediate_size_per_partition: int,
                   params_dtype: torch.dtype, **extra_weight_attrs):

    layer.num_experts = num_experts
    layer.params_dtype = params_dtype

    w13_weight = torch.nn.Parameter(
        torch.empty(
            num_experts,
            2 * intermediate_size_per_partition,
            # 2 fp4 items are packed in the input dimension
            hidden_size // 2,
            requires_grad=False,
            dtype=torch.uint8),
        requires_grad=False)
    layer.register_parameter("w13_weight_packed", w13_weight)
    set_weight_attrs(w13_weight, extra_weight_attrs)

    w2_weight = torch.nn.Parameter(
        torch.empty(
            num_experts,
            hidden_size,
            # 2 fp4 items are packed in the input dimension
            intermediate_size_per_partition // 2,
            dtype=torch.uint8),
        requires_grad=False)
    layer.register_parameter("w2_weight_packed", w2_weight)
    set_weight_attrs(w2_weight, extra_weight_attrs)

    # Weight Scales
    w13_weight_scale = torch.nn.Parameter(
        torch.empty(
            num_experts,
            2 * intermediate_size_per_partition,
            # 2 fp4 items are packed in the input dimension
            hidden_size // self.group_size,
            dtype=torch.float8_e4m3fn),
        requires_grad=False)
    layer.register_parameter("w13_weight_scale", w13_weight_scale)
    extra_weight_attrs.update(
        {"quant_method": FusedMoeWeightScaleSupported.GROUP.value})
    set_weight_attrs(w13_weight_scale, extra_weight_attrs)

    w2_weight_scale = torch.nn.Parameter(
        torch.empty(
            num_experts,
            hidden_size,
            # 2 fp4 items are packed in the input dimension
            intermediate_size_per_partition // self.group_size,
            dtype=torch.float8_e4m3fn),
        requires_grad=False)
    layer.register_parameter("w2_weight_scale", w2_weight_scale)
    extra_weight_attrs.update(
        {"quant_method": FusedMoeWeightScaleSupported.GROUP.value})
    set_weight_attrs(w2_weight_scale, extra_weight_attrs)

    # Weight Global Scales
    w13_weight_scale_2 = torch.nn.Parameter(torch.empty(
        num_experts, 2, dtype=torch.float32),
                                            requires_grad=False)
    layer.register_parameter("w13_weight_global_scale", w13_weight_scale_2)
    extra_weight_attrs.update(
        {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value})
    set_weight_attrs(w13_weight_scale_2, extra_weight_attrs)

    w2_weight_scale_2 = torch.nn.Parameter(torch.empty(
        num_experts, dtype=torch.float32),
                                           requires_grad=False)
    layer.register_parameter("w2_weight_global_scale", w2_weight_scale_2)
    extra_weight_attrs.update(
        {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value})
    set_weight_attrs(w2_weight_scale_2, extra_weight_attrs)

    # Input Global Scales
    w13_input_scale = torch.nn.Parameter(torch.empty(num_experts,
                                                     2,
                                                     dtype=torch.float32),
                                         requires_grad=False)
    layer.register_parameter("w13_input_global_scale", w13_input_scale)
    extra_weight_attrs.update(
        {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value})
    set_weight_attrs(w13_input_scale, extra_weight_attrs)

    w2_input_scale = torch.nn.Parameter(torch.empty(num_experts,
                                                    dtype=torch.float32),
                                        requires_grad=False)
    layer.register_parameter("w2_input_global_scale", w2_input_scale)
    extra_weight_attrs.update(
        {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value})
    set_weight_attrs(w2_input_scale, extra_weight_attrs)

process_weights_after_loading 

process_weights_after_loading(layer: Module) -> None
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
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
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:

    # From packed to weight
    layer.w13_weight = torch.nn.Parameter(layer.w13_weight_packed.data,
                                          requires_grad=False)

    layer.w2_weight = torch.nn.Parameter(layer.w2_weight_packed.data,
                                         requires_grad=False)

    if not torch.allclose(layer.w13_weight_global_scale[:, 0],
                          layer.w13_weight_global_scale[:, 1]):
        logger.warning_once(
            "w1_weight_global_scale must match w3_weight_global_scale. "
            "Accuracy may be affected.")

    # Take inverse of global scale saved to disk
    layer.w13_weight_scale_2 = torch.nn.Parameter(
        1 / layer.w13_weight_global_scale[:, 0], requires_grad=False)

    layer.w2_weight_scale_2 = torch.nn.Parameter(
        1 / layer.w2_weight_global_scale.data, requires_grad=False)

    if self.use_marlin:
        prepare_moe_fp4_layer_for_marlin(layer)
        return

    # swizzle weight scales
    layer.w13_blockscale_swizzled = torch.nn.Parameter(
        self.swizzle_blockscale(layer.w13_weight_scale),
        requires_grad=False)

    layer.w2_blockscale_swizzled = torch.nn.Parameter(
        self.swizzle_blockscale(layer.w2_weight_scale),
        requires_grad=False)

    # w13
    w13_input_global_scale = layer.w13_input_global_scale.max(
        dim=1).values.to(torch.float32)

    layer.g1_alphas = torch.nn.Parameter(
        ((1 / w13_input_global_scale) * layer.w13_weight_scale_2),
        requires_grad=False)

    layer.w13_input_scale_quant = torch.nn.Parameter(
        (w13_input_global_scale), requires_grad=False)

    # w2
    layer.g2_alphas = torch.nn.Parameter(
        ((1 / layer.w2_input_global_scale) * layer.w2_weight_scale_2).to(
            torch.float32),
        requires_grad=False)

    layer.w2_input_scale_quant = torch.nn.Parameter(
        (layer.w2_input_global_scale), requires_grad=False)

swizzle_blockscale 

swizzle_blockscale(scale: tensor)
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
def swizzle_blockscale(self, scale: torch.tensor):
    assert (scale.dtype == torch.float8_e4m3fn)
    # Pad and blockwise interleave weight_scale
    scale_ndim = scale.ndim
    if scale.ndim == 2:
        scale = scale.unsqueeze(0)
    assert scale.ndim == 3
    B, M, K = scale.shape
    round_up_multiple = lambda x, m: (x + m - 1) // m * m
    M_padded = round_up_multiple(M, 128)
    K_padded = round_up_multiple(K, 4)
    padded_scale = torch.zeros((B, M_padded, K_padded), dtype=scale.dtype)
    padded_scale[:B, :M, :K] = scale
    batches, rows, cols = padded_scale.shape
    assert rows % 128 == 0
    assert cols % 4 == 0
    padded_scale = padded_scale.reshape(batches, rows // 128, 4, 32,
                                        cols // 4, 4)
    swizzled_scale = padded_scale.permute((0, 1, 4, 3, 2, 5))
    swizzled_scale = swizzled_scale.contiguous().cuda()
    return (swizzled_scale.reshape(M, K)
            if scale_ndim == 2 else swizzled_scale.reshape(B, M, K))

CompressedTensorsW8A8Fp8MoECutlassMethod

Bases: CompressedTensorsMoEMethod

Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
class CompressedTensorsW8A8Fp8MoECutlassMethod(CompressedTensorsMoEMethod):

    def __init__(
            self,
            quant_config: "CompressedTensorsConfig"  # type: ignore # noqa E501
    ):
        self.quant_config = quant_config
        self.weight_quant = self.quant_config.target_scheme_map["Linear"].get(
            "weights")
        self.input_quant = self.quant_config.target_scheme_map["Linear"].get(
            "input_activations")

        per_tensor = (self.weight_quant.strategy == QuantizationStrategy.TENSOR
                      and self.input_quant.strategy
                      == QuantizationStrategy.TENSOR)
        per_channel = (
            self.weight_quant.strategy == QuantizationStrategy.CHANNEL
            and self.input_quant.strategy == QuantizationStrategy.TOKEN)
        if not (per_tensor or per_channel):
            raise ValueError(
                "For FP8 Fused MoE layers, we require per tensor "
                "or channelwise, dynamic per token quantization. Found "
                f"{self.weight_quant}, {self.input_quant}")

        self.static_input_scales = not self.input_quant.dynamic
        if self.static_input_scales and per_channel:
            raise ValueError(
                "For FP8 Fused MoE layer, we require either per tensor or "
                "channelwise, dynamic per token quantization.")

        from vllm.model_executor.layers.fused_moe.cutlass_moe import (
            cutlass_moe_fp8)
        self.fused_experts = cutlass_moe_fp8  # type: ignore
        self.disable_expert_map = False

    def create_weights(self, layer: torch.nn.Module, num_experts: int,
                       hidden_size: int, intermediate_size_per_partition: int,
                       params_dtype: torch.dtype, **extra_weight_attrs):

        params_dtype = torch.float8_e4m3fn

        # WEIGHTS
        w13_weight = torch.nn.Parameter(torch.empty(
            num_experts,
            2 * intermediate_size_per_partition,
            hidden_size,
            dtype=params_dtype),
                                        requires_grad=False)
        layer.register_parameter("w13_weight", w13_weight)
        set_weight_attrs(w13_weight, extra_weight_attrs)

        w2_weight = torch.nn.Parameter(torch.empty(
            num_experts,
            hidden_size,
            intermediate_size_per_partition,
            dtype=params_dtype),
                                       requires_grad=False)
        layer.register_parameter("w2_weight", w2_weight)
        set_weight_attrs(w2_weight, extra_weight_attrs)

        # WEIGHT_SCALES
        if self.weight_quant.strategy == QuantizationStrategy.TENSOR:
            # Allocate 2 scales for w1 and w3 respectively.
            # They are combined to a single scale after weight loading.
            w13_weight_scale = torch.nn.Parameter(torch.ones(
                num_experts, 2, dtype=torch.float32),
                                                  requires_grad=False)
            layer.register_parameter("w13_weight_scale", w13_weight_scale)
            w2_weight_scale = torch.nn.Parameter(torch.ones(
                num_experts, dtype=torch.float32),
                                                 requires_grad=False)
            layer.register_parameter("w2_weight_scale", w2_weight_scale)
            # Add PER-TENSOR quantization for FusedMoE.weight_loader.
            extra_weight_attrs.update(
                {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value})
            set_weight_attrs(w13_weight_scale, extra_weight_attrs)
            set_weight_attrs(w2_weight_scale, extra_weight_attrs)

        elif self.weight_quant.strategy == QuantizationStrategy.CHANNEL:
            w13_weight_scale = torch.nn.Parameter(torch.ones(
                num_experts,
                2 * intermediate_size_per_partition,
                1,
                dtype=torch.float32),
                                                  requires_grad=False)
            layer.register_parameter("w13_weight_scale", w13_weight_scale)
            w2_weight_scale = torch.nn.Parameter(torch.ones(
                num_experts, hidden_size, 1, dtype=torch.float32),
                                                 requires_grad=False)
            layer.register_parameter("w2_weight_scale", w2_weight_scale)
            # Add PER-CHANNEL quantization for FusedMoE.weight_loader.
            extra_weight_attrs.update(
                {"quant_method": FusedMoeWeightScaleSupported.CHANNEL.value})
            set_weight_attrs(w13_weight_scale, extra_weight_attrs)
            set_weight_attrs(w2_weight_scale, extra_weight_attrs)

        # INPUT_SCALES
        if self.static_input_scales:
            w13_input_scale = torch.nn.Parameter(torch.ones(
                num_experts, dtype=torch.float32),
                                                 requires_grad=False)
            layer.register_parameter("w13_input_scale", w13_input_scale)
            set_weight_attrs(w13_input_scale, extra_weight_attrs)

            w2_input_scale = torch.nn.Parameter(torch.ones(
                num_experts, dtype=torch.float32),
                                                requires_grad=False)
            layer.register_parameter("w2_input_scale", w2_input_scale)
            set_weight_attrs(w2_input_scale, extra_weight_attrs)
        else:
            layer.w13_input_scale = None
            layer.w2_input_scale = None

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        # Fp8 moe kernels require a single activation scale.
        # We take the max of all the scales in case they differ.
        if self.static_input_scales:
            assert self.input_quant.strategy == QuantizationStrategy.TENSOR
            if (layer.w13_input_scale is None or layer.w2_input_scale is None):
                raise ValueError(
                    "QuantConfig has static quantization, but found "
                    "activation scales are None.")
            if (not all_close_1d(layer.w13_input_scale)
                    or not all_close_1d(layer.w2_input_scale)):
                logger.warning_once(
                    "Found input_scales that are not equal for "
                    "fp8 MoE layer. Using the maximum across experts "
                    "for each layer.")
            layer.w13_input_scale = torch.nn.Parameter(
                layer.w13_input_scale.max(), requires_grad=False)
            layer.w2_input_scale = torch.nn.Parameter(
                layer.w2_input_scale.max(), requires_grad=False)

        # For Per-TENSOR case, Fp8 moe kernel needs single weight scale
        # for w13 per expert. Use max then dequant and requant each expert.
        if self.weight_quant.strategy == QuantizationStrategy.TENSOR:
            assert layer.w13_weight_scale is not None
            shard_size = layer.intermediate_size_per_partition
            max_w13_scales = layer.w13_weight_scale.max(dim=1).values
            for expert_id in range(layer.local_num_experts):
                start = 0
                for shard_id in range(2):
                    dq_weight = per_tensor_dequantize(
                        layer.w13_weight[expert_id][start:start +
                                                    shard_size, :],
                        layer.w13_weight_scale[expert_id][shard_id])
                    layer.w13_weight[expert_id][
                        start:start + shard_size, :], _ = ops.scaled_fp8_quant(
                            dq_weight, max_w13_scales[expert_id])
                    start += shard_size
            layer.w13_weight_scale = torch.nn.Parameter(max_w13_scales,
                                                        requires_grad=False)

    def select_gemm_impl(
        self,
        prepare_finalize: FusedMoEPrepareAndFinalize,
        moe: FusedMoEConfig,
    ) -> FusedMoEPermuteExpertsUnpermute:
        from vllm.model_executor.layers.fused_moe import CutlassExpertsFp8

        use_batched_format = (prepare_finalize.activation_format ==
                              FusedMoEActivationFormat.BatchedExperts)

        num_dispatchers = prepare_finalize.num_dispatchers()

        num_experts = (moe.num_local_experts
                       if use_batched_format else moe.num_experts)

        logger.debug("CutlassExpertsFp8(%s)", self.__class__.__name__)

        experts = CutlassExpertsFp8(
            num_experts,
            moe.in_dtype,
            self.input_quant.strategy == QuantizationStrategy.TOKEN,
            self.weight_quant.strategy == QuantizationStrategy.CHANNEL,
            num_dispatchers=num_dispatchers,
            use_batched_format=use_batched_format,
        )

        self.disable_expert_map = (num_dispatchers > 1
                                   or not experts.supports_expert_map())

        return experts

    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 "
                "`CompressedTensorsW8A8Fp8MoECutlassMethod` yet.")

        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,
            indices_type=self.topk_indices_dtype,
        )

        return self.fused_experts(
            x,
            layer.w13_weight,
            layer.w2_weight,
            topk_weights,
            topk_ids,
            activation=activation,
            global_num_experts=global_num_experts,
            expert_map=None if self.disable_expert_map else expert_map,
            w1_scale=layer.w13_weight_scale,
            w2_scale=layer.w2_weight_scale,
            a1_scale=layer.w13_input_scale,
            a2_scale=layer.w2_input_scale,
        )

disable_expert_map instance-attribute 

disable_expert_map = False

fused_experts instance-attribute 

fused_experts = cutlass_moe_fp8

input_quant instance-attribute 

input_quant = get('input_activations')

quant_config instance-attribute 

quant_config = quant_config

static_input_scales instance-attribute 

static_input_scales = not dynamic

weight_quant instance-attribute 

weight_quant = get('weights')

__init__ 

__init__(quant_config: CompressedTensorsConfig)
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
def __init__(
        self,
        quant_config: "CompressedTensorsConfig"  # type: ignore # noqa E501
):
    self.quant_config = quant_config
    self.weight_quant = self.quant_config.target_scheme_map["Linear"].get(
        "weights")
    self.input_quant = self.quant_config.target_scheme_map["Linear"].get(
        "input_activations")

    per_tensor = (self.weight_quant.strategy == QuantizationStrategy.TENSOR
                  and self.input_quant.strategy
                  == QuantizationStrategy.TENSOR)
    per_channel = (
        self.weight_quant.strategy == QuantizationStrategy.CHANNEL
        and self.input_quant.strategy == QuantizationStrategy.TOKEN)
    if not (per_tensor or per_channel):
        raise ValueError(
            "For FP8 Fused MoE layers, we require per tensor "
            "or channelwise, dynamic per token quantization. Found "
            f"{self.weight_quant}, {self.input_quant}")

    self.static_input_scales = not self.input_quant.dynamic
    if self.static_input_scales and per_channel:
        raise ValueError(
            "For FP8 Fused MoE layer, we require either per tensor or "
            "channelwise, dynamic per token quantization.")

    from vllm.model_executor.layers.fused_moe.cutlass_moe import (
        cutlass_moe_fp8)
    self.fused_experts = cutlass_moe_fp8  # type: ignore
    self.disable_expert_map = False

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/compressed_tensors/compressed_tensors_moe.py 
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
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 "
            "`CompressedTensorsW8A8Fp8MoECutlassMethod` yet.")

    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,
        indices_type=self.topk_indices_dtype,
    )

    return self.fused_experts(
        x,
        layer.w13_weight,
        layer.w2_weight,
        topk_weights,
        topk_ids,
        activation=activation,
        global_num_experts=global_num_experts,
        expert_map=None if self.disable_expert_map else expert_map,
        w1_scale=layer.w13_weight_scale,
        w2_scale=layer.w2_weight_scale,
        a1_scale=layer.w13_input_scale,
        a2_scale=layer.w2_input_scale,
    )

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/compressed_tensors/compressed_tensors_moe.py 
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
def create_weights(self, layer: torch.nn.Module, num_experts: int,
                   hidden_size: int, intermediate_size_per_partition: int,
                   params_dtype: torch.dtype, **extra_weight_attrs):

    params_dtype = torch.float8_e4m3fn

    # WEIGHTS
    w13_weight = torch.nn.Parameter(torch.empty(
        num_experts,
        2 * intermediate_size_per_partition,
        hidden_size,
        dtype=params_dtype),
                                    requires_grad=False)
    layer.register_parameter("w13_weight", w13_weight)
    set_weight_attrs(w13_weight, extra_weight_attrs)

    w2_weight = torch.nn.Parameter(torch.empty(
        num_experts,
        hidden_size,
        intermediate_size_per_partition,
        dtype=params_dtype),
                                   requires_grad=False)
    layer.register_parameter("w2_weight", w2_weight)
    set_weight_attrs(w2_weight, extra_weight_attrs)

    # WEIGHT_SCALES
    if self.weight_quant.strategy == QuantizationStrategy.TENSOR:
        # Allocate 2 scales for w1 and w3 respectively.
        # They are combined to a single scale after weight loading.
        w13_weight_scale = torch.nn.Parameter(torch.ones(
            num_experts, 2, dtype=torch.float32),
                                              requires_grad=False)
        layer.register_parameter("w13_weight_scale", w13_weight_scale)
        w2_weight_scale = torch.nn.Parameter(torch.ones(
            num_experts, dtype=torch.float32),
                                             requires_grad=False)
        layer.register_parameter("w2_weight_scale", w2_weight_scale)
        # Add PER-TENSOR quantization for FusedMoE.weight_loader.
        extra_weight_attrs.update(
            {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value})
        set_weight_attrs(w13_weight_scale, extra_weight_attrs)
        set_weight_attrs(w2_weight_scale, extra_weight_attrs)

    elif self.weight_quant.strategy == QuantizationStrategy.CHANNEL:
        w13_weight_scale = torch.nn.Parameter(torch.ones(
            num_experts,
            2 * intermediate_size_per_partition,
            1,
            dtype=torch.float32),
                                              requires_grad=False)
        layer.register_parameter("w13_weight_scale", w13_weight_scale)
        w2_weight_scale = torch.nn.Parameter(torch.ones(
            num_experts, hidden_size, 1, dtype=torch.float32),
                                             requires_grad=False)
        layer.register_parameter("w2_weight_scale", w2_weight_scale)
        # Add PER-CHANNEL quantization for FusedMoE.weight_loader.
        extra_weight_attrs.update(
            {"quant_method": FusedMoeWeightScaleSupported.CHANNEL.value})
        set_weight_attrs(w13_weight_scale, extra_weight_attrs)
        set_weight_attrs(w2_weight_scale, extra_weight_attrs)

    # INPUT_SCALES
    if self.static_input_scales:
        w13_input_scale = torch.nn.Parameter(torch.ones(
            num_experts, dtype=torch.float32),
                                             requires_grad=False)
        layer.register_parameter("w13_input_scale", w13_input_scale)
        set_weight_attrs(w13_input_scale, extra_weight_attrs)

        w2_input_scale = torch.nn.Parameter(torch.ones(
            num_experts, dtype=torch.float32),
                                            requires_grad=False)
        layer.register_parameter("w2_input_scale", w2_input_scale)
        set_weight_attrs(w2_input_scale, extra_weight_attrs)
    else:
        layer.w13_input_scale = None
        layer.w2_input_scale = None

process_weights_after_loading 

process_weights_after_loading(layer: Module) -> None
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
    # Fp8 moe kernels require a single activation scale.
    # We take the max of all the scales in case they differ.
    if self.static_input_scales:
        assert self.input_quant.strategy == QuantizationStrategy.TENSOR
        if (layer.w13_input_scale is None or layer.w2_input_scale is None):
            raise ValueError(
                "QuantConfig has static quantization, but found "
                "activation scales are None.")
        if (not all_close_1d(layer.w13_input_scale)
                or not all_close_1d(layer.w2_input_scale)):
            logger.warning_once(
                "Found input_scales that are not equal for "
                "fp8 MoE layer. Using the maximum across experts "
                "for each layer.")
        layer.w13_input_scale = torch.nn.Parameter(
            layer.w13_input_scale.max(), requires_grad=False)
        layer.w2_input_scale = torch.nn.Parameter(
            layer.w2_input_scale.max(), requires_grad=False)

    # For Per-TENSOR case, Fp8 moe kernel needs single weight scale
    # for w13 per expert. Use max then dequant and requant each expert.
    if self.weight_quant.strategy == QuantizationStrategy.TENSOR:
        assert layer.w13_weight_scale is not None
        shard_size = layer.intermediate_size_per_partition
        max_w13_scales = layer.w13_weight_scale.max(dim=1).values
        for expert_id in range(layer.local_num_experts):
            start = 0
            for shard_id in range(2):
                dq_weight = per_tensor_dequantize(
                    layer.w13_weight[expert_id][start:start +
                                                shard_size, :],
                    layer.w13_weight_scale[expert_id][shard_id])
                layer.w13_weight[expert_id][
                    start:start + shard_size, :], _ = ops.scaled_fp8_quant(
                        dq_weight, max_w13_scales[expert_id])
                start += shard_size
        layer.w13_weight_scale = torch.nn.Parameter(max_w13_scales,
                                                    requires_grad=False)

select_gemm_impl 

select_gemm_impl(
    prepare_finalize: FusedMoEPrepareAndFinalize,
    moe: FusedMoEConfig,
) -> FusedMoEPermuteExpertsUnpermute
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
def select_gemm_impl(
    self,
    prepare_finalize: FusedMoEPrepareAndFinalize,
    moe: FusedMoEConfig,
) -> FusedMoEPermuteExpertsUnpermute:
    from vllm.model_executor.layers.fused_moe import CutlassExpertsFp8

    use_batched_format = (prepare_finalize.activation_format ==
                          FusedMoEActivationFormat.BatchedExperts)

    num_dispatchers = prepare_finalize.num_dispatchers()

    num_experts = (moe.num_local_experts
                   if use_batched_format else moe.num_experts)

    logger.debug("CutlassExpertsFp8(%s)", self.__class__.__name__)

    experts = CutlassExpertsFp8(
        num_experts,
        moe.in_dtype,
        self.input_quant.strategy == QuantizationStrategy.TOKEN,
        self.weight_quant.strategy == QuantizationStrategy.CHANNEL,
        num_dispatchers=num_dispatchers,
        use_batched_format=use_batched_format,
    )

    self.disable_expert_map = (num_dispatchers > 1
                               or not experts.supports_expert_map())

    return experts

CompressedTensorsW8A8Fp8MoEMethod

Bases: CompressedTensorsMoEMethod

Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
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
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):

    def __init__(
            self,
            quant_config: "CompressedTensorsConfig"  # type: ignore # noqa E501
    ):
        self.quant_config = quant_config
        self.weight_quant = self.quant_config.target_scheme_map["Linear"].get(
            "weights")
        self.input_quant = self.quant_config.target_scheme_map["Linear"].get(
            "input_activations")

        per_tensor = (self.weight_quant.strategy == QuantizationStrategy.TENSOR
                      and self.input_quant.strategy
                      == QuantizationStrategy.TENSOR)
        per_channel = (
            self.weight_quant.strategy == QuantizationStrategy.CHANNEL
            and self.input_quant.strategy == QuantizationStrategy.TOKEN)
        if not (per_tensor or per_channel):
            raise ValueError(
                "For FP8 Fused MoE layers, we require per tensor "
                "or channelwise, dynamic per token quantization. Found "
                f"{self.weight_quant}, {self.input_quant}")

        self.static_input_scales = not self.input_quant.dynamic
        if self.static_input_scales and per_channel:
            raise ValueError(
                "For FP8 Fused MoE layer, we require either per tensor or "
                "channelwise, dynamic per token quantization.")

        # For GPUs that lack FP8 hardware support, we can leverage the Marlin
        # kernel for fast weight-only FP8 quantization
        self.use_marlin = (not current_platform.has_device_capability(89)
                           or envs.VLLM_TEST_FORCE_FP8_MARLIN)
        # Disable marlin for rocm
        if current_platform.is_rocm():
            self.use_marlin = False
        from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
            is_rocm_aiter_moe_enabled)

        self.rocm_aiter_moe_enabled = is_rocm_aiter_moe_enabled()

    def create_weights(self, layer: torch.nn.Module, num_experts: int,
                       hidden_size: int, intermediate_size_per_partition: int,
                       params_dtype: torch.dtype, **extra_weight_attrs):

        layer.intermediate_size_per_partition = intermediate_size_per_partition
        layer.hidden_size = hidden_size
        layer.num_experts = num_experts
        layer.orig_dtype = params_dtype
        layer.weight_block_size = None

        params_dtype = torch.float8_e4m3fn

        # WEIGHTS
        w13_weight = torch.nn.Parameter(torch.empty(
            num_experts,
            2 * intermediate_size_per_partition,
            hidden_size,
            dtype=params_dtype),
                                        requires_grad=False)
        layer.register_parameter("w13_weight", w13_weight)
        set_weight_attrs(w13_weight, extra_weight_attrs)

        w2_weight = torch.nn.Parameter(torch.empty(
            num_experts,
            hidden_size,
            intermediate_size_per_partition,
            dtype=params_dtype),
                                       requires_grad=False)
        layer.register_parameter("w2_weight", w2_weight)
        set_weight_attrs(w2_weight, extra_weight_attrs)

        # WEIGHT_SCALES
        if self.weight_quant.strategy == QuantizationStrategy.TENSOR:
            # Allocate 2 scales for w1 and w3 respectively.
            # They are combined to a single scale after weight loading.
            w13_weight_scale = torch.nn.Parameter(torch.ones(
                num_experts, 2, dtype=torch.float32),
                                                  requires_grad=False)
            layer.register_parameter("w13_weight_scale", w13_weight_scale)
            w2_weight_scale = torch.nn.Parameter(torch.ones(
                num_experts, dtype=torch.float32),
                                                 requires_grad=False)
            layer.register_parameter("w2_weight_scale", w2_weight_scale)
            # Add PER-TENSOR quantization for FusedMoE.weight_loader.
            extra_weight_attrs.update(
                {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value})
            set_weight_attrs(w13_weight_scale, extra_weight_attrs)
            set_weight_attrs(w2_weight_scale, extra_weight_attrs)

        elif self.weight_quant.strategy == QuantizationStrategy.CHANNEL:
            w13_weight_scale = torch.nn.Parameter(torch.ones(
                num_experts,
                2 * intermediate_size_per_partition,
                1,
                dtype=torch.float32),
                                                  requires_grad=False)
            layer.register_parameter("w13_weight_scale", w13_weight_scale)
            w2_weight_scale = torch.nn.Parameter(torch.ones(
                num_experts, hidden_size, 1, dtype=torch.float32),
                                                 requires_grad=False)
            layer.register_parameter("w2_weight_scale", w2_weight_scale)
            # Add PER-CHANNEL quantization for FusedMoE.weight_loader.
            extra_weight_attrs.update(
                {"quant_method": FusedMoeWeightScaleSupported.CHANNEL.value})
            set_weight_attrs(w13_weight_scale, extra_weight_attrs)
            set_weight_attrs(w2_weight_scale, extra_weight_attrs)

        # INPUT_SCALES
        if self.static_input_scales:
            w13_input_scale = torch.nn.Parameter(torch.ones(
                num_experts, dtype=torch.float32),
                                                 requires_grad=False)
            layer.register_parameter("w13_input_scale", w13_input_scale)
            set_weight_attrs(w13_input_scale, extra_weight_attrs)

            w2_input_scale = torch.nn.Parameter(torch.ones(
                num_experts, dtype=torch.float32),
                                                requires_grad=False)
            layer.register_parameter("w2_input_scale", w2_input_scale)
            set_weight_attrs(w2_input_scale, extra_weight_attrs)
        else:
            layer.w13_input_scale = None
            layer.w2_input_scale = None

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        # Fp8 moe kernels require a single activation scale.
        # We take the max of all the scales in case they differ.
        if self.static_input_scales:
            assert self.input_quant.strategy == QuantizationStrategy.TENSOR
            if (layer.w13_input_scale is None or layer.w2_input_scale is None):
                raise ValueError(
                    "QuantConfig has static quantization, but found "
                    "activation scales are None.")
            if (not all_close_1d(layer.w13_input_scale)
                    or not all_close_1d(layer.w2_input_scale)):
                logger.warning_once(
                    "Found input_scales that are not equal for "
                    "fp8 MoE layer. Using the maximum across experts "
                    "for each layer.")
            layer.w13_input_scale = torch.nn.Parameter(
                layer.w13_input_scale.max(), requires_grad=False)
            layer.w2_input_scale = torch.nn.Parameter(
                layer.w2_input_scale.max(), requires_grad=False)

        if current_platform.is_fp8_fnuz():
            # Normalize the weights and scales
            w13_weight, w13_weight_scale, w13_input_scale = \
                normalize_e4m3fn_to_e4m3fnuz(
                    layer.w13_weight, layer.w13_weight_scale,
                    layer.w13_input_scale)
            w2_weight, w2_weight_scale, w2_input_scale = \
                normalize_e4m3fn_to_e4m3fnuz(
                    layer.w2_weight, layer.w2_weight_scale,
                    layer.w2_input_scale)
            # Reset the parameter
            layer.w13_weight = torch.nn.Parameter(w13_weight,
                                                  requires_grad=False)
            layer.w13_weight_scale = torch.nn.Parameter(w13_weight_scale,
                                                        requires_grad=False)
            if w13_input_scale is not None:
                layer.w13_input_scale = torch.nn.Parameter(w13_input_scale,
                                                           requires_grad=False)
            layer.w2_weight = torch.nn.Parameter(w2_weight,
                                                 requires_grad=False)
            layer.w2_weight_scale = torch.nn.Parameter(w2_weight_scale,
                                                       requires_grad=False)
            if w2_input_scale is not None:
                layer.w2_input_scale = torch.nn.Parameter(w2_input_scale,
                                                          requires_grad=False)

        # For Per-TENSOR case, Fp8 moe kernel needs single weight scale
        # for w13 per expert. Use max then dequant and requant each expert.
        if self.weight_quant.strategy == QuantizationStrategy.TENSOR:
            assert layer.w13_weight_scale is not None
            shard_size = layer.intermediate_size_per_partition
            max_w13_scales = layer.w13_weight_scale.max(dim=1).values
            for expert_id in range(layer.local_num_experts):
                start = 0
                for shard_id in range(2):
                    dq_weight = per_tensor_dequantize(
                        layer.w13_weight[expert_id][start:start +
                                                    shard_size, :],
                        layer.w13_weight_scale[expert_id][shard_id])
                    layer.w13_weight[expert_id][
                        start:start + shard_size, :], _ = ops.scaled_fp8_quant(
                            dq_weight, max_w13_scales[expert_id])
                    start += shard_size
            layer.w13_weight_scale = torch.nn.Parameter(max_w13_scales,
                                                        requires_grad=False)

        # Property to determine if AITER is used
        if self.rocm_aiter_moe_enabled:
            from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (  # noqa E501
                rocm_aiter_fused_experts, shuffle_weights)

            # reshaping weights is required for aiter moe kernel.
            shuffled_w13, shuffled_w2 = shuffle_weights(
                layer.w13_weight.data, layer.w2_weight.data)

            layer.w13_weight = torch.nn.Parameter(shuffled_w13,
                                                  requires_grad=False)
            layer.w2_weight = torch.nn.Parameter(shuffled_w2,
                                                 requires_grad=False)

            self.rocm_aiter_fused_experts_func = rocm_aiter_fused_experts
        elif self.use_marlin:
            prepare_moe_fp8_layer_for_marlin(layer, False)
            # Activations not quantized for marlin.
            del layer.w13_input_scale
            del layer.w2_input_scale
            self.fused_experts_func = None
        else:
            from vllm.model_executor.layers.fused_moe import fused_experts
            self.fused_experts_func = fused_experts

    def select_gemm_impl(
        self,
        prepare_finalize: FusedMoEPrepareAndFinalize,
        moe: FusedMoEConfig,
    ) -> FusedMoEPermuteExpertsUnpermute:
        from vllm.model_executor.layers.fused_moe import TritonExperts
        from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
            BatchedTritonExperts)

        assert not self.rocm_aiter_moe_enabled and not self.use_marlin

        logger.debug("BatchedTritonExperts(%s)", self.__class__.__name__)

        if (prepare_finalize.activation_format ==
                FusedMoEActivationFormat.BatchedExperts):
            max_num_tokens_per_rank = prepare_finalize.max_num_tokens_per_rank(
            )
            assert max_num_tokens_per_rank is not None

            return BatchedTritonExperts(
                max_num_tokens=max_num_tokens_per_rank,
                num_dispatchers=prepare_finalize.num_dispatchers(),
                use_fp8_w8a8=True,
                block_shape=self.quant_config.weight_block_size,
                per_act_token_quant=(
                    self.input_quant.strategy == QuantizationStrategy.TOKEN),
            )
        else:
            return TritonExperts(
                use_fp8_w8a8=True,
                block_shape=self.quant_config.weight_block_size,
                per_act_token_quant=(
                    self.input_quant.strategy == QuantizationStrategy.TOKEN),
            )

    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 "
                "`CompressedTensorsW8A8Fp8MoEMethod` yet.")

        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,
            indices_type=self.topk_indices_dtype,
        )

        if self.rocm_aiter_moe_enabled:
            return self.rocm_aiter_fused_experts_func(
                hidden_states=x,
                w1=layer.w13_weight,
                w2=layer.w2_weight,
                topk_weights=topk_weights,
                topk_ids=topk_ids,
                activation=activation,
                apply_router_weight_on_input=apply_router_weight_on_input,
                use_fp8_w8a8=True,
                per_channel_quant=self.weight_quant.strategy ==
                QuantizationStrategy.CHANNEL,
                w1_scale=layer.w13_weight_scale,
                w2_scale=layer.w2_weight_scale,
                a1_scale=layer.w13_input_scale,
                a2_scale=layer.w2_input_scale,
                expert_map=expert_map)
        if self.use_marlin:
            assert activation == "silu", (
                f"{activation} not supported for Marlin MoE.")
            return torch.ops.vllm.fused_marlin_moe(
                x,
                layer.w13_weight,
                layer.w2_weight,
                layer.w13_weight_scale,
                layer.w2_weight_scale,
                router_logits,
                topk_weights,
                topk_ids,
                quant_type_id=scalar_types.float8_e4m3fn.id,
                apply_router_weight_on_input=apply_router_weight_on_input,
                global_num_experts=global_num_experts,
                expert_map=expert_map)

        assert self.fused_experts_func is not None

        return self.fused_experts_func(
            hidden_states=x,
            w1=layer.w13_weight,
            w2=layer.w2_weight,
            topk_weights=topk_weights,
            topk_ids=topk_ids,
            inplace=True,
            activation=activation,
            apply_router_weight_on_input=apply_router_weight_on_input,
            use_fp8_w8a8=True,
            per_channel_quant=self.weight_quant.strategy ==
            QuantizationStrategy.CHANNEL,
            global_num_experts=global_num_experts,
            expert_map=expert_map,
            w1_scale=layer.w13_weight_scale,
            w2_scale=layer.w2_weight_scale,
            a1_scale=layer.w13_input_scale,
            a2_scale=layer.w2_input_scale)

input_quant instance-attribute 

input_quant = get('input_activations')

quant_config instance-attribute 

quant_config = quant_config

rocm_aiter_moe_enabled instance-attribute 

rocm_aiter_moe_enabled = is_rocm_aiter_moe_enabled()

static_input_scales instance-attribute 

static_input_scales = not dynamic

use_marlin instance-attribute 

use_marlin = (
    not has_device_capability(89)
    or VLLM_TEST_FORCE_FP8_MARLIN
)

weight_quant instance-attribute 

weight_quant = get('weights')

__init__ 

__init__(quant_config: CompressedTensorsConfig)
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
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
def __init__(
        self,
        quant_config: "CompressedTensorsConfig"  # type: ignore # noqa E501
):
    self.quant_config = quant_config
    self.weight_quant = self.quant_config.target_scheme_map["Linear"].get(
        "weights")
    self.input_quant = self.quant_config.target_scheme_map["Linear"].get(
        "input_activations")

    per_tensor = (self.weight_quant.strategy == QuantizationStrategy.TENSOR
                  and self.input_quant.strategy
                  == QuantizationStrategy.TENSOR)
    per_channel = (
        self.weight_quant.strategy == QuantizationStrategy.CHANNEL
        and self.input_quant.strategy == QuantizationStrategy.TOKEN)
    if not (per_tensor or per_channel):
        raise ValueError(
            "For FP8 Fused MoE layers, we require per tensor "
            "or channelwise, dynamic per token quantization. Found "
            f"{self.weight_quant}, {self.input_quant}")

    self.static_input_scales = not self.input_quant.dynamic
    if self.static_input_scales and per_channel:
        raise ValueError(
            "For FP8 Fused MoE layer, we require either per tensor or "
            "channelwise, dynamic per token quantization.")

    # For GPUs that lack FP8 hardware support, we can leverage the Marlin
    # kernel for fast weight-only FP8 quantization
    self.use_marlin = (not current_platform.has_device_capability(89)
                       or envs.VLLM_TEST_FORCE_FP8_MARLIN)
    # Disable marlin for rocm
    if current_platform.is_rocm():
        self.use_marlin = False
    from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (
        is_rocm_aiter_moe_enabled)

    self.rocm_aiter_moe_enabled = is_rocm_aiter_moe_enabled()

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/compressed_tensors/compressed_tensors_moe.py 
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
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 "
            "`CompressedTensorsW8A8Fp8MoEMethod` yet.")

    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,
        indices_type=self.topk_indices_dtype,
    )

    if self.rocm_aiter_moe_enabled:
        return self.rocm_aiter_fused_experts_func(
            hidden_states=x,
            w1=layer.w13_weight,
            w2=layer.w2_weight,
            topk_weights=topk_weights,
            topk_ids=topk_ids,
            activation=activation,
            apply_router_weight_on_input=apply_router_weight_on_input,
            use_fp8_w8a8=True,
            per_channel_quant=self.weight_quant.strategy ==
            QuantizationStrategy.CHANNEL,
            w1_scale=layer.w13_weight_scale,
            w2_scale=layer.w2_weight_scale,
            a1_scale=layer.w13_input_scale,
            a2_scale=layer.w2_input_scale,
            expert_map=expert_map)
    if self.use_marlin:
        assert activation == "silu", (
            f"{activation} not supported for Marlin MoE.")
        return torch.ops.vllm.fused_marlin_moe(
            x,
            layer.w13_weight,
            layer.w2_weight,
            layer.w13_weight_scale,
            layer.w2_weight_scale,
            router_logits,
            topk_weights,
            topk_ids,
            quant_type_id=scalar_types.float8_e4m3fn.id,
            apply_router_weight_on_input=apply_router_weight_on_input,
            global_num_experts=global_num_experts,
            expert_map=expert_map)

    assert self.fused_experts_func is not None

    return self.fused_experts_func(
        hidden_states=x,
        w1=layer.w13_weight,
        w2=layer.w2_weight,
        topk_weights=topk_weights,
        topk_ids=topk_ids,
        inplace=True,
        activation=activation,
        apply_router_weight_on_input=apply_router_weight_on_input,
        use_fp8_w8a8=True,
        per_channel_quant=self.weight_quant.strategy ==
        QuantizationStrategy.CHANNEL,
        global_num_experts=global_num_experts,
        expert_map=expert_map,
        w1_scale=layer.w13_weight_scale,
        w2_scale=layer.w2_weight_scale,
        a1_scale=layer.w13_input_scale,
        a2_scale=layer.w2_input_scale)

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/compressed_tensors/compressed_tensors_moe.py 
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
def create_weights(self, layer: torch.nn.Module, num_experts: int,
                   hidden_size: int, intermediate_size_per_partition: int,
                   params_dtype: torch.dtype, **extra_weight_attrs):

    layer.intermediate_size_per_partition = intermediate_size_per_partition
    layer.hidden_size = hidden_size
    layer.num_experts = num_experts
    layer.orig_dtype = params_dtype
    layer.weight_block_size = None

    params_dtype = torch.float8_e4m3fn

    # WEIGHTS
    w13_weight = torch.nn.Parameter(torch.empty(
        num_experts,
        2 * intermediate_size_per_partition,
        hidden_size,
        dtype=params_dtype),
                                    requires_grad=False)
    layer.register_parameter("w13_weight", w13_weight)
    set_weight_attrs(w13_weight, extra_weight_attrs)

    w2_weight = torch.nn.Parameter(torch.empty(
        num_experts,
        hidden_size,
        intermediate_size_per_partition,
        dtype=params_dtype),
                                   requires_grad=False)
    layer.register_parameter("w2_weight", w2_weight)
    set_weight_attrs(w2_weight, extra_weight_attrs)

    # WEIGHT_SCALES
    if self.weight_quant.strategy == QuantizationStrategy.TENSOR:
        # Allocate 2 scales for w1 and w3 respectively.
        # They are combined to a single scale after weight loading.
        w13_weight_scale = torch.nn.Parameter(torch.ones(
            num_experts, 2, dtype=torch.float32),
                                              requires_grad=False)
        layer.register_parameter("w13_weight_scale", w13_weight_scale)
        w2_weight_scale = torch.nn.Parameter(torch.ones(
            num_experts, dtype=torch.float32),
                                             requires_grad=False)
        layer.register_parameter("w2_weight_scale", w2_weight_scale)
        # Add PER-TENSOR quantization for FusedMoE.weight_loader.
        extra_weight_attrs.update(
            {"quant_method": FusedMoeWeightScaleSupported.TENSOR.value})
        set_weight_attrs(w13_weight_scale, extra_weight_attrs)
        set_weight_attrs(w2_weight_scale, extra_weight_attrs)

    elif self.weight_quant.strategy == QuantizationStrategy.CHANNEL:
        w13_weight_scale = torch.nn.Parameter(torch.ones(
            num_experts,
            2 * intermediate_size_per_partition,
            1,
            dtype=torch.float32),
                                              requires_grad=False)
        layer.register_parameter("w13_weight_scale", w13_weight_scale)
        w2_weight_scale = torch.nn.Parameter(torch.ones(
            num_experts, hidden_size, 1, dtype=torch.float32),
                                             requires_grad=False)
        layer.register_parameter("w2_weight_scale", w2_weight_scale)
        # Add PER-CHANNEL quantization for FusedMoE.weight_loader.
        extra_weight_attrs.update(
            {"quant_method": FusedMoeWeightScaleSupported.CHANNEL.value})
        set_weight_attrs(w13_weight_scale, extra_weight_attrs)
        set_weight_attrs(w2_weight_scale, extra_weight_attrs)

    # INPUT_SCALES
    if self.static_input_scales:
        w13_input_scale = torch.nn.Parameter(torch.ones(
            num_experts, dtype=torch.float32),
                                             requires_grad=False)
        layer.register_parameter("w13_input_scale", w13_input_scale)
        set_weight_attrs(w13_input_scale, extra_weight_attrs)

        w2_input_scale = torch.nn.Parameter(torch.ones(
            num_experts, dtype=torch.float32),
                                            requires_grad=False)
        layer.register_parameter("w2_input_scale", w2_input_scale)
        set_weight_attrs(w2_input_scale, extra_weight_attrs)
    else:
        layer.w13_input_scale = None
        layer.w2_input_scale = None

process_weights_after_loading 

process_weights_after_loading(layer: Module) -> None
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
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
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
    # Fp8 moe kernels require a single activation scale.
    # We take the max of all the scales in case they differ.
    if self.static_input_scales:
        assert self.input_quant.strategy == QuantizationStrategy.TENSOR
        if (layer.w13_input_scale is None or layer.w2_input_scale is None):
            raise ValueError(
                "QuantConfig has static quantization, but found "
                "activation scales are None.")
        if (not all_close_1d(layer.w13_input_scale)
                or not all_close_1d(layer.w2_input_scale)):
            logger.warning_once(
                "Found input_scales that are not equal for "
                "fp8 MoE layer. Using the maximum across experts "
                "for each layer.")
        layer.w13_input_scale = torch.nn.Parameter(
            layer.w13_input_scale.max(), requires_grad=False)
        layer.w2_input_scale = torch.nn.Parameter(
            layer.w2_input_scale.max(), requires_grad=False)

    if current_platform.is_fp8_fnuz():
        # Normalize the weights and scales
        w13_weight, w13_weight_scale, w13_input_scale = \
            normalize_e4m3fn_to_e4m3fnuz(
                layer.w13_weight, layer.w13_weight_scale,
                layer.w13_input_scale)
        w2_weight, w2_weight_scale, w2_input_scale = \
            normalize_e4m3fn_to_e4m3fnuz(
                layer.w2_weight, layer.w2_weight_scale,
                layer.w2_input_scale)
        # Reset the parameter
        layer.w13_weight = torch.nn.Parameter(w13_weight,
                                              requires_grad=False)
        layer.w13_weight_scale = torch.nn.Parameter(w13_weight_scale,
                                                    requires_grad=False)
        if w13_input_scale is not None:
            layer.w13_input_scale = torch.nn.Parameter(w13_input_scale,
                                                       requires_grad=False)
        layer.w2_weight = torch.nn.Parameter(w2_weight,
                                             requires_grad=False)
        layer.w2_weight_scale = torch.nn.Parameter(w2_weight_scale,
                                                   requires_grad=False)
        if w2_input_scale is not None:
            layer.w2_input_scale = torch.nn.Parameter(w2_input_scale,
                                                      requires_grad=False)

    # For Per-TENSOR case, Fp8 moe kernel needs single weight scale
    # for w13 per expert. Use max then dequant and requant each expert.
    if self.weight_quant.strategy == QuantizationStrategy.TENSOR:
        assert layer.w13_weight_scale is not None
        shard_size = layer.intermediate_size_per_partition
        max_w13_scales = layer.w13_weight_scale.max(dim=1).values
        for expert_id in range(layer.local_num_experts):
            start = 0
            for shard_id in range(2):
                dq_weight = per_tensor_dequantize(
                    layer.w13_weight[expert_id][start:start +
                                                shard_size, :],
                    layer.w13_weight_scale[expert_id][shard_id])
                layer.w13_weight[expert_id][
                    start:start + shard_size, :], _ = ops.scaled_fp8_quant(
                        dq_weight, max_w13_scales[expert_id])
                start += shard_size
        layer.w13_weight_scale = torch.nn.Parameter(max_w13_scales,
                                                    requires_grad=False)

    # Property to determine if AITER is used
    if self.rocm_aiter_moe_enabled:
        from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (  # noqa E501
            rocm_aiter_fused_experts, shuffle_weights)

        # reshaping weights is required for aiter moe kernel.
        shuffled_w13, shuffled_w2 = shuffle_weights(
            layer.w13_weight.data, layer.w2_weight.data)

        layer.w13_weight = torch.nn.Parameter(shuffled_w13,
                                              requires_grad=False)
        layer.w2_weight = torch.nn.Parameter(shuffled_w2,
                                             requires_grad=False)

        self.rocm_aiter_fused_experts_func = rocm_aiter_fused_experts
    elif self.use_marlin:
        prepare_moe_fp8_layer_for_marlin(layer, False)
        # Activations not quantized for marlin.
        del layer.w13_input_scale
        del layer.w2_input_scale
        self.fused_experts_func = None
    else:
        from vllm.model_executor.layers.fused_moe import fused_experts
        self.fused_experts_func = fused_experts

select_gemm_impl 

select_gemm_impl(
    prepare_finalize: FusedMoEPrepareAndFinalize,
    moe: FusedMoEConfig,
) -> FusedMoEPermuteExpertsUnpermute
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
def select_gemm_impl(
    self,
    prepare_finalize: FusedMoEPrepareAndFinalize,
    moe: FusedMoEConfig,
) -> FusedMoEPermuteExpertsUnpermute:
    from vllm.model_executor.layers.fused_moe import TritonExperts
    from vllm.model_executor.layers.fused_moe.fused_batched_moe import (
        BatchedTritonExperts)

    assert not self.rocm_aiter_moe_enabled and not self.use_marlin

    logger.debug("BatchedTritonExperts(%s)", self.__class__.__name__)

    if (prepare_finalize.activation_format ==
            FusedMoEActivationFormat.BatchedExperts):
        max_num_tokens_per_rank = prepare_finalize.max_num_tokens_per_rank(
        )
        assert max_num_tokens_per_rank is not None

        return BatchedTritonExperts(
            max_num_tokens=max_num_tokens_per_rank,
            num_dispatchers=prepare_finalize.num_dispatchers(),
            use_fp8_w8a8=True,
            block_shape=self.quant_config.weight_block_size,
            per_act_token_quant=(
                self.input_quant.strategy == QuantizationStrategy.TOKEN),
        )
    else:
        return TritonExperts(
            use_fp8_w8a8=True,
            block_shape=self.quant_config.weight_block_size,
            per_act_token_quant=(
                self.input_quant.strategy == QuantizationStrategy.TOKEN),
        )

CompressedTensorsW8A8Int8MoEMethod

Bases: CompressedTensorsMoEMethod

Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
 950
 951
 952
 953
 954
 955
 956
 957
 958
 959
 960
 961
 962
 963
 964
 965
 966
 967
 968
 969
 970
 971
 972
 973
 974
 975
 976
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
1027
1028
1029
1030
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
class CompressedTensorsW8A8Int8MoEMethod(CompressedTensorsMoEMethod):

    def __init__(
            self,
            quant_config: "CompressedTensorsConfig"  # type: ignore # noqa E501
    ):
        self.quant_config = quant_config
        self.weight_quant = self.quant_config.target_scheme_map["Linear"].get(
            "weights")
        self.input_quant = self.quant_config.target_scheme_map["Linear"].get(
            "input_activations")

        per_channel = (
            self.weight_quant.strategy == QuantizationStrategy.CHANNEL
            and self.input_quant.strategy == QuantizationStrategy.TOKEN)
        if not per_channel:
            raise ValueError(
                "For INT8 Fused MoE layers, we require channelwise, "
                "dynamic per token quantization. Found "
                f"{self.weight_quant}, {self.input_quant}")

        self.static_input_scales = not self.input_quant.dynamic
        if self.static_input_scales:
            raise ValueError(
                "For INT8 Fused MoE layers, we require channelwise, "
                "dynamic per token quantization. Found static input scales.")

    def create_weights(self, layer: torch.nn.Module, num_experts: int,
                       hidden_size: int, intermediate_size_per_partition: int,
                       params_dtype: torch.dtype, **extra_weight_attrs):

        params_dtype = torch.int8

        # WEIGHTS
        w13_weight = torch.nn.Parameter(torch.empty(
            num_experts,
            2 * intermediate_size_per_partition,
            hidden_size,
            dtype=params_dtype),
                                        requires_grad=False)
        layer.register_parameter("w13_weight", w13_weight)
        set_weight_attrs(w13_weight, extra_weight_attrs)

        w2_weight = torch.nn.Parameter(torch.empty(
            num_experts,
            hidden_size,
            intermediate_size_per_partition,
            dtype=params_dtype),
                                       requires_grad=False)
        layer.register_parameter("w2_weight", w2_weight)
        set_weight_attrs(w2_weight, extra_weight_attrs)

        # WEIGHT_SCALES
        assert self.weight_quant.strategy == QuantizationStrategy.CHANNEL
        w13_weight_scale = torch.nn.Parameter(torch.ones(
            num_experts,
            2 * intermediate_size_per_partition,
            1,
            dtype=torch.float32),
                                              requires_grad=False)
        layer.register_parameter("w13_weight_scale", w13_weight_scale)
        w2_weight_scale = torch.nn.Parameter(torch.ones(num_experts,
                                                        hidden_size,
                                                        1,
                                                        dtype=torch.float32),
                                             requires_grad=False)
        layer.register_parameter("w2_weight_scale", w2_weight_scale)
        # Add PER-CHANNEL quantization for FusedMoE.weight_loader.
        extra_weight_attrs.update(
            {"quant_method": FusedMoeWeightScaleSupported.CHANNEL.value})
        set_weight_attrs(w13_weight_scale, extra_weight_attrs)
        set_weight_attrs(w2_weight_scale, extra_weight_attrs)

        # INPUT_SCALES
        assert not self.static_input_scales
        layer.w13_input_scale = None
        layer.w2_input_scale = None

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        pass

    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 "
                "`CompressedTensorsW8A8Int8MoEMethod` yet.")

        from vllm.model_executor.layers.fused_moe import fused_experts

        topk_weights, topk_ids = FusedMoE.select_experts(
            hidden_states=x,
            router_logits=router_logits,
            use_grouped_topk=use_grouped_topk,
            top_k=top_k,
            renormalize=renormalize,
            topk_group=topk_group,
            num_expert_group=num_expert_group,
            custom_routing_function=custom_routing_function,
            scoring_func=scoring_func,
            e_score_correction_bias=e_score_correction_bias)

        return fused_experts(
            hidden_states=x,
            w1=layer.w13_weight,
            w2=layer.w2_weight,
            topk_weights=topk_weights,
            topk_ids=topk_ids,
            inplace=True,
            activation=activation,
            apply_router_weight_on_input=apply_router_weight_on_input,
            use_int8_w8a8=True,
            per_channel_quant=True,
            global_num_experts=global_num_experts,
            expert_map=expert_map,
            w1_scale=layer.w13_weight_scale,
            w2_scale=layer.w2_weight_scale,
            a1_scale=layer.w13_input_scale,
            a2_scale=layer.w2_input_scale)

input_quant instance-attribute 

input_quant = get('input_activations')

quant_config instance-attribute 

quant_config = quant_config

static_input_scales instance-attribute 

static_input_scales = not dynamic

weight_quant instance-attribute 

weight_quant = get('weights')

__init__ 

__init__(quant_config: CompressedTensorsConfig)
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
def __init__(
        self,
        quant_config: "CompressedTensorsConfig"  # type: ignore # noqa E501
):
    self.quant_config = quant_config
    self.weight_quant = self.quant_config.target_scheme_map["Linear"].get(
        "weights")
    self.input_quant = self.quant_config.target_scheme_map["Linear"].get(
        "input_activations")

    per_channel = (
        self.weight_quant.strategy == QuantizationStrategy.CHANNEL
        and self.input_quant.strategy == QuantizationStrategy.TOKEN)
    if not per_channel:
        raise ValueError(
            "For INT8 Fused MoE layers, we require channelwise, "
            "dynamic per token quantization. Found "
            f"{self.weight_quant}, {self.input_quant}")

    self.static_input_scales = not self.input_quant.dynamic
    if self.static_input_scales:
        raise ValueError(
            "For INT8 Fused MoE layers, we require channelwise, "
            "dynamic per token quantization. Found static input scales.")

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/compressed_tensors/compressed_tensors_moe.py 
1031
1032
1033
1034
1035
1036
1037
1038
1039
1040
1041
1042
1043
1044
1045
1046
1047
1048
1049
1050
1051
1052
1053
1054
1055
1056
1057
1058
1059
1060
1061
1062
1063
1064
1065
1066
1067
1068
1069
1070
1071
1072
1073
1074
1075
1076
1077
1078
1079
1080
1081
1082
1083
1084
1085
1086
1087
1088
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 "
            "`CompressedTensorsW8A8Int8MoEMethod` yet.")

    from vllm.model_executor.layers.fused_moe import fused_experts

    topk_weights, topk_ids = FusedMoE.select_experts(
        hidden_states=x,
        router_logits=router_logits,
        use_grouped_topk=use_grouped_topk,
        top_k=top_k,
        renormalize=renormalize,
        topk_group=topk_group,
        num_expert_group=num_expert_group,
        custom_routing_function=custom_routing_function,
        scoring_func=scoring_func,
        e_score_correction_bias=e_score_correction_bias)

    return fused_experts(
        hidden_states=x,
        w1=layer.w13_weight,
        w2=layer.w2_weight,
        topk_weights=topk_weights,
        topk_ids=topk_ids,
        inplace=True,
        activation=activation,
        apply_router_weight_on_input=apply_router_weight_on_input,
        use_int8_w8a8=True,
        per_channel_quant=True,
        global_num_experts=global_num_experts,
        expert_map=expert_map,
        w1_scale=layer.w13_weight_scale,
        w2_scale=layer.w2_weight_scale,
        a1_scale=layer.w13_input_scale,
        a2_scale=layer.w2_input_scale)

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/compressed_tensors/compressed_tensors_moe.py 
 977
 978
 979
 980
 981
 982
 983
 984
 985
 986
 987
 988
 989
 990
 991
 992
 993
 994
 995
 996
 997
 998
 999
1000
1001
1002
1003
1004
1005
1006
1007
1008
1009
1010
1011
1012
1013
1014
1015
1016
1017
1018
1019
1020
1021
1022
1023
1024
1025
1026
def create_weights(self, layer: torch.nn.Module, num_experts: int,
                   hidden_size: int, intermediate_size_per_partition: int,
                   params_dtype: torch.dtype, **extra_weight_attrs):

    params_dtype = torch.int8

    # WEIGHTS
    w13_weight = torch.nn.Parameter(torch.empty(
        num_experts,
        2 * intermediate_size_per_partition,
        hidden_size,
        dtype=params_dtype),
                                    requires_grad=False)
    layer.register_parameter("w13_weight", w13_weight)
    set_weight_attrs(w13_weight, extra_weight_attrs)

    w2_weight = torch.nn.Parameter(torch.empty(
        num_experts,
        hidden_size,
        intermediate_size_per_partition,
        dtype=params_dtype),
                                   requires_grad=False)
    layer.register_parameter("w2_weight", w2_weight)
    set_weight_attrs(w2_weight, extra_weight_attrs)

    # WEIGHT_SCALES
    assert self.weight_quant.strategy == QuantizationStrategy.CHANNEL
    w13_weight_scale = torch.nn.Parameter(torch.ones(
        num_experts,
        2 * intermediate_size_per_partition,
        1,
        dtype=torch.float32),
                                          requires_grad=False)
    layer.register_parameter("w13_weight_scale", w13_weight_scale)
    w2_weight_scale = torch.nn.Parameter(torch.ones(num_experts,
                                                    hidden_size,
                                                    1,
                                                    dtype=torch.float32),
                                         requires_grad=False)
    layer.register_parameter("w2_weight_scale", w2_weight_scale)
    # Add PER-CHANNEL quantization for FusedMoE.weight_loader.
    extra_weight_attrs.update(
        {"quant_method": FusedMoeWeightScaleSupported.CHANNEL.value})
    set_weight_attrs(w13_weight_scale, extra_weight_attrs)
    set_weight_attrs(w2_weight_scale, extra_weight_attrs)

    # INPUT_SCALES
    assert not self.static_input_scales
    layer.w13_input_scale = None
    layer.w2_input_scale = None

process_weights_after_loading 

process_weights_after_loading(layer: Module) -> None
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
1028
1029
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
    pass

CompressedTensorsWNA16MarlinMoEMethod

Bases: CompressedTensorsMoEMethod

Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
1091
1092
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
1117
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
1243
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
1329
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
class CompressedTensorsWNA16MarlinMoEMethod(CompressedTensorsMoEMethod):

    def __init__(
            self,
            quant_config: "CompressedTensorsConfig"  # type: ignore # noqa E501
    ):
        self.quant_config = quant_config
        # TODO: @dsikka: refactor this to use schemes as other kernels
        # are supported + check if the layer is being ignored.
        config = self.quant_config.target_scheme_map["Linear"].get("weights")
        self.num_bits = config.num_bits
        self.packed_factor = 32 // config.num_bits
        self.strategy = config.strategy
        self.group_size = config.group_size
        self.actorder = config.actorder
        assert config.symmetric, (
            "Only symmetric quantization is supported for MoE")

        if not (self.quant_config.quant_format
                == CompressionFormat.pack_quantized.value
                and self.num_bits in WNA16_SUPPORTED_BITS):
            raise ValueError("For Fused MoE layers, only ",
                             f"{CompressionFormat.pack_quantized.value} ",
                             "is supported for the following bits: ",
                             f"{WNA16_SUPPORTED_BITS}")
        self.quant_type = WNA16_SUPPORTED_TYPES_MAP[self.num_bits]

    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):

        intermediate_size_full = extra_weight_attrs.pop(
            "intermediate_size_full")

        # Will transpose the loaded weight along the
        # intermediate and hidden dim sizes. Will
        # shard for TP along the transposed dims
        extra_weight_attrs.update({
            "is_transposed": True,
            "quant_method": self.strategy
        })
        w13_weight = torch.nn.Parameter(torch.empty(
            num_experts,
            hidden_size // self.packed_factor,
            2 * intermediate_size_per_partition,
            dtype=torch.int32),
                                        requires_grad=False)
        layer.register_parameter("w13_weight_packed", w13_weight)
        set_weight_attrs(w13_weight, extra_weight_attrs)

        w2_weight = torch.nn.Parameter(torch.empty(
            num_experts,
            intermediate_size_per_partition // self.packed_factor,
            hidden_size,
            dtype=torch.int32),
                                       requires_grad=False)
        layer.register_parameter("w2_weight_packed", w2_weight)
        set_weight_attrs(w2_weight, extra_weight_attrs)

        # In the case where we have actorder/g_idx,
        # we do not partition the w2 scales
        load_full_w2 = self.actorder and self.group_size != -1
        w2_scales_size = (intermediate_size_full
                          if load_full_w2 else intermediate_size_per_partition)

        self.is_k_full = (not self.actorder) or (
            intermediate_size_per_partition == intermediate_size_full)

        if self.strategy == "channel":
            num_groups_w2 = num_groups_w13 = 1
            self.group_size = -1
        else:
            num_groups_w2 = w2_scales_size // self.group_size
            num_groups_w13 = hidden_size // self.group_size

        w13_scale = torch.nn.Parameter(torch.ones(
            num_experts,
            num_groups_w13,
            2 * intermediate_size_per_partition,
            dtype=params_dtype),
                                       requires_grad=False)
        layer.register_parameter("w13_weight_scale", w13_scale)
        set_weight_attrs(w13_scale, extra_weight_attrs)

        w2_scale = torch.nn.Parameter(torch.ones(num_experts,
                                                 num_groups_w2,
                                                 hidden_size,
                                                 dtype=params_dtype),
                                      requires_grad=False)
        layer.register_parameter("w2_weight_scale", w2_scale)
        set_weight_attrs(w2_scale, extra_weight_attrs)
        set_weight_attrs(w2_scale, {"load_full_w2": load_full_w2})

        w2_weight_shape = torch.nn.Parameter(torch.empty(num_experts, 2),
                                             requires_grad=False)
        layer.register_parameter("w2_weight_shape", w2_weight_shape)
        set_weight_attrs(w2_weight_shape, extra_weight_attrs)
        w13_weight_shape = torch.nn.Parameter(torch.empty(num_experts, 2),
                                              requires_grad=False)

        layer.register_parameter("w13_weight_shape", w13_weight_shape)
        set_weight_attrs(w13_weight_shape, extra_weight_attrs)

        w13_g_idx = torch.nn.Parameter(
            torch.empty(
                num_experts,
                hidden_size,
                dtype=torch.int32,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w13_weight_g_idx", w13_g_idx)
        set_weight_attrs(w13_g_idx, extra_weight_attrs)

        w2_g_idx = torch.nn.Parameter(
            torch.empty(
                num_experts,
                intermediate_size_per_partition,
                dtype=torch.int32,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w2_weight_g_idx", w2_g_idx)
        set_weight_attrs(w2_g_idx, extra_weight_attrs)

        w13_g_idx_sort_indices = torch.nn.Parameter(
            torch.empty(
                num_experts,
                hidden_size,
                dtype=torch.int32,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w13_g_idx_sort_indices",
                                 w13_g_idx_sort_indices)
        set_weight_attrs(w13_g_idx_sort_indices, extra_weight_attrs)

        w2_g_idx_sort_indices = torch.nn.Parameter(
            torch.empty(
                num_experts,
                intermediate_size_per_partition,
                dtype=torch.int32,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w2_g_idx_sort_indices",
                                 w2_g_idx_sort_indices)
        set_weight_attrs(w2_g_idx_sort_indices, extra_weight_attrs)

        layer.a13_scale = None
        layer.a2_scale = None
        layer.marlin_state = GPTQMarlinState.REPACK

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        num_experts = layer.w13_weight_g_idx.shape[0]
        device = layer.w13_weight_g_idx.device

        # when running models with grouped act order,
        # resort to g_idx values provided in checkpoint
        if self.actorder == "group":
            w13_g_idx_sort_indices = torch.empty_like(layer.w13_weight_g_idx)
            w2_g_idx_sort_indices = torch.empty_like(layer.w2_weight_g_idx)
            w13_sorted_g_idx = torch.empty_like(layer.w13_weight_g_idx)
            w2_sorted_g_idx = torch.empty_like(layer.w2_weight_g_idx)

            for e in range(num_experts):
                w13_g_idx_sort_indices[e] = torch.argsort(
                    layer.w13_weight_g_idx[e]).to(torch.int32)
                w2_g_idx_sort_indices[e] = torch.argsort(
                    layer.w2_weight_g_idx[e]).to(torch.int32)
                w13_sorted_g_idx[e] = layer.w13_weight_g_idx[e][
                    w13_g_idx_sort_indices[e]]
                w2_sorted_g_idx[e] = layer.w2_weight_g_idx[e][
                    w2_g_idx_sort_indices[e]]

            replace_parameter(layer, "w13_weight_g_idx", w13_sorted_g_idx)
            replace_parameter(layer, "w2_weight_g_idx", w2_sorted_g_idx)
            replace_parameter(layer, "w13_g_idx_sort_indices",
                              w13_g_idx_sort_indices)
            replace_parameter(layer, "w2_g_idx_sort_indices",
                              w2_g_idx_sort_indices)

        else:
            layer.w13_weight_g_idx = torch.nn.Parameter(
                torch.empty((num_experts, 0), dtype=torch.int32,
                            device=device),
                requires_grad=False,
            )
            layer.w2_weight_g_idx = torch.nn.Parameter(
                torch.empty((num_experts, 0), dtype=torch.int32,
                            device=device),
                requires_grad=False,
            )
            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.gptq_marlin_moe_repack(
            layer.w13_weight_packed,
            layer.w13_g_idx_sort_indices,
            layer.w13_weight_packed.shape[1] * self.packed_factor,
            layer.w13_weight_packed.shape[2],
            self.num_bits,
        )
        replace_parameter(layer, "w13_weight_packed", marlin_w13_qweight)
        marlin_w2_qweight = ops.gptq_marlin_moe_repack(
            layer.w2_weight_packed,
            layer.w2_g_idx_sort_indices,
            layer.w2_weight_packed.shape[1] * self.packed_factor,
            layer.w2_weight_packed.shape[2],
            self.num_bits,
        )
        replace_parameter(layer, "w2_weight_packed", marlin_w2_qweight)
        # Repack scales
        marlin_w13_scales = marlin_moe_permute_scales(
            s=layer.w13_weight_scale,
            size_k=layer.w13_weight_packed.shape[2],
            size_n=layer.w13_weight_scale.shape[2],
            group_size=self.group_size,
        )
        replace_parameter(layer, "w13_weight_scale", marlin_w13_scales)
        marlin_w2_scales = marlin_moe_permute_scales(
            s=layer.w2_weight_scale,
            size_k=layer.w2_weight_scale.shape[1] *
            (self.group_size if self.group_size != -1 else self.packed_factor),
            size_n=layer.w2_weight_scale.shape[2],
            group_size=self.group_size,
        )
        replace_parameter(layer, "w2_weight_scale", marlin_w2_scales)

        layer.workspace = marlin_make_workspace_new(device, 4)

    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 "
                "`CompressedTensorsWNA16MarlinMoEMethod` yet.")

        assert activation == "silu", (
            f"{activation} not supported for Marlin MoE.")

        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_weight_packed,
            layer.w2_weight_packed,
            layer.w13_weight_scale,
            layer.w2_weight_scale,
            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,
            g_idx1=layer.w13_weight_g_idx,
            g_idx2=layer.w2_weight_g_idx,
            sort_indices1=layer.w13_g_idx_sort_indices,
            sort_indices2=layer.w2_g_idx_sort_indices,
            workspace=layer.workspace,
            is_k_full=self.is_k_full)

actorder instance-attribute 

actorder = actorder

group_size instance-attribute 

group_size = group_size

num_bits instance-attribute 

num_bits = num_bits

packed_factor instance-attribute 

packed_factor = 32 // num_bits

quant_config instance-attribute 

quant_config = quant_config

quant_type instance-attribute 

quant_type = WNA16_SUPPORTED_TYPES_MAP[num_bits]

strategy instance-attribute 

strategy = strategy

__init__ 

__init__(quant_config: CompressedTensorsConfig)
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
1093
1094
1095
1096
1097
1098
1099
1100
1101
1102
1103
1104
1105
1106
1107
1108
1109
1110
1111
1112
1113
1114
1115
1116
def __init__(
        self,
        quant_config: "CompressedTensorsConfig"  # type: ignore # noqa E501
):
    self.quant_config = quant_config
    # TODO: @dsikka: refactor this to use schemes as other kernels
    # are supported + check if the layer is being ignored.
    config = self.quant_config.target_scheme_map["Linear"].get("weights")
    self.num_bits = config.num_bits
    self.packed_factor = 32 // config.num_bits
    self.strategy = config.strategy
    self.group_size = config.group_size
    self.actorder = config.actorder
    assert config.symmetric, (
        "Only symmetric quantization is supported for MoE")

    if not (self.quant_config.quant_format
            == CompressionFormat.pack_quantized.value
            and self.num_bits in WNA16_SUPPORTED_BITS):
        raise ValueError("For Fused MoE layers, only ",
                         f"{CompressionFormat.pack_quantized.value} ",
                         "is supported for the following bits: ",
                         f"{WNA16_SUPPORTED_BITS}")
    self.quant_type = WNA16_SUPPORTED_TYPES_MAP[self.num_bits]

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/compressed_tensors/compressed_tensors_moe.py 
1330
1331
1332
1333
1334
1335
1336
1337
1338
1339
1340
1341
1342
1343
1344
1345
1346
1347
1348
1349
1350
1351
1352
1353
1354
1355
1356
1357
1358
1359
1360
1361
1362
1363
1364
1365
1366
1367
1368
1369
1370
1371
1372
1373
1374
1375
1376
1377
1378
1379
1380
1381
1382
1383
1384
1385
1386
1387
1388
1389
1390
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 "
            "`CompressedTensorsWNA16MarlinMoEMethod` yet.")

    assert activation == "silu", (
        f"{activation} not supported for Marlin MoE.")

    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_weight_packed,
        layer.w2_weight_packed,
        layer.w13_weight_scale,
        layer.w2_weight_scale,
        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,
        g_idx1=layer.w13_weight_g_idx,
        g_idx2=layer.w2_weight_g_idx,
        sort_indices1=layer.w13_g_idx_sort_indices,
        sort_indices2=layer.w2_g_idx_sort_indices,
        workspace=layer.workspace,
        is_k_full=self.is_k_full)

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/compressed_tensors/compressed_tensors_moe.py 
1118
1119
1120
1121
1122
1123
1124
1125
1126
1127
1128
1129
1130
1131
1132
1133
1134
1135
1136
1137
1138
1139
1140
1141
1142
1143
1144
1145
1146
1147
1148
1149
1150
1151
1152
1153
1154
1155
1156
1157
1158
1159
1160
1161
1162
1163
1164
1165
1166
1167
1168
1169
1170
1171
1172
1173
1174
1175
1176
1177
1178
1179
1180
1181
1182
1183
1184
1185
1186
1187
1188
1189
1190
1191
1192
1193
1194
1195
1196
1197
1198
1199
1200
1201
1202
1203
1204
1205
1206
1207
1208
1209
1210
1211
1212
1213
1214
1215
1216
1217
1218
1219
1220
1221
1222
1223
1224
1225
1226
1227
1228
1229
1230
1231
1232
1233
1234
1235
1236
1237
1238
1239
1240
1241
1242
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):

    intermediate_size_full = extra_weight_attrs.pop(
        "intermediate_size_full")

    # Will transpose the loaded weight along the
    # intermediate and hidden dim sizes. Will
    # shard for TP along the transposed dims
    extra_weight_attrs.update({
        "is_transposed": True,
        "quant_method": self.strategy
    })
    w13_weight = torch.nn.Parameter(torch.empty(
        num_experts,
        hidden_size // self.packed_factor,
        2 * intermediate_size_per_partition,
        dtype=torch.int32),
                                    requires_grad=False)
    layer.register_parameter("w13_weight_packed", w13_weight)
    set_weight_attrs(w13_weight, extra_weight_attrs)

    w2_weight = torch.nn.Parameter(torch.empty(
        num_experts,
        intermediate_size_per_partition // self.packed_factor,
        hidden_size,
        dtype=torch.int32),
                                   requires_grad=False)
    layer.register_parameter("w2_weight_packed", w2_weight)
    set_weight_attrs(w2_weight, extra_weight_attrs)

    # In the case where we have actorder/g_idx,
    # we do not partition the w2 scales
    load_full_w2 = self.actorder and self.group_size != -1
    w2_scales_size = (intermediate_size_full
                      if load_full_w2 else intermediate_size_per_partition)

    self.is_k_full = (not self.actorder) or (
        intermediate_size_per_partition == intermediate_size_full)

    if self.strategy == "channel":
        num_groups_w2 = num_groups_w13 = 1
        self.group_size = -1
    else:
        num_groups_w2 = w2_scales_size // self.group_size
        num_groups_w13 = hidden_size // self.group_size

    w13_scale = torch.nn.Parameter(torch.ones(
        num_experts,
        num_groups_w13,
        2 * intermediate_size_per_partition,
        dtype=params_dtype),
                                   requires_grad=False)
    layer.register_parameter("w13_weight_scale", w13_scale)
    set_weight_attrs(w13_scale, extra_weight_attrs)

    w2_scale = torch.nn.Parameter(torch.ones(num_experts,
                                             num_groups_w2,
                                             hidden_size,
                                             dtype=params_dtype),
                                  requires_grad=False)
    layer.register_parameter("w2_weight_scale", w2_scale)
    set_weight_attrs(w2_scale, extra_weight_attrs)
    set_weight_attrs(w2_scale, {"load_full_w2": load_full_w2})

    w2_weight_shape = torch.nn.Parameter(torch.empty(num_experts, 2),
                                         requires_grad=False)
    layer.register_parameter("w2_weight_shape", w2_weight_shape)
    set_weight_attrs(w2_weight_shape, extra_weight_attrs)
    w13_weight_shape = torch.nn.Parameter(torch.empty(num_experts, 2),
                                          requires_grad=False)

    layer.register_parameter("w13_weight_shape", w13_weight_shape)
    set_weight_attrs(w13_weight_shape, extra_weight_attrs)

    w13_g_idx = torch.nn.Parameter(
        torch.empty(
            num_experts,
            hidden_size,
            dtype=torch.int32,
        ),
        requires_grad=False,
    )
    layer.register_parameter("w13_weight_g_idx", w13_g_idx)
    set_weight_attrs(w13_g_idx, extra_weight_attrs)

    w2_g_idx = torch.nn.Parameter(
        torch.empty(
            num_experts,
            intermediate_size_per_partition,
            dtype=torch.int32,
        ),
        requires_grad=False,
    )
    layer.register_parameter("w2_weight_g_idx", w2_g_idx)
    set_weight_attrs(w2_g_idx, extra_weight_attrs)

    w13_g_idx_sort_indices = torch.nn.Parameter(
        torch.empty(
            num_experts,
            hidden_size,
            dtype=torch.int32,
        ),
        requires_grad=False,
    )
    layer.register_parameter("w13_g_idx_sort_indices",
                             w13_g_idx_sort_indices)
    set_weight_attrs(w13_g_idx_sort_indices, extra_weight_attrs)

    w2_g_idx_sort_indices = torch.nn.Parameter(
        torch.empty(
            num_experts,
            intermediate_size_per_partition,
            dtype=torch.int32,
        ),
        requires_grad=False,
    )
    layer.register_parameter("w2_g_idx_sort_indices",
                             w2_g_idx_sort_indices)
    set_weight_attrs(w2_g_idx_sort_indices, extra_weight_attrs)

    layer.a13_scale = None
    layer.a2_scale = None
    layer.marlin_state = GPTQMarlinState.REPACK

process_weights_after_loading 

process_weights_after_loading(layer: Module) -> None
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
1244
1245
1246
1247
1248
1249
1250
1251
1252
1253
1254
1255
1256
1257
1258
1259
1260
1261
1262
1263
1264
1265
1266
1267
1268
1269
1270
1271
1272
1273
1274
1275
1276
1277
1278
1279
1280
1281
1282
1283
1284
1285
1286
1287
1288
1289
1290
1291
1292
1293
1294
1295
1296
1297
1298
1299
1300
1301
1302
1303
1304
1305
1306
1307
1308
1309
1310
1311
1312
1313
1314
1315
1316
1317
1318
1319
1320
1321
1322
1323
1324
1325
1326
1327
1328
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
    num_experts = layer.w13_weight_g_idx.shape[0]
    device = layer.w13_weight_g_idx.device

    # when running models with grouped act order,
    # resort to g_idx values provided in checkpoint
    if self.actorder == "group":
        w13_g_idx_sort_indices = torch.empty_like(layer.w13_weight_g_idx)
        w2_g_idx_sort_indices = torch.empty_like(layer.w2_weight_g_idx)
        w13_sorted_g_idx = torch.empty_like(layer.w13_weight_g_idx)
        w2_sorted_g_idx = torch.empty_like(layer.w2_weight_g_idx)

        for e in range(num_experts):
            w13_g_idx_sort_indices[e] = torch.argsort(
                layer.w13_weight_g_idx[e]).to(torch.int32)
            w2_g_idx_sort_indices[e] = torch.argsort(
                layer.w2_weight_g_idx[e]).to(torch.int32)
            w13_sorted_g_idx[e] = layer.w13_weight_g_idx[e][
                w13_g_idx_sort_indices[e]]
            w2_sorted_g_idx[e] = layer.w2_weight_g_idx[e][
                w2_g_idx_sort_indices[e]]

        replace_parameter(layer, "w13_weight_g_idx", w13_sorted_g_idx)
        replace_parameter(layer, "w2_weight_g_idx", w2_sorted_g_idx)
        replace_parameter(layer, "w13_g_idx_sort_indices",
                          w13_g_idx_sort_indices)
        replace_parameter(layer, "w2_g_idx_sort_indices",
                          w2_g_idx_sort_indices)

    else:
        layer.w13_weight_g_idx = torch.nn.Parameter(
            torch.empty((num_experts, 0), dtype=torch.int32,
                        device=device),
            requires_grad=False,
        )
        layer.w2_weight_g_idx = torch.nn.Parameter(
            torch.empty((num_experts, 0), dtype=torch.int32,
                        device=device),
            requires_grad=False,
        )
        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.gptq_marlin_moe_repack(
        layer.w13_weight_packed,
        layer.w13_g_idx_sort_indices,
        layer.w13_weight_packed.shape[1] * self.packed_factor,
        layer.w13_weight_packed.shape[2],
        self.num_bits,
    )
    replace_parameter(layer, "w13_weight_packed", marlin_w13_qweight)
    marlin_w2_qweight = ops.gptq_marlin_moe_repack(
        layer.w2_weight_packed,
        layer.w2_g_idx_sort_indices,
        layer.w2_weight_packed.shape[1] * self.packed_factor,
        layer.w2_weight_packed.shape[2],
        self.num_bits,
    )
    replace_parameter(layer, "w2_weight_packed", marlin_w2_qweight)
    # Repack scales
    marlin_w13_scales = marlin_moe_permute_scales(
        s=layer.w13_weight_scale,
        size_k=layer.w13_weight_packed.shape[2],
        size_n=layer.w13_weight_scale.shape[2],
        group_size=self.group_size,
    )
    replace_parameter(layer, "w13_weight_scale", marlin_w13_scales)
    marlin_w2_scales = marlin_moe_permute_scales(
        s=layer.w2_weight_scale,
        size_k=layer.w2_weight_scale.shape[1] *
        (self.group_size if self.group_size != -1 else self.packed_factor),
        size_n=layer.w2_weight_scale.shape[2],
        group_size=self.group_size,
    )
    replace_parameter(layer, "w2_weight_scale", marlin_w2_scales)

    layer.workspace = marlin_make_workspace_new(device, 4)

CompressedTensorsWNA16MoEMethod

Bases: CompressedTensorsMoEMethod

Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
1393
1394
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
1421
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
1536
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
1553
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):

    def __init__(
            self,
            quant_config: "CompressedTensorsConfig"  # type: ignore # noqa E501
    ):
        self.quant_config = quant_config
        # TODO: @dsikka: refactor this to use schemes as other kernels
        # are supported + check if the layer is being ignored.
        config = self.quant_config.target_scheme_map["Linear"].get("weights")
        self.num_bits = config.num_bits
        self.packed_factor = 32 // config.num_bits
        self.strategy = config.strategy
        # channelwise is not supported by this kernel
        assert config.strategy == "group"
        self.group_size = config.group_size
        # grouped actorder isn't supported by this kernel
        assert config.actorder != "group"
        assert config.symmetric, (
            "Only symmetric quantization is supported for MoE")

        if not (self.quant_config.quant_format
                == CompressionFormat.pack_quantized.value
                and self.num_bits in WNA16_SUPPORTED_BITS):
            raise ValueError("For Fused MoE layers, only ",
                             f"{CompressionFormat.pack_quantized.value} ",
                             "is supported for the following bits: ",
                             f"{WNA16_SUPPORTED_BITS}")

    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):

        # Will transpose the loaded weight along the
        # intermediate and hidden dim sizes. Will
        # shard for TP along the transposed dims
        extra_weight_attrs.update({
            "is_transposed": True,
            "quant_method": self.strategy
        })
        w13_weight = torch.nn.Parameter(torch.empty(
            num_experts,
            hidden_size // self.packed_factor,
            2 * intermediate_size_per_partition,
            dtype=torch.int32),
                                        requires_grad=False)
        layer.register_parameter("w13_weight_packed", w13_weight)
        set_weight_attrs(w13_weight, extra_weight_attrs)

        w2_weight = torch.nn.Parameter(torch.empty(
            num_experts,
            intermediate_size_per_partition // self.packed_factor,
            hidden_size,
            dtype=torch.int32),
                                       requires_grad=False)
        layer.register_parameter("w2_weight_packed", w2_weight)
        set_weight_attrs(w2_weight, extra_weight_attrs)

        w2_scales_size = intermediate_size_per_partition

        if self.strategy == "channel":
            num_groups_w2 = num_groups_w13 = 1
            self.group_size = -1
        else:
            num_groups_w2 = w2_scales_size // self.group_size
            num_groups_w13 = hidden_size // self.group_size

        w13_scale = torch.nn.Parameter(torch.ones(
            num_experts,
            num_groups_w13,
            2 * intermediate_size_per_partition,
            dtype=params_dtype),
                                       requires_grad=False)
        layer.register_parameter("w13_weight_scale", w13_scale)
        set_weight_attrs(w13_scale, extra_weight_attrs)

        w2_scale = torch.nn.Parameter(torch.ones(num_experts,
                                                 num_groups_w2,
                                                 hidden_size,
                                                 dtype=params_dtype),
                                      requires_grad=False)
        layer.register_parameter("w2_weight_scale", w2_scale)
        set_weight_attrs(w2_scale, extra_weight_attrs)
        set_weight_attrs(w2_scale, {"load_full_w2": False})

        w2_weight_shape = torch.nn.Parameter(torch.empty(num_experts, 2),
                                             requires_grad=False)
        layer.register_parameter("w2_weight_shape", w2_weight_shape)
        set_weight_attrs(w2_weight_shape, extra_weight_attrs)
        w13_weight_shape = torch.nn.Parameter(torch.empty(num_experts, 2),
                                              requires_grad=False)

        layer.register_parameter("w13_weight_shape", w13_weight_shape)
        set_weight_attrs(w13_weight_shape, extra_weight_attrs)

        w13_g_idx = torch.nn.Parameter(
            torch.empty(
                num_experts,
                hidden_size,
                dtype=torch.int32,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w13_weight_g_idx", w13_g_idx)
        set_weight_attrs(w13_g_idx, extra_weight_attrs)

        w2_g_idx = torch.nn.Parameter(
            torch.empty(
                num_experts,
                intermediate_size_per_partition,
                dtype=torch.int32,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w2_weight_g_idx", w2_g_idx)
        set_weight_attrs(w2_g_idx, extra_weight_attrs)

        w13_g_idx_sort_indices = torch.nn.Parameter(
            torch.empty(
                num_experts,
                hidden_size,
                dtype=torch.int32,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w13_g_idx_sort_indices",
                                 w13_g_idx_sort_indices)
        set_weight_attrs(w13_g_idx_sort_indices, extra_weight_attrs)

        w2_g_idx_sort_indices = torch.nn.Parameter(
            torch.empty(
                num_experts,
                intermediate_size_per_partition,
                dtype=torch.int32,
            ),
            requires_grad=False,
        )
        layer.register_parameter("w2_g_idx_sort_indices",
                                 w2_g_idx_sort_indices)
        set_weight_attrs(w2_g_idx_sort_indices, extra_weight_attrs)

        layer.a13_scale = None
        layer.a2_scale = None

    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        # Reconfigure packed weights and scales to match moe_wna16 format
        layer.w13_weight_packed = torch.nn.Parameter(
            layer.w13_weight_packed.transpose(1, 2).contiguous().view(
                torch.uint8),
            requires_grad=False)
        layer.w2_weight_packed = torch.nn.Parameter(
            layer.w2_weight_packed.transpose(1,
                                             2).contiguous().view(torch.uint8),
            requires_grad=False)
        layer.w13_weight_scale = torch.nn.Parameter(
            layer.w13_weight_scale.transpose(1, 2).contiguous(),
            requires_grad=False)
        layer.w2_weight_scale = torch.nn.Parameter(
            layer.w2_weight_scale.transpose(1, 2).contiguous(),
            requires_grad=False)

    def apply(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        router_logits: torch.Tensor,
        top_k: int,
        renormalize: bool,
        use_grouped_topk: bool = False,
        topk_group: Optional[int] = None,
        num_expert_group: Optional[int] = None,
        global_num_experts: int = -1,
        expert_map: Optional[torch.Tensor] = None,
        custom_routing_function: Optional[Callable] = None,
        scoring_func: str = "softmax",
        e_score_correction_bias: Optional[torch.Tensor] = None,
        apply_router_weight_on_input: bool = False,
        activation: str = "silu",
        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 "
                                      "`CompressedTensorsWNA16MoEMethod` yet.")

        from vllm.model_executor.layers.fused_moe import fused_experts

        topk_weights, topk_ids = FusedMoE.select_experts(
            hidden_states=x,
            router_logits=router_logits,
            use_grouped_topk=use_grouped_topk,
            top_k=top_k,
            renormalize=renormalize,
            topk_group=topk_group,
            num_expert_group=num_expert_group,
            custom_routing_function=custom_routing_function,
            scoring_func=scoring_func,
            e_score_correction_bias=e_score_correction_bias)

        return fused_experts(
            x,
            layer.w13_weight_packed,
            layer.w2_weight_packed,
            topk_weights=topk_weights,
            topk_ids=topk_ids,
            inplace=True,
            activation=activation,
            use_int4_w4a16=self.num_bits == 4,
            use_int8_w8a16=self.num_bits == 8,
            global_num_experts=global_num_experts,
            apply_router_weight_on_input=apply_router_weight_on_input,
            expert_map=expert_map,
            w1_scale=layer.w13_weight_scale,
            w2_scale=layer.w2_weight_scale,
            w1_zp=None,
            w2_zp=None,
            block_shape=[0, self.group_size])

group_size instance-attribute 

group_size = group_size

num_bits instance-attribute 

num_bits = num_bits

packed_factor instance-attribute 

packed_factor = 32 // num_bits

quant_config instance-attribute 

quant_config = quant_config

strategy instance-attribute 

strategy = strategy

__init__ 

__init__(quant_config: CompressedTensorsConfig)
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
1395
1396
1397
1398
1399
1400
1401
1402
1403
1404
1405
1406
1407
1408
1409
1410
1411
1412
1413
1414
1415
1416
1417
1418
1419
1420
def __init__(
        self,
        quant_config: "CompressedTensorsConfig"  # type: ignore # noqa E501
):
    self.quant_config = quant_config
    # TODO: @dsikka: refactor this to use schemes as other kernels
    # are supported + check if the layer is being ignored.
    config = self.quant_config.target_scheme_map["Linear"].get("weights")
    self.num_bits = config.num_bits
    self.packed_factor = 32 // config.num_bits
    self.strategy = config.strategy
    # channelwise is not supported by this kernel
    assert config.strategy == "group"
    self.group_size = config.group_size
    # grouped actorder isn't supported by this kernel
    assert config.actorder != "group"
    assert config.symmetric, (
        "Only symmetric quantization is supported for MoE")

    if not (self.quant_config.quant_format
            == CompressionFormat.pack_quantized.value
            and self.num_bits in WNA16_SUPPORTED_BITS):
        raise ValueError("For Fused MoE layers, only ",
                         f"{CompressionFormat.pack_quantized.value} ",
                         "is supported for the following bits: ",
                         f"{WNA16_SUPPORTED_BITS}")

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/compressed_tensors/compressed_tensors_moe.py 
1554
1555
1556
1557
1558
1559
1560
1561
1562
1563
1564
1565
1566
1567
1568
1569
1570
1571
1572
1573
1574
1575
1576
1577
1578
1579
1580
1581
1582
1583
1584
1585
1586
1587
1588
1589
1590
1591
1592
1593
1594
1595
1596
1597
1598
1599
1600
1601
1602
1603
1604
1605
1606
1607
1608
1609
1610
1611
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 "
                                  "`CompressedTensorsWNA16MoEMethod` yet.")

    from vllm.model_executor.layers.fused_moe import fused_experts

    topk_weights, topk_ids = FusedMoE.select_experts(
        hidden_states=x,
        router_logits=router_logits,
        use_grouped_topk=use_grouped_topk,
        top_k=top_k,
        renormalize=renormalize,
        topk_group=topk_group,
        num_expert_group=num_expert_group,
        custom_routing_function=custom_routing_function,
        scoring_func=scoring_func,
        e_score_correction_bias=e_score_correction_bias)

    return fused_experts(
        x,
        layer.w13_weight_packed,
        layer.w2_weight_packed,
        topk_weights=topk_weights,
        topk_ids=topk_ids,
        inplace=True,
        activation=activation,
        use_int4_w4a16=self.num_bits == 4,
        use_int8_w8a16=self.num_bits == 8,
        global_num_experts=global_num_experts,
        apply_router_weight_on_input=apply_router_weight_on_input,
        expert_map=expert_map,
        w1_scale=layer.w13_weight_scale,
        w2_scale=layer.w2_weight_scale,
        w1_zp=None,
        w2_zp=None,
        block_shape=[0, self.group_size])

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/compressed_tensors/compressed_tensors_moe.py 
1422
1423
1424
1425
1426
1427
1428
1429
1430
1431
1432
1433
1434
1435
1436
1437
1438
1439
1440
1441
1442
1443
1444
1445
1446
1447
1448
1449
1450
1451
1452
1453
1454
1455
1456
1457
1458
1459
1460
1461
1462
1463
1464
1465
1466
1467
1468
1469
1470
1471
1472
1473
1474
1475
1476
1477
1478
1479
1480
1481
1482
1483
1484
1485
1486
1487
1488
1489
1490
1491
1492
1493
1494
1495
1496
1497
1498
1499
1500
1501
1502
1503
1504
1505
1506
1507
1508
1509
1510
1511
1512
1513
1514
1515
1516
1517
1518
1519
1520
1521
1522
1523
1524
1525
1526
1527
1528
1529
1530
1531
1532
1533
1534
1535
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):

    # Will transpose the loaded weight along the
    # intermediate and hidden dim sizes. Will
    # shard for TP along the transposed dims
    extra_weight_attrs.update({
        "is_transposed": True,
        "quant_method": self.strategy
    })
    w13_weight = torch.nn.Parameter(torch.empty(
        num_experts,
        hidden_size // self.packed_factor,
        2 * intermediate_size_per_partition,
        dtype=torch.int32),
                                    requires_grad=False)
    layer.register_parameter("w13_weight_packed", w13_weight)
    set_weight_attrs(w13_weight, extra_weight_attrs)

    w2_weight = torch.nn.Parameter(torch.empty(
        num_experts,
        intermediate_size_per_partition // self.packed_factor,
        hidden_size,
        dtype=torch.int32),
                                   requires_grad=False)
    layer.register_parameter("w2_weight_packed", w2_weight)
    set_weight_attrs(w2_weight, extra_weight_attrs)

    w2_scales_size = intermediate_size_per_partition

    if self.strategy == "channel":
        num_groups_w2 = num_groups_w13 = 1
        self.group_size = -1
    else:
        num_groups_w2 = w2_scales_size // self.group_size
        num_groups_w13 = hidden_size // self.group_size

    w13_scale = torch.nn.Parameter(torch.ones(
        num_experts,
        num_groups_w13,
        2 * intermediate_size_per_partition,
        dtype=params_dtype),
                                   requires_grad=False)
    layer.register_parameter("w13_weight_scale", w13_scale)
    set_weight_attrs(w13_scale, extra_weight_attrs)

    w2_scale = torch.nn.Parameter(torch.ones(num_experts,
                                             num_groups_w2,
                                             hidden_size,
                                             dtype=params_dtype),
                                  requires_grad=False)
    layer.register_parameter("w2_weight_scale", w2_scale)
    set_weight_attrs(w2_scale, extra_weight_attrs)
    set_weight_attrs(w2_scale, {"load_full_w2": False})

    w2_weight_shape = torch.nn.Parameter(torch.empty(num_experts, 2),
                                         requires_grad=False)
    layer.register_parameter("w2_weight_shape", w2_weight_shape)
    set_weight_attrs(w2_weight_shape, extra_weight_attrs)
    w13_weight_shape = torch.nn.Parameter(torch.empty(num_experts, 2),
                                          requires_grad=False)

    layer.register_parameter("w13_weight_shape", w13_weight_shape)
    set_weight_attrs(w13_weight_shape, extra_weight_attrs)

    w13_g_idx = torch.nn.Parameter(
        torch.empty(
            num_experts,
            hidden_size,
            dtype=torch.int32,
        ),
        requires_grad=False,
    )
    layer.register_parameter("w13_weight_g_idx", w13_g_idx)
    set_weight_attrs(w13_g_idx, extra_weight_attrs)

    w2_g_idx = torch.nn.Parameter(
        torch.empty(
            num_experts,
            intermediate_size_per_partition,
            dtype=torch.int32,
        ),
        requires_grad=False,
    )
    layer.register_parameter("w2_weight_g_idx", w2_g_idx)
    set_weight_attrs(w2_g_idx, extra_weight_attrs)

    w13_g_idx_sort_indices = torch.nn.Parameter(
        torch.empty(
            num_experts,
            hidden_size,
            dtype=torch.int32,
        ),
        requires_grad=False,
    )
    layer.register_parameter("w13_g_idx_sort_indices",
                             w13_g_idx_sort_indices)
    set_weight_attrs(w13_g_idx_sort_indices, extra_weight_attrs)

    w2_g_idx_sort_indices = torch.nn.Parameter(
        torch.empty(
            num_experts,
            intermediate_size_per_partition,
            dtype=torch.int32,
        ),
        requires_grad=False,
    )
    layer.register_parameter("w2_g_idx_sort_indices",
                             w2_g_idx_sort_indices)
    set_weight_attrs(w2_g_idx_sort_indices, extra_weight_attrs)

    layer.a13_scale = None
    layer.a2_scale = None

process_weights_after_loading 

process_weights_after_loading(layer: Module) -> None
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
1537
1538
1539
1540
1541
1542
1543
1544
1545
1546
1547
1548
1549
1550
1551
1552
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
    # Reconfigure packed weights and scales to match moe_wna16 format
    layer.w13_weight_packed = torch.nn.Parameter(
        layer.w13_weight_packed.transpose(1, 2).contiguous().view(
            torch.uint8),
        requires_grad=False)
    layer.w2_weight_packed = torch.nn.Parameter(
        layer.w2_weight_packed.transpose(1,
                                         2).contiguous().view(torch.uint8),
        requires_grad=False)
    layer.w13_weight_scale = torch.nn.Parameter(
        layer.w13_weight_scale.transpose(1, 2).contiguous(),
        requires_grad=False)
    layer.w2_weight_scale = torch.nn.Parameter(
        layer.w2_weight_scale.transpose(1, 2).contiguous(),
        requires_grad=False)

GPTQMarlinState

Bases: Enum

Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py 
41
42
43
class GPTQMarlinState(Enum):
    REPACK = enum.auto()
    READY = enum.auto()

READY class-attribute instance-attribute 

READY = auto()

REPACK class-attribute instance-attribute 

REPACK = auto()