article_summaries¶
news_recommendation.article_summaries
Summarize each retained article as a whole with Qwen (greedy decoding, frozen prompt).
python -m news_recommendation.article_summaries run --output runs/article-summaries
The configuration uses two GPUs. Ordinary workers skip inputs over 18,000 tokens. The long stage splits the layers over
two GPUs for every article still missing, and recovery retries what remains with a larger output cap. A completion
must end with EOS; one that hits the cap is retried alone with twice the cap. Each stage appends to
batches(prepared, single=False)
¶
Consecutive length-sorted inputs, fewer per batch as they grow, never mixing very different lengths.
Source code in src/news_recommendation/article_summaries.py
21 22 23 24 25 26 27 28 29 30 31 32 | |
prepare(config, output)
¶
Copy the config and prompt into output and deal the retained articles into one shard per GPU.
Source code in src/news_recommendation/article_summaries.py
35 36 37 38 39 40 41 42 43 44 45 46 | |
worker(output, stage, worker_id=0)
¶
Summarize one shard (stage "worker"), or every article still missing (stages "long" and "recovery").
Source code in src/news_recommendation/article_summaries.py
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 | |
main()
¶
Source code in src/news_recommendation/article_summaries.py
125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 | |