From 098e1776ba72805bbf8f68a44b10f499413e5189 Mon Sep 17 00:00:00 2001 From: hxer7963 Date: Thu, 28 Mar 2024 09:12:54 +0800 Subject: [PATCH] [Model] Add support for xverse (#3610) Co-authored-by: willhe Co-authored-by: root --- README.md | 1 + vllm/model_executor/models/__init__.py | 1 + vllm/model_executor/models/xverse.py | 372 +++++++++++++++++++++++++ 3 files changed, 374 insertions(+) create mode 100644 vllm/model_executor/models/xverse.py diff --git a/README.md b/README.md index 20a2f53e18084..bf8fbd4173949 100644 --- a/README.md +++ b/README.md @@ -91,6 +91,7 @@ vLLM seamlessly supports many Hugging Face models, including the following archi - Qwen2 (`Qwen/Qwen2-7B-beta`, `Qwen/Qwen-7B-Chat-beta`, etc.) - StableLM(`stabilityai/stablelm-3b-4e1t`, `stabilityai/stablelm-base-alpha-7b-v2`, etc.) - Starcoder2(`bigcode/starcoder2-3b`, `bigcode/starcoder2-7b`, `bigcode/starcoder2-15b`, etc.) +- Xverse (`xverse/XVERSE-7B-Chat`, `xverse/XVERSE-13B-Chat`, `xverse/XVERSE-65B-Chat`, etc.) - Yi (`01-ai/Yi-6B`, `01-ai/Yi-34B`, etc.) Install vLLM with pip or [from source](https://vllm.readthedocs.io/en/latest/getting_started/installation.html#build-from-source): diff --git a/vllm/model_executor/models/__init__.py b/vllm/model_executor/models/__init__.py index 5195b6fac93bf..79ddb4736e25c 100755 --- a/vllm/model_executor/models/__init__.py +++ b/vllm/model_executor/models/__init__.py @@ -51,6 +51,7 @@ "StableLMEpochForCausalLM": ("stablelm", "StablelmForCausalLM"), "StableLmForCausalLM": ("stablelm", "StablelmForCausalLM"), "Starcoder2ForCausalLM": ("starcoder2", "Starcoder2ForCausalLM"), + "XverseForCausalLM": ("xverse", "XverseForCausalLM"), } # Models not supported by ROCm. diff --git a/vllm/model_executor/models/xverse.py b/vllm/model_executor/models/xverse.py new file mode 100644 index 0000000000000..83d2ddb2bcf35 --- /dev/null +++ b/vllm/model_executor/models/xverse.py @@ -0,0 +1,372 @@ +# coding=utf-8 +# Adapted from +# https://huggingface.co/xverse/XVERSE-7B/blob/main/modeling_xverse.py +# Copyright 2022 EleutherAI and the HuggingFace Inc. team. All rights reserved. +# +# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX +# and OPT implementations in this library. It has been modified from its +# original forms to accommodate minor architectural differences compared +# to GPT-NeoX and OPT used by the Meta AI team that trained the model. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""Inference-only Xverse model compatible with HuggingFace weights.""" +from typing import Any, Dict, List, Optional, Tuple + +import torch +from torch import nn +from transformers import PretrainedConfig + +from vllm.attention import Attention, AttentionMetadata +from vllm.config import LoRAConfig +from vllm.model_executor.layers.activation import SiluAndMul +from vllm.model_executor.layers.layernorm import RMSNorm +from vllm.model_executor.layers.linear import (LinearMethodBase, + MergedColumnParallelLinear, + QKVParallelLinear, + RowParallelLinear) +from vllm.model_executor.layers.logits_processor import LogitsProcessor +from vllm.model_executor.layers.rotary_embedding import get_rope +from vllm.model_executor.layers.sampler import Sampler +from vllm.model_executor.layers.vocab_parallel_embedding import ( + ParallelLMHead, VocabParallelEmbedding) +from vllm.model_executor.parallel_utils.parallel_state import ( + get_tensor_model_parallel_world_size) +from vllm.model_executor.sampling_metadata import SamplingMetadata +from vllm.model_executor.weight_utils import (default_weight_loader, + hf_model_weights_iterator) +from vllm.sequence import SamplerOutput + + +class XverseMLP(nn.Module): + + def __init__( + self, + hidden_size: int, + intermediate_size: int, + hidden_act: str, + linear_method: Optional[LinearMethodBase] = None, + ) -> None: + super().__init__() + self.gate_up_proj = MergedColumnParallelLinear( + hidden_size, [intermediate_size] * 2, + bias=False, + linear_method=linear_method) + self.down_proj = RowParallelLinear(intermediate_size, + hidden_size, + bias=False, + linear_method=linear_method) + if hidden_act != "silu": + raise ValueError(f"Unsupported activation: {hidden_act}. " + "Only silu is supported for now.") + self.act_fn = SiluAndMul() + + def forward(self, x): + gate, _ = self.gate_up_proj(x) + x = self.act_fn(gate) + x, _ = self.down_proj(x) + return x + + +class XverseAttention(nn.Module): + + def __init__( + self, + hidden_size: int, + num_heads: int, + num_kv_heads: int, + rope_theta: float = 10000, + rope_scaling: Optional[Dict[str, Any]] = None, + max_position_embeddings: int = 8192, + linear_method: Optional[LinearMethodBase] = None, + bias: bool = False, + sliding_window: Optional[int] = None, + ) -> None: + super().__init__() + self.hidden_size = hidden_size + tp_size = get_tensor_model_parallel_world_size() + self.total_num_heads = num_heads + assert self.total_num_heads % tp_size == 0 + self.num_heads = self.total_num_heads // tp_size + self.total_num_kv_heads = num_kv_heads + # partition the KV heads across multiple tensor parallel GPUs. + assert self.total_num_kv_heads % tp_size == 0 + self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size) + self.head_dim = hidden_size // self.total_num_heads + self.q_size = self.num_heads * self.head_dim + self.kv_size = self.num_kv_heads * self.head_dim + self.scaling = self.head_dim**-0.5 + self.rope_theta = rope_theta + self.max_position_embeddings = max_position_embeddings + + self.qkv_proj = QKVParallelLinear( + hidden_size, + self.head_dim, + self.total_num_heads, + self.total_num_kv_heads, + bias=bias, + linear_method=linear_method, + ) + self.o_proj = RowParallelLinear( + self.total_num_heads * self.head_dim, + hidden_size, + bias=bias, + linear_method=linear_method, + ) + + self.rotary_emb = get_rope( + self.head_dim, + rotary_dim=self.head_dim, + max_position=max_position_embeddings, + base=rope_theta, + rope_scaling=rope_scaling, + ) + self.attn = Attention(self.num_heads, + self.head_dim, + self.scaling, + num_kv_heads=self.num_kv_heads, + sliding_window=sliding_window) + + def forward( + self, + positions: torch.Tensor, + hidden_states: torch.Tensor, + kv_cache: torch.Tensor, + attn_metadata: AttentionMetadata, + ) -> torch.Tensor: + qkv, _ = self.qkv_proj(hidden_states) + q, k, v = qkv.split([self.q_size, self.kv_size, self.kv_size], dim=-1) + q, k = self.rotary_emb(positions, q, k) + attn_output = self.attn(q, k, v, kv_cache, attn_metadata) + output, _ = self.o_proj(attn_output) + return output + + +class XverseDecoderLayer(nn.Module): + + def __init__( + self, + config: PretrainedConfig, + linear_method: Optional[LinearMethodBase] = None, + ) -> None: + super().__init__() + self.hidden_size = config.hidden_size + rope_theta = getattr(config, "rope_theta", 10000) + rope_scaling = getattr(config, "rope_scaling", None) + max_position_embeddings = getattr(config, "max_position_embeddings", + 8192) + sliding_window = getattr(config, "sliding_window", None) + self.self_attn = XverseAttention( + hidden_size=self.hidden_size, + num_heads=config.num_attention_heads, + num_kv_heads=getattr(config, "num_key_value_heads", + config.num_attention_heads), + rope_theta=rope_theta, + rope_scaling=rope_scaling, + max_position_embeddings=max_position_embeddings, + linear_method=linear_method, + bias=getattr(config, "bias", False), + sliding_window=sliding_window, + ) + self.mlp = XverseMLP( + hidden_size=self.hidden_size, + intermediate_size=config.intermediate_size, + hidden_act=config.hidden_act, + linear_method=linear_method, + ) + self.input_layernorm = RMSNorm(config.hidden_size, + eps=config.rms_norm_eps) + self.post_attention_layernorm = RMSNorm(config.hidden_size, + eps=config.rms_norm_eps) + + def forward( + self, + positions: torch.Tensor, + hidden_states: torch.Tensor, + kv_cache: torch.Tensor, + attn_metadata: AttentionMetadata, + residual: Optional[torch.Tensor], + ) -> Tuple[torch.Tensor, torch.Tensor]: + # Self Attention + if residual is None: + residual = hidden_states + hidden_states = self.input_layernorm(hidden_states) + else: + hidden_states, residual = self.input_layernorm( + hidden_states, residual) + hidden_states = self.self_attn( + positions=positions, + hidden_states=hidden_states, + kv_cache=kv_cache, + attn_metadata=attn_metadata, + ) + + # Fully Connected + hidden_states, residual = self.post_attention_layernorm( + hidden_states, residual) + hidden_states = self.mlp(hidden_states) + return hidden_states, residual + + +class XverseModel(nn.Module): + + def __init__( + self, + config: PretrainedConfig, + linear_method: Optional[LinearMethodBase] = None, + lora_config: Optional[LoRAConfig] = None, + ) -> None: + super().__init__() + self.config = config + self.padding_idx = config.pad_token_id + lora_vocab = (lora_config.lora_extra_vocab_size * + (lora_config.max_loras or 1)) if lora_config else 0 + self.vocab_size = config.vocab_size + lora_vocab + self.org_vocab_size = config.vocab_size + self.embed_tokens = VocabParallelEmbedding( + self.vocab_size, + config.hidden_size, + org_num_embeddings=config.vocab_size, + ) + self.layers = nn.ModuleList([ + XverseDecoderLayer(config, linear_method) + for _ in range(config.num_hidden_layers) + ]) + self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps) + + def forward( + self, + input_ids: torch.Tensor, + positions: torch.Tensor, + kv_caches: List[torch.Tensor], + attn_metadata: AttentionMetadata, + ) -> torch.Tensor: + hidden_states = self.embed_tokens(input_ids) + residual = None + for i in range(len(self.layers)): + layer = self.layers[i] + hidden_states, residual = layer( + positions, + hidden_states, + kv_caches[i], + attn_metadata, + residual, + ) + hidden_states, _ = self.norm(hidden_states, residual) + return hidden_states + + +class XverseForCausalLM(nn.Module): + packed_modules_mapping = { + "qkv_proj": [ + "q_proj", + "k_proj", + "v_proj", + ], + "gate_up_proj": [ + "gate_proj", + "up_proj", + ], + } + + # LoRA specific attributes + supported_lora_modules = [ + "qkv_proj", + "o_proj", + "gate_up_proj", + "down_proj", + "embed_tokens", + "lm_head", + ] + embedding_modules = { + "embed_tokens": "input_embeddings", + "lm_head": "output_embeddings", + } + embedding_padding_modules = ["lm_head"] + + def __init__( + self, + config: PretrainedConfig, + linear_method: Optional[LinearMethodBase] = None, + lora_config=None, + ) -> None: + super().__init__() + self.config = config + self.linear_method = linear_method + self.model = XverseModel(config, linear_method) + self.lm_head = ParallelLMHead(config.vocab_size, config.hidden_size) + self.logits_processor = LogitsProcessor(config.vocab_size) + self.sampler = Sampler() + + def forward( + self, + input_ids: torch.Tensor, + positions: torch.Tensor, + kv_caches: List[torch.Tensor], + attn_metadata: AttentionMetadata, + ) -> torch.Tensor: + hidden_states = self.model(input_ids, positions, kv_caches, + attn_metadata) + return hidden_states + + def compute_logits(self, hidden_states: torch.Tensor, + sampling_metadata: SamplingMetadata) -> torch.Tensor: + logits = self.logits_processor(self.lm_head.weight, hidden_states, + sampling_metadata) + return logits + + def sample( + self, + logits: torch.Tensor, + sampling_metadata: SamplingMetadata, + ) -> Optional[SamplerOutput]: + next_tokens = self.sampler(logits, sampling_metadata) + return next_tokens + + def load_weights(self, + model_name_or_path: str, + cache_dir: Optional[str] = None, + load_format: str = "auto", + revision: Optional[str] = None): + stacked_params_mapping = [ + ("qkv_proj", "q_proj", "q"), + ("qkv_proj", "k_proj", "k"), + ("qkv_proj", "v_proj", "v"), + ("gate_up_proj", "gate_proj", 0), + ("gate_up_proj", "up_proj", 1), + ] + params_dict = dict(self.named_parameters()) + for name, loaded_weight in hf_model_weights_iterator( + model_name_or_path, cache_dir, load_format, revision): + if ("rotary_emb.inv_freq" in name + or "rotary_emb.cos_cached" in name + or "rotary_emb.sin_cached" in name): + continue + for (param_name, weight_name, shard_id) in stacked_params_mapping: + if weight_name not in name: + continue + name = name.replace(weight_name, param_name) + # Skip loading extra bias for GPTQ models. + if name.endswith(".bias") and name not in params_dict: + continue + param = params_dict[name] + weight_loader = param.weight_loader + weight_loader(param, loaded_weight, shard_id) + break + else: + # Skip loading extra bias for GPTQ models. + if name.endswith(".bias") and name not in params_dict: + continue + param = params_dict[name] + weight_loader = getattr(param, "weight_loader", + default_weight_loader) + weight_loader(param, loaded_weight)