A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 16, 17:00 UTC
[Bugfix] Fix internlm2 tensor parallel inference (#8055)
Isotr0py committed
Sep 2, 2024 at 23:48 UTC
dd2a6a82e3f41b4673b1dbb24b2e99230ea96981
1 file changed
+34
-13
vllm/model_executor/models/internlm2.py
+34
-13
@@ -1,4 +1,5 @@
1
# -*- coding: utf-8 -*-
2
+from functools import partial
3
from typing import Any, Dict, Iterable, List, Optional, Tuple
4
5
import torch
@@ -7,7 +8,10 @@ from transformers import PretrainedConfig
8
9
from vllm.attention import Attention, AttentionMetadata
10
from vllm.config import CacheConfig
10
-from vllm.distributed import get_tensor_model_parallel_world_size
11
+from vllm.distributed import (get_tensor_model_parallel_rank,
12
+ get_tensor_model_parallel_world_size,
13
+ split_tensor_along_last_dim,
14
+ tensor_model_parallel_all_gather)
15
from vllm.model_executor.layers.activation import SiluAndMul
16
from vllm.model_executor.layers.layernorm import RMSNorm
17
from vllm.model_executor.layers.linear import (MergedColumnParallelLinear,
@@ -70,20 +74,21 @@ class InternLM2Attention(nn.Module):
74
) -> None:
75
super().__init__()
76
self.hidden_size = hidden_size
73
- tp_size = get_tensor_model_parallel_world_size()
77
+ self.tp_size = get_tensor_model_parallel_world_size()
78
+ self.tp_rank = get_tensor_model_parallel_rank()
79
self.total_num_heads = num_heads
75
- assert self.total_num_heads % tp_size == 0
76
- self.num_heads = self.total_num_heads // tp_size
80
+ assert self.total_num_heads % self.tp_size == 0
81
+ self.num_heads = self.total_num_heads // self.tp_size
82
self.total_num_kv_heads = num_kv_heads
78
- if self.total_num_kv_heads >= tp_size:
83
+ if self.total_num_kv_heads >= self.tp_size:
84
# Number of KV heads is greater than TP size, so we partition
85
# the KV heads across multiple tensor parallel GPUs.
81
- assert self.total_num_kv_heads % tp_size == 0
86
+ assert self.total_num_kv_heads % self.tp_size == 0
87
else:
88
# Number of KV heads is less than TP size, so we replicate
89
# the KV heads across multiple tensor parallel GPUs.
85
- assert tp_size % self.total_num_kv_heads == 0
86
- self.num_kv_heads = max(1, self.total_num_kv_heads // tp_size)
90
+ assert self.tp_size % self.total_num_kv_heads == 0
91
+ self.num_kv_heads = max(1, self.total_num_kv_heads // self.tp_size)
92
self.head_dim = hidden_size // self.total_num_heads
93
self.q_size = self.num_heads * self.head_dim
94
self.kv_size = self.num_kv_heads * self.head_dim
@@ -122,11 +127,27 @@ class InternLM2Attention(nn.Module):
127
quant_config=quant_config)
128
129
def split_qkv(self, qkv: torch.Tensor):
125
- qkv = qkv.view(-1, self.num_kv_heads, self.key_value_groups + 2, 128)
126
- q, k, v = torch.split(qkv, [self.key_value_groups, 1, 1], dim=2)
127
- q = q.reshape(-1, self.q_size)
128
- k = k.reshape(-1, self.kv_size)
129
- v = v.reshape(-1, self.kv_size)
130
+ seq_len = qkv.shape[0]
131
+ if self.tp_size > 1:
132
+ qkv_map = [self.q_size, self.kv_size, self.kv_size] * self.tp_size
133
+ qkv = tensor_model_parallel_all_gather(qkv)
134
+ qkv = torch.split(qkv, qkv_map, dim=-1)
135
+ qkv = qkv[::3] + qkv[1::3] + qkv[2::3]
136
+ qkv = torch.cat(qkv, dim=-1)
137
+
138
+ qkv = qkv.view(seq_len, self.total_num_kv_heads,
139
+ self.key_value_groups + 2, self.head_dim)
140
+ q, k, v = torch.split(qkv, [self.key_value_groups, 1, 1], dim=-2)
141
+ q = q.reshape(seq_len, self.q_size * self.tp_size)
142
+ k = k.reshape(seq_len, self.kv_size * self.tp_size)
143
+ v = v.reshape(seq_len, self.kv_size * self.tp_size)
144
+
145
+ if self.tp_size > 1:
146
+ splitter = partial(split_tensor_along_last_dim,
147
+ num_partitions=self.tp_size)
148
+ q = splitter(q)[self.tp_rank]
149
+ k = splitter(k)[self.tp_rank]
150
+ v = splitter(v)[self.tp_rank]
151
return q, k, v
152
153
def forward(