|
| 1 | +# SPDX-License-Identifier: Apache-2.0 |
| 2 | +# SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | +"""Inference-only Qwen3Next MTP model.""" |
| 4 | +import torch |
| 5 | +from vllm.compilation.decorators import support_torch_compile |
| 6 | +from vllm.config import VllmConfig |
| 7 | +from vllm.model_executor.layers.linear import ColumnParallelLinear |
| 8 | +from vllm.model_executor.layers.logits_processor import LogitsProcessor |
| 9 | +from vllm.model_executor.layers.vocab_parallel_embedding import ( |
| 10 | + DEFAULT_VOCAB_PADDING_SIZE, ParallelLMHead, VocabParallelEmbedding) |
| 11 | +from vllm.model_executor.models.interfaces import SupportsPP |
| 12 | +from vllm.model_executor.models.qwen3_next_mtp import ( |
| 13 | + Qwen3NextMTP, Qwen3NextMultiTokenPredictor) |
| 14 | +from vllm.model_executor.models.utils import ( |
| 15 | + make_empty_intermediate_tensors_factory, maybe_prefix) |
| 16 | +from vllm.transformers_utils.configs import Qwen3NextConfig |
| 17 | + |
| 18 | +from vllm_ascend.models.qwen3_next import (CustomQwen3NextDecoderLayer, |
| 19 | + Qwen3NextRMSNorm) |
| 20 | + |
| 21 | + |
| 22 | +@support_torch_compile |
| 23 | +class CustomQwen3NextMultiTokenPredictor(Qwen3NextMultiTokenPredictor): |
| 24 | + |
| 25 | + def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 26 | + super(Qwen3NextMultiTokenPredictor, self).__init__() |
| 27 | + |
| 28 | + model_config = vllm_config.model_config |
| 29 | + quant_config = vllm_config.quant_config |
| 30 | + lora_config = vllm_config.lora_config |
| 31 | + config: Qwen3NextConfig = model_config.hf_config |
| 32 | + |
| 33 | + self.config = config |
| 34 | + lora_vocab = ((lora_config.lora_extra_vocab_size * |
| 35 | + (lora_config.max_loras or 1)) if lora_config else 0) |
| 36 | + self.vocab_size = config.vocab_size + lora_vocab |
| 37 | + self.org_vocab_size = config.vocab_size |
| 38 | + |
| 39 | + self.mtp_start_layer_idx = config.num_hidden_layers |
| 40 | + self.num_mtp_layers = getattr(config, "num_nextn_predict_layers", 1) |
| 41 | + |
| 42 | + self.embed_tokens = VocabParallelEmbedding( |
| 43 | + self.vocab_size, |
| 44 | + config.hidden_size, |
| 45 | + org_num_embeddings=config.vocab_size, |
| 46 | + ) |
| 47 | + |
| 48 | + self.fc = ColumnParallelLinear(self.config.hidden_size * 2, |
| 49 | + self.config.hidden_size, |
| 50 | + gather_output=True, |
| 51 | + bias=False, |
| 52 | + return_bias=False, |
| 53 | + quant_config=quant_config, |
| 54 | + prefix=f'{prefix}.fc') |
| 55 | + |
| 56 | + # use old version mtp layer name to avoid a exception in vllm |
| 57 | + self.layers = torch.nn.ModuleList( |
| 58 | + CustomQwen3NextDecoderLayer( |
| 59 | + vllm_config, |
| 60 | + layer_type="full_attention", |
| 61 | + prefix=f'{prefix}.layers.{self.mtp_start_layer_idx + idx}', |
| 62 | + ) for idx in range(self.num_mtp_layers)) |
| 63 | + |
| 64 | + self.make_empty_intermediate_tensors = ( |
| 65 | + make_empty_intermediate_tensors_factory( |
| 66 | + ["hidden_states", "residual"], config.hidden_size)) |
| 67 | + |
| 68 | + self.norm = Qwen3NextRMSNorm(config.hidden_size, |
| 69 | + eps=config.rms_norm_eps) |
| 70 | + self.pre_fc_norm_hidden = Qwen3NextRMSNorm(config.hidden_size, |
| 71 | + eps=config.rms_norm_eps) |
| 72 | + self.pre_fc_norm_embedding = Qwen3NextRMSNorm(config.hidden_size, |
| 73 | + eps=config.rms_norm_eps) |
| 74 | + |
| 75 | + |
| 76 | +@support_torch_compile |
| 77 | +class CustomQwen3NextMTP(Qwen3NextMTP, SupportsPP): |
| 78 | + packed_modules_mapping = { |
| 79 | + "qkv_proj": [ |
| 80 | + "q_proj", |
| 81 | + "k_proj", |
| 82 | + "v_proj", |
| 83 | + ], |
| 84 | + "gate_up_proj": ["up_proj", "down_proj"] |
| 85 | + } |
| 86 | + |
| 87 | + def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): |
| 88 | + config = vllm_config.model_config.hf_config |
| 89 | + self.vllm_config = vllm_config |
| 90 | + cache_config = vllm_config.cache_config |
| 91 | + assert not cache_config.enable_prefix_caching, \ |
| 92 | + "Qwen3NextMTP currently does not support prefix caching" |
| 93 | + |
| 94 | + self.quant_config = vllm_config.quant_config |
| 95 | + |
| 96 | + super(Qwen3NextMTP, self).__init__() |
| 97 | + self.config = config |
| 98 | + self.model = CustomQwen3NextMultiTokenPredictor( |
| 99 | + vllm_config=vllm_config, prefix=maybe_prefix(prefix, "model")) |
| 100 | + self.unpadded_vocab_size = config.vocab_size |
| 101 | + self.lm_head = ParallelLMHead(self.unpadded_vocab_size, |
| 102 | + config.hidden_size, |
| 103 | + org_num_embeddings=config.vocab_size, |
| 104 | + padding_size=DEFAULT_VOCAB_PADDING_SIZE, |
| 105 | + prefix=maybe_prefix(prefix, "lm_head")) |
| 106 | + self.logits_processor = LogitsProcessor(self.unpadded_vocab_size, |
| 107 | + config.vocab_size) |
| 108 | + self.make_empty_intermediate_tensors = ( |
| 109 | + self.model.make_empty_intermediate_tensors) |
0 commit comments