Skip to content

vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors_moe

__all__ module-attribute

__all__ = [
    "CompressedTensorsMoEMethod",
    "CompressedTensorsW8A8Fp8MoEMethod",
    "CompressedTensorsW8A8Int8MoEMethod",
    "CompressedTensorsWNA16MarlinMoEMethod",
    "CompressedTensorsWNA16MoEMethod",
    "CompressedTensorsW4A4MoeMethod",
    "CompressedTensorsW4A8Int8MoEMethod",
]

logger module-attribute

logger = init_logger(__name__)

CompressedTensorsMoEMethod

Bases: FusedMoEMethodBase

Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
class CompressedTensorsMoEMethod(FusedMoEMethodBase):
    def __init_(self, moe: FusedMoEConfig):
        super().__init__(moe)

    @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.
        # Check if a using "Linear" to select schemes
        if "Linear" in quant_config.target_scheme_map:
            matched_target = "Linear"
        else:
            # May have instead defined the linear layers in the fused model

            fused_layers = ["re:.*down_proj.*", "re:.*gate_proj.*", "re:.*up_proj.*"]
            current_scheme = None
            for fused_layer in fused_layers:
                # Check if one of the fused layers are defined in quant_config
                matched_target = find_matched_target(
                    layer_name=fused_layer,
                    module=layer,
                    targets=quant_config.target_scheme_map.keys(),
                    fused_mapping=quant_config.packed_modules_mapping,
                )

                # Only valid if down_proj, gate_proj, and up_proj
                # are mapped to the same quant scheme in the quant_config
                if current_scheme is None:
                    current_scheme = quant_config.target_scheme_map.get(matched_target)
                else:
                    assert current_scheme == quant_config.target_scheme_map.get(
                        matched_target
                    )

        weight_quant = quant_config.target_scheme_map[matched_target].get("weights")
        input_quant = quant_config.target_scheme_map[matched_target].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 == 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, layer.moe_config)
            else:
                logger.info_once("Using CompressedTensorsWNA16MarlinMoEMethod")
                return CompressedTensorsWNA16MarlinMoEMethod(
                    quant_config, layer.moe_config
                )
        elif quant_config._is_fp4a4_nvfp4(weight_quant, input_quant):
            return CompressedTensorsW4A4MoeMethod(layer.moe_config)
        elif (
            quant_config._is_fp8_w8a8_sm90(weight_quant, input_quant)
            or quant_config._is_fp8_w8a8_sm100(weight_quant, input_quant)
            or quant_config._is_fp8_w8a8(weight_quant, input_quant)
        ):
            return CompressedTensorsW8A8Fp8MoEMethod(quant_config, layer.moe_config)
        elif quant_config._is_dynamic_token_w8a8(weight_quant, input_quant):
            return CompressedTensorsW8A8Int8MoEMethod(quant_config, layer.moe_config)
        elif quant_config._is_dynamic_token_w4a8_int(weight_quant, input_quant):
            return CompressedTensorsW4A8Int8MoEMethod(quant_config, layer.moe_config)
        else:
            raise RuntimeError(
                f"Unsupported FusedMoe scheme: {weight_quant}, {input_quant}"
            )

__init_

__init_(moe: FusedMoEConfig)
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def __init_(self, moe: FusedMoEConfig):
    super().__init__(moe)

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
@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.
    # Check if a using "Linear" to select schemes
    if "Linear" in quant_config.target_scheme_map:
        matched_target = "Linear"
    else:
        # May have instead defined the linear layers in the fused model

        fused_layers = ["re:.*down_proj.*", "re:.*gate_proj.*", "re:.*up_proj.*"]
        current_scheme = None
        for fused_layer in fused_layers:
            # Check if one of the fused layers are defined in quant_config
            matched_target = find_matched_target(
                layer_name=fused_layer,
                module=layer,
                targets=quant_config.target_scheme_map.keys(),
                fused_mapping=quant_config.packed_modules_mapping,
            )

            # Only valid if down_proj, gate_proj, and up_proj
            # are mapped to the same quant scheme in the quant_config
            if current_scheme is None:
                current_scheme = quant_config.target_scheme_map.get(matched_target)
            else:
                assert current_scheme == quant_config.target_scheme_map.get(
                    matched_target
                )

    weight_quant = quant_config.target_scheme_map[matched_target].get("weights")
    input_quant = quant_config.target_scheme_map[matched_target].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 == 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, layer.moe_config)
        else:
            logger.info_once("Using CompressedTensorsWNA16MarlinMoEMethod")
            return CompressedTensorsWNA16MarlinMoEMethod(
                quant_config, layer.moe_config
            )
    elif quant_config._is_fp4a4_nvfp4(weight_quant, input_quant):
        return CompressedTensorsW4A4MoeMethod(layer.moe_config)
    elif (
        quant_config._is_fp8_w8a8_sm90(weight_quant, input_quant)
        or quant_config._is_fp8_w8a8_sm100(weight_quant, input_quant)
        or quant_config._is_fp8_w8a8(weight_quant, input_quant)
    ):
        return CompressedTensorsW8A8Fp8MoEMethod(quant_config, layer.moe_config)
    elif quant_config._is_dynamic_token_w8a8(weight_quant, input_quant):
        return CompressedTensorsW8A8Int8MoEMethod(quant_config, layer.moe_config)
    elif quant_config._is_dynamic_token_w4a8_int(weight_quant, input_quant):
        return CompressedTensorsW4A8Int8MoEMethod(quant_config, layer.moe_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
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
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
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
class CompressedTensorsW4A4MoeMethod(CompressedTensorsMoEMethod):
    def __init__(self, moe: FusedMoEConfig):
        from vllm.model_executor.layers.quantization.utils.nvfp4_moe_support import (  # noqa: E501
            detect_nvfp4_moe_support,
        )

        super().__init__(moe)
        _nvfp4 = detect_nvfp4_moe_support(self.__class__.__name__)
        self.cutlass_nvfp4_supported = _nvfp4.cutlass_supported
        self.allow_flashinfer = _nvfp4.allow_flashinfer
        self.use_marlin = _nvfp4.use_marlin
        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 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
        )

        # reorder GEMM1 weights and block scales for FlashInfer CUTLASS kernel.
        if self.allow_flashinfer:
            w, s = reorder_w1w3_to_w3w1(
                layer.w13_weight.data, layer.w13_weight_scale.data, dim=-2
            )
            layer.w13_weight = torch.nn.Parameter(w, requires_grad=False)
            layer.w13_weight_scale = torch.nn.Parameter(s, 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_weight_scale = torch.nn.Parameter(
            swizzle_blockscale(layer.w13_weight_scale), requires_grad=False
        )

        layer.w2_weight_scale = torch.nn.Parameter(
            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 maybe_make_prepare_finalize(self) -> mk.FusedMoEPrepareAndFinalize | None:
        if self.use_marlin:
            return None
        elif not self.allow_flashinfer:
            return super().maybe_make_prepare_finalize()

        prepare_finalize = build_flashinfer_fp4_cutlass_moe_prepare_finalize(self.moe)
        logger.debug_once("%s", prepare_finalize.__class__.__name__)
        return prepare_finalize

    def select_gemm_impl(
        self,
        prepare_finalize: mk.FusedMoEPrepareAndFinalize,
        layer: torch.nn.Module,
    ) -> mk.FusedMoEPermuteExpertsUnpermute:
        assert self.moe_quant_config is not None
        """Return the appropriate GEMM experts implementation."""
        experts = select_nvfp4_gemm_impl(
            self.moe,
            self.moe_quant_config,
            allow_flashinfer=self.allow_flashinfer,
        )
        logger.debug_once("Using %s", experts.__class__.__name__)
        return experts

    def get_fused_moe_quant_config(
        self, layer: torch.nn.Module
    ) -> FusedMoEQuantConfig | None:
        if self.use_marlin:
            return None

        return nvfp4_moe_quant_config(
            g1_alphas=layer.g1_alphas,
            g2_alphas=layer.g2_alphas,
            a1_gscale=layer.w13_input_scale_quant,
            a2_gscale=layer.w2_input_scale_quant,
            w1_scale=layer.w13_weight_scale,
            w2_scale=layer.w2_weight_scale,
        )

    def apply(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        router_logits: torch.Tensor,
        top_k: int,
        renormalize: bool,
        use_grouped_topk: bool = False,
        topk_group: int | None = None,
        num_expert_group: int | None = None,
        global_num_experts: int = -1,
        expert_map: torch.Tensor | None = None,
        custom_routing_function: Callable | None = None,
        scoring_func: str = "softmax",
        routed_scaling_factor: float = 1.0,
        e_score_correction_bias: torch.Tensor | None = None,
        apply_router_weight_on_input: bool = False,
        activation: str = "silu",
        enable_eplb: bool = False,
        expert_load_view: torch.Tensor | None = None,
        logical_to_physical_map: torch.Tensor | None = None,
        logical_replica_count: torch.Tensor | None = None,
    ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
        if enable_eplb:
            raise NotImplementedError(
                "EPLB not supported for `CompressedTensorsW4A4MoeMethod` yet."
            )
        assert activation == "silu", "Only SiLU activation is supported."

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

        #
        # Note: the order here is important. self.fused_experts can override
        # flashinfer cutlass, cutlass fp4 or fused_experts but not marlin.
        #
        if self.use_marlin:
            assert self.fused_experts is None
            return fused_marlin_moe(
                x,
                layer.w13_weight,
                layer.w2_weight,
                None,
                None,
                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,
                workspace=layer.workspace,
            )

        elif self.fused_experts is not None:
            assert is_valid_flashinfer_cutlass_fused_moe(
                x, layer.w13_weight, layer.w2_weight
            ), "Flashinfer CUTLASS Fused MoE not applicable!"

            return self.fused_experts(
                hidden_states=x,
                w1=layer.w13_weight,
                w2=layer.w2_weight,
                topk_weights=topk_weights,
                topk_ids=topk_ids,
                inplace=False,  # TODO(shuw): fix later, now output is high prec
                activation=activation,
                global_num_experts=global_num_experts,
                expert_map=expert_map,
                apply_router_weight_on_input=apply_router_weight_on_input,
            )

        # FlashInfer fused experts path
        elif self.allow_flashinfer:
            from vllm.model_executor.layers.fused_moe.flashinfer_cutlass_moe import (  # noqa: E501
                flashinfer_cutlass_moe_fp4,
            )

            assert is_valid_flashinfer_cutlass_fused_moe(
                x, layer.w13_weight, layer.w2_weight
            ), "Flashinfer CUTLASS Fused MoE not applicable!"

            assert self.moe_quant_config is not None

            return flashinfer_cutlass_moe_fp4(
                hidden_states=x,
                w1=layer.w13_weight,
                w2=layer.w2_weight,
                topk_weights=topk_weights,
                topk_ids=topk_ids,
                quant_config=self.moe_quant_config,
                inplace=False,  # TODO(shuw): fix later, now output is high prec
                activation=activation,
                global_num_experts=global_num_experts,
                expert_map=expert_map,
                apply_router_weight_on_input=apply_router_weight_on_input,
            )
        else:
            from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp4

            assert expert_map is None, (
                "Expert Parallelism / expert_map "
                "is currently not supported for "
                "CompressedTensorsW4A4MoeMethod."
            )
            assert self.moe_quant_config is not None

            # 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,
                w2_fp4=layer.w2_weight,
                topk_weights=topk_weights,
                topk_ids=topk_ids,
                quant_config=self.moe_quant_config,
                apply_router_weight_on_input=apply_router_weight_on_input,
                # TODO(bnell): derive these from arguments
                m=x.shape[0],
                n=layer.w2_weight.shape[2] * 2,
                k=x.shape[1],
                e=layer.w13_weight.shape[0],
            ).to(x.dtype)

allow_flashinfer instance-attribute

allow_flashinfer = allow_flashinfer

cutlass_nvfp4_supported instance-attribute

cutlass_nvfp4_supported = cutlass_supported

group_size instance-attribute

group_size = 16

use_marlin instance-attribute

use_marlin = use_marlin

__init__

__init__(moe: FusedMoEConfig)
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def __init__(self, moe: FusedMoEConfig):
    from vllm.model_executor.layers.quantization.utils.nvfp4_moe_support import (  # noqa: E501
        detect_nvfp4_moe_support,
    )

    super().__init__(moe)
    _nvfp4 = detect_nvfp4_moe_support(self.__class__.__name__)
    self.cutlass_nvfp4_supported = _nvfp4.cutlass_supported
    self.allow_flashinfer = _nvfp4.allow_flashinfer
    self.use_marlin = _nvfp4.use_marlin
    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: int | None = None,
    num_expert_group: int | None = None,
    global_num_experts: int = -1,
    expert_map: Tensor | None = None,
    custom_routing_function: Callable | None = None,
    scoring_func: str = "softmax",
    routed_scaling_factor: float = 1.0,
    e_score_correction_bias: Tensor | None = None,
    apply_router_weight_on_input: bool = False,
    activation: str = "silu",
    enable_eplb: bool = False,
    expert_load_view: Tensor | None = None,
    logical_to_physical_map: Tensor | None = None,
    logical_replica_count: Tensor | None = None,
) -> Tensor | tuple[Tensor, Tensor]
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def apply(
    self,
    layer: torch.nn.Module,
    x: torch.Tensor,
    router_logits: torch.Tensor,
    top_k: int,
    renormalize: bool,
    use_grouped_topk: bool = False,
    topk_group: int | None = None,
    num_expert_group: int | None = None,
    global_num_experts: int = -1,
    expert_map: torch.Tensor | None = None,
    custom_routing_function: Callable | None = None,
    scoring_func: str = "softmax",
    routed_scaling_factor: float = 1.0,
    e_score_correction_bias: torch.Tensor | None = None,
    apply_router_weight_on_input: bool = False,
    activation: str = "silu",
    enable_eplb: bool = False,
    expert_load_view: torch.Tensor | None = None,
    logical_to_physical_map: torch.Tensor | None = None,
    logical_replica_count: torch.Tensor | None = None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
    if enable_eplb:
        raise NotImplementedError(
            "EPLB not supported for `CompressedTensorsW4A4MoeMethod` yet."
        )
    assert activation == "silu", "Only SiLU activation is supported."

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

    #
    # Note: the order here is important. self.fused_experts can override
    # flashinfer cutlass, cutlass fp4 or fused_experts but not marlin.
    #
    if self.use_marlin:
        assert self.fused_experts is None
        return fused_marlin_moe(
            x,
            layer.w13_weight,
            layer.w2_weight,
            None,
            None,
            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,
            workspace=layer.workspace,
        )

    elif self.fused_experts is not None:
        assert is_valid_flashinfer_cutlass_fused_moe(
            x, layer.w13_weight, layer.w2_weight
        ), "Flashinfer CUTLASS Fused MoE not applicable!"

        return self.fused_experts(
            hidden_states=x,
            w1=layer.w13_weight,
            w2=layer.w2_weight,
            topk_weights=topk_weights,
            topk_ids=topk_ids,
            inplace=False,  # TODO(shuw): fix later, now output is high prec
            activation=activation,
            global_num_experts=global_num_experts,
            expert_map=expert_map,
            apply_router_weight_on_input=apply_router_weight_on_input,
        )

    # FlashInfer fused experts path
    elif self.allow_flashinfer:
        from vllm.model_executor.layers.fused_moe.flashinfer_cutlass_moe import (  # noqa: E501
            flashinfer_cutlass_moe_fp4,
        )

        assert is_valid_flashinfer_cutlass_fused_moe(
            x, layer.w13_weight, layer.w2_weight
        ), "Flashinfer CUTLASS Fused MoE not applicable!"

        assert self.moe_quant_config is not None

        return flashinfer_cutlass_moe_fp4(
            hidden_states=x,
            w1=layer.w13_weight,
            w2=layer.w2_weight,
            topk_weights=topk_weights,
            topk_ids=topk_ids,
            quant_config=self.moe_quant_config,
            inplace=False,  # TODO(shuw): fix later, now output is high prec
            activation=activation,
            global_num_experts=global_num_experts,
            expert_map=expert_map,
            apply_router_weight_on_input=apply_router_weight_on_input,
        )
    else:
        from vllm.model_executor.layers.fused_moe.cutlass_moe import cutlass_moe_fp4

        assert expert_map is None, (
            "Expert Parallelism / expert_map "
            "is currently not supported for "
            "CompressedTensorsW4A4MoeMethod."
        )
        assert self.moe_quant_config is not None

        # 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,
            w2_fp4=layer.w2_weight,
            topk_weights=topk_weights,
            topk_ids=topk_ids,
            quant_config=self.moe_quant_config,
            apply_router_weight_on_input=apply_router_weight_on_input,
            # TODO(bnell): derive these from arguments
            m=x.shape[0],
            n=layer.w2_weight.shape[2] * 2,
            k=x.shape[1],
            e=layer.w13_weight.shape[0],
        ).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
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)

get_fused_moe_quant_config

get_fused_moe_quant_config(
    layer: Module,
) -> FusedMoEQuantConfig | None
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def get_fused_moe_quant_config(
    self, layer: torch.nn.Module
) -> FusedMoEQuantConfig | None:
    if self.use_marlin:
        return None

    return nvfp4_moe_quant_config(
        g1_alphas=layer.g1_alphas,
        g2_alphas=layer.g2_alphas,
        a1_gscale=layer.w13_input_scale_quant,
        a2_gscale=layer.w2_input_scale_quant,
        w1_scale=layer.w13_weight_scale,
        w2_scale=layer.w2_weight_scale,
    )

maybe_make_prepare_finalize

maybe_make_prepare_finalize() -> (
    FusedMoEPrepareAndFinalize | None
)
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def maybe_make_prepare_finalize(self) -> mk.FusedMoEPrepareAndFinalize | None:
    if self.use_marlin:
        return None
    elif not self.allow_flashinfer:
        return super().maybe_make_prepare_finalize()

    prepare_finalize = build_flashinfer_fp4_cutlass_moe_prepare_finalize(self.moe)
    logger.debug_once("%s", prepare_finalize.__class__.__name__)
    return prepare_finalize

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
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
    )

    # reorder GEMM1 weights and block scales for FlashInfer CUTLASS kernel.
    if self.allow_flashinfer:
        w, s = reorder_w1w3_to_w3w1(
            layer.w13_weight.data, layer.w13_weight_scale.data, dim=-2
        )
        layer.w13_weight = torch.nn.Parameter(w, requires_grad=False)
        layer.w13_weight_scale = torch.nn.Parameter(s, 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_weight_scale = torch.nn.Parameter(
        swizzle_blockscale(layer.w13_weight_scale), requires_grad=False
    )

    layer.w2_weight_scale = torch.nn.Parameter(
        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
    )

select_gemm_impl

select_gemm_impl(
    prepare_finalize: FusedMoEPrepareAndFinalize,
    layer: Module,
) -> FusedMoEPermuteExpertsUnpermute
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def select_gemm_impl(
    self,
    prepare_finalize: mk.FusedMoEPrepareAndFinalize,
    layer: torch.nn.Module,
) -> mk.FusedMoEPermuteExpertsUnpermute:
    assert self.moe_quant_config is not None
    """Return the appropriate GEMM experts implementation."""
    experts = select_nvfp4_gemm_impl(
        self.moe,
        self.moe_quant_config,
        allow_flashinfer=self.allow_flashinfer,
    )
    logger.debug_once("Using %s", experts.__class__.__name__)
    return experts

CompressedTensorsW4A8Int8MoEMethod

Bases: CompressedTensorsMoEMethod

CPU-only MoE method using dynamic 4-bit matmul kernels on Arm Platform - Weights: int4 (stored as int8 values in [-8,7], packed to uint8 nibbles) - Scales: Fp32 for Channelwise , bf16 for groupwise quantization - Bias: Same data type as original weights - Activations: FP32/Bf16 dynamic per-token (A8 Int), quantized inside the kernel

Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
1938
1939
1940
1941
1942
1943
1944
1945
1946
1947
1948
1949
1950
1951
1952
1953
1954
1955
1956
1957
1958
1959
1960
1961
1962
1963
1964
1965
1966
1967
1968
1969
1970
1971
1972
1973
1974
1975
1976
1977
1978
1979
1980
1981
1982
1983
1984
1985
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
1996
1997
1998
1999
2000
2001
2002
2003
2004
2005
2006
2007
2008
2009
2010
2011
2012
2013
2014
2015
2016
2017
2018
2019
2020
2021
2022
2023
2024
2025
2026
2027
2028
2029
2030
2031
2032
2033
2034
2035
2036
2037
2038
2039
2040
2041
2042
2043
2044
2045
2046
2047
2048
2049
2050
2051
2052
2053
2054
2055
2056
2057
2058
2059
2060
2061
2062
2063
2064
2065
2066
2067
2068
2069
2070
2071
2072
2073
2074
2075
2076
2077
2078
2079
2080
2081
2082
2083
2084
2085
2086
2087
2088
2089
2090
2091
2092
2093
2094
2095
2096
2097
2098
2099
2100
2101
2102
2103
2104
2105
2106
2107
2108
2109
2110
2111
2112
2113
2114
2115
2116
2117
2118
2119
2120
2121
2122
2123
2124
2125
2126
2127
2128
2129
2130
2131
2132
2133
2134
2135
2136
2137
2138
2139
2140
2141
2142
2143
2144
2145
2146
2147
2148
2149
2150
2151
2152
2153
2154
2155
2156
2157
2158
2159
2160
2161
2162
2163
2164
2165
2166
2167
2168
2169
2170
2171
2172
2173
2174
2175
2176
2177
2178
2179
2180
2181
2182
2183
2184
2185
2186
2187
2188
2189
2190
2191
2192
2193
2194
2195
2196
2197
2198
2199
2200
2201
2202
2203
2204
2205
2206
2207
2208
2209
2210
2211
2212
2213
2214
2215
2216
2217
2218
2219
2220
2221
2222
2223
2224
2225
2226
2227
2228
2229
2230
2231
2232
2233
2234
2235
2236
2237
2238
2239
2240
2241
2242
2243
2244
2245
2246
2247
2248
2249
2250
2251
2252
2253
2254
2255
2256
2257
2258
2259
2260
2261
2262
2263
2264
2265
2266
2267
2268
class CompressedTensorsW4A8Int8MoEMethod(CompressedTensorsMoEMethod):
    """
    CPU-only MoE method using dynamic 4-bit matmul kernels on Arm Platform
    - Weights: int4 (stored as int8 values in [-8,7], packed to uint8 nibbles)
    - Scales: Fp32 for Channelwise , bf16 for groupwise quantization
    - Bias: Same data type as original weights
    - Activations: FP32/Bf16 dynamic per-token (A8 Int),
      quantized inside the kernel
    """

    def __init__(
        self,
        quant_config: "CompressedTensorsConfig",  # type: ignore # noqa E501
        moe: FusedMoEConfig,
    ):
        super().__init__(moe)
        self.has_bias = self.moe.has_bias
        self.quant_config = quant_config

        # Validate scheme: weights=W4 (channel or group),
        # activations=dynamic TOKEN (A8)
        wq = self.quant_config.target_scheme_map["Linear"].get("weights")
        aq = self.quant_config.target_scheme_map["Linear"].get("input_activations")

        # Must be dynamic per-token activations
        if aq.strategy != QuantizationStrategy.TOKEN or not aq.dynamic:
            raise ValueError(
                "W4A8-int MoE needs dynamic per-token activation quantization."
            )

        # Weight can be channel-wise (group_size=None) or group-wise
        self.group_size = wq.group_size if (wq.group_size is not None) else -1
        if wq.num_bits != 4:
            raise ValueError("This method only supports 4-bit weights (num_bits=4).")

        # CPU only
        if not current_platform.is_cpu():
            raise ValueError("CompressedTensorsW4A8Int8MoEMethod is CPU-only.")

        # Arm: check _dyn ops availability
        if current_platform.get_cpu_architecture() == CpuArchEnum.ARM:
            try:
                _ = torch.ops.aten._dyn_quant_matmul_4bit
                _ = torch.ops.aten._dyn_quant_pack_4bit_weight
            except AttributeError as err:
                raise RuntimeError(
                    f"""PyTorch {torch.__version__} lacks _dyn_quant_* 4bit ops;
                    install a newer build."""
                ) from err
        self.static_input_scales = False  # always dynamic per token

    # ---- parameter creation ----
    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,
    ):
        # Shapes per local rank (TP/EP):
        #   w13: [E, 2*I_local, H]  int8  (int4 values in [-8,7])
        #   w2 : [E, H, I_local]    int8
        # Scales:
        #   channel-wise: group_size=-1 -> per-output-row, single scale per row
        #   group-wise  : group_size=g   ->
        #   per-output-row, (in_features/g) scales

        E = num_experts
        H = hidden_size
        IN = intermediate_size_per_partition
        g = self.group_size

        # Per-row scale columns
        def _n_scale_cols(in_features: int) -> int:
            return 1 if g == -1 else (in_features // g)

        # Register unpacked int4-as-int8 weights the loader will fill.
        w13 = torch.nn.Parameter(
            torch.empty(E, 2 * IN, H, dtype=torch.int8), requires_grad=False
        )
        set_weight_attrs(w13, extra_weight_attrs)
        layer.register_parameter("w13_weight", w13)

        w2 = torch.nn.Parameter(
            torch.empty(E, H, IN, dtype=torch.int8), requires_grad=False
        )
        set_weight_attrs(w2, extra_weight_attrs)
        layer.register_parameter("w2_weight", w2)

        # Register scales
        # KleidiAI groupwise kernels accepts float32 scales
        # KleidiAI groupwise kernels accepts bfloat16 scales
        scale_dtype = torch.float32 if g == -1 else torch.bfloat16

        w13_s = torch.nn.Parameter(
            torch.ones(E, 2 * IN, _n_scale_cols(H), dtype=scale_dtype),
            requires_grad=False,
        )
        set_weight_attrs(
            w13_s,
            {"quant_method": "channel" if g == -1 else "group", **extra_weight_attrs},
        )
        layer.register_parameter("w13_weight_scale", w13_s)

        w2_s = torch.nn.Parameter(
            torch.ones(E, H, _n_scale_cols(IN), dtype=scale_dtype), requires_grad=False
        )
        set_weight_attrs(
            w2_s,
            {"quant_method": "channel" if g == -1 else "group", **extra_weight_attrs},
        )
        layer.register_parameter("w2_weight_scale", w2_s)

        if self.has_bias:
            w13_bias = torch.nn.Parameter(
                torch.zeros(E, 2 * IN, dtype=params_dtype), requires_grad=False
            )
            layer.register_parameter("w13_bias", w13_bias)
            set_weight_attrs(w13_bias, extra_weight_attrs)

            w2_bias = torch.nn.Parameter(
                torch.zeros(num_experts, hidden_size, dtype=params_dtype),
                requires_grad=False,
            )
            layer.register_parameter("w2_bias", w2_bias)
            set_weight_attrs(w2_bias, extra_weight_attrs)

        # Placeholders for packed weights (will be replaced after packing)
        layer.register_parameter(
            "w13_weight_packed", torch.nn.Parameter(torch.empty(0), requires_grad=False)
        )
        set_weight_attrs(layer.w13_weight_packed, extra_weight_attrs)

        layer.register_parameter(
            "w2_weight_packed", torch.nn.Parameter(torch.empty(0), requires_grad=False)
        )
        set_weight_attrs(layer.w2_weight_packed, extra_weight_attrs)

        # dims for 4 bit fused matmuls
        layer.w13_in_features = H
        layer.w13_out_features = 2 * IN
        layer.w2_in_features = IN
        layer.w2_out_features = H
        layer.group_size = g

    # post-load packing to dyn-4bit KleidiAI kernel's format
    def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
        E = layer.w13_weight.shape[0]
        H = layer.w13_in_features
        I2 = layer.w13_out_features
        IN = layer.w2_in_features
        g = layer.group_size

        def _pack_matrix(
            int4_as_int8_2d: torch.Tensor,
            scales_2d: torch.Tensor,
            bias_1d: torch.Tensor | None,
            in_features: int,
            out_features: int,
        ) -> torch.Tensor:
            # int4 values are stored as int8 in [-8,7].
            # Shift to unsigned nibble and pack pairs along input-dim.
            tmp = int4_as_int8_2d.add(8)  # [out, in]
            uint8_nibbles = ((tmp[:, 1::2] << 4) | tmp[:, ::2]).to(
                torch.uint8
            )  # [out, in//2]

            # KleidiAI groupwise kernels accepts float32 scales
            # KleidiAI groupwise kernels accepts bfloat16 scales
            scale_dtype = torch.float32 if g == -1 else torch.bfloat16
            scales = scales_2d.to(scale_dtype)
            bias = None if bias_1d is None else bias_1d.to(torch.float32)
            return torch.ops.aten._dyn_quant_pack_4bit_weight(
                uint8_nibbles,
                scales,
                bias,
                g if g != -1 else in_features,
                in_features,
                out_features,
            )

        # Pack per expert
        w13_packed_list = []
        w2_packed_list = []

        has_w13_bias = hasattr(layer, "w13_bias") and layer.w13_bias is not None
        has_w2_bias = hasattr(layer, "w2_bias") and layer.w2_bias is not None

        for e in range(E):
            w13_packed_list.append(
                _pack_matrix(
                    layer.w13_weight[e],  # [2I, H]
                    layer.w13_weight_scale[e],  # [2I, H/g or 1]
                    layer.w13_bias[e] if has_w13_bias else None,  # [2I]
                    H,
                    I2,
                )
            )
            w2_packed_list.append(
                _pack_matrix(
                    # w2 shape is [H, IN]; we need [out, in] == [H, IN].
                    layer.w2_weight[e],  # [H, IN]
                    layer.w2_weight_scale[e],  # [H, IN/g or 1]
                    layer.w2_bias[e] if has_w2_bias else None,  # [H]
                    IN,
                    layer.w2_out_features,  # in_features=IN, out_features=H
                )
            )

        # each packed tensor has identical shape per expert; stack on dim 0
        w13_packed = torch.stack(w13_packed_list, dim=0)
        w2_packed = torch.stack(w2_packed_list, dim=0)

        replace_parameter(
            layer,
            "w13_weight_packed",
            torch.nn.Parameter(w13_packed, requires_grad=False),
        )
        replace_parameter(
            layer,
            "w2_weight_packed",
            torch.nn.Parameter(w2_packed, requires_grad=False),
        )

        # free raw tensors/scales/bias now that they're packed into the payload.
        replace_parameter(
            layer, "w13_weight", torch.nn.Parameter(torch.empty(0), requires_grad=False)
        )
        replace_parameter(
            layer, "w2_weight", torch.nn.Parameter(torch.empty(0), requires_grad=False)
        )
        replace_parameter(
            layer,
            "w13_weight_scale",
            torch.nn.Parameter(torch.empty(0), requires_grad=False),
        )
        replace_parameter(
            layer,
            "w2_weight_scale",
            torch.nn.Parameter(torch.empty(0), requires_grad=False),
        )
        if has_w13_bias:
            replace_parameter(
                layer,
                "w13_bias",
                torch.nn.Parameter(torch.empty(0), requires_grad=False),
            )
        if has_w2_bias:
            replace_parameter(
                layer,
                "w2_bias",
                torch.nn.Parameter(torch.empty(0), requires_grad=False),
            )

    def get_fused_moe_quant_config(
        self, layer: torch.nn.Module
    ) -> FusedMoEQuantConfig | None:
        # CPU dynamic 4-bit MoE path does not use modular kernels or
        # fused_experts; quant config is not needed.
        return None

    def apply(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        router_logits: torch.Tensor,
        top_k: int,
        renormalize: bool,
        use_grouped_topk: bool = False,
        topk_group: int | None = None,
        num_expert_group: int | None = None,
        global_num_experts: int = -1,
        expert_map: torch.Tensor | None = None,
        custom_routing_function: Callable | None = None,
        scoring_func: str = "softmax",
        routed_scaling_factor: float = 1.0,
        e_score_correction_bias: torch.Tensor | None = None,
        apply_router_weight_on_input: bool = False,
        activation: str = "silu",
        enable_eplb: bool = False,
        expert_load_view: torch.Tensor | None = None,
        logical_to_physical_map: torch.Tensor | None = None,
        logical_replica_count: torch.Tensor | None = None,
    ) -> torch.Tensor:
        assert not enable_eplb, "EPLB not supported for W4A8-int MoE yet."
        assert activation in ("silu", "swigluoai", "swiglu"), (
            "Only SiLU/SwiGLUGU/SwiGLUUG are supported."
        )
        assert expert_map is None, """expert_map/EP not implemented
        for CPU dyn-4bit MoE."""

        def _act_kind(s: str) -> int:
            # 0 = SwiGLU_Gu (SiLU(g)*u), 1 = SwiGLU_Ug (SiLU(u)*g), 2 = SiLU
            if s == "swiglu":
                return 0
            if s == "swigluoai":
                return 1
            if s == "silu":
                return 2
            raise ValueError(f"Unknown activation '{s}'")

        # Apply topk softmax on router output
        topk_weights, topk_ids = select_experts(
            hidden_states=x,
            router_logits=router_logits,
            use_grouped_topk=use_grouped_topk,
            top_k=top_k,
            renormalize=renormalize,
            topk_group=topk_group,
            num_expert_group=num_expert_group,
            custom_routing_function=custom_routing_function,
            scoring_func=scoring_func,
            routed_scaling_factor=routed_scaling_factor,
            e_score_correction_bias=e_score_correction_bias,
        )

        return torch.ops._C.dynamic_4bit_int_moe(
            x,
            topk_ids.to(torch.long),
            topk_weights,
            layer.w13_weight_packed,
            layer.w2_weight_packed,
            layer.w2_out_features,
            layer.w2_in_features,
            layer.w13_out_features,
            layer.group_size,
            apply_router_weight_on_input,
            int(_act_kind(activation)),
        )

group_size instance-attribute

group_size = group_size if group_size is not None else -1

has_bias instance-attribute

has_bias = has_bias

quant_config instance-attribute

quant_config = quant_config

static_input_scales instance-attribute

static_input_scales = False

__init__

__init__(
    quant_config: CompressedTensorsConfig,
    moe: FusedMoEConfig,
)
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def __init__(
    self,
    quant_config: "CompressedTensorsConfig",  # type: ignore # noqa E501
    moe: FusedMoEConfig,
):
    super().__init__(moe)
    self.has_bias = self.moe.has_bias
    self.quant_config = quant_config

    # Validate scheme: weights=W4 (channel or group),
    # activations=dynamic TOKEN (A8)
    wq = self.quant_config.target_scheme_map["Linear"].get("weights")
    aq = self.quant_config.target_scheme_map["Linear"].get("input_activations")

    # Must be dynamic per-token activations
    if aq.strategy != QuantizationStrategy.TOKEN or not aq.dynamic:
        raise ValueError(
            "W4A8-int MoE needs dynamic per-token activation quantization."
        )

    # Weight can be channel-wise (group_size=None) or group-wise
    self.group_size = wq.group_size if (wq.group_size is not None) else -1
    if wq.num_bits != 4:
        raise ValueError("This method only supports 4-bit weights (num_bits=4).")

    # CPU only
    if not current_platform.is_cpu():
        raise ValueError("CompressedTensorsW4A8Int8MoEMethod is CPU-only.")

    # Arm: check _dyn ops availability
    if current_platform.get_cpu_architecture() == CpuArchEnum.ARM:
        try:
            _ = torch.ops.aten._dyn_quant_matmul_4bit
            _ = torch.ops.aten._dyn_quant_pack_4bit_weight
        except AttributeError as err:
            raise RuntimeError(
                f"""PyTorch {torch.__version__} lacks _dyn_quant_* 4bit ops;
                install a newer build."""
            ) from err
    self.static_input_scales = False  # always dynamic per token

apply

apply(
    layer: Module,
    x: Tensor,
    router_logits: Tensor,
    top_k: int,
    renormalize: bool,
    use_grouped_topk: bool = False,
    topk_group: int | None = None,
    num_expert_group: int | None = None,
    global_num_experts: int = -1,
    expert_map: Tensor | None = None,
    custom_routing_function: Callable | None = None,
    scoring_func: str = "softmax",
    routed_scaling_factor: float = 1.0,
    e_score_correction_bias: Tensor | None = None,
    apply_router_weight_on_input: bool = False,
    activation: str = "silu",
    enable_eplb: bool = False,
    expert_load_view: Tensor | None = None,
    logical_to_physical_map: Tensor | None = None,
    logical_replica_count: Tensor | None = None,
) -> Tensor
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def apply(
    self,
    layer: torch.nn.Module,
    x: torch.Tensor,
    router_logits: torch.Tensor,
    top_k: int,
    renormalize: bool,
    use_grouped_topk: bool = False,
    topk_group: int | None = None,
    num_expert_group: int | None = None,
    global_num_experts: int = -1,
    expert_map: torch.Tensor | None = None,
    custom_routing_function: Callable | None = None,
    scoring_func: str = "softmax",
    routed_scaling_factor: float = 1.0,
    e_score_correction_bias: torch.Tensor | None = None,
    apply_router_weight_on_input: bool = False,
    activation: str = "silu",
    enable_eplb: bool = False,
    expert_load_view: torch.Tensor | None = None,
    logical_to_physical_map: torch.Tensor | None = None,
    logical_replica_count: torch.Tensor | None = None,
) -> torch.Tensor:
    assert not enable_eplb, "EPLB not supported for W4A8-int MoE yet."
    assert activation in ("silu", "swigluoai", "swiglu"), (
        "Only SiLU/SwiGLUGU/SwiGLUUG are supported."
    )
    assert expert_map is None, """expert_map/EP not implemented
    for CPU dyn-4bit MoE."""

    def _act_kind(s: str) -> int:
        # 0 = SwiGLU_Gu (SiLU(g)*u), 1 = SwiGLU_Ug (SiLU(u)*g), 2 = SiLU
        if s == "swiglu":
            return 0
        if s == "swigluoai":
            return 1
        if s == "silu":
            return 2
        raise ValueError(f"Unknown activation '{s}'")

    # Apply topk softmax on router output
    topk_weights, topk_ids = select_experts(
        hidden_states=x,
        router_logits=router_logits,
        use_grouped_topk=use_grouped_topk,
        top_k=top_k,
        renormalize=renormalize,
        topk_group=topk_group,
        num_expert_group=num_expert_group,
        custom_routing_function=custom_routing_function,
        scoring_func=scoring_func,
        routed_scaling_factor=routed_scaling_factor,
        e_score_correction_bias=e_score_correction_bias,
    )

    return torch.ops._C.dynamic_4bit_int_moe(
        x,
        topk_ids.to(torch.long),
        topk_weights,
        layer.w13_weight_packed,
        layer.w2_weight_packed,
        layer.w2_out_features,
        layer.w2_in_features,
        layer.w13_out_features,
        layer.group_size,
        apply_router_weight_on_input,
        int(_act_kind(activation)),
    )

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
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,
):
    # Shapes per local rank (TP/EP):
    #   w13: [E, 2*I_local, H]  int8  (int4 values in [-8,7])
    #   w2 : [E, H, I_local]    int8
    # Scales:
    #   channel-wise: group_size=-1 -> per-output-row, single scale per row
    #   group-wise  : group_size=g   ->
    #   per-output-row, (in_features/g) scales

    E = num_experts
    H = hidden_size
    IN = intermediate_size_per_partition
    g = self.group_size

    # Per-row scale columns
    def _n_scale_cols(in_features: int) -> int:
        return 1 if g == -1 else (in_features // g)

    # Register unpacked int4-as-int8 weights the loader will fill.
    w13 = torch.nn.Parameter(
        torch.empty(E, 2 * IN, H, dtype=torch.int8), requires_grad=False
    )
    set_weight_attrs(w13, extra_weight_attrs)
    layer.register_parameter("w13_weight", w13)

    w2 = torch.nn.Parameter(
        torch.empty(E, H, IN, dtype=torch.int8), requires_grad=False
    )
    set_weight_attrs(w2, extra_weight_attrs)
    layer.register_parameter("w2_weight", w2)

    # Register scales
    # KleidiAI groupwise kernels accepts float32 scales
    # KleidiAI groupwise kernels accepts bfloat16 scales
    scale_dtype = torch.float32 if g == -1 else torch.bfloat16

    w13_s = torch.nn.Parameter(
        torch.ones(E, 2 * IN, _n_scale_cols(H), dtype=scale_dtype),
        requires_grad=False,
    )
    set_weight_attrs(
        w13_s,
        {"quant_method": "channel" if g == -1 else "group", **extra_weight_attrs},
    )
    layer.register_parameter("w13_weight_scale", w13_s)

    w2_s = torch.nn.Parameter(
        torch.ones(E, H, _n_scale_cols(IN), dtype=scale_dtype), requires_grad=False
    )
    set_weight_attrs(
        w2_s,
        {"quant_method": "channel" if g == -1 else "group", **extra_weight_attrs},
    )
    layer.register_parameter("w2_weight_scale", w2_s)

    if self.has_bias:
        w13_bias = torch.nn.Parameter(
            torch.zeros(E, 2 * IN, dtype=params_dtype), requires_grad=False
        )
        layer.register_parameter("w13_bias", w13_bias)
        set_weight_attrs(w13_bias, extra_weight_attrs)

        w2_bias = torch.nn.Parameter(
            torch.zeros(num_experts, hidden_size, dtype=params_dtype),
            requires_grad=False,
        )
        layer.register_parameter("w2_bias", w2_bias)
        set_weight_attrs(w2_bias, extra_weight_attrs)

    # Placeholders for packed weights (will be replaced after packing)
    layer.register_parameter(
        "w13_weight_packed", torch.nn.Parameter(torch.empty(0), requires_grad=False)
    )
    set_weight_attrs(layer.w13_weight_packed, extra_weight_attrs)

    layer.register_parameter(
        "w2_weight_packed", torch.nn.Parameter(torch.empty(0), requires_grad=False)
    )
    set_weight_attrs(layer.w2_weight_packed, extra_weight_attrs)

    # dims for 4 bit fused matmuls
    layer.w13_in_features = H
    layer.w13_out_features = 2 * IN
    layer.w2_in_features = IN
    layer.w2_out_features = H
    layer.group_size = g

get_fused_moe_quant_config

get_fused_moe_quant_config(
    layer: Module,
) -> FusedMoEQuantConfig | None
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def get_fused_moe_quant_config(
    self, layer: torch.nn.Module
) -> FusedMoEQuantConfig | None:
    # CPU dynamic 4-bit MoE path does not use modular kernels or
    # fused_experts; quant config is not needed.
    return 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
def process_weights_after_loading(self, layer: torch.nn.Module) -> None:
    E = layer.w13_weight.shape[0]
    H = layer.w13_in_features
    I2 = layer.w13_out_features
    IN = layer.w2_in_features
    g = layer.group_size

    def _pack_matrix(
        int4_as_int8_2d: torch.Tensor,
        scales_2d: torch.Tensor,
        bias_1d: torch.Tensor | None,
        in_features: int,
        out_features: int,
    ) -> torch.Tensor:
        # int4 values are stored as int8 in [-8,7].
        # Shift to unsigned nibble and pack pairs along input-dim.
        tmp = int4_as_int8_2d.add(8)  # [out, in]
        uint8_nibbles = ((tmp[:, 1::2] << 4) | tmp[:, ::2]).to(
            torch.uint8
        )  # [out, in//2]

        # KleidiAI groupwise kernels accepts float32 scales
        # KleidiAI groupwise kernels accepts bfloat16 scales
        scale_dtype = torch.float32 if g == -1 else torch.bfloat16
        scales = scales_2d.to(scale_dtype)
        bias = None if bias_1d is None else bias_1d.to(torch.float32)
        return torch.ops.aten._dyn_quant_pack_4bit_weight(
            uint8_nibbles,
            scales,
            bias,
            g if g != -1 else in_features,
            in_features,
            out_features,
        )

    # Pack per expert
    w13_packed_list = []
    w2_packed_list = []

    has_w13_bias = hasattr(layer, "w13_bias") and layer.w13_bias is not None
    has_w2_bias = hasattr(layer, "w2_bias") and layer.w2_bias is not None

    for e in range(E):
        w13_packed_list.append(
            _pack_matrix(
                layer.w13_weight[e],  # [2I, H]
                layer.w13_weight_scale[e],  # [2I, H/g or 1]
                layer.w13_bias[e] if has_w13_bias else None,  # [2I]
                H,
                I2,
            )
        )
        w2_packed_list.append(
            _pack_matrix(
                # w2 shape is [H, IN]; we need [out, in] == [H, IN].
                layer.w2_weight[e],  # [H, IN]
                layer.w2_weight_scale[e],  # [H, IN/g or 1]
                layer.w2_bias[e] if has_w2_bias else None,  # [H]
                IN,
                layer.w2_out_features,  # in_features=IN, out_features=H
            )
        )

    # each packed tensor has identical shape per expert; stack on dim 0
    w13_packed = torch.stack(w13_packed_list, dim=0)
    w2_packed = torch.stack(w2_packed_list, dim=0)

    replace_parameter(
        layer,
        "w13_weight_packed",
        torch.nn.Parameter(w13_packed, requires_grad=False),
    )
    replace_parameter(
        layer,
        "w2_weight_packed",
        torch.nn.Parameter(w2_packed, requires_grad=False),
    )

    # free raw tensors/scales/bias now that they're packed into the payload.
    replace_parameter(
        layer, "w13_weight", torch.nn.Parameter(torch.empty(0), requires_grad=False)
    )
    replace_parameter(
        layer, "w2_weight", torch.nn.Parameter(torch.empty(0), requires_grad=False)
    )
    replace_parameter(
        layer,
        "w13_weight_scale",
        torch.nn.Parameter(torch.empty(0), requires_grad=False),
    )
    replace_parameter(
        layer,
        "w2_weight_scale",
        torch.nn.Parameter(torch.empty(0), requires_grad=False),
    )
    if has_w13_bias:
        replace_parameter(
            layer,
            "w13_bias",
            torch.nn.Parameter(torch.empty(0), requires_grad=False),
        )
    if has_w2_bias:
        replace_parameter(
            layer,
            "w2_bias",
            torch.nn.Parameter(torch.empty(0), requires_grad=False),
        )

CompressedTensorsW8A8Fp8MoEMethod

Bases: CompressedTensorsMoEMethod

Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
 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
 707
 708
 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
 948
 949
 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
1089
1090
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
class CompressedTensorsW8A8Fp8MoEMethod(CompressedTensorsMoEMethod):
    def __init__(
        self,
        quant_config: "CompressedTensorsConfig",  # type: ignore # noqa E501
        moe: FusedMoEConfig,
    ):
        super().__init__(moe)
        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):
            assert self.weight_quant.strategy == QuantizationStrategy.BLOCK
            self.weight_block_size = self.weight_quant.block_structure
            assert self.weight_quant.dynamic is not None
        else:
            self.weight_block_size = None
        self.block_quant = self.weight_block_size is not None

        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
            and not self.block_quant
        )
        # 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()

        # cutlass path
        self.is_fp8_w8a8_sm100 = quant_config._is_fp8_w8a8_sm100(
            self.weight_quant, self.input_quant
        )
        self.use_cutlass = not self.block_quant and (
            quant_config._is_fp8_w8a8_sm90(self.weight_quant, self.input_quant)
            or self.is_fp8_w8a8_sm100
        )
        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,
    ):
        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

        if self.block_quant:
            assert self.weight_block_size is not None
            layer.weight_block_size = self.weight_block_size
            tp_size = get_tensor_model_parallel_world_size()
            block_n, block_k = (
                self.weight_block_size[0],
                self.weight_block_size[1],
            )
            # NOTE: To ensure proper alignment of the block-wise quantization
            # scales, the output_size of the weights for both the gate and up
            # layers must be divisible by block_n.
            # Required by column parallel or enabling merged weights
            if intermediate_size_per_partition % block_n != 0:
                raise ValueError(
                    f"The output_size of gate's and up's weight = "
                    f"{intermediate_size_per_partition} is not divisible by "
                    f"weight quantization block_n = {block_n}."
                )
            if tp_size > 1 and intermediate_size_per_partition % block_k != 0:
                # Required by row parallel
                raise ValueError(
                    f"The input_size of down's weight = "
                    f"{intermediate_size_per_partition} is not divisible by "
                    f"weight quantization block_k = {block_k}."
                )

        # 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)

        elif self.weight_quant.strategy == QuantizationStrategy.BLOCK:
            w13_weight_scale = torch.nn.Parameter(
                torch.ones(
                    num_experts,
                    2 * ((intermediate_size_per_partition + block_n - 1) // block_n),
                    (hidden_size + block_k - 1) // block_k,
                    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 + block_n - 1) // block_n,
                    (intermediate_size_per_partition + block_k - 1) // block_k,
                    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.BLOCK.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)

        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

        if self.use_cutlass:
            assert self.weight_quant.strategy != QuantizationStrategy.BLOCK
            device = layer.w13_weight.device
            # ab_strides1 and c_strides2 are the same
            self.ab_strides1_c_strides2 = torch.full(
                (layer.local_num_experts,),
                layer.hidden_size,
                device=device,
                dtype=torch.int64,
            )
            self.ab_strides2 = torch.full(
                (layer.local_num_experts,),
                layer.intermediate_size_per_partition,
                device=device,
                dtype=torch.int64,
            )
            self.c_strides1 = torch.full(
                (layer.local_num_experts,),
                2 * layer.intermediate_size_per_partition,
                device=device,
                dtype=torch.int64,
            )

        if is_deep_gemm_e8m0_used() and self.block_quant:
            assert layer.weight_block_size is not None
            # Re-quantise the expert weights so their scales are UE8M0.
            block_sz = tuple(layer.weight_block_size)
            requant_weight_ue8m0_inplace(
                layer.w13_weight.data,
                layer.w13_weight_scale.data,
                block_sz,
            )
            requant_weight_ue8m0_inplace(
                layer.w2_weight.data,
                layer.w2_weight_scale.data,
                block_sz,
            )

            # Ensure column-major TMA alignment expected by DeepGEMM.
            if expert_weight_is_col_major(layer.w13_weight_scale):
                layer.w13_weight_scale = get_col_major_tma_aligned_tensor(
                    layer.w13_weight_scale
                )
            if expert_weight_is_col_major(layer.w2_weight_scale):
                layer.w2_weight_scale = get_col_major_tma_aligned_tensor(
                    layer.w2_weight_scale
                )

    def maybe_make_prepare_finalize(self) -> mk.FusedMoEPrepareAndFinalize | None:
        if self.use_marlin or self.rocm_aiter_moe_enabled:
            return None
        else:
            return super().maybe_make_prepare_finalize()

    def select_gemm_impl(
        self,
        prepare_finalize: mk.FusedMoEPrepareAndFinalize,
        layer: torch.nn.Module,
    ) -> FusedMoEPermuteExpertsUnpermute:
        # cutlass path
        assert self.moe_quant_config is not None
        if self.use_cutlass:
            from vllm.model_executor.layers.fused_moe import (
                CutlassBatchedExpertsFp8,
                CutlassExpertsFp8,
            )

            experts: FusedMoEPermuteExpertsUnpermute

            num_dispatchers = prepare_finalize.num_dispatchers()

            if (
                prepare_finalize.activation_format
                == FusedMoEActivationFormat.BatchedExperts
            ):
                logger.debug("CutlassBatchedExpertsFp8(%s)", self.__class__.__name__)
                experts = CutlassBatchedExpertsFp8(
                    self.moe.num_local_experts,
                    num_dispatchers,
                    self.moe.in_dtype,
                    ab_strides1=self.ab_strides1_c_strides2,
                    ab_strides2=self.ab_strides2,
                    c_strides1=self.c_strides1,
                    c_strides2=self.ab_strides1_c_strides2,
                    quant_config=self.moe_quant_config,
                )
            else:
                logger.debug("CutlassExpertsFp8(%s)", self.__class__.__name__)
                experts = CutlassExpertsFp8(
                    self.moe.in_dtype,
                    ab_strides1=self.ab_strides1_c_strides2,
                    ab_strides2=self.ab_strides2,
                    c_strides1=self.c_strides1,
                    c_strides2=self.ab_strides1_c_strides2,
                    quant_config=self.moe_quant_config,
                )

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

            return experts

        # triton path
        from vllm.model_executor.layers.fused_moe.batched_triton_or_deep_gemm_moe import (  # noqa: E501
            BatchedTritonOrDeepGemmExperts,
        )
        from vllm.model_executor.layers.fused_moe.triton_deep_gemm_moe import (
            TritonOrDeepGemmExperts,
        )

        assert not self.rocm_aiter_moe_enabled and not self.use_marlin

        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

            logger.debug("BatchedTritonExperts(%s)", self.__class__.__name__)
            return BatchedTritonOrDeepGemmExperts(
                max_num_tokens=max_num_tokens_per_rank,
                num_dispatchers=prepare_finalize.num_dispatchers(),
                quant_config=self.moe_quant_config,
            )
        else:
            logger.debug("TritonOrDeepGemmExperts(%s)", self.__class__.__name__)
            return TritonOrDeepGemmExperts(self.moe_quant_config, allow_deep_gemm=True)

    def get_fused_moe_quant_config(
        self, layer: torch.nn.Module
    ) -> FusedMoEQuantConfig | None:
        if self.use_marlin:
            return None

        per_act_token = self.input_quant.strategy == QuantizationStrategy.TOKEN
        per_channel_quant = self.weight_quant.strategy == QuantizationStrategy.CHANNEL

        return fp8_w8a8_moe_quant_config(
            w1_scale=layer.w13_weight_scale,
            w2_scale=layer.w2_weight_scale,
            a1_scale=layer.w13_input_scale,
            a2_scale=layer.w2_input_scale,
            per_act_token_quant=per_act_token,
            per_out_ch_quant=per_channel_quant,
            block_shape=layer.weight_block_size,
        )

    def apply(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        router_logits: torch.Tensor,
        top_k: int,
        renormalize: bool,
        use_grouped_topk: bool = False,
        topk_group: int | None = None,
        num_expert_group: int | None = None,
        global_num_experts: int = -1,
        expert_map: torch.Tensor | None = None,
        custom_routing_function: Callable | None = None,
        scoring_func: str = "softmax",
        routed_scaling_factor: float = 1.0,
        e_score_correction_bias: torch.Tensor | None = None,
        apply_router_weight_on_input: bool = False,
        activation: str = "silu",
        enable_eplb: bool = False,
        expert_load_view: torch.Tensor | None = None,
        logical_to_physical_map: torch.Tensor | None = None,
        logical_replica_count: torch.Tensor | None = None,
    ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
        if enable_eplb:
            raise NotImplementedError(
                "EPLB not supported for `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,
            routed_scaling_factor=routed_scaling_factor,
            e_score_correction_bias=e_score_correction_bias,
            indices_type=self.topk_indices_dtype,
        )

        per_act_token = self.input_quant.strategy == QuantizationStrategy.TOKEN
        per_channel_quant = self.weight_quant.strategy == QuantizationStrategy.CHANNEL

        #
        # Note: the order here is important. self.fused_experts can override
        # cutlass fp8 or fused_experts but not marlin or rocm.
        #
        if self.use_marlin:
            assert activation == "silu", f"{activation} not supported for Marlin MoE."
            assert self.fused_experts is None
            return fused_marlin_moe(
                x,
                layer.w13_weight,
                layer.w2_weight,
                None,
                None,
                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,
                workspace=layer.workspace,
            )

        elif self.rocm_aiter_moe_enabled:
            from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (  # noqa E501
                rocm_aiter_fused_experts,
            )

            assert per_act_token == per_channel_quant
            assert self.moe_quant_config is not None
            assert self.fused_experts is None
            return rocm_aiter_fused_experts(
                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,
                expert_map=expert_map,
                quant_config=self.moe_quant_config,
            )

        elif self.fused_experts is not None:
            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,
            )

        # cutlass path
        elif self.use_cutlass:
            assert self.moe_quant_config is not None

            # small-batch fallback on SM100
            if self.is_fp8_w8a8_sm100 and topk_ids.shape[0] <= 8:
                from vllm.model_executor.layers.fused_moe import fused_experts

                assert per_act_token == per_channel_quant
                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,
                    global_num_experts=global_num_experts,
                    expert_map=None if self.disable_expert_map else expert_map,
                    quant_config=self.moe_quant_config,
                )
            else:
                from vllm.model_executor.layers.fused_moe.cutlass_moe import (
                    cutlass_moe_fp8,
                )

                assert per_act_token == per_channel_quant
                assert self.moe_quant_config is not None
                return cutlass_moe_fp8(
                    x,
                    layer.w13_weight,
                    layer.w2_weight,
                    topk_weights,
                    topk_ids,
                    quant_config=self.moe_quant_config,
                    activation=activation,
                    global_num_experts=global_num_experts,
                    expert_map=None if self.disable_expert_map else expert_map,
                    ab_strides1=self.ab_strides1_c_strides2,
                    ab_strides2=self.ab_strides2,
                    c_strides1=self.c_strides1,
                    c_strides2=self.ab_strides1_c_strides2,
                )

        else:
            from vllm.model_executor.layers.fused_moe import fused_experts

            assert per_act_token == per_channel_quant
            assert self.moe_quant_config is not None
            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,
                global_num_experts=global_num_experts,
                expert_map=expert_map,
                quant_config=self.moe_quant_config,
            )

block_quant instance-attribute

block_quant = weight_block_size is not None

disable_expert_map instance-attribute

disable_expert_map = False

input_quant instance-attribute

input_quant = get('input_activations')

is_fp8_w8a8_sm100 instance-attribute

is_fp8_w8a8_sm100 = _is_fp8_w8a8_sm100(
    weight_quant, input_quant
)

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_cutlass instance-attribute

use_cutlass = not block_quant and (
    _is_fp8_w8a8_sm90(weight_quant, input_quant)
    or is_fp8_w8a8_sm100
)

use_marlin instance-attribute

use_marlin = (
    not has_device_capability(89)
    or VLLM_TEST_FORCE_FP8_MARLIN
    and not block_quant
)

weight_block_size instance-attribute

weight_block_size = block_structure

weight_quant instance-attribute

weight_quant = get('weights')

__init__

__init__(
    quant_config: CompressedTensorsConfig,
    moe: FusedMoEConfig,
)
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def __init__(
    self,
    quant_config: "CompressedTensorsConfig",  # type: ignore # noqa E501
    moe: FusedMoEConfig,
):
    super().__init__(moe)
    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):
        assert self.weight_quant.strategy == QuantizationStrategy.BLOCK
        self.weight_block_size = self.weight_quant.block_structure
        assert self.weight_quant.dynamic is not None
    else:
        self.weight_block_size = None
    self.block_quant = self.weight_block_size is not None

    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
        and not self.block_quant
    )
    # 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()

    # cutlass path
    self.is_fp8_w8a8_sm100 = quant_config._is_fp8_w8a8_sm100(
        self.weight_quant, self.input_quant
    )
    self.use_cutlass = not self.block_quant and (
        quant_config._is_fp8_w8a8_sm90(self.weight_quant, self.input_quant)
        or self.is_fp8_w8a8_sm100
    )
    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: int | None = None,
    num_expert_group: int | None = None,
    global_num_experts: int = -1,
    expert_map: Tensor | None = None,
    custom_routing_function: Callable | None = None,
    scoring_func: str = "softmax",
    routed_scaling_factor: float = 1.0,
    e_score_correction_bias: Tensor | None = None,
    apply_router_weight_on_input: bool = False,
    activation: str = "silu",
    enable_eplb: bool = False,
    expert_load_view: Tensor | None = None,
    logical_to_physical_map: Tensor | None = None,
    logical_replica_count: Tensor | None = None,
) -> Tensor | tuple[Tensor, Tensor]
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def apply(
    self,
    layer: torch.nn.Module,
    x: torch.Tensor,
    router_logits: torch.Tensor,
    top_k: int,
    renormalize: bool,
    use_grouped_topk: bool = False,
    topk_group: int | None = None,
    num_expert_group: int | None = None,
    global_num_experts: int = -1,
    expert_map: torch.Tensor | None = None,
    custom_routing_function: Callable | None = None,
    scoring_func: str = "softmax",
    routed_scaling_factor: float = 1.0,
    e_score_correction_bias: torch.Tensor | None = None,
    apply_router_weight_on_input: bool = False,
    activation: str = "silu",
    enable_eplb: bool = False,
    expert_load_view: torch.Tensor | None = None,
    logical_to_physical_map: torch.Tensor | None = None,
    logical_replica_count: torch.Tensor | None = None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
    if enable_eplb:
        raise NotImplementedError(
            "EPLB not supported for `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,
        routed_scaling_factor=routed_scaling_factor,
        e_score_correction_bias=e_score_correction_bias,
        indices_type=self.topk_indices_dtype,
    )

    per_act_token = self.input_quant.strategy == QuantizationStrategy.TOKEN
    per_channel_quant = self.weight_quant.strategy == QuantizationStrategy.CHANNEL

    #
    # Note: the order here is important. self.fused_experts can override
    # cutlass fp8 or fused_experts but not marlin or rocm.
    #
    if self.use_marlin:
        assert activation == "silu", f"{activation} not supported for Marlin MoE."
        assert self.fused_experts is None
        return fused_marlin_moe(
            x,
            layer.w13_weight,
            layer.w2_weight,
            None,
            None,
            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,
            workspace=layer.workspace,
        )

    elif self.rocm_aiter_moe_enabled:
        from vllm.model_executor.layers.fused_moe.rocm_aiter_fused_moe import (  # noqa E501
            rocm_aiter_fused_experts,
        )

        assert per_act_token == per_channel_quant
        assert self.moe_quant_config is not None
        assert self.fused_experts is None
        return rocm_aiter_fused_experts(
            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,
            expert_map=expert_map,
            quant_config=self.moe_quant_config,
        )

    elif self.fused_experts is not None:
        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,
        )

    # cutlass path
    elif self.use_cutlass:
        assert self.moe_quant_config is not None

        # small-batch fallback on SM100
        if self.is_fp8_w8a8_sm100 and topk_ids.shape[0] <= 8:
            from vllm.model_executor.layers.fused_moe import fused_experts

            assert per_act_token == per_channel_quant
            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,
                global_num_experts=global_num_experts,
                expert_map=None if self.disable_expert_map else expert_map,
                quant_config=self.moe_quant_config,
            )
        else:
            from vllm.model_executor.layers.fused_moe.cutlass_moe import (
                cutlass_moe_fp8,
            )

            assert per_act_token == per_channel_quant
            assert self.moe_quant_config is not None
            return cutlass_moe_fp8(
                x,
                layer.w13_weight,
                layer.w2_weight,
                topk_weights,
                topk_ids,
                quant_config=self.moe_quant_config,
                activation=activation,
                global_num_experts=global_num_experts,
                expert_map=None if self.disable_expert_map else expert_map,
                ab_strides1=self.ab_strides1_c_strides2,
                ab_strides2=self.ab_strides2,
                c_strides1=self.c_strides1,
                c_strides2=self.ab_strides1_c_strides2,
            )

    else:
        from vllm.model_executor.layers.fused_moe import fused_experts

        assert per_act_token == per_channel_quant
        assert self.moe_quant_config is not None
        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,
            global_num_experts=global_num_experts,
            expert_map=expert_map,
            quant_config=self.moe_quant_config,
        )

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
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

    if self.block_quant:
        assert self.weight_block_size is not None
        layer.weight_block_size = self.weight_block_size
        tp_size = get_tensor_model_parallel_world_size()
        block_n, block_k = (
            self.weight_block_size[0],
            self.weight_block_size[1],
        )
        # NOTE: To ensure proper alignment of the block-wise quantization
        # scales, the output_size of the weights for both the gate and up
        # layers must be divisible by block_n.
        # Required by column parallel or enabling merged weights
        if intermediate_size_per_partition % block_n != 0:
            raise ValueError(
                f"The output_size of gate's and up's weight = "
                f"{intermediate_size_per_partition} is not divisible by "
                f"weight quantization block_n = {block_n}."
            )
        if tp_size > 1 and intermediate_size_per_partition % block_k != 0:
            # Required by row parallel
            raise ValueError(
                f"The input_size of down's weight = "
                f"{intermediate_size_per_partition} is not divisible by "
                f"weight quantization block_k = {block_k}."
            )

    # 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)

    elif self.weight_quant.strategy == QuantizationStrategy.BLOCK:
        w13_weight_scale = torch.nn.Parameter(
            torch.ones(
                num_experts,
                2 * ((intermediate_size_per_partition + block_n - 1) // block_n),
                (hidden_size + block_k - 1) // block_k,
                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 + block_n - 1) // block_n,
                (intermediate_size_per_partition + block_k - 1) // block_k,
                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.BLOCK.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

get_fused_moe_quant_config

get_fused_moe_quant_config(
    layer: Module,
) -> FusedMoEQuantConfig | None
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def get_fused_moe_quant_config(
    self, layer: torch.nn.Module
) -> FusedMoEQuantConfig | None:
    if self.use_marlin:
        return None

    per_act_token = self.input_quant.strategy == QuantizationStrategy.TOKEN
    per_channel_quant = self.weight_quant.strategy == QuantizationStrategy.CHANNEL

    return fp8_w8a8_moe_quant_config(
        w1_scale=layer.w13_weight_scale,
        w2_scale=layer.w2_weight_scale,
        a1_scale=layer.w13_input_scale,
        a2_scale=layer.w2_input_scale,
        per_act_token_quant=per_act_token,
        per_out_ch_quant=per_channel_quant,
        block_shape=layer.weight_block_size,
    )

maybe_make_prepare_finalize

maybe_make_prepare_finalize() -> (
    FusedMoEPrepareAndFinalize | None
)
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def maybe_make_prepare_finalize(self) -> mk.FusedMoEPrepareAndFinalize | None:
    if self.use_marlin or self.rocm_aiter_moe_enabled:
        return None
    else:
        return super().maybe_make_prepare_finalize()

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
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)

    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

    if self.use_cutlass:
        assert self.weight_quant.strategy != QuantizationStrategy.BLOCK
        device = layer.w13_weight.device
        # ab_strides1 and c_strides2 are the same
        self.ab_strides1_c_strides2 = torch.full(
            (layer.local_num_experts,),
            layer.hidden_size,
            device=device,
            dtype=torch.int64,
        )
        self.ab_strides2 = torch.full(
            (layer.local_num_experts,),
            layer.intermediate_size_per_partition,
            device=device,
            dtype=torch.int64,
        )
        self.c_strides1 = torch.full(
            (layer.local_num_experts,),
            2 * layer.intermediate_size_per_partition,
            device=device,
            dtype=torch.int64,
        )

    if is_deep_gemm_e8m0_used() and self.block_quant:
        assert layer.weight_block_size is not None
        # Re-quantise the expert weights so their scales are UE8M0.
        block_sz = tuple(layer.weight_block_size)
        requant_weight_ue8m0_inplace(
            layer.w13_weight.data,
            layer.w13_weight_scale.data,
            block_sz,
        )
        requant_weight_ue8m0_inplace(
            layer.w2_weight.data,
            layer.w2_weight_scale.data,
            block_sz,
        )

        # Ensure column-major TMA alignment expected by DeepGEMM.
        if expert_weight_is_col_major(layer.w13_weight_scale):
            layer.w13_weight_scale = get_col_major_tma_aligned_tensor(
                layer.w13_weight_scale
            )
        if expert_weight_is_col_major(layer.w2_weight_scale):
            layer.w2_weight_scale = get_col_major_tma_aligned_tensor(
                layer.w2_weight_scale
            )

select_gemm_impl

select_gemm_impl(
    prepare_finalize: FusedMoEPrepareAndFinalize,
    layer: Module,
) -> FusedMoEPermuteExpertsUnpermute
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def select_gemm_impl(
    self,
    prepare_finalize: mk.FusedMoEPrepareAndFinalize,
    layer: torch.nn.Module,
) -> FusedMoEPermuteExpertsUnpermute:
    # cutlass path
    assert self.moe_quant_config is not None
    if self.use_cutlass:
        from vllm.model_executor.layers.fused_moe import (
            CutlassBatchedExpertsFp8,
            CutlassExpertsFp8,
        )

        experts: FusedMoEPermuteExpertsUnpermute

        num_dispatchers = prepare_finalize.num_dispatchers()

        if (
            prepare_finalize.activation_format
            == FusedMoEActivationFormat.BatchedExperts
        ):
            logger.debug("CutlassBatchedExpertsFp8(%s)", self.__class__.__name__)
            experts = CutlassBatchedExpertsFp8(
                self.moe.num_local_experts,
                num_dispatchers,
                self.moe.in_dtype,
                ab_strides1=self.ab_strides1_c_strides2,
                ab_strides2=self.ab_strides2,
                c_strides1=self.c_strides1,
                c_strides2=self.ab_strides1_c_strides2,
                quant_config=self.moe_quant_config,
            )
        else:
            logger.debug("CutlassExpertsFp8(%s)", self.__class__.__name__)
            experts = CutlassExpertsFp8(
                self.moe.in_dtype,
                ab_strides1=self.ab_strides1_c_strides2,
                ab_strides2=self.ab_strides2,
                c_strides1=self.c_strides1,
                c_strides2=self.ab_strides1_c_strides2,
                quant_config=self.moe_quant_config,
            )

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

        return experts

    # triton path
    from vllm.model_executor.layers.fused_moe.batched_triton_or_deep_gemm_moe import (  # noqa: E501
        BatchedTritonOrDeepGemmExperts,
    )
    from vllm.model_executor.layers.fused_moe.triton_deep_gemm_moe import (
        TritonOrDeepGemmExperts,
    )

    assert not self.rocm_aiter_moe_enabled and not self.use_marlin

    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

        logger.debug("BatchedTritonExperts(%s)", self.__class__.__name__)
        return BatchedTritonOrDeepGemmExperts(
            max_num_tokens=max_num_tokens_per_rank,
            num_dispatchers=prepare_finalize.num_dispatchers(),
            quant_config=self.moe_quant_config,
        )
    else:
        logger.debug("TritonOrDeepGemmExperts(%s)", self.__class__.__name__)
        return TritonOrDeepGemmExperts(self.moe_quant_config, allow_deep_gemm=True)

CompressedTensorsW8A8Int8MoEMethod

Bases: CompressedTensorsMoEMethod

Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
class CompressedTensorsW8A8Int8MoEMethod(CompressedTensorsMoEMethod):
    def __init__(
        self,
        quant_config: "CompressedTensorsConfig",  # type: ignore # noqa E501
        moe: FusedMoEConfig,
    ):
        super().__init__(moe)
        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 get_fused_moe_quant_config(
        self, layer: torch.nn.Module
    ) -> FusedMoEQuantConfig | None:
        return int8_w8a8_moe_quant_config(
            w1_scale=layer.w13_weight_scale,
            w2_scale=layer.w2_weight_scale,
            a1_scale=layer.w13_input_scale,
            a2_scale=layer.w2_input_scale,
            per_act_token_quant=True,
        )

    def apply(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        router_logits: torch.Tensor,
        top_k: int,
        renormalize: bool,
        use_grouped_topk: bool = False,
        topk_group: int | None = None,
        num_expert_group: int | None = None,
        global_num_experts: int = -1,
        expert_map: torch.Tensor | None = None,
        custom_routing_function: Callable | None = None,
        scoring_func: str = "softmax",
        routed_scaling_factor: float = 1.0,
        e_score_correction_bias: torch.Tensor | None = None,
        apply_router_weight_on_input: bool = False,
        activation: str = "silu",
        enable_eplb: bool = False,
        expert_load_view: torch.Tensor | None = None,
        logical_to_physical_map: torch.Tensor | None = None,
        logical_replica_count: torch.Tensor | None = None,
    ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
        assert self.fused_experts is None

        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,
            routed_scaling_factor=routed_scaling_factor,
            e_score_correction_bias=e_score_correction_bias,
            indices_type=self.topk_indices_dtype,
        )

        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,
            global_num_experts=global_num_experts,
            expert_map=expert_map,
            quant_config=self.moe_quant_config,
        )

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,
    moe: FusedMoEConfig,
)
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def __init__(
    self,
    quant_config: "CompressedTensorsConfig",  # type: ignore # noqa E501
    moe: FusedMoEConfig,
):
    super().__init__(moe)
    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: int | None = None,
    num_expert_group: int | None = None,
    global_num_experts: int = -1,
    expert_map: Tensor | None = None,
    custom_routing_function: Callable | None = None,
    scoring_func: str = "softmax",
    routed_scaling_factor: float = 1.0,
    e_score_correction_bias: Tensor | None = None,
    apply_router_weight_on_input: bool = False,
    activation: str = "silu",
    enable_eplb: bool = False,
    expert_load_view: Tensor | None = None,
    logical_to_physical_map: Tensor | None = None,
    logical_replica_count: Tensor | None = None,
) -> Tensor | tuple[Tensor, Tensor]
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def apply(
    self,
    layer: torch.nn.Module,
    x: torch.Tensor,
    router_logits: torch.Tensor,
    top_k: int,
    renormalize: bool,
    use_grouped_topk: bool = False,
    topk_group: int | None = None,
    num_expert_group: int | None = None,
    global_num_experts: int = -1,
    expert_map: torch.Tensor | None = None,
    custom_routing_function: Callable | None = None,
    scoring_func: str = "softmax",
    routed_scaling_factor: float = 1.0,
    e_score_correction_bias: torch.Tensor | None = None,
    apply_router_weight_on_input: bool = False,
    activation: str = "silu",
    enable_eplb: bool = False,
    expert_load_view: torch.Tensor | None = None,
    logical_to_physical_map: torch.Tensor | None = None,
    logical_replica_count: torch.Tensor | None = None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
    assert self.fused_experts is None

    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,
        routed_scaling_factor=routed_scaling_factor,
        e_score_correction_bias=e_score_correction_bias,
        indices_type=self.topk_indices_dtype,
    )

    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,
        global_num_experts=global_num_experts,
        expert_map=expert_map,
        quant_config=self.moe_quant_config,
    )

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
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

get_fused_moe_quant_config

get_fused_moe_quant_config(
    layer: Module,
) -> FusedMoEQuantConfig | None
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def get_fused_moe_quant_config(
    self, layer: torch.nn.Module
) -> FusedMoEQuantConfig | None:
    return int8_w8a8_moe_quant_config(
        w1_scale=layer.w13_weight_scale,
        w2_scale=layer.w2_weight_scale,
        a1_scale=layer.w13_input_scale,
        a2_scale=layer.w2_input_scale,
        per_act_token_quant=True,
    )

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
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
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
1391
1392
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
1612
1613
1614
1615
1616
1617
1618
1619
1620
1621
1622
1623
1624
1625
1626
1627
1628
1629
1630
1631
1632
1633
1634
1635
1636
1637
1638
1639
1640
1641
1642
1643
1644
1645
1646
1647
1648
1649
1650
1651
1652
1653
1654
1655
1656
1657
1658
1659
1660
1661
1662
1663
1664
1665
1666
1667
1668
1669
1670
1671
1672
1673
1674
1675
1676
1677
1678
1679
class CompressedTensorsWNA16MarlinMoEMethod(CompressedTensorsMoEMethod):
    def __init__(
        self,
        quant_config: "CompressedTensorsConfig",  # type: ignore # noqa E501
        moe: FusedMoEConfig,
    ):
        super().__init__(moe)
        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 get_fused_moe_quant_config(
        self, layer: torch.nn.Module
    ) -> FusedMoEQuantConfig | None:
        return None

    def apply(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        router_logits: torch.Tensor,
        top_k: int,
        renormalize: bool,
        use_grouped_topk: bool = False,
        topk_group: int | None = None,
        num_expert_group: int | None = None,
        global_num_experts: int = -1,
        expert_map: torch.Tensor | None = None,
        custom_routing_function: Callable | None = None,
        scoring_func: str = "softmax",
        routed_scaling_factor: float = 1.0,
        e_score_correction_bias: torch.Tensor | None = None,
        apply_router_weight_on_input: bool = False,
        activation: str = "silu",
        enable_eplb: bool = False,
        expert_load_view: torch.Tensor | None = None,
        logical_to_physical_map: torch.Tensor | None = None,
        logical_replica_count: torch.Tensor | None = None,
    ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
        assert self.fused_experts is None

        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,
            routed_scaling_factor=routed_scaling_factor,
            e_score_correction_bias=e_score_correction_bias,
            indices_type=self.topk_indices_dtype,
        )

        return fused_marlin_moe(
            x,
            layer.w13_weight_packed,
            layer.w2_weight_packed,
            None,
            None,
            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,
    moe: FusedMoEConfig,
)
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def __init__(
    self,
    quant_config: "CompressedTensorsConfig",  # type: ignore # noqa E501
    moe: FusedMoEConfig,
):
    super().__init__(moe)
    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: int | None = None,
    num_expert_group: int | None = None,
    global_num_experts: int = -1,
    expert_map: Tensor | None = None,
    custom_routing_function: Callable | None = None,
    scoring_func: str = "softmax",
    routed_scaling_factor: float = 1.0,
    e_score_correction_bias: Tensor | None = None,
    apply_router_weight_on_input: bool = False,
    activation: str = "silu",
    enable_eplb: bool = False,
    expert_load_view: Tensor | None = None,
    logical_to_physical_map: Tensor | None = None,
    logical_replica_count: Tensor | None = None,
) -> Tensor | tuple[Tensor, Tensor]
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def apply(
    self,
    layer: torch.nn.Module,
    x: torch.Tensor,
    router_logits: torch.Tensor,
    top_k: int,
    renormalize: bool,
    use_grouped_topk: bool = False,
    topk_group: int | None = None,
    num_expert_group: int | None = None,
    global_num_experts: int = -1,
    expert_map: torch.Tensor | None = None,
    custom_routing_function: Callable | None = None,
    scoring_func: str = "softmax",
    routed_scaling_factor: float = 1.0,
    e_score_correction_bias: torch.Tensor | None = None,
    apply_router_weight_on_input: bool = False,
    activation: str = "silu",
    enable_eplb: bool = False,
    expert_load_view: torch.Tensor | None = None,
    logical_to_physical_map: torch.Tensor | None = None,
    logical_replica_count: torch.Tensor | None = None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
    assert self.fused_experts is None

    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,
        routed_scaling_factor=routed_scaling_factor,
        e_score_correction_bias=e_score_correction_bias,
        indices_type=self.topk_indices_dtype,
    )

    return fused_marlin_moe(
        x,
        layer.w13_weight_packed,
        layer.w2_weight_packed,
        None,
        None,
        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
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

get_fused_moe_quant_config

get_fused_moe_quant_config(
    layer: Module,
) -> FusedMoEQuantConfig | None
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def get_fused_moe_quant_config(
    self, layer: torch.nn.Module
) -> FusedMoEQuantConfig | None:
    return 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
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
class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
    def __init__(
        self,
        quant_config: "CompressedTensorsConfig",  # type: ignore # noqa E501
        moe: FusedMoEConfig,
    ):
        super().__init__(moe)
        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 get_fused_moe_quant_config(
        self, layer: torch.nn.Module
    ) -> FusedMoEQuantConfig | None:
        assert self.num_bits == 4 or self.num_bits == 8
        config_builder = (
            int4_w4a16_moe_quant_config
            if self.num_bits == 4
            else int8_w8a16_moe_quant_config
        )

        return config_builder(
            w1_scale=layer.w13_weight_scale,
            w2_scale=layer.w2_weight_scale,
            w1_zp=None,
            w2_zp=None,
            block_shape=[0, self.group_size],
        )

    def apply(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        router_logits: torch.Tensor,
        top_k: int,
        renormalize: bool,
        use_grouped_topk: bool = False,
        topk_group: int | None = None,
        num_expert_group: int | None = None,
        global_num_experts: int = -1,
        expert_map: torch.Tensor | None = None,
        custom_routing_function: Callable | None = None,
        scoring_func: str = "softmax",
        routed_scaling_factor: float = 1.0,
        e_score_correction_bias: torch.Tensor | None = None,
        apply_router_weight_on_input: bool = False,
        activation: str = "silu",
        enable_eplb: bool = False,
        expert_load_view: torch.Tensor | None = None,
        logical_to_physical_map: torch.Tensor | None = None,
        logical_replica_count: torch.Tensor | None = None,
    ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
        assert self.fused_experts is None

        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,
            routed_scaling_factor=routed_scaling_factor,
            e_score_correction_bias=e_score_correction_bias,
            indices_type=self.topk_indices_dtype,
        )

        return fused_experts(
            x,
            layer.w13_weight_packed,
            layer.w2_weight_packed,
            topk_weights=topk_weights,
            topk_ids=topk_ids,
            inplace=True,
            activation=activation,
            apply_router_weight_on_input=apply_router_weight_on_input,
            global_num_experts=global_num_experts,
            expert_map=expert_map,
            quant_config=self.moe_quant_config,
        )

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,
    moe: FusedMoEConfig,
)
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def __init__(
    self,
    quant_config: "CompressedTensorsConfig",  # type: ignore # noqa E501
    moe: FusedMoEConfig,
):
    super().__init__(moe)
    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: int | None = None,
    num_expert_group: int | None = None,
    global_num_experts: int = -1,
    expert_map: Tensor | None = None,
    custom_routing_function: Callable | None = None,
    scoring_func: str = "softmax",
    routed_scaling_factor: float = 1.0,
    e_score_correction_bias: Tensor | None = None,
    apply_router_weight_on_input: bool = False,
    activation: str = "silu",
    enable_eplb: bool = False,
    expert_load_view: Tensor | None = None,
    logical_to_physical_map: Tensor | None = None,
    logical_replica_count: Tensor | None = None,
) -> Tensor | tuple[Tensor, Tensor]
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def apply(
    self,
    layer: torch.nn.Module,
    x: torch.Tensor,
    router_logits: torch.Tensor,
    top_k: int,
    renormalize: bool,
    use_grouped_topk: bool = False,
    topk_group: int | None = None,
    num_expert_group: int | None = None,
    global_num_experts: int = -1,
    expert_map: torch.Tensor | None = None,
    custom_routing_function: Callable | None = None,
    scoring_func: str = "softmax",
    routed_scaling_factor: float = 1.0,
    e_score_correction_bias: torch.Tensor | None = None,
    apply_router_weight_on_input: bool = False,
    activation: str = "silu",
    enable_eplb: bool = False,
    expert_load_view: torch.Tensor | None = None,
    logical_to_physical_map: torch.Tensor | None = None,
    logical_replica_count: torch.Tensor | None = None,
) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
    assert self.fused_experts is None

    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,
        routed_scaling_factor=routed_scaling_factor,
        e_score_correction_bias=e_score_correction_bias,
        indices_type=self.topk_indices_dtype,
    )

    return fused_experts(
        x,
        layer.w13_weight_packed,
        layer.w2_weight_packed,
        topk_weights=topk_weights,
        topk_ids=topk_ids,
        inplace=True,
        activation=activation,
        apply_router_weight_on_input=apply_router_weight_on_input,
        global_num_experts=global_num_experts,
        expert_map=expert_map,
        quant_config=self.moe_quant_config,
    )

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
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

get_fused_moe_quant_config

get_fused_moe_quant_config(
    layer: Module,
) -> FusedMoEQuantConfig | None
Source code in vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_moe.py
def get_fused_moe_quant_config(
    self, layer: torch.nn.Module
) -> FusedMoEQuantConfig | None:
    assert self.num_bits == 4 or self.num_bits == 8
    config_builder = (
        int4_w4a16_moe_quant_config
        if self.num_bits == 4
        else int8_w8a16_moe_quant_config
    )

    return config_builder(
        w1_scale=layer.w13_weight_scale,
        w2_scale=layer.w2_weight_scale,
        w1_zp=None,
        w2_zp=None,
        block_shape=[0, self.group_size],
    )

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
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
class GPTQMarlinState(Enum):
    REPACK = enum.auto()
    READY = enum.auto()

READY class-attribute instance-attribute

READY = auto()

REPACK class-attribute instance-attribute

REPACK = auto()