Hugging Face's logo

Respair
/
NeMo_Canary

NeMo
TensorBoard
Model card Files
xet
Metrics Community
NeMo_Canary
File size: 5,983 Bytes 
b386992
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
 
# Long Sequence Performance

- The table below shows the pre-training performance of the LLAMA2-7B and LLAMA3-8B models on H100 and B200 GPUs, respectively, with CP (context parallelism), 
and compares it against the results without CP at various input sequence lengths. 
The detailed model-parallel configurations and the achieved performance are shown in the training results with CP. 
In non-CP training runs, we use the most performant model- and data-parallel configurations without CP given the memory capacity constraint of the each GPU system.

## LLAMA3-8B (FP8) - B200

  - Container: [NeMo25.04.rc2](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/nemo/tags)
  - System: DGX-B200

<table>
  <thead>
    <tr>
      <th rowspan="2" class="top-border">SeqLen (K)</th>
      <th rowspan="2" class="top-border"># of GPUs</th>
      <th rowspan="2" class="top-border">Batch Size</th>
      <th rowspan="1" class="top-border">Without CP</th>
      <th colspan="5" class="top-border">With CP</th>
      <th rowspan="2" class="top-border">Speedup with CP/without CP</th>
    </tr>
    <tr>
      <th>TFLOPS / GPU</th>
      <th>TP</th>
      <th>PP</th>
      <th>DP</th>
      <th>CP</th>
      <th>TFLOPS / GPU</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>8</td>
      <td>8</td>
      <td>512</td>
      <td>1,671</td>
      <td>1</td>
      <td>1</td>
      <td>2</td>
      <td>1</td>
      <td>1,671</td>
      <td class="speedup">1.00</td>
    </tr>
    <tr>
      <td>16</td>
      <td>16</td>
      <td>256</td>
      <td>1,717</td>
      <td>1</td>
      <td>1</td>
      <td>4</td>
      <td>1</td>
      <td>1,717</td>
      <td class="speedup">1.00</td>
    </tr>
    <tr>
      <td>32</td>
      <td>32</td>
      <td>128</td>
      <td>1,549</td>
      <td>1</td>
      <td>1</td>
      <td>4</td>
      <td>2</td>
      <td>1,624</td>
      <td class="speedup">1.05</td>
    </tr>
    <tr>
      <td>64</td>
      <td>64</td>
      <td>64</td>
      <td>1,481</td>
      <td>1</td>
      <td>1</td>
      <td>4</td>
      <td>4</td>
      <td>1,600</td>
      <td class="speedup">1.08</td>
    </tr>
    <tr>
      <td>128</td>
      <td>128</td>
      <td>32</td>
      <td>1,438</td>
      <td>2</td>
      <td>1</td>
      <td>4</td>
      <td>4</td>
      <td>1,588</td>
      <td class="speedup">1.10</td>
    </tr>
    <tr>
      <td>256</td>
      <td>256</td>
      <td>16</td>
      <td>1,162</td>
      <td>4</td>
      <td>1</td>
      <td>4</td>
      <td>4</td>
      <td>1,590</td>
      <td class="speedup">1.37</td>
    </tr>
    <tr>
      <td>512</td>
      <td>512</td>
      <td>8</td>
      <td>607</td>
      <td>4</td>
      <td>1</td>
      <td>4</td>
      <td>8</td>
      <td>1,619</td>
      <td class="speedup">2.67</td>
    </tr>
    <tr>
      <td>1024</td>
      <td>1024</td>
      <td>4</td>
      <td>-<sup>1)</sup></td>
      <td>4</td>
      <td>1</td>
      <td>4</td>
      <td>16</td>
      <td>1,608</td>
      <td class="speedup">-</td>
    </tr>
  </tbody>
</table>

<sup> 1) Since the maximum TP size is limited by the number of query groups (8 in LLAMA3-8B), 
even with full activation recomputation it is impossible to run the LLAMA3-8B model on a 1024K token sequence without CP due to the GPU memory constraints.</sup>

## LLAMA2-7B (FP8) - H100

  - Container: [NeMo24.03.01.framework](https://catalog.ngc.nvidia.com/orgs/nvidia/containers/nemo/tags)
  - System: DGX-H100


<table>
  <thead>
    <tr>
      <th rowspan="2" class="top-border">SeqLen (K)</th>
      <th rowspan="2" class="top-border"># of GPUs</th>
      <th rowspan="2" class="top-border">Batch Size</th>
      <th rowspan="1" class="top-border">Without CP</th>
      <th colspan="5" class="top-border">With CP</th>
      <th rowspan="2" class="top-border">Speedup with CP/without CP</th>
    </tr>
    <tr>
      <th>TFLOPS / GPU</th>
      <th>TP</th>
      <th>PP</th>
      <th>DP</th>
      <th>CP</th>
      <th>TFLOPS / GPU</th>
    </tr>
  </thead>
  <tbody>
    <tr>
      <td>4</td>
      <td>4</td>
      <td>1024</td>
      <td>768</td>
      <td>1</td>
      <td>1</td>
      <td>4</td>
      <td>1</td>
      <td>768</td>
      <td class="speedup">1.00</td>
    </tr>
    <tr>
      <td>8</td>
      <td>8</td>
      <td>512</td>
      <td>730</td>
      <td>1</td>
      <td>2</td>
      <td>4</td>
      <td>1</td>
      <td>730</td>
      <td class="speedup">1.00</td>
    </tr>
    <tr>
      <td>16</td>
      <td>16</td>
      <td>256</td>
      <td>660</td>
      <td>2</td>
      <td>1</td>
      <td>8</td>
      <td>1</td>
      <td>660</td>
      <td class="speedup">1.00</td>
    </tr>
    <tr>
      <td>32</td>
      <td>32</td>
      <td>128</td>
      <td>595</td>
      <td>2</td>
      <td>1</td>
      <td>8</td>
      <td>2</td>
      <td>610</td>
      <td class="speedup">1.03</td>
    </tr>
    <tr>
      <td>64</td>
      <td>64</td>
      <td>64</td>
      <td>534</td>
      <td>4</td>
      <td>1</td>
      <td>8</td>
      <td>2</td>
      <td>574</td>
      <td class="speedup">1.07</td>
    </tr>
    <tr>
      <td>128</td>
      <td>128</td>
      <td>32</td>
      <td>424</td>
      <td>4</td>
      <td>1</td>
      <td>8</td>
      <td>4</td>
      <td>555</td>
      <td class="speedup">1.31</td>
    </tr>
    <tr>
      <td>256</td>
      <td>256</td>
      <td>16</td>
      <td>392</td>
      <td>4</td>
      <td>1</td>
      <td>8</td>
      <td>8</td>
      <td>549</td>
      <td class="speedup">1.40</td>
    </tr>
    <tr>
      <td>512</td>
      <td>512</td>
      <td>8</td>
      <td>104</td>
      <td>8</td>
      <td>1</td>
      <td>4</td>
      <td>16</td>
      <td>549</td>
      <td class="speedup">5.28</td>
    </tr>
    <tr>
      <td>1024</td>
      <td>1024</td>
      <td>4</td>
      <td>26.5</td>
      <td>8</td>
      <td>1</td>
      <td>4</td>
      <td>32</td>
      <td>536</td>
      <td class="speedup">20.23</td>
    </tr>
  </tbody>
</table>