From 77766933730388101ede10797f5699698c04897a Mon Sep 17 00:00:00 2001
From: jiaxingli <43110891+li126com@users.noreply.github.com>
Date: Tue, 21 Nov 2023 19:20:02 +0800
Subject: [PATCH] feat(doc): add GPU memory info for 7B & 20B models (#507)
* unitest_only_forward
* memory_test
* doc fix
* doc fix
---
doc/en/train_performance.md | 64 +++++++++++++++++++++++++++++++++++++
doc/train_performance.md | 63 ++++++++++++++++++++++++++++++++++++
2 files changed, 127 insertions(+)
diff --git a/doc/en/train_performance.md b/doc/en/train_performance.md
index 9c77d9e..91b87de 100644
--- a/doc/en/train_performance.md
+++ b/doc/en/train_performance.md
@@ -91,3 +91,67 @@ When `Activation Ckpt` is turned off, the test results are as shown in the table
+
+### GPU Memory Usage Test
+Test configuration:
+| Configuration | Description |
+| :-------: | :-----: |
+| branch | develop |
+| tag | v0.2.1dev20231121 |
+| GPU | A800 |
+| Checkpoint| True |
+| micro_bsz | 1 |
+| micro_num | 4 |
+| dtype | bfloat16|
+
+```python
+# InternLM/configs/7B_sft.py
+data = dict(
+ # micro_num means the number of micro_batch contained in one gradient update
+ micro_num=4,
+ # packed_length = micro_bsz * SEQ_LEN
+ micro_bsz=1,
+ ...
+)
+
+model = dict(
+ checkpoint=True,
+ dtype="torch.bfloat16",
+ ...
+)
+
+parallel = dict(
+ zero1=dict(size=8, fsdp=False),
+ tensor=1,
+ pipeline=dict(size=1, interleaved_overlap=True),
+ sequence_parallel=False,
+)
+```
+
+Pre-training & Fine-tuning test:
+|model|Number of GPU|zero1|tp|pp|fsdp|GPU Memory (GB)|
+|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
+| 7B | 3 | -1 | 1 | 3 |False| 75 |
+| 7B | 3 | -1 | 1 | 1 |True | 72 |
+| 7B | 4 | -1 | 4 | 1 |True | 52 |
+| 7B | 4 | -1 | 4 | 1 |False| 61 |
+| 7B | 4 | -1 | 1 | 4 |False| 69 |
+| 7B | 4 | -1 | 1 | 1 |True | 56 |
+| 7B | 5 | -1 | 1 | 1 |True | 49 |
+| 7B | 5 | -1 | 1 | 5 |False| 62 |
+| 7B | 6 | -1 | 1 | 1 |True | 39 |
+| 7B | 6 | -1 | 2 | 1 |True | 38 |
+| 7B | 6 | -1 | 1 | 6 |False| 56 |
+| 20B | 8 | -1 | 1 | 1 |True | 78 |
+| 20B | 8 | -1 | 8 | 1 |True | 71 |
+| 20B | 16 | -1 | 1 | 1 |True | 40 |
+| 20B | 16 | -1 | 8 | 1 |True | 39 |
+| 20B | 16 | -1 | 1 | 16 |False| 52 |
+
+
+Web_demo test:
+
+|model|GPU|GPU Memory (GB)|System Memory (MB)|
+|:-:|:-:|:-:|:-:|
+| 7B | A800 | 14.5 | 2465 |
+| 20B | A800 | 39 | 9547 |
diff --git a/doc/train_performance.md b/doc/train_performance.md
index 239e20f..c5e2a7c 100644
--- a/doc/train_performance.md
+++ b/doc/train_performance.md
@@ -88,3 +88,66 @@ InternLM中`zero1`的配置决定了优化器状态的分配范围。
+### 显存占用测试
+测试配置：
+| 配置 | 描述 |
+| :-------: | :-----: |
+| 分支 | develop |
+| tag | v0.2.1dev20231121 |
+| 显卡 | A800 |
+| 重计算 | True |
+| micro_bsz | 1 |
+| micro_num | 4 |
+| dtype | bfloat16|
+
+```python
+# InternLM/configs/7B_sft.py
+data = dict(
+ # micro_num means the number of micro_batch contained in one gradient update
+ micro_num=4,
+ # packed_length = micro_bsz * SEQ_LEN
+ micro_bsz=1,
+ ...
+)
+
+model = dict(
+ checkpoint=True,
+ dtype="torch.bfloat16",
+ ...
+)
+
+parallel = dict(
+ zero1=dict(size=8, fsdp=False),
+ tensor=1,
+ pipeline=dict(size=1, interleaved_overlap=True),
+ sequence_parallel=False,
+)
+```
+
+预训练和微调测试：
+|模型|卡数|zero1|tp|pp|fsdp|显存（GB）|
+|:-:|:-:|:-:|:-:|:-:|:-:|:-:|
+| 7B | 3 | -1 | 1 | 3 |False| 75 |
+| 7B | 3 | -1 | 1 | 1 |True | 72 |
+| 7B | 4 | -1 | 4 | 1 |True | 52 |
+| 7B | 4 | -1 | 4 | 1 |False| 61 |
+| 7B | 4 | -1 | 1 | 4 |False| 69 |
+| 7B | 4 | -1 | 1 | 1 |True | 56 |
+| 7B | 5 | -1 | 1 | 1 |True | 49 |
+| 7B | 5 | -1 | 1 | 5 |False| 62 |
+| 7B | 6 | -1 | 1 | 1 |True | 39 |
+| 7B | 6 | -1 | 2 | 1 |True | 38 |
+| 7B | 6 | -1 | 1 | 6 |False| 56 |
+| 20B | 8 | -1 | 1 | 1 |True | 78 |
+| 20B | 8 | -1 | 8 | 1 |True | 71 |
+| 20B | 16 | -1 | 1 | 1 |True | 40 |
+| 20B | 16 | -1 | 8 | 1 |True | 39 |
+| 20B | 16 | -1 | 1 | 16 |False| 52 |
+
+
+Web_demo 测试:
+
+|模型|显卡|显存（GB）|内存（MB）|
+|:-:|:-:|:-:|:-:|
+| 7B | A800 | 14.5 | 2465 |
+| 20B | A800 | 39 | 9547 |
\ No newline at end of file