https://303030.org:443/index.php?action=history&feed=atom&title=Let%27s_reproduce_GPT-2_%28124M%29 2026-10-18T01:23:46Z تاريخ التعديل لهذه الصفحة في الويكي MediaWiki 1.41.0 https://303030.org:443/index.php?title=Let%27s_reproduce_GPT-2_(124M)&diff=19&oldid=prev 2024-06-26T19:21:30Z

<p></p> <table style="background-color: #fff; color: #202122;" data-mw="interface"> <col class="diff-marker" /> <col class="diff-content" /> <col class="diff-marker" /> <col class="diff-content" /> <tr class="diff-title" lang="ar"> <td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">→ مراجعة أقدم</td> <td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">مراجعة ١٤:٢١، ٢٦ يونيو ٢٠٢٤</td> </tr><tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l6">سطر ٦:</td> <td colspan="2" class="diff-lineno">سطر ٦:</td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;/html&gt;</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;/html&gt;</div></td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;hr&gt;</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;hr&gt;</div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;html&gt;</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; </ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 305,330 views  Jun 10, 2024</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; We reproduce the GPT-2 (124M) from scratch. This video covers the whole process: First we build the GPT-2 network, then we optimize its training to be really fast, then we set up the training run following the GPT-2 and GPT-3 paper and their hyperparameters, then we hit run, and come back the next morning to see our results, and enjoy some amusing model generations. Keep in mind that in some places this video builds on the knowledge from earlier videos in the Zero to Hero Playlist (see my channel). You could also see this video as building my nanoGPT repo, which by the end is about 90% similar.</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; </ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; Links:</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; - build-nanogpt GitHub repo, with all the changes in this video as individual commits: https://github.com/karpathy/build-nan...</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; - nanoGPT repo: https://github.com/karpathy/nanoGPT</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; - llm.c repo: https://github.com/karpathy/llm.c</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; - my website: https://karpathy.ai</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; - my twitter:  / karpathy  </ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; - our Discord channel:  / discord  </ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; </ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; Supplementary links:</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; - Attention is All You Need paper: https://arxiv.org/abs/1706.03762</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; - OpenAI GPT-3 paper: https://arxiv.org/abs/2005.14165 - OpenAI GPT-2 paper: https://d4mucfpksywv.cloudfront.net/b... The GPU I'm training the model on is from Lambda GPU Cloud, I think the best and easiest way to spin up an on-demand GPU instance in the cloud that you can ssh to: https://lambdalabs.com </ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; </ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; Chapters:</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 00:00:00 intro: Let’s reproduce GPT-2 (124M)</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 00:03:39 exploring the GPT-2 (124M) OpenAI checkpoint</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 00:13:47 SECTION 1: implementing the GPT-2 nn.Module</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 00:28:08 loading the huggingface/GPT-2 parameters</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 00:31:00 implementing the forward pass to get logits</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 00:33:31 sampling init, prefix tokens, tokenization</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 00:37:02 sampling loop</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 00:41:47 sample, auto-detect the device</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 00:45:50 let’s train: data batches (B,T) → logits (B,T,C)</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 00:52:53 cross entropy loss</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 00:56:42 optimization loop: overfit a single batch</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 01:02:00 data loader lite</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 01:06:14 parameter sharing wte and lm_head</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 01:13:47 model initialization: std 0.02, residual init</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 01:22:18 SECTION 2: Let’s make it fast. GPUs, mixed precision, 1000ms</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 01:28:14 Tensor Cores, timing the code, TF32 precision, 333ms</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 01:39:38 float16, gradient scalers, bfloat16, 300ms</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 01:48:15 torch.compile, Python overhead, kernel fusion, 130ms</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 02:00:18 flash attention, 96ms</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 02:06:54 nice/ugly numbers. vocab size 50257 → 50304, 93ms</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 02:14:55 SECTION 3: hyperpamaters, AdamW, gradient clipping</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 02:21:06 learning rate scheduler: warmup + cosine decay</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 02:26:21 batch size schedule, weight decay, FusedAdamW, 90ms</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 02:34:09 gradient accumulation</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 02:46:52 distributed data parallel (DDP)</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 03:10:21 datasets used in GPT-2, GPT-3, FineWeb (EDU)</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 03:23:10 validation data split, validation loss, sampling revive</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 03:28:23 evaluation: HellaSwag, starting the run</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 03:43:05 SECTION 4: results in the morning! GPT-2, GPT-3 repro</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 03:56:21 shoutout to llm.c, equivalent but faster code in raw C/CUDA</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; 03:59:39 summary, phew, build-nanogpt github repo</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; </ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; Corrections:</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; I will post all errata and followups to the build-nanogpt GitHub repo (link above)</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; </ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; SuperThanks:</ins></div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;br&gt; I experimentally enabled them on my channel yesterday. Totally optional and only use if rich. All revenue goes to to supporting my work in AI + Education.</ins></div></td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr> <tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div> </div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div>&lt;/<ins style="font-weight: bold; text-decoration: none;">html</ins>&gt;</div></td></tr> <tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div><del style="font-weight: bold; text-decoration: none;">305,330 views  Jun 10, 2024</del></div></td><td colspan="2" class="diff-side-added"></td></tr> <tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div> </div></td><td colspan="2" class="diff-side-added"></td></tr> <tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div><del style="font-weight: bold; text-decoration: none;">We reproduce the GPT-2 (124M) from scratch. This video covers the whole process: First we build the GPT-2 network, then we optimize its training to be really fast, then we set up the training run following the GPT-2 and GPT-3 paper and their hyperparameters, then we hit run, and come back the next morning to see our results, and enjoy some amusing model generations. Keep in mind that in some places this video builds on the knowledge from earlier videos in the Zero to Hero Playlist (see my channel). You could also see this video as building my nanoGPT repo, which by the end is about 90% similar. Links: - build-nanogpt GitHub repo, with all the changes in this video as individual commits: </del>&lt;<del style="font-weight: bold; text-decoration: none;">nowiki&gt;https://github.com/karpathy/build-nan&lt;/nowiki&gt;... - nanoGPT repo: &lt;nowiki&gt;https:</del>/<del style="font-weight: bold; text-decoration: none;">/github.com/karpathy/nanoGPT&lt;/nowiki&gt; - llm.c repo: &lt;nowiki</del>&gt;<del style="font-weight: bold; text-decoration: none;">https://github.com/karpathy/llm.c&lt;/nowiki&gt; - my website: &lt;nowiki&gt;https://karpathy.ai&lt;/nowiki&gt; - my twitter:   / karpathy  - our Discord channel:   / discord  Supplementary links: - Attention is All You Need paper: &lt;nowiki&gt;https://arxiv.org/abs/1706.03762&lt;/nowiki&gt; - OpenAI GPT-3 paper: &lt;nowiki&gt;https://arxiv.org/abs/2005.14165&lt;/nowiki&gt; - OpenAI GPT-2 paper: &lt;nowiki&gt;https://d4mucfpksywv.cloudfront.net/b&lt;/nowiki&gt;... The GPU I'm training the model on is from Lambda GPU Cloud, I think the best and easiest way to spin up an on-demand GPU instance in the cloud that you can ssh to: &lt;nowiki&gt;https://lambdalabs.com&lt;/nowiki&gt; Chapters: 00:00:00 intro: Let’s reproduce GPT-2 (124M) 00:03:39 exploring the GPT-2 (124M) OpenAI checkpoint 00:13:47 SECTION 1: implementing the GPT-2 nn.Module 00:28:08 loading the huggingface/GPT-2 parameters 00:31:00 implementing the forward pass to get logits 00:33:31 sampling init, prefix tokens, tokenization 00:37:02 sampling loop 00:41:47 sample, auto-detect the device 00:45:50 let’s train: data batches (B,T) → logits (B,T,C) 00:52:53 cross entropy loss 00:56:42 optimization loop: overfit a single batch 01:02:00 data loader lite 01:06:14 parameter sharing wte and lm_head 01:13:47 model initialization: std 0.02, residual init 01:22:18 SECTION 2: Let’s make it fast. GPUs, mixed precision, 1000ms 01:28:14 Tensor Cores, timing the code, TF32 precision, 333ms 01:39:38 float16, gradient scalers, bfloat16, 300ms 01:48:15 torch.compile, Python overhead, kernel fusion, 130ms 02:00:18 flash attention, 96ms 02:06:54 nice/ugly numbers. vocab size 50257 → 50304, 93ms 02:14:55 SECTION 3: hyperpamaters, AdamW, gradient clipping 02:21:06 learning rate scheduler: warmup + cosine decay 02:26:21 batch size schedule, weight decay, FusedAdamW, 90ms 02:34:09 gradient accumulation 02:46:52 distributed data parallel (DDP) 03:10:21 datasets used in GPT-2, GPT-3, FineWeb (EDU) 03:23:10 validation data split, validation loss, sampling revive 03:28:23 evaluation: HellaSwag, starting the run 03:43:05 SECTION 4: results in the morning! GPT-2, GPT-3 repro 03:56:21 shoutout to llm.c, equivalent but faster code in raw C/CUDA 03:59:39 summary, phew, build-nanogpt github repo Corrections: I will post all errata and followups to the build-nanogpt GitHub repo (link above) SuperThanks: I experimentally enabled them on my channel yesterday. Totally optional and only use if rich. All revenue goes to to supporting my work in AI + Education.</del></div></td><td colspan="2" class="diff-side-added"></td></tr> <!-- diff cache key ph30_wiki:diff:1.41:old-18:rev-19:php=table --> </table>

Admin https://303030.org:443/index.php?title=Let%27s_reproduce_GPT-2_(124M)&diff=18&oldid=prev 2024-06-26T19:16:34Z

<p></p> <table style="background-color: #fff; color: #202122;" data-mw="interface"> <col class="diff-marker" /> <col class="diff-content" /> <col class="diff-marker" /> <col class="diff-content" /> <tr class="diff-title" lang="ar"> <td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">→ مراجعة أقدم</td> <td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">مراجعة ١٤:١٦، ٢٦ يونيو ٢٠٢٤</td> </tr><tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l3">سطر ٣:</td> <td colspan="2" class="diff-lineno">سطر ٣:</td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;hr&gt;</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;hr&gt;</div></td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;html&gt;</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;html&gt;</div></td></tr> <tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div>&lt;iframe width="<del style="font-weight: bold; text-decoration: none;">1522</del>" height="593" src="https://www.youtube.com/embed/l8pRSuU81PU" title="Let&amp;#39;s reproduce GPT-2 (124M)" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen&gt;&lt;/iframe&gt;</div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div>&lt;iframe width="<ins style="font-weight: bold; text-decoration: none;">100%</ins>" height="593" src="https://www.youtube.com/embed/l8pRSuU81PU" title="Let&amp;#39;s reproduce GPT-2 (124M)" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen&gt;&lt;/iframe&gt;</div></td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;/html&gt;</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;/html&gt;</div></td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;hr&gt;</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;hr&gt;</div></td></tr> <!-- diff cache key ph30_wiki:diff:1.41:old-17:rev-18:php=table --> </table>

Admin https://303030.org:443/index.php?title=Let%27s_reproduce_GPT-2_(124M)&diff=17&oldid=prev 2024-06-26T19:16:15Z

<p></p> <table style="background-color: #fff; color: #202122;" data-mw="interface"> <col class="diff-marker" /> <col class="diff-content" /> <col class="diff-marker" /> <col class="diff-content" /> <tr class="diff-title" lang="ar"> <td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">→ مراجعة أقدم</td> <td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">مراجعة ١٤:١٦، ٢٦ يونيو ٢٠٢٤</td> </tr><tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l2">سطر ٢:</td> <td colspan="2" class="diff-lineno">سطر ٢:</td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;hr&gt;</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;hr&gt;</div></td></tr> <tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div> </div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;html&gt;</ins></div></td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;iframe width=&quot;1522&quot; height=&quot;593&quot; src=&quot;https://www.youtube.com/embed/l8pRSuU81PU&quot; title=&quot;Let&amp;#39;s reproduce GPT-2 (124M)&quot; frameborder=&quot;0&quot; allow=&quot;accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share&quot; referrerpolicy=&quot;strict-origin-when-cross-origin&quot; allowfullscreen&gt;&lt;/iframe&gt;</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;iframe width=&quot;1522&quot; height=&quot;593&quot; src=&quot;https://www.youtube.com/embed/l8pRSuU81PU&quot; title=&quot;Let&amp;#39;s reproduce GPT-2 (124M)&quot; frameborder=&quot;0&quot; allow=&quot;accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share&quot; referrerpolicy=&quot;strict-origin-when-cross-origin&quot; allowfullscreen&gt;&lt;/iframe&gt;</div></td></tr> <tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div> </div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;/html&gt;</ins></div></td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;hr&gt;</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;hr&gt;</div></td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr> <!-- diff cache key ph30_wiki:diff:1.41:old-16:rev-17:php=table --> </table>

Admin https://303030.org:443/index.php?title=Let%27s_reproduce_GPT-2_(124M)&diff=16&oldid=prev 2024-06-26T19:11:02Z

<p></p> <table style="background-color: #fff; color: #202122;" data-mw="interface"> <col class="diff-marker" /> <col class="diff-content" /> <col class="diff-marker" /> <col class="diff-content" /> <tr class="diff-title" lang="ar"> <td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">→ مراجعة أقدم</td> <td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">مراجعة ١٤:١١، ٢٦ يونيو ٢٠٢٤</td> </tr><tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l1">سطر ١:</td> <td colspan="2" class="diff-lineno">سطر ١:</td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>https://www.youtube.com/watch?v=l8pRSuU81PU</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>https://www.youtube.com/watch?v=l8pRSuU81PU</div></td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;hr&gt;</ins></div></td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;iframe width=&quot;1522&quot; height=&quot;593&quot; src=&quot;https://www.youtube.com/embed/l8pRSuU81PU&quot; title=&quot;Let&amp;#39;s reproduce GPT-2 (124M)&quot; frameborder=&quot;0&quot; allow=&quot;accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share&quot; referrerpolicy=&quot;strict-origin-when-cross-origin&quot; allowfullscreen&gt;&lt;/iframe&gt;</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;iframe width=&quot;1522&quot; height=&quot;593&quot; src=&quot;https://www.youtube.com/embed/l8pRSuU81PU&quot; title=&quot;Let&amp;#39;s reproduce GPT-2 (124M)&quot; frameborder=&quot;0&quot; allow=&quot;accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share&quot; referrerpolicy=&quot;strict-origin-when-cross-origin&quot; allowfullscreen&gt;&lt;/iframe&gt;</div></td></tr> <!-- diff cache key ph30_wiki:diff:1.41:old-15:rev-16:php=table --> </table>

Admin https://303030.org:443/index.php?title=Let%27s_reproduce_GPT-2_(124M)&diff=15&oldid=prev 2024-06-26T19:09:08Z

<p></p> <table style="background-color: #fff; color: #202122;" data-mw="interface"> <col class="diff-marker" /> <col class="diff-content" /> <col class="diff-marker" /> <col class="diff-content" /> <tr class="diff-title" lang="ar"> <td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">→ مراجعة أقدم</td> <td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">مراجعة ١٤:٠٩، ٢٦ يونيو ٢٠٢٤</td> </tr><tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l4">سطر ٤:</td> <td colspan="2" class="diff-lineno">سطر ٤:</td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;iframe width=&quot;1522&quot; height=&quot;593&quot; src=&quot;https://www.youtube.com/embed/l8pRSuU81PU&quot; title=&quot;Let&amp;#39;s reproduce GPT-2 (124M)&quot; frameborder=&quot;0&quot; allow=&quot;accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share&quot; referrerpolicy=&quot;strict-origin-when-cross-origin&quot; allowfullscreen&gt;&lt;/iframe&gt;</div></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><div>&lt;iframe width=&quot;1522&quot; height=&quot;593&quot; src=&quot;https://www.youtube.com/embed/l8pRSuU81PU&quot; title=&quot;Let&amp;#39;s reproduce GPT-2 (124M)&quot; frameborder=&quot;0&quot; allow=&quot;accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share&quot; referrerpolicy=&quot;strict-origin-when-cross-origin&quot; allowfullscreen&gt;&lt;/iframe&gt;</div></td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div><ins style="font-weight: bold; text-decoration: none;">&lt;hr&gt;</ins></div></td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr> <!-- diff cache key ph30_wiki:diff:1.41:old-14:rev-15:php=table --> </table>

Admin https://303030.org:443/index.php?title=Let%27s_reproduce_GPT-2_(124M)&diff=14&oldid=prev 2024-06-26T19:07:56Z

<p></p> <table style="background-color: #fff; color: #202122;" data-mw="interface"> <col class="diff-marker" /> <col class="diff-content" /> <col class="diff-marker" /> <col class="diff-content" /> <tr class="diff-title" lang="ar"> <td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">→ مراجعة أقدم</td> <td colspan="2" style="background-color: #fff; color: #202122; text-align: center;">مراجعة ١٤:٠٧، ٢٦ يونيو ٢٠٢٤</td> </tr><tr><td colspan="2" class="diff-lineno" id="mw-diff-left-l2">سطر ٢:</td> <td colspan="2" class="diff-lineno">سطر ٢:</td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr> <tr><td class="diff-marker" data-marker="−"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #ffe49c; vertical-align: top; white-space: pre-wrap;"><div>&lt;iframe width="1522" height="593" src="https://www.youtube.com/embed/l8pRSuU81PU" title="Let&amp;<del style="font-weight: bold; text-decoration: none;">amp;</del>#39;s reproduce GPT-2 (124M)" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen&gt;&lt;/iframe&gt;</div></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div>&lt;iframe width="1522" height="593" src="https://www.youtube.com/embed/l8pRSuU81PU" title="Let&amp;#39;s reproduce GPT-2 (124M)" frameborder="0" allow="accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share" referrerpolicy="strict-origin-when-cross-origin" allowfullscreen&gt;&lt;/iframe&gt;</div></td></tr> <tr><td colspan="2" class="diff-side-deleted"></td><td class="diff-marker" data-marker="+"></td><td style="color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #a3d3ff; vertical-align: top; white-space: pre-wrap;"><div> </div></td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr> <tr><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td><td class="diff-marker"></td><td style="background-color: #f8f9fa; color: #202122; font-size: 88%; border-style: solid; border-width: 1px 1px 1px 4px; border-radius: 0.33em; border-color: #eaecf0; vertical-align: top; white-space: pre-wrap;"><br></td></tr> <!-- diff cache key ph30_wiki:diff:1.41:old-13:rev-14:php=table --> </table>

Admin https://303030.org:443/index.php?title=Let%27s_reproduce_GPT-2_(124M)&diff=13&oldid=prev 2024-06-26T19:06:58Z

<p>أنشأ الصفحة ب&#039;https://www.youtube.com/watch?v=l8pRSuU81PU &lt;iframe width=&quot;1522&quot; height=&quot;593&quot; src=&quot;https://www.youtube.com/embed/l8pRSuU81PU&quot; title=&quot;Let&amp;#39;s reproduce GPT-2 (124M)&quot; frameborder=&quot;0&quot; allow=&quot;accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share&quot; referrerpolicy=&quot;strict-origin-when-cross-origin&quot; allowfullscreen&gt;&lt;/iframe&gt; 305,330 views Jun 10, 2024 We reproduce the GPT-2 (124M) from scratch. This video covers the...&#039;</p> <p><b>صفحة جديدة</b></p><div>https://www.youtube.com/watch?v=l8pRSuU81PU<br /> <br /> <br /> &lt;iframe width=&quot;1522&quot; height=&quot;593&quot; src=&quot;https://www.youtube.com/embed/l8pRSuU81PU&quot; title=&quot;Let&amp;amp;#39;s reproduce GPT-2 (124M)&quot; frameborder=&quot;0&quot; allow=&quot;accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share&quot; referrerpolicy=&quot;strict-origin-when-cross-origin&quot; allowfullscreen&gt;&lt;/iframe&gt;<br /> <br /> <br /> 305,330 views Jun 10, 2024<br /> <br /> We reproduce the GPT-2 (124M) from scratch. This video covers the whole process: First we build the GPT-2 network, then we optimize its training to be really fast, then we set up the training run following the GPT-2 and GPT-3 paper and their hyperparameters, then we hit run, and come back the next morning to see our results, and enjoy some amusing model generations. Keep in mind that in some places this video builds on the knowledge from earlier videos in the Zero to Hero Playlist (see my channel). You could also see this video as building my nanoGPT repo, which by the end is about 90% similar. Links: - build-nanogpt GitHub repo, with all the changes in this video as individual commits: &lt;nowiki&gt;https://github.com/karpathy/build-nan&lt;/nowiki&gt;... - nanoGPT repo: &lt;nowiki&gt;https://github.com/karpathy/nanoGPT&lt;/nowiki&gt; - llm.c repo: &lt;nowiki&gt;https://github.com/karpathy/llm.c&lt;/nowiki&gt; - my website: &lt;nowiki&gt;https://karpathy.ai&lt;/nowiki&gt; - my twitter:   / karpathy  - our Discord channel:   / discord  Supplementary links: - Attention is All You Need paper: &lt;nowiki&gt;https://arxiv.org/abs/1706.03762&lt;/nowiki&gt; - OpenAI GPT-3 paper: &lt;nowiki&gt;https://arxiv.org/abs/2005.14165&lt;/nowiki&gt; - OpenAI GPT-2 paper: &lt;nowiki&gt;https://d4mucfpksywv.cloudfront.net/b&lt;/nowiki&gt;... The GPU I&#039;m training the model on is from Lambda GPU Cloud, I think the best and easiest way to spin up an on-demand GPU instance in the cloud that you can ssh to: &lt;nowiki&gt;https://lambdalabs.com&lt;/nowiki&gt; Chapters: 00:00:00 intro: Let’s reproduce GPT-2 (124M) 00:03:39 exploring the GPT-2 (124M) OpenAI checkpoint 00:13:47 SECTION 1: implementing the GPT-2 nn.Module 00:28:08 loading the huggingface/GPT-2 parameters 00:31:00 implementing the forward pass to get logits 00:33:31 sampling init, prefix tokens, tokenization 00:37:02 sampling loop 00:41:47 sample, auto-detect the device 00:45:50 let’s train: data batches (B,T) → logits (B,T,C) 00:52:53 cross entropy loss 00:56:42 optimization loop: overfit a single batch 01:02:00 data loader lite 01:06:14 parameter sharing wte and lm_head 01:13:47 model initialization: std 0.02, residual init 01:22:18 SECTION 2: Let’s make it fast. GPUs, mixed precision, 1000ms 01:28:14 Tensor Cores, timing the code, TF32 precision, 333ms 01:39:38 float16, gradient scalers, bfloat16, 300ms 01:48:15 torch.compile, Python overhead, kernel fusion, 130ms 02:00:18 flash attention, 96ms 02:06:54 nice/ugly numbers. vocab size 50257 → 50304, 93ms 02:14:55 SECTION 3: hyperpamaters, AdamW, gradient clipping 02:21:06 learning rate scheduler: warmup + cosine decay 02:26:21 batch size schedule, weight decay, FusedAdamW, 90ms 02:34:09 gradient accumulation 02:46:52 distributed data parallel (DDP) 03:10:21 datasets used in GPT-2, GPT-3, FineWeb (EDU) 03:23:10 validation data split, validation loss, sampling revive 03:28:23 evaluation: HellaSwag, starting the run 03:43:05 SECTION 4: results in the morning! GPT-2, GPT-3 repro 03:56:21 shoutout to llm.c, equivalent but faster code in raw C/CUDA 03:59:39 summary, phew, build-nanogpt github repo Corrections: I will post all errata and followups to the build-nanogpt GitHub repo (link above) SuperThanks: I experimentally enabled them on my channel yesterday. Totally optional and only use if rich. All revenue goes to to supporting my work in AI + Education.</div>

Admin