|
| 1 | +from collections.abc import Iterable |
| 2 | +from typing import Optional, Union |
| 3 | + |
| 4 | +import torch |
| 5 | +from torch import nn |
| 6 | +from transformers import Qwen3Config |
| 7 | +from vllm.compilation.decorators import support_torch_compile |
| 8 | +from vllm.config import CacheConfig, VllmConfig |
| 9 | +from vllm.distributed import get_pp_group |
| 10 | +from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 11 | +from vllm.model_executor.layers.quantization import QuantizationConfig |
| 12 | +from vllm.model_executor.layers.vocab_parallel_embedding import ParallelLMHead |
| 13 | +from vllm.model_executor.models.interfaces import SupportsLoRA, SupportsPP |
| 14 | +from vllm.model_executor.models.qwen2 import Qwen2Model |
| 15 | +from vllm.model_executor.models.qwen3 import Qwen3DecoderLayer |
| 16 | +from vllm.model_executor.models.utils import (AutoWeightsLoader, |
| 17 | + PPMissingLayer, maybe_prefix) |
| 18 | +from vllm.model_executor.sampling_metadata import SamplingMetadata |
| 19 | +from vllm.sequence import IntermediateTensors |
| 20 | + |
| 21 | +from vllm_ascend.ops.layernorm import AddRMSNormW8A8Quant |
| 22 | + |
| 23 | + |
| 24 | +class CustomQwen3DecoderLayer(Qwen3DecoderLayer): |
| 25 | + |
| 26 | + def __init__( |
| 27 | + self, |
| 28 | + config: Qwen3Config, |
| 29 | + cache_config: Optional[CacheConfig] = None, |
| 30 | + quant_config: Optional[QuantizationConfig] = None, |
| 31 | + prefix: str = "", |
| 32 | + ) -> None: |
| 33 | + super().__init__(config=config, |
| 34 | + cache_config=cache_config, |
| 35 | + quant_config=quant_config, |
| 36 | + prefix=prefix) |
| 37 | + if quant_config is None: |
| 38 | + return |
| 39 | + |
| 40 | + from vllm_ascend.quantization.quant_config import AscendQuantConfig |
| 41 | + from vllm_ascend.quantization.w8a8 import AscendW8A8LinearMethod |
| 42 | + |
| 43 | + assert isinstance(quant_config, AscendQuantConfig), \ |
| 44 | + "Expected quant_config to be an instance of AscendQuantConfig" |
| 45 | + |
| 46 | + if isinstance(self.self_attn.qkv_proj.quant_method.quant_method, |
| 47 | + AscendW8A8LinearMethod): |
| 48 | + self.input_layernorm = AddRMSNormW8A8Quant( |
| 49 | + config.hidden_size, |
| 50 | + layer=self.self_attn.qkv_proj, |
| 51 | + eps=config.rms_norm_eps) |
| 52 | + if isinstance(self.mlp.gate_up_proj.quant_method.quant_method, |
| 53 | + AscendW8A8LinearMethod): |
| 54 | + self.post_attention_layernorm = AddRMSNormW8A8Quant( |
| 55 | + config.hidden_size, |
| 56 | + layer=self.mlp.gate_up_proj, |
| 57 | + eps=config.rms_norm_eps) |
| 58 | + |
| 59 | + |
| 60 | +ALL_DECODER_LAYER_TYPES = { |
| 61 | + "attention": CustomQwen3DecoderLayer, |
| 62 | +} |
| 63 | + |
| 64 | + |
| 65 | +@support_torch_compile( |
| 66 | + dynamic_arg_dims={ |
| 67 | + "input_ids": 0, |
| 68 | + # positions is of shape (3, seq_len) if mrope is enabled for qwen2-vl, |
| 69 | + # otherwise (seq_len, ). |
| 70 | + "positions": -1, |
| 71 | + "intermediate_tensors": 0, |
| 72 | + "inputs_embeds": 0, |
| 73 | + }) |
| 74 | +class CustomQwen3Model(Qwen2Model): |
| 75 | + |
| 76 | + def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 77 | + super().__init__(vllm_config=vllm_config, |
| 78 | + prefix=prefix, |
| 79 | + decoder_layer_type=CustomQwen3DecoderLayer) |
| 80 | + |
| 81 | + |
| 82 | +class CustomQwen3ForCausalLM(nn.Module, SupportsLoRA, SupportsPP): |
| 83 | + # add `CustomQwen3Model` to init self.model |
| 84 | + packed_modules_mapping = { |
| 85 | + "qkv_proj": [ |
| 86 | + "q_proj", |
| 87 | + "k_proj", |
| 88 | + "v_proj", |
| 89 | + ], |
| 90 | + "gate_up_proj": [ |
| 91 | + "gate_proj", |
| 92 | + "up_proj", |
| 93 | + ], |
| 94 | + } |
| 95 | + |
| 96 | + def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 97 | + super().__init__() |
| 98 | + config = vllm_config.model_config.hf_config |
| 99 | + quant_config = vllm_config.quant_config |
| 100 | + lora_config = vllm_config.lora_config |
| 101 | + |
| 102 | + self.config = config |
| 103 | + self.lora_config = lora_config |
| 104 | + |
| 105 | + self.quant_config = quant_config |
| 106 | + self.model = CustomQwen3Model(vllm_config=vllm_config, |
| 107 | + prefix=maybe_prefix(prefix, "model")) |
| 108 | + |
| 109 | + if get_pp_group().is_last_rank: |
| 110 | + if config.tie_word_embeddings: |
| 111 | + self.lm_head = self.model.embed_tokens |
| 112 | + else: |
| 113 | + self.lm_head = ParallelLMHead(config.vocab_size, |
| 114 | + config.hidden_size, |
| 115 | + quant_config=quant_config, |
| 116 | + prefix=maybe_prefix( |
| 117 | + prefix, "lm_head")) |
| 118 | + else: |
| 119 | + self.lm_head = PPMissingLayer() |
| 120 | + |
| 121 | + self.logits_processor = LogitsProcessor(config.vocab_size) |
| 122 | + |
| 123 | + self.make_empty_intermediate_tensors = ( |
| 124 | + self.model.make_empty_intermediate_tensors) |
| 125 | + |
| 126 | + def get_input_embeddings(self, input_ids: torch.Tensor) -> torch.Tensor: |
| 127 | + return self.model.get_input_embeddings(input_ids) |
| 128 | + |
| 129 | + def forward( |
| 130 | + self, |
| 131 | + input_ids: torch.Tensor, |
| 132 | + positions: torch.Tensor, |
| 133 | + intermediate_tensors: Optional[IntermediateTensors] = None, |
| 134 | + inputs_embeds: Optional[torch.Tensor] = None, |
| 135 | + ) -> Union[torch.Tensor, IntermediateTensors]: |
| 136 | + hidden_states = self.model(input_ids, positions, intermediate_tensors, |
| 137 | + inputs_embeds) |
| 138 | + return hidden_states |
| 139 | + |
| 140 | + def compute_logits( |
| 141 | + self, |
| 142 | + hidden_states: torch.Tensor, |
| 143 | + sampling_metadata: SamplingMetadata, |
| 144 | + ) -> Optional[torch.Tensor]: |
| 145 | + logits = self.logits_processor(self.lm_head, hidden_states, |
| 146 | + sampling_metadata) |
| 147 | + return logits |
| 148 | + |
| 149 | + def load_weights(self, weights: Iterable[tuple[str, |
| 150 | + torch.Tensor]]) -> set[str]: |
| 151 | + loader = AutoWeightsLoader( |
| 152 | + self, |
| 153 | + skip_prefixes=(["lm_head."] |
| 154 | + if self.config.tie_word_embeddings else None), |
| 155 | + ) |
| 156 | + return loader.load_weights(weights) |
0 commit comments