|
29 | 29 |
|
30 | 30 | import torch |
31 | 31 | import torch.distributed as dist |
32 | | -import torch_npu # noqa: F401 |
| 32 | +import torch_npu |
33 | 33 | import vllm.envs as envs |
34 | 34 | from torch import nn |
35 | 35 | from transformers import PretrainedConfig |
|
40 | 40 | get_tp_group, tensor_model_parallel_all_reduce) |
41 | 41 | from vllm.distributed.parallel_state import get_dp_group |
42 | 42 | from vllm.forward_context import get_forward_context |
| 43 | +from vllm.model_executor.layers.activation import SiluAndMul |
43 | 44 | from vllm.model_executor.layers.layernorm import RMSNorm |
44 | 45 | from vllm.model_executor.layers.linear import (ColumnParallelLinear, |
| 46 | + MergedColumnParallelLinear, |
45 | 47 | ReplicatedLinear, |
46 | | - RowParallelLinear) |
| 48 | + RowParallelLinear, |
| 49 | + UnquantizedLinearMethod) |
47 | 50 | from vllm.model_executor.layers.logits_processor import LogitsProcessor |
48 | 51 | from vllm.model_executor.layers.quantization import QuantizationConfig |
49 | 52 | from vllm.model_executor.layers.rotary_embedding import get_rope |
|
64 | 67 |
|
65 | 68 | import vllm_ascend.envs as envs_ascend |
66 | 69 | from vllm_ascend.ascend_config import get_ascend_config |
67 | | -from vllm_ascend.models.deepseek_v2 import CustomDeepseekV2MLP |
68 | 70 | from vllm_ascend.multistream.base import MSEventKey |
69 | 71 | from vllm_ascend.multistream.context import ( |
70 | 72 | advance_step_multistream_layer_context, get_multistream_comm_context, |
|
76 | 78 | make_multistream_metadata_ds) |
77 | 79 | from vllm_ascend.multistream.ms_split import compute_split_seq_index |
78 | 80 | from vllm_ascend.ops.fused_moe import AscendFusedMoE |
| 81 | +from vllm_ascend.quantization.w8a8_dynamic import AscendW8A8DynamicLinearMethod |
79 | 82 | from vllm_ascend.utils import dispose_tensor |
80 | 83 |
|
81 | 84 | VLLM_ASCEND_ENABLE_DBO: bool = envs_ascend.VLLM_ASCEND_ENABLE_DBO |
82 | 85 | VLLM_ENABLE_MC2: bool = envs_ascend.VLLM_ENABLE_MC2 |
83 | 86 |
|
84 | 87 |
|
85 | | -class CustomDeepseekDBOMLP(CustomDeepseekV2MLP): |
| 88 | +class CustomDeepseekDBOMLP(nn.Module): |
| 89 | + |
| 90 | + def __init__( |
| 91 | + self, |
| 92 | + hidden_size: int, |
| 93 | + intermediate_size: int, |
| 94 | + hidden_act: str, |
| 95 | + quant_config: Optional[QuantizationConfig] = None, |
| 96 | + reduce_results: bool = True, |
| 97 | + prefix: str = "", |
| 98 | + ) -> None: |
| 99 | + super().__init__() |
| 100 | + self.gate_up_proj = MergedColumnParallelLinear( |
| 101 | + hidden_size, [intermediate_size] * 2, |
| 102 | + bias=False, |
| 103 | + quant_config=quant_config, |
| 104 | + prefix=f"{prefix}.gate_up_proj") |
| 105 | + self.down_proj = RowParallelLinear(intermediate_size, |
| 106 | + hidden_size, |
| 107 | + bias=False, |
| 108 | + quant_config=quant_config, |
| 109 | + reduce_results=reduce_results, |
| 110 | + prefix=f"{prefix}.down_proj") |
| 111 | + if hidden_act != "silu": |
| 112 | + raise ValueError(f"Unsupported activation: {hidden_act}. " |
| 113 | + "Only silu is supported for now.") |
| 114 | + self.act_fn = SiluAndMul() |
| 115 | + |
| 116 | + # NOTE: `torch_npu.npu_dequant_swiglu_quant` can only be enabled in dynamic quant |
| 117 | + self.is_dynamic_quant = not isinstance( |
| 118 | + self.gate_up_proj.quant_method, |
| 119 | + UnquantizedLinearMethod) and isinstance( |
| 120 | + self.gate_up_proj.quant_method.quant_method, |
| 121 | + AscendW8A8DynamicLinearMethod) |
| 122 | + |
| 123 | + def forward(self, x): |
| 124 | + if self.is_dynamic_quant: |
| 125 | + x, dynamic_scale = torch_npu.npu_dynamic_quant(x) |
| 126 | + x = torch_npu.npu_quant_matmul( |
| 127 | + x, |
| 128 | + self.gate_up_proj.weight, |
| 129 | + self.gate_up_proj.weight_scale, |
| 130 | + output_dtype=torch.int32, |
| 131 | + ) |
| 132 | + x, dynamic_scale = torch_npu.npu_dequant_swiglu_quant( |
| 133 | + x=x, |
| 134 | + weight_scale=self.gate_up_proj.weight_scale_fp32, |
| 135 | + activation_scale=dynamic_scale, |
| 136 | + bias=None, |
| 137 | + quant_scale=None, |
| 138 | + quant_offset=None, |
| 139 | + group_index=None, |
| 140 | + activate_left=True, |
| 141 | + quant_mode=1) |
| 142 | + x = torch_npu.npu_quant_matmul( |
| 143 | + x, |
| 144 | + self.down_proj.weight, |
| 145 | + self.down_proj.weight_scale, |
| 146 | + pertoken_scale=dynamic_scale, |
| 147 | + output_dtype=torch.bfloat16, |
| 148 | + ) |
| 149 | + if self.down_proj.reduce_results and self.down_proj.tp_size > 1: |
| 150 | + x = tensor_model_parallel_all_reduce(x) |
| 151 | + return x |
| 152 | + gate_up, _ = self.gate_up_proj(x) |
| 153 | + x = self.act_fn(gate_up) |
| 154 | + x, _ = self.down_proj(x) |
| 155 | + return x |
86 | 156 |
|
87 | 157 | def _forward_ms_mlp(self, x): |
88 | 158 | current_ms_metadata = get_multistream_comm_context() |
89 | 159 | assert current_ms_metadata is not None |
| 160 | + if self.is_dynamic_quant: |
| 161 | + x, dynamic_scale = torch_npu.npu_dynamic_quant(x) |
| 162 | + x = torch_npu.npu_quant_matmul( |
| 163 | + x, |
| 164 | + self.gate_up_proj.weight, |
| 165 | + self.gate_up_proj.weight_scale, |
| 166 | + output_dtype=torch.int32, |
| 167 | + ) |
| 168 | + x, dynamic_scale = torch_npu.npu_dequant_swiglu_quant( |
| 169 | + x=x, |
| 170 | + weight_scale=self.gate_up_proj.weight_scale_fp32, |
| 171 | + activation_scale=dynamic_scale, |
| 172 | + bias=None, |
| 173 | + quant_scale=None, |
| 174 | + quant_offset=None, |
| 175 | + group_index=None, |
| 176 | + activate_left=True, |
| 177 | + quant_mode=1) |
| 178 | + x = torch_npu.npu_quant_matmul( |
| 179 | + x, |
| 180 | + self.down_proj.weight, |
| 181 | + self.down_proj.weight_scale, |
| 182 | + pertoken_scale=dynamic_scale, |
| 183 | + output_dtype=torch.bfloat16, |
| 184 | + ) |
| 185 | + if self.down_proj.reduce_results and self.down_proj.tp_size > 1: |
| 186 | + current_ms_metadata.before_comm_event.record() |
| 187 | + with torch.npu.stream(current_ms_metadata.comm_stream): |
| 188 | + current_ms_metadata.before_comm_event.wait() |
| 189 | + x = tensor_model_parallel_all_reduce(x) |
| 190 | + current_ms_metadata.after_comm_event.record() |
| 191 | + return x |
90 | 192 | gate_up, _ = self.gate_up_proj(x) |
91 | 193 | x = self.act_fn(gate_up) |
92 | 194 | current_ms_metadata.before_comm_event.record() |
|
0 commit comments