After spending few hours on reading the code shared on github, I found that maybe the main point is in the file 'featurizers.h' function 'FeaturizeHloInstruction'
\nFEATURIZE_FUNCTION_PAIR_2ARG(\n FeaturizeHloInstruction, HloInstructionProto, HloComputationProto, {\n \n \n ..\nBut when I add some code in it, for example, I pick out the code out->push_back(FEATURIZE_SWITCH( \"indices_are_sorted\", source1.indices_are_sorted() ? 1. : 0.)); from the 'if' section. I got an error \nF tpu_graphs/process_data/xla/hlo_encoder.cc:144] Check failed: values_added_ < out_values_->NumElements() (5646480 vs. 5646480)\n
After spending few hours on reading the code shared on github, I found that maybe the main point is in the file 'featurizers.h' function 'FeaturizeHloInstruction'
\nFEATURIZE_FUNCTION_PAIR_2ARG(\n FeaturizeHloInstruction, HloInstructionProto, HloComputationProto, {\n \n \n ..\nBut when I add some code in it, for example, I pick out the code out->push_back(FEATURIZE_SWITCH( \"indices_are_sorted\", source1.indices_are_sorted() ? 1. : 0.)); from the 'if' section. I got an error \nF tpu_graphs/process_data/xla/hlo_encoder.cc:144] Check failed: values_added_ < out_values_->NumElements() (5646480 vs. 5646480)\n
You also need to modify the node feature size to match the change here: https://github.com/google-research-datasets/tpu_graphs/blob/main/tpu_graphs/process_data/xla/hlo_opcode.h#L33
", "rawMarkdown": "You also need to modify the node feature size to match the change here: https://github.com/google-research-datasets/tpu_graphs/blob/main/tpu_graphs/process_data/xla/hlo_opcode.h#L33", "votes": null }, { "id": "2493259", "postDate": "10/23/2023 08:57:33", "content": "Hi @tonyarobertson, I first came across you guys during the LLM competition and, honestly, I felt a bit gutted about how things turned out for you there. However, I've seen that the Avengers have bounced back remarkably in just a few days in this competition. Massive congrats on that! Did you manage to integrate some new features into your strategy?
", "rawMarkdown": "Hi @tonyarobertson, I first came across you guys during the LLM competition and, honestly, I felt a bit gutted about how things turned out for you there. However, I've seen that the Avengers have bounced back remarkably in just a few days in this competition. Massive congrats on that! Did you manage to integrate some new features into your strategy?", "votes": null }, { "id": "2493694", "postDate": "10/23/2023 14:26:12", "content": "Thanks to your kind word. We have not tried to train any model based on additional features. Maybe we will try it later.
", "rawMarkdown": "Thanks to your kind word. We have not tried to train any model based on additional features. Maybe we will try it later.", "votes": null }, { "id": "2493709", "postDate": "10/23/2023 14:30:25", "content": "Thanks. We are still in the process of boosting LB scores for the original features. But new features have been extracted and may be used in the future.
", "rawMarkdown": "Thanks. We are still in the process of boosting LB scores for the original features. But new features have been extracted and may be used in the future.", "votes": null } ], "comments": [ { "id": 2492829, "author_name": "mangpophothilimthana", "author_url": "", "post_date": "10/22/2023 23:35:26", "content": "You also need to modify the node feature size to match the change here: https://github.com/google-research-datasets/tpu_graphs/blob/main/tpu_graphs/process_data/xla/hlo_opcode.h#L33
", "votes": null, "replies": [ { "id": 2493709, "author_name": "rdxsun", "author_url": "", "post_date": "10/23/2023 14:30:25", "content": "Thanks. We are still in the process of boosting LB scores for the original features. But new features have been extracted and may be used in the future.
", "votes": null, "replies": [] } ] }, { "id": 2493259, "author_name": "chitra077", "author_url": "", "post_date": "10/23/2023 08:57:33", "content": "Hi @tonyarobertson, I first came across you guys during the LLM competition and, honestly, I felt a bit gutted about how things turned out for you there. However, I've seen that the Avengers have bounced back remarkably in just a few days in this competition. Massive congrats on that! Did you manage to integrate some new features into your strategy?
", "votes": null, "replies": [ { "id": 2493694, "author_name": "tonyarobertson", "author_url": "", "post_date": "10/23/2023 14:26:12", "content": "Thanks to your kind word. We have not tried to train any model based on additional features. Maybe we will try it later.
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "2488816": "After spending few hours on reading the code shared on [github](https://github.com/google-research-datasets/tpu_graphs), I found that maybe the main point is in the file 'featurizers.h' function 'FeaturizeHloInstruction'\n```\nFEATURIZE_FUNCTION_PAIR_2ARG(\n FeaturizeHloInstruction, HloInstructionProto, HloComputationProto, {\n // Add a feature to indicate whether or not this instruction is the\n // computation's root instruction\n .....\n```\nBut when I add some code in it, for example, I pick out the code `out->push_back(FEATURIZE_SWITCH( \"indices_are_sorted\", source1.indices_are_sorted() ? 1. : 0.));` from the 'if' section. I got an error `\nF tpu_graphs/process_data/xla/hlo_encoder.cc:144] Check failed: values_added_ < out_values_->NumElements() (5646480 vs. 5646480)\n`", "2492829": "You also need to modify the node feature size to match the change here: https://github.com/google-research-datasets/tpu_graphs/blob/main/tpu_graphs/process_data/xla/hlo_opcode.h#L33", "2493259": "Hi @tonyarobertson, I first came across you guys during the LLM competition and, honestly, I felt a bit gutted about how things turned out for you there. However, I've seen that the Avengers have bounced back remarkably in just a few days in this competition. Massive congrats on that! Did you manage to integrate some new features into your strategy?", "2493694": "Thanks to your kind word. We have not tried to train any model based on additional features. Maybe we will try it later.", "2493709": "Thanks. We are still in the process of boosting LB scores for the original features. But new features have been extracted and may be used in the future." }, "source": "meta" }