Skip to content

vllm.model_executor.layers.fused_moe.cpu_fused_moe

CPUFusedMOE

Source code in vllm/model_executor/layers/fused_moe/cpu_fused_moe.py 
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
class CPUFusedMOE:

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

    def __call__(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        use_grouped_topk: bool,
        top_k: int,
        router_logits: torch.Tensor,
        renormalize: bool,
        topk_group: Optional[int] = None,
        num_expert_group: Optional[int] = None,
        global_num_experts: int = -1,
        expert_map: Optional[torch.Tensor] = None,
        custom_routing_function: Optional[Callable] = None,
        scoring_func: str = "softmax",
        e_score_correction_bias: Optional[torch.Tensor] = None,
        apply_router_weight_on_input: bool = False,
        activation: str = "silu",
    ) -> torch.Tensor:
        assert activation == "silu", f"{activation} is not supported."
        assert not apply_router_weight_on_input
        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,
            e_score_correction_bias=e_score_correction_bias,
        )

        # Ref code from https://github.com/sgl-project/sglang/blob/716e682721397df103f347d22da8bd46c6016dab/python/sglang/srt/layers/moe/fused_moe_native.py#L53
        len_experts = global_num_experts

        cnts = topk_ids.new_zeros((topk_ids.shape[0], len_experts))
        cnts.scatter_(1, topk_ids.to(torch.int64), 1)
        tokens_per_expert = cnts.sum(dim=0)
        idxs = topk_ids.view(-1).argsort()

        sorted_tokens = x[idxs // topk_ids.shape[1]]
        tokens_per_expert = tokens_per_expert.cpu().numpy()

        outputs = []
        start_idx = 0
        for i, num_tokens in enumerate(tokens_per_expert):
            end_idx = start_idx + num_tokens
            if num_tokens == 0:
                continue
            tokens_for_this_expert = sorted_tokens[start_idx:end_idx]

            layer_w13_weight = layer.w13_weight[i]
            layer_w2_weight = layer.w2_weight[i]

            gate_up = F.linear(tokens_for_this_expert, layer_w13_weight)
            gate_up = silu_and_mul(gate_up)
            expert_out = F.linear(gate_up, layer_w2_weight)
            outputs.append(expert_out)
            start_idx = end_idx

        outs = torch.cat(outputs,
                         dim=0) if len(outputs) else sorted_tokens.new_empty(0)
        new_x = torch.empty_like(outs)

        new_x[idxs] = outs
        final_out = (new_x.view(
            *topk_ids.shape, -1).type(topk_weights.dtype).mul_(
                topk_weights.unsqueeze(dim=-1)).sum(dim=1).type(new_x.dtype))
        return final_out

__call__ 

__call__(
    layer: Module,
    x: Tensor,
    use_grouped_topk: bool,
    top_k: int,
    router_logits: Tensor,
    renormalize: bool,
    topk_group: Optional[int] = None,
    num_expert_group: Optional[int] = None,
    global_num_experts: int = -1,
    expert_map: Optional[Tensor] = None,
    custom_routing_function: Optional[Callable] = None,
    scoring_func: str = "softmax",
    e_score_correction_bias: Optional[Tensor] = None,
    apply_router_weight_on_input: bool = False,
    activation: str = "silu",
) -> Tensor
Source code in vllm/model_executor/layers/fused_moe/cpu_fused_moe.py 
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
def __call__(
    self,
    layer: torch.nn.Module,
    x: torch.Tensor,
    use_grouped_topk: bool,
    top_k: int,
    router_logits: torch.Tensor,
    renormalize: bool,
    topk_group: Optional[int] = None,
    num_expert_group: Optional[int] = None,
    global_num_experts: int = -1,
    expert_map: Optional[torch.Tensor] = None,
    custom_routing_function: Optional[Callable] = None,
    scoring_func: str = "softmax",
    e_score_correction_bias: Optional[torch.Tensor] = None,
    apply_router_weight_on_input: bool = False,
    activation: str = "silu",
) -> torch.Tensor:
    assert activation == "silu", f"{activation} is not supported."
    assert not apply_router_weight_on_input
    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,
        e_score_correction_bias=e_score_correction_bias,
    )

    # Ref code from https://github.com/sgl-project/sglang/blob/716e682721397df103f347d22da8bd46c6016dab/python/sglang/srt/layers/moe/fused_moe_native.py#L53
    len_experts = global_num_experts

    cnts = topk_ids.new_zeros((topk_ids.shape[0], len_experts))
    cnts.scatter_(1, topk_ids.to(torch.int64), 1)
    tokens_per_expert = cnts.sum(dim=0)
    idxs = topk_ids.view(-1).argsort()

    sorted_tokens = x[idxs // topk_ids.shape[1]]
    tokens_per_expert = tokens_per_expert.cpu().numpy()

    outputs = []
    start_idx = 0
    for i, num_tokens in enumerate(tokens_per_expert):
        end_idx = start_idx + num_tokens
        if num_tokens == 0:
            continue
        tokens_for_this_expert = sorted_tokens[start_idx:end_idx]

        layer_w13_weight = layer.w13_weight[i]
        layer_w2_weight = layer.w2_weight[i]

        gate_up = F.linear(tokens_for_this_expert, layer_w13_weight)
        gate_up = silu_and_mul(gate_up)
        expert_out = F.linear(gate_up, layer_w2_weight)
        outputs.append(expert_out)
        start_idx = end_idx

    outs = torch.cat(outputs,
                     dim=0) if len(outputs) else sorted_tokens.new_empty(0)
    new_x = torch.empty_like(outs)

    new_x[idxs] = outs
    final_out = (new_x.view(
        *topk_ids.shape, -1).type(topk_weights.dtype).mul_(
            topk_weights.unsqueeze(dim=-1)).sum(dim=1).type(new_x.dtype))
    return final_out

__init__ 

__init__(layer: Module) -> None
Source code in vllm/model_executor/layers/fused_moe/cpu_fused_moe.py 
213
214
def __init__(self, layer: torch.nn.Module) -> None:
    pass

IPEXFusedMOE

Source code in vllm/model_executor/layers/fused_moe/cpu_fused_moe.py 
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
class IPEXFusedMOE:

    def __init__(self, layer: torch.nn.Module) -> None:
        import intel_extension_for_pytorch as ipex
        layer.ipex_fusion = ipex.llm.modules.GatedMLPMOE(
            layer.w13_weight,
            layer.w2_weight,
            use_prepack=envs.VLLM_CPU_MOE_PREPACK,
        )

    def __call__(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        use_grouped_topk: bool,
        top_k: int,
        router_logits: torch.Tensor,
        renormalize: bool,
        topk_group: Optional[int] = None,
        num_expert_group: Optional[int] = None,
        global_num_experts: int = -1,
        expert_map: Optional[torch.Tensor] = None,
        custom_routing_function: Optional[Callable] = None,
        scoring_func: str = "softmax",
        e_score_correction_bias: Optional[torch.Tensor] = None,
        apply_router_weight_on_input: bool = False,
        activation: str = "silu",
    ) -> torch.Tensor:
        assert activation == "silu", f"{activation} is not supported."
        assert not apply_router_weight_on_input
        return layer.ipex_fusion(
            x,
            use_grouped_topk,
            top_k,
            router_logits,
            renormalize,
            topk_group,
            num_expert_group,
            custom_routing_function,
            scoring_func,
            e_score_correction_bias,
        )

__call__ 

__call__(
    layer: Module,
    x: Tensor,
    use_grouped_topk: bool,
    top_k: int,
    router_logits: Tensor,
    renormalize: bool,
    topk_group: Optional[int] = None,
    num_expert_group: Optional[int] = None,
    global_num_experts: int = -1,
    expert_map: Optional[Tensor] = None,
    custom_routing_function: Optional[Callable] = None,
    scoring_func: str = "softmax",
    e_score_correction_bias: Optional[Tensor] = None,
    apply_router_weight_on_input: bool = False,
    activation: str = "silu",
) -> Tensor
Source code in vllm/model_executor/layers/fused_moe/cpu_fused_moe.py 
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
def __call__(
    self,
    layer: torch.nn.Module,
    x: torch.Tensor,
    use_grouped_topk: bool,
    top_k: int,
    router_logits: torch.Tensor,
    renormalize: bool,
    topk_group: Optional[int] = None,
    num_expert_group: Optional[int] = None,
    global_num_experts: int = -1,
    expert_map: Optional[torch.Tensor] = None,
    custom_routing_function: Optional[Callable] = None,
    scoring_func: str = "softmax",
    e_score_correction_bias: Optional[torch.Tensor] = None,
    apply_router_weight_on_input: bool = False,
    activation: str = "silu",
) -> torch.Tensor:
    assert activation == "silu", f"{activation} is not supported."
    assert not apply_router_weight_on_input
    return layer.ipex_fusion(
        x,
        use_grouped_topk,
        top_k,
        router_logits,
        renormalize,
        topk_group,
        num_expert_group,
        custom_routing_function,
        scoring_func,
        e_score_correction_bias,
    )

__init__ 

__init__(layer: Module) -> None
Source code in vllm/model_executor/layers/fused_moe/cpu_fused_moe.py 
112
113
114
115
116
117
118
def __init__(self, layer: torch.nn.Module) -> None:
    import intel_extension_for_pytorch as ipex
    layer.ipex_fusion = ipex.llm.modules.GatedMLPMOE(
        layer.w13_weight,
        layer.w2_weight,
        use_prepack=envs.VLLM_CPU_MOE_PREPACK,
    )

SGLFusedMOE

Source code in vllm/model_executor/layers/fused_moe/cpu_fused_moe.py 
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
class SGLFusedMOE:

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

    def __call__(
        self,
        layer: torch.nn.Module,
        x: torch.Tensor,
        use_grouped_topk: bool,
        top_k: int,
        router_logits: torch.Tensor,
        renormalize: bool,
        topk_group: Optional[int] = None,
        num_expert_group: Optional[int] = None,
        global_num_experts: int = -1,
        expert_map: Optional[torch.Tensor] = None,
        custom_routing_function: Optional[Callable] = None,
        scoring_func: str = "softmax",
        e_score_correction_bias: Optional[torch.Tensor] = None,
        apply_router_weight_on_input: bool = False,
        activation: str = "silu",
    ) -> torch.Tensor:
        assert activation == "silu", f"{activation} is not supported."
        assert not apply_router_weight_on_input
        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,
            e_score_correction_bias=e_score_correction_bias,
        )

        torch.ops._C.fused_experts_cpu(
            x,
            layer.w13_weight,
            layer.w2_weight,
            topk_weights,
            topk_ids,
            True,
            False,
            False,
            None,
            None,
            None,
            None,
            None,
            True,
        )
        return x

__call__ 

__call__(
    layer: Module,
    x: Tensor,
    use_grouped_topk: bool,
    top_k: int,
    router_logits: Tensor,
    renormalize: bool,
    topk_group: Optional[int] = None,
    num_expert_group: Optional[int] = None,
    global_num_experts: int = -1,
    expert_map: Optional[Tensor] = None,
    custom_routing_function: Optional[Callable] = None,
    scoring_func: str = "softmax",
    e_score_correction_bias: Optional[Tensor] = None,
    apply_router_weight_on_input: bool = False,
    activation: str = "silu",
) -> Tensor
Source code in vllm/model_executor/layers/fused_moe/cpu_fused_moe.py 
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
def __call__(
    self,
    layer: torch.nn.Module,
    x: torch.Tensor,
    use_grouped_topk: bool,
    top_k: int,
    router_logits: torch.Tensor,
    renormalize: bool,
    topk_group: Optional[int] = None,
    num_expert_group: Optional[int] = None,
    global_num_experts: int = -1,
    expert_map: Optional[torch.Tensor] = None,
    custom_routing_function: Optional[Callable] = None,
    scoring_func: str = "softmax",
    e_score_correction_bias: Optional[torch.Tensor] = None,
    apply_router_weight_on_input: bool = False,
    activation: str = "silu",
) -> torch.Tensor:
    assert activation == "silu", f"{activation} is not supported."
    assert not apply_router_weight_on_input
    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,
        e_score_correction_bias=e_score_correction_bias,
    )

    torch.ops._C.fused_experts_cpu(
        x,
        layer.w13_weight,
        layer.w2_weight,
        topk_weights,
        topk_ids,
        True,
        False,
        False,
        None,
        None,
        None,
        None,
        None,
        True,
    )
    return x

__init__ 

__init__(layer: Module) -> None
Source code in vllm/model_executor/layers/fused_moe/cpu_fused_moe.py 
156
157
def __init__(self, layer: torch.nn.Module) -> None:
    pass

grouped_topk 

grouped_topk(
    hidden_states: Tensor,
    gating_output: Tensor,
    topk: int,
    renormalize: bool,
    num_expert_group: int = 0,
    topk_group: int = 0,
    scoring_func: str = "softmax",
    e_score_correction_bias: Optional[Tensor] = None,
) -> tuple[Tensor, Tensor]
Source code in vllm/model_executor/layers/fused_moe/cpu_fused_moe.py 
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
def grouped_topk(
    hidden_states: torch.Tensor,
    gating_output: torch.Tensor,
    topk: int,
    renormalize: bool,
    num_expert_group: int = 0,
    topk_group: int = 0,
    scoring_func: str = "softmax",
    e_score_correction_bias: Optional[torch.Tensor] = None
) -> tuple[torch.Tensor, torch.Tensor]:
    assert hidden_states.shape[0] == gating_output.shape[0], (
        "Number of tokens mismatch")

    gating_output = gating_output.float()
    if scoring_func == "softmax":
        scores = torch.softmax(gating_output, dim=-1)
    elif scoring_func == "sigmoid":
        scores = gating_output.sigmoid()
    else:
        raise ValueError(f"Unsupported scoring function: {scoring_func}")

    num_token = scores.shape[0]
    if e_score_correction_bias is not None:
        original_scores = scores
        scores = scores + e_score_correction_bias.unsqueeze(0)
        group_scores = (scores.view(num_token, num_expert_group,
                                    -1).topk(2, dim=-1)[0].sum(dim=-1))
    else:
        group_scores = scores.view(num_token, num_expert_group,
                                   -1).max(dim=-1).values  # [n, n_group]
    group_idx = torch.topk(group_scores, k=topk_group, dim=-1,
                           sorted=False)[1]  # [n, top_k_group]
    group_mask = torch.zeros_like(group_scores)  # [n, n_group]
    group_mask.scatter_(1, group_idx, 1)  # [n, n_group]
    score_mask = group_mask.unsqueeze(-1).expand(
        num_token, num_expert_group,
        scores.shape[-1] // num_expert_group).reshape(num_token, -1)  # [n, e]
    tmp_scores = scores.masked_fill(~score_mask.bool(),
                                    float("-inf"))  # [n, e]

    if e_score_correction_bias is not None:
        topk_ids = torch.topk(tmp_scores, k=topk, dim=-1, sorted=False)[1]
        topk_weights = original_scores.gather(1, topk_ids)
    else:
        topk_weights, topk_ids = torch.topk(tmp_scores,
                                            k=topk,
                                            dim=-1,
                                            sorted=False)

    if renormalize:
        topk_weights = topk_weights / topk_weights.sum(dim=-1, keepdim=True)

    return topk_weights, topk_ids.to(torch.int32)

select_experts 

select_experts(
    hidden_states: Tensor,
    router_logits: Tensor,
    top_k: int,
    use_grouped_topk: bool,
    renormalize: bool,
    topk_group: Optional[int] = None,
    num_expert_group: Optional[int] = None,
    custom_routing_function: Optional[Callable] = None,
    scoring_func: str = "softmax",
    e_score_correction_bias: Optional[Tensor] = None,
) -> tuple[Tensor, Tensor]
Source code in vllm/model_executor/layers/fused_moe/cpu_fused_moe.py 
 71
 72
 73
 74
 75
 76
 77
 78
 79
 80
 81
 82
 83
 84
 85
 86
 87
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
def select_experts(
    hidden_states: torch.Tensor,
    router_logits: torch.Tensor,
    top_k: int,
    use_grouped_topk: bool,
    renormalize: bool,
    topk_group: Optional[int] = None,
    num_expert_group: Optional[int] = None,
    custom_routing_function: Optional[Callable] = None,
    scoring_func: str = "softmax",
    e_score_correction_bias: Optional[torch.Tensor] = None,
) -> tuple[torch.Tensor, torch.Tensor]:
    if use_grouped_topk:
        assert topk_group is not None
        assert num_expert_group is not None
        return grouped_topk(hidden_states=hidden_states,
                            gating_output=router_logits,
                            topk=top_k,
                            renormalize=renormalize,
                            num_expert_group=num_expert_group,
                            topk_group=topk_group,
                            scoring_func=scoring_func,
                            e_score_correction_bias=e_score_correction_bias)
    elif custom_routing_function is None:
        assert scoring_func == "softmax"
        topk_weights = torch.nn.functional.softmax(router_logits,
                                                   dim=1,
                                                   dtype=torch.float32)
        topk_weights, topk_ids = torch.topk(topk_weights, top_k, dim=-1)
        if renormalize:
            topk_weights /= topk_weights.sum(dim=-1, keepdim=True)
        return topk_weights, topk_ids.to(torch.int32)
    else:
        return custom_routing_function(hidden_states=hidden_states,
                                       gating_output=router_logits,
                                       topk=top_k,
                                       renormalize=renormalize)

silu_and_mul 

silu_and_mul(x: Tensor) -> Tensor
Source code in vllm/model_executor/layers/fused_moe/cpu_fused_moe.py 
11
12
13
def silu_and_mul(x: torch.Tensor) -> torch.Tensor:
    d = x.shape[-1] // 2
    return F.silu(x[..., :d]) * x[..., d:]