Datasets:

---

renwei2024

/

nemo-github-issues

like 0

[ { "id": 4556669155, "node_id": "LA_kwDOC_bI7s8AAAABD5lA4w", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/cherry-pick", "name": "cherry-pick", "color": "ededed", "default": false, "description": null }, { "id": 4811886210, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ogg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/NLP", "name": "NLP", "color": "CBF09A", "default": false, "description": "" } ]

[]

[ "jenkins" ]

[ { "id": 4811886210, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ogg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/NLP", "name": "NLP", "color": "CBF09A", "default": false, "description": "" } ]

[]

[ "jenkins", "jenkins" ]

[]

[]

[ "@prathwishshetty \r\nI have raised a PR for this at https://github.com/NVIDIA/NeMo/pull/8498\r\nLet's wait to get it merged soon. Thanks for raising PR", "This can be closed as PR gets merged", "Thank you for the quick PR!" ]

[ { "id": 1485815557, "node_id": "MDU6TGFiZWwxNDg1ODE1NTU3", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/bug", "name": "bug", "color": "d73a4a", "default": true, "description": "Something isn't working" } ]

[ { "login": "dimapihtar", "id": 37850217, "node_id": "MDQ6VXNlcjM3ODUwMjE3", "avatar_url": "https://avatars.githubusercontent.com/u/37850217?v=4", "gravatar_id": "", "url": "https://api.github.com/users/dimapihtar", "html_url": "https://github.com/dimapihtar", "followers_url": "https://api.github.com/users/dimapihtar/followers", "following_url": "https://api.github.com/users/dimapihtar/following{/other_user}", "gists_url": "https://api.github.com/users/dimapihtar/gists{/gist_id}", "starred_url": "https://api.github.com/users/dimapihtar/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/dimapihtar/subscriptions", "organizations_url": "https://api.github.com/users/dimapihtar/orgs", "repos_url": "https://api.github.com/users/dimapihtar/repos", "events_url": "https://api.github.com/users/dimapihtar/events{/privacy}", "received_events_url": "https://api.github.com/users/dimapihtar/received_events", "type": "User", "user_view_type": "public", "site_admin": false } ]

[ "Same issue in FSDP+SFT.\r\nThe loss is 3-4 times higher than TP, PP, and DP cases. I doubted fused_adam at first (currently, FSDP cannot apply distributed_fused_adam) but, fused_adam worked well except for FSDP.\r\n\r\n### Environment details\r\n\r\nDocker Container: `nvcr.io/nvidia/nemo:24.01.01.framework`\r\n\r\n```\r\n#!/bin/sh\r\n\r\nexport HYDRA_FULL_ERROR=1\r\n\r\nexport WANDB=False\r\nexport PJ_NAME=\"databricks-dolly-15k-ja\"\r\nexport EXP_DIR=\"/raid/results/\"${PJ_NAME}\r\nexport EXP_NAME=\"Swallow-MS-7b-v0.1_SFT\"\r\nexport MODEL=\"/workspace/models/Swallow-MS-7b-v0.1/Swallow-MS-7b-v0.1.nemo\"\r\nexport TP_SIZE=1\r\nexport PP_SIZE=1\r\nexport TRAIN_DS=\"[/workspace/data/\"${PJ_NAME}\"/train.jsonl]\"\r\nexport VALID_DS=\"[/workspace/data/\"${PJ_NAME}\"/valid.jsonl]\"\r\nexport TEST_DS=\"[/workspace/data/\"${PJ_NAME}\"/test.jsonl]\"\r\n\r\ntorchrun --nproc_per_node=8 /opt/NeMo/examples/nlp/language_modeling/tuning/megatron_gpt_finetuning.py \\\r\n trainer.precision=bf16-mixed \\\r\n trainer.devices=8 \\\r\n trainer.num_nodes=1 \\\r\n trainer.max_epochs=5 \\\r\n trainer.max_steps=938 \\\r\n trainer.val_check_interval=0.1 \\\r\n model.peft.peft_scheme=null \\\r\n model.restore_from_path=${MODEL} \\\r\n model.global_batch_size=64 \\\r\n model.micro_batch_size=4 \\\r\n model.tensor_model_parallel_size=${TP_SIZE} \\\r\n model.pipeline_model_parallel_size=${PP_SIZE} \\\r\n model.megatron_amp_O2=False \\\r\n model.fsdp=True \\\r\n model.fsdp_sharding_strategy='full' \\\r\n model.fsdp_grad_reduce_dtype='bf16' \\\r\n model.fsdp_sharded_checkpoint=False \\\r\n model.fsdp_use_orig_params=False \\\r\n model.optim.name=fused_adam \\\r\n model.optim.lr=2e-5 \\\r\n model.answer_only_loss=True \\\r\n model.data.train_ds.file_names=${TRAIN_DS} \\\r\n model.data.train_ds.concat_sampling_probabilities=\"[1]\" \\\r\n model.data.train_ds.max_seq_length=2048 \\\r\n model.data.train_ds.num_workers=0 \\\r\n model.data.validation_ds.file_names=${VALID_DS} \\\r\n model.data.validation_ds.max_seq_length=2048 \\\r\n model.data.validation_ds.num_workers=0 \\\r\n model.data.validation_ds.metric.name=loss \\\r\n model.data.test_ds.file_names=${TEST_DS} \\\r\n exp_manager.explicit_log_dir=${EXP_DIR} \\\r\n exp_manager.name=${EXP_NAME} \\\r\n exp_manager.create_wandb_logger=${WANDB} \\\r\n exp_manager.wandb_logger_kwargs.project=${PJ_NAME} \\\r\n exp_manager.wandb_logger_kwargs.name=${EXP_NAME} \\\r\n exp_manager.checkpoint_callback_params.save_nemo_on_train_end=True \\\r\n exp_manager.resume_if_exists=True \\\r\n exp_manager.resume_ignore_no_checkpoint=True \\\r\n exp_manager.create_checkpoint_callback=True \\\r\n exp_manager.checkpoint_callback_params.monitor=validation_loss \\\r\n exp_manager.early_stopping_callback_params.patience=5 \\\r\n ++cluster_type=BCP\r\n```", "Hi. Looks like this issue is fixed in 24.03 release container. I can't reproduce this divergence with 24.03.", "> Hi. Looks like this issue is fixed in 24.03 release container. I can't reproduce this divergence with 24.03.\r\n\r\nDid you test FSDP + SFT or FSDP + LoRA?", "@dimapihtar I tried the FSDP + LoRA tuning with the same experiment but on the 24.03 container and it still has the same train loss issue. ", "@dimapihtar FSDP+SFT on the 24.03 container also has the same training loss issue.", "I also have very significant loss differences in pre-training when comparing FSDP vs. DDP, with FSDP being worse. I'm using FP32 gradient reduction with FSDP and the original params (i.e., `model.fsdp_grad_reduce_dtype=32` and `model.fsdp_use_orig_params=True`).\r\nNote, this difference does not appear when using FSDP without sharding (i.e., `sharding_strategy=ShardingStrategy.NO_SHARD` (this has no config setting ATM)) or with `fsdp_use_orig_params=False`.", "The problem is a double all-reduce on the sharded parameter gradients when setting `fsdp_use_orig_params=True`, which can be fixed for the case of `tensor_model_parallel_size=1` by (in `megatron_gpt_model.py`):\r\n```patch\r\n self.megatron_timer_start('gradient_allreduce', log_level=1)\r\n- if self.use_fsdp:\r\n+ if self.use_fsdp and self.cfg.get('fsdp_use_orig_params', False):\r\n+ assert self.cfg.get('tensor_model_parallel_size', 1) == 1, 'fix only works for TP=1'\r\n+ elif self.use_fsdp:\r\n # Reduce the gradients omitted from FSDP-sharding\r\n self.allreduce_fsdp_sharding_omitted_gradients()\r\n elif self.with_distributed_adam:\r\n```\r\n\r\nI still need to find a nice way to fix it with `tensor_model_parallel_size > 1`.", "The fix for `tensor_model_parallel_size >= 1` can be implemented very simply by accessing a private variable of the FSDP module.\r\nAgain in `megatron_gpt_model.py`, and the patch from the previous comment should not be included, so this is the proposed solution for all cases:\r\n```patch\r\n grads = []\r\n- for param in self.model.parameters():\r\n- if not isinstance(param, torch.distributed.fsdp.FlatParameter) and param.requires_grad:\r\n+ for param in self.model._ignored_params:\r\n+ if param.requires_grad and param.grad is not None:\r\n grad = param.grad\r\n grads.append(grad.data)\r\n```" ]

[ { "id": 1923495043, "node_id": "MDU6TGFiZWwxOTIzNDk1MDQz", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/core", "name": "core", "color": "f25e3a", "default": false, "description": "Changes to NeMo Core" }, { "id": 4418253970, "node_id": "LA_kwDOC_bI7s8AAAABB1k0kg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/TTS", "name": "TTS", "color": "49AF28", "default": false, "description": "" }, { "id": 4811884691, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ikw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/ASR", "name": "ASR", "color": "F3B33E", "default": false, "description": "" }, { "id": 4811886210, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ogg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/NLP", "name": "NLP", "color": "CBF09A", "default": false, "description": "" }, { "id": 4811888028, "node_id": "LA_kwDOC_bI7s8AAAABHs-VnA", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/Speaker%20Tasks", "name": "Speaker Tasks", "color": "871AD4", "default": false, "description": "" }, { "id": 4840216855, "node_id": "LA_kwDOC_bI7s8AAAABIH_ZFw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/CI", "name": "CI", "color": "1D76DB", "default": false, "description": "" }, { "id": 4847373924, "node_id": "LA_kwDOC_bI7s8AAAABIO0OZA", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/common", "name": "common", "color": "c5def5", "default": false, "description": "" } ]

[]

[]

[ { "id": 1485815557, "node_id": "MDU6TGFiZWwxNDg1ODE1NTU3", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/bug", "name": "bug", "color": "d73a4a", "default": true, "description": "Something isn't working" }, { "id": 4585431311, "node_id": "LA_kwDOC_bI7s8AAAABEVAhDw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/stale", "name": "stale", "color": "ededed", "default": false, "description": null } ]

[]

[ "What batch size are you using for training ? Can you try grad accumulation to make your global batch size at least 128 ?", "Thanks @titu1994 , I am using:\r\naccumulate_grad_batches=18\r\nbatch_size=15\r\n2 GPUs.\r\n So if I understand correctly, the global batch size = 18 * 15 * 2 = 540, right?", "Yes that much should be enough. It seems you're training from scratch, can you use a higher learning rate of 0.001 and see if it's learning something useful? \r\n\r\nBoth plots above are just learning blank tokens, nothing more ", "> Yes that much should be enough. It seems you're training from scratch, can you use a higher learning rate of 0.001 and see if it's learning something useful?\r\n\r\nYeah, I train from scratch this is an essential for the project.\r\nI'll try, I actually only tried to reduce the learning rate, not increase it.\r\n\r\n> Both plots above are just learning blank tokens, nothing more\r\n\r\nWhat do you mean, I do have some good predictions in the logs, maybe I am missing something:\r\n```\r\n[NeMo I 2024-02-22 00:35:33 wer:319] reference:i have a strong feeling about this meeting\r\n[NeMo I 2024-02-22 00:35:33 wer:320] predicted:i have a strong feeling about this meeting\r\n\r\n[NeMo I 2024-02-22 00:35:36 wer:319] reference:where can i find gifts for my daughter near by\r\n[NeMo I 2024-02-22 00:35:36 wer:320] predicted:onhe day fi so duhter near by\r\n\r\n[NeMo I 2024-02-22 00:35:41 wer:319] reference:come on mia follow me he said\r\n[NeMo I 2024-02-22 00:35:41 wer:320] predicted:em ci said\r\n```\r\n\r\nThanks a lot for your help @titu1994 ! I would appreciate any other ideas that you have.\r\n", "Your training batch wer is ~ 1, that would mean your model is learning nothing yet other than blank tokens. The text here - is it from later in training ?", "Thanks @titu1994 I do manage to get some better results, but after some training the model just collapses:\r\n<img width=\"438\" alt=\"image\" src=\"https://github.com/NVIDIA/NeMo/assets/8983713/f0c20ab5-31d4-4958-ab67-ecc2c1360a26\">\r\n\r\nFunny thing is that it depends on luck, some time if I resume from a previous checkpoint the model sometimes manage to continue converging (and then maybe collapse again and I try to continue from checkpoint once more).\r\n<img width=\"461\" alt=\"image\" src=\"https://github.com/NVIDIA/NeMo/assets/8983713/c9ce2239-7566-4ede-ae83-ba3f84f1b2c3\">\r\nanything that I can add to increase stability ?\r\nThanks", "This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days.", "This issue was closed because it has been inactive for 7 days since being marked as stale." ]

[ { "id": 4418253970, "node_id": "LA_kwDOC_bI7s8AAAABB1k0kg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/TTS", "name": "TTS", "color": "49AF28", "default": false, "description": "" } ]

[]

[ "jenkins", "jenkins", "jenkins", "jenkins" ]

[ { "id": 1923495043, "node_id": "MDU6TGFiZWwxOTIzNDk1MDQz", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/core", "name": "core", "color": "f25e3a", "default": false, "description": "Changes to NeMo Core" }, { "id": 4418253970, "node_id": "LA_kwDOC_bI7s8AAAABB1k0kg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/TTS", "name": "TTS", "color": "49AF28", "default": false, "description": "" }, { "id": 4811884691, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ikw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/ASR", "name": "ASR", "color": "F3B33E", "default": false, "description": "" }, { "id": 4811886210, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ogg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/NLP", "name": "NLP", "color": "CBF09A", "default": false, "description": "" }, { "id": 4840216855, "node_id": "LA_kwDOC_bI7s8AAAABIH_ZFw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/CI", "name": "CI", "color": "1D76DB", "default": false, "description": "" }, { "id": 4847373924, "node_id": "LA_kwDOC_bI7s8AAAABIO0OZA", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/common", "name": "common", "color": "c5def5", "default": false, "description": "" } ]

[]

[ "jenkins" ]

[ { "id": 4556669155, "node_id": "LA_kwDOC_bI7s8AAAABD5lA4w", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/cherry-pick", "name": "cherry-pick", "color": "ededed", "default": false, "description": null }, { "id": 4811886210, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ogg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/NLP", "name": "NLP", "color": "CBF09A", "default": false, "description": "" } ]

[]

[ "jenkins" ]

[ { "id": 4811886210, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ogg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/NLP", "name": "NLP", "color": "CBF09A", "default": false, "description": "" } ]

[]

[ "jenkins", "jenkins", "jenkins" ]

[ { "id": 1923495043, "node_id": "MDU6TGFiZWwxOTIzNDk1MDQz", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/core", "name": "core", "color": "f25e3a", "default": false, "description": "Changes to NeMo Core" } ]

[]

[ "This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.", "@stevehuang52 this problem basically happens when we try to resume 2nd time right ? ", "> @stevehuang52 this problem basically happens when we try to resume 2nd time right ?\r\n\r\n@athitten Yes, it happens when we resume for the 2nd time, the first resume is fine", "jenkins", "jenkins", "jenkins", "jenkins", "@athitten seems that Jenkins passed without issues. Could you please review the PR again and see if there's other changes or tests needed? Thanks~" ]

[ { "id": 4811886210, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ogg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/NLP", "name": "NLP", "color": "CBF09A", "default": false, "description": "" } ]

[]

[ "jenkins", "jenkins", "jenkins" ]

[]

[]

[ "jenkins", "jenkins", "@GNroy FYI: I am able to build a container from `nvcr.io/nvidia/pytorch:23.09-py3` using this commit, but `nvcr.io/nvidia/pytorch:23.10-py3` and higher fails.\r\nLocally I can install k2 with pytorch 2.2 (compiled with Cuda 12.1)" ]

[ { "id": 4585431311, "node_id": "LA_kwDOC_bI7s8AAAABEVAhDw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/stale", "name": "stale", "color": "ededed", "default": false, "description": null }, { "id": 4811884691, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ikw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/ASR", "name": "ASR", "color": "F3B33E", "default": false, "description": "" }, { "id": 4847373924, "node_id": "LA_kwDOC_bI7s8AAAABIO0OZA", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/common", "name": "common", "color": "c5def5", "default": false, "description": "" } ]

[]

[ "~Note: requires some fixes in lhotse, will bump the requirements as soon as I release it.~ EDIT: done", "jenkins", "jenkins", "This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.", "This PR was closed because it has been inactive for 7 days since being marked as stale.", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "Everything passed in jenkins, merging" ]

[ { "id": 4840216855, "node_id": "LA_kwDOC_bI7s8AAAABIH_ZFw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/CI", "name": "CI", "color": "1D76DB", "default": false, "description": "" } ]

[]

[ "This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.", "This PR was closed because it has been inactive for 7 days since being marked as stale.", "@titu1994, @ericharper To make this finally work we need a separate token with minimal permissions:\r\n- Read access to metadata\r\n- Read and Write access to actions and code\r\n\r\nThe token will have fewer permissions than the default `GITHUB_TOKEN`, so I do not think that using custom token introduces any new issues.\r\n\r\nThe reason why we need a custom token is that `GITHUB_TOKEN` cannot trigger workflows, see https://github.com/orgs/community/discussions/25702 for details.\r\n\r\nI tested the solution on the separate repository, and everything works (code reformatted, pushed if needed, after push the workflows are triggered).", "Test changes from fork: #8996 " ]

[ { "id": 1923495043, "node_id": "MDU6TGFiZWwxOTIzNDk1MDQz", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/core", "name": "core", "color": "f25e3a", "default": false, "description": "Changes to NeMo Core" }, { "id": 4418253970, "node_id": "LA_kwDOC_bI7s8AAAABB1k0kg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/TTS", "name": "TTS", "color": "49AF28", "default": false, "description": "" }, { "id": 4811884691, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ikw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/ASR", "name": "ASR", "color": "F3B33E", "default": false, "description": "" }, { "id": 4811886210, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ogg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/NLP", "name": "NLP", "color": "CBF09A", "default": false, "description": "" }, { "id": 4811888028, "node_id": "LA_kwDOC_bI7s8AAAABHs-VnA", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/Speaker%20Tasks", "name": "Speaker Tasks", "color": "871AD4", "default": false, "description": "" }, { "id": 4840216855, "node_id": "LA_kwDOC_bI7s8AAAABIH_ZFw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/CI", "name": "CI", "color": "1D76DB", "default": false, "description": "" }, { "id": 4847373924, "node_id": "LA_kwDOC_bI7s8AAAABIO0OZA", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/common", "name": "common", "color": "c5def5", "default": false, "description": "" } ]

[]

[]

[ { "id": 1485815557, "node_id": "MDU6TGFiZWwxNDg1ODE1NTU3", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/bug", "name": "bug", "color": "d73a4a", "default": true, "description": "Something isn't working" }, { "id": 4585431311, "node_id": "LA_kwDOC_bI7s8AAAABEVAhDw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/stale", "name": "stale", "color": "ededed", "default": false, "description": null } ]

[]

[ "This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days.", "This issue was closed because it has been inactive for 7 days since being marked as stale." ]

[ { "id": 4585431311, "node_id": "LA_kwDOC_bI7s8AAAABEVAhDw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/stale", "name": "stale", "color": "ededed", "default": false, "description": null } ]

[]

[ "This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days.", "This issue was closed because it has been inactive for 7 days since being marked as stale." ]

[ { "id": 1485815557, "node_id": "MDU6TGFiZWwxNDg1ODE1NTU3", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/bug", "name": "bug", "color": "d73a4a", "default": true, "description": "Something isn't working" }, { "id": 4585431311, "node_id": "LA_kwDOC_bI7s8AAAABEVAhDw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/stale", "name": "stale", "color": "ededed", "default": false, "description": null } ]

[ { "login": "krishnacpuvvada", "id": 93558329, "node_id": "U_kgDOBZOWOQ", "avatar_url": "https://avatars.githubusercontent.com/u/93558329?v=4", "gravatar_id": "", "url": "https://api.github.com/users/krishnacpuvvada", "html_url": "https://github.com/krishnacpuvvada", "followers_url": "https://api.github.com/users/krishnacpuvvada/followers", "following_url": "https://api.github.com/users/krishnacpuvvada/following{/other_user}", "gists_url": "https://api.github.com/users/krishnacpuvvada/gists{/gist_id}", "starred_url": "https://api.github.com/users/krishnacpuvvada/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/krishnacpuvvada/subscriptions", "organizations_url": "https://api.github.com/users/krishnacpuvvada/orgs", "repos_url": "https://api.github.com/users/krishnacpuvvada/repos", "events_url": "https://api.github.com/users/krishnacpuvvada/events{/privacy}", "received_events_url": "https://api.github.com/users/krishnacpuvvada/received_events", "type": "User", "user_view_type": "public", "site_admin": false } ]

[ "I'm in the same situation. I get : \r\n\r\n```\r\n transcript = asr_model.transcribe(batch, batch_size=1)\r\n File \"/home/ubuntu/miniconda3/envs/nemo202403/lib/python3.10/site-packages/torch/utils/_contextlib.py\", line 115, in decorate_context\r\n return func(*args, **kwargs)\r\n File \"/home/ubuntu/miniconda3/envs/nemo202403/lib/python3.10/site-packages/nemo/collections/asr/models/rnnt_models.py\", line 286, in transcribe\r\n fp.write(json.dumps(entry) + '\\n')\r\n File \"/home/ubuntu/miniconda3/envs/nemo202403/lib/python3.10/json/__init__.py\", line 231, in dumps\r\n return _default_encoder.encode(obj)\r\n File \"/home/ubuntu/miniconda3/envs/nemo202403/lib/python3.10/json/encoder.py\", line 199, in encode\r\n chunks = self.iterencode(o, _one_shot=True)\r\n File \"/home/ubuntu/miniconda3/envs/nemo202403/lib/python3.10/json/encoder.py\", line 257, in iterencode\r\n return _iterencode(o, 0)\r\n File \"/home/ubuntu/miniconda3/envs/nemo202403/lib/python3.10/json/encoder.py\", line 179, in default\r\n raise TypeError(f'Object of type {o.__class__.__name__} '\r\nTypeError: Object of type Tensor is not JSON serializable\r\n```\r\n", "@jprobichaud\r\n\r\nThe only way i got it working is by writing temporary files i can use. (json manifest + audio file)\r\n\r\nYou can have a look at my code here:\r\nhttps://github.com/Sharrnah/whispering/blob/main/Models/STT/nemo_canary.py#L189\r\n", "Canary at this moment does not support audio as a tensor. We are working on improving this support in the coming release, but as the model is new, it will take a little bit of time to mature to full support of all of ASR domains features ", "Fyi @pzelasko @krishnacpuvvada ", "@titu1994 Do other models also not support ndarray or tensor as inputs currently?", "Mlat other ASR models should support it, Multi Task models are still being developed ", "This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days.", "This issue was closed because it has been inactive for 7 days since being marked as stale." ]

[ { "id": 4811884691, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ikw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/ASR", "name": "ASR", "color": "F3B33E", "default": false, "description": "" } ]

[]

[ "jenkins", "This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.", "jenkins", "jenkins" ]

[ { "id": 4585431311, "node_id": "LA_kwDOC_bI7s8AAAABEVAhDw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/stale", "name": "stale", "color": "ededed", "default": false, "description": null }, { "id": 4811884691, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ikw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/ASR", "name": "ASR", "color": "F3B33E", "default": false, "description": "" } ]

[]

[ "@dkurt \r\nThank you for preparing this PR.\r\nLet me test this and review it.\r\nNeed to verify on multiple samples. ", "@tango4j, any updates?", "This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.", "@tango4j , friendly reminder. Maybe I can run some tests on my side?", "This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.", "@tango4j, reminder", "This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.", "This PR was closed because it has been inactive for 7 days since being marked as stale." ]

[ { "id": 4811886210, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ogg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/NLP", "name": "NLP", "color": "CBF09A", "default": false, "description": "" }, { "id": 4847373924, "node_id": "LA_kwDOC_bI7s8AAAABIO0OZA", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/common", "name": "common", "color": "c5def5", "default": false, "description": "" } ]

[]

[ "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins" ]

[ { "id": 4811886210, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ogg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/NLP", "name": "NLP", "color": "CBF09A", "default": false, "description": "" } ]

[]

[]

[ { "id": 4585431311, "node_id": "LA_kwDOC_bI7s8AAAABEVAhDw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/stale", "name": "stale", "color": "ededed", "default": false, "description": null }, { "id": 4811886210, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ogg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/NLP", "name": "NLP", "color": "CBF09A", "default": false, "description": "" } ]

[]

[ "jenkins", "jenkins", "jenkins", "jenkins", "This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.", "I need to get back to this.\r\n\r\nOn Wed, Mar 20, 2024 at 6:44 PM github-actions[bot] <\r\n***@***.***> wrote:\r\n\r\n> This PR is stale because it has been open for 14 days with no activity.\r\n> Remove stale label or comment or update or this will be closed in 7 days.\r\n>\r\n> —\r\n> Reply to this email directly, view it on GitHub\r\n> <https://github.com/NVIDIA/NeMo/pull/8466#issuecomment-2011033150>, or\r\n> unsubscribe\r\n> <https://github.com/notifications/unsubscribe-auth/AADQCBXYXVSTXGBJXW6C7DDYZI3QHAVCNFSM6AAAAABDSCNISGVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDAMJRGAZTGMJVGA>\r\n> .\r\n> You are receiving this because you authored the thread.Message ID:\r\n> ***@***.***>\r\n>\r\n", "This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.", "This PR was closed because it has been inactive for 7 days since being marked as stale." ]

[]

[]

[ "jenkins" ]

[ { "id": 4847373924, "node_id": "LA_kwDOC_bI7s8AAAABIO0OZA", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/common", "name": "common", "color": "c5def5", "default": false, "description": "" } ]

[]

[ "jenkins", "Closing this in favor of https://github.com/NVIDIA/NeMo/pull/8449" ]

[ { "id": 1485815557, "node_id": "MDU6TGFiZWwxNDg1ODE1NTU3", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/bug", "name": "bug", "color": "d73a4a", "default": true, "description": "Something isn't working" }, { "id": 4585431311, "node_id": "LA_kwDOC_bI7s8AAAABEVAhDw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/stale", "name": "stale", "color": "ededed", "default": false, "description": null } ]

[]

[ "Solved the issue, by providing the ckpt path, could I ask why it doesn't work using \".nemo\" files?", "This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days.", "This issue was closed because it has been inactive for 7 days since being marked as stale.", "I've also met this problem and solved by providing `.ckpt`, not the `.nemo`. \r\nI wonder what is the difference between both files." ]

[ { "id": 4811884691, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ikw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/ASR", "name": "ASR", "color": "F3B33E", "default": false, "description": "" } ]

[]

[ "This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.", "@artbataev thank you so much for the review!\r\n\r\nAbout your suggestions/comments:\r\n\r\n1. Done\r\n2. Done\r\n3. I leaved installing graphviz optional. The installation takes 5-10 minutes during which the window may become glitchy.\r\n4. I made a fix, but it works only on my remote desktop. I do not know why Google Colab does not display the right colors. I removed comments about colors.\r\n5. Done\r\n6. Done\r\n7. Done \r\n", "jenkins" ]

[ { "id": 4585431311, "node_id": "LA_kwDOC_bI7s8AAAABEVAhDw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/stale", "name": "stale", "color": "ededed", "default": false, "description": null } ]

[]

[ "This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.", "This PR was closed because it has been inactive for 7 days since being marked as stale." ]

[ { "id": 1485815557, "node_id": "MDU6TGFiZWwxNDg1ODE1NTU3", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/bug", "name": "bug", "color": "d73a4a", "default": true, "description": "Something isn't working" }, { "id": 4585431311, "node_id": "LA_kwDOC_bI7s8AAAABEVAhDw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/stale", "name": "stale", "color": "ededed", "default": false, "description": null } ]

[ { "login": "dimapihtar", "id": 37850217, "node_id": "MDQ6VXNlcjM3ODUwMjE3", "avatar_url": "https://avatars.githubusercontent.com/u/37850217?v=4", "gravatar_id": "", "url": "https://api.github.com/users/dimapihtar", "html_url": "https://github.com/dimapihtar", "followers_url": "https://api.github.com/users/dimapihtar/followers", "following_url": "https://api.github.com/users/dimapihtar/following{/other_user}", "gists_url": "https://api.github.com/users/dimapihtar/gists{/gist_id}", "starred_url": "https://api.github.com/users/dimapihtar/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/dimapihtar/subscriptions", "organizations_url": "https://api.github.com/users/dimapihtar/orgs", "repos_url": "https://api.github.com/users/dimapihtar/repos", "events_url": "https://api.github.com/users/dimapihtar/events{/privacy}", "received_events_url": "https://api.github.com/users/dimapihtar/received_events", "type": "User", "user_view_type": "public", "site_admin": false } ]

[ "@dimapihtar @ericharper \r\nSame issue occurs when trying to load the distributed checkpoint for continued training / sft.\r\n\r\nLoading distributed checkpoint with a single A100 works fine, with gbs=1,tp=1,pp=1,mbs=1. \r\nWhen scaling to 8 gpus, and changing gbs to 8, loading the checkpoint fails.\r\n\r\nHas the team been able to reproduce this internally?\r\nCurrently using the following commits:\r\nMegatron-LM ad53b1e38689a0ceed75ade7821f4e6c7554abb4\r\nNeMo 9b64e390b534d4eb5ad7f28502bcfe4c7f0c6c39\r\nTransformerEngine: da30634a6c9ccdbb6c587b6c93b1860e4b038204\r\n\r\n```\r\n[0]: model = cls.load_from_checkpoint(checkpoint_path=checkpoint_path, trainer=trainer, hparams_file=f.name)\r\n[0]: File \"/workspace/src/3rdparty/NeMo/build/__editable__.nemo_toolkit-1.23.0rc0-py3-none-any/nemo/collections/nlp/models/nlp_model.py\", line 397, in load_from_checkpoint\r\n[0]: checkpoint = dist_checkpointing.load(sharded_state_dict=checkpoint, checkpoint_dir=checkpoint_dir)\r\n[0]: File \"/workspace/src/3rdparty/Megatron-LM/build/__editable__.megatron_core-0.5.0rc0-cp310-cp310-linux_x86_64/megatron/core/dist_checkpointing/serialization.py\", line 99, in load\r\n[0]: validate_sharding_integrity(nested_values(sharded_state_dict))\r\n[0]: File \"/workspace/src/3rdparty/Megatron-LM/build/__editable__.megatron_core-0.5.0rc0-cp310-cp310-linux_x86_64/megatron/core/dist_checkpointing/serialization.py\", line 346, in validate_sharding_integrity\r\n[0]: _validate_sharding_for_key(shardings)\r\n[0]: File \"/workspace/src/3rdparty/Megatron-LM/build/__editable__.megatron_core-0.5.0rc0-cp310-cp310-linux_x86_64/megatron/core/dist_checkpointing/serialization.py\", line 384, in _validate_sharding_for_key\r\n[0]: raise CheckpointingException(f'Invalid access pattern for {rank_sharding[0][1]}')\r\n[0]:megatron.core.dist_checkpointing.core.CheckpointingException: Invalid access pattern for ShardedTensor(key='model.embedding.word_embeddings.weight')\r\n```", "Some debug logs as well, let me know if anything else could be useful to include:\r\n\r\n```\r\n> rank_sharing[0][1]\r\n\r\n[(0, ShardedTensor(key='model.embedding.word_embeddings.weight', data=None, dtype=torch.float32, local_shape=(50304, 768), global_shape=(50304, 768), global_offset=(0, 0), axis_fragmentations=(1, 1), replica_id=(0, 0, 0), prepend_axis_num=0, allow_shape_mismatch=True, flattened_range=None)), (1, ShardedTensor(key='model.embedding.word_embeddings.weight', data=None, dtype=torch.float32, local_shape=(50304, 768), global_shape=(50304, 768), global_offset=(0, 0), axis_fragmentations=(1, 1), replica_id=(0, 0, 0), prepend_axis_num=0, allow_shape_mismatch=True, flattened_range=None)), (2, ShardedTensor(key='model.embedding.word_embeddings.weight', data=None, dtype=torch.float32, local_shape=(50304, 768), global_shape=(50304, 768), global_offset=(0, 0), axis_fragmentations=(1, 1), replica_id=(0, 0, 0), prepend_axis_num=0, allow_shape_mismatch=True, flattened_range=None)), (3, ShardedTensor(key='model.embedding.word_embeddings.weight', data=None, dtype=torch.float32, local_shape=(50304, 768), global_shape=(50304, 768), global_offset=(0, 0), axis_fragmentations=(1, 1), replica_id=(0, 0, 0), prepend_axis_num=0, allow_shape_mismatch=True, flattened_range=None)), (4, ShardedTensor(key='model.embedding.word_embeddings.weight', data=None, dtype=torch.float32, local_shape=(50304, 768), global_shape=(50304, 768), global_offset=(0, 0), axis_fragmentations=(1, 1), replica_id=(0, 0, 0), prepend_axis_num=0, allow_shape_mismatch=True, flattened_range=None)), (5, ShardedTensor(key='model.embedding.word_embeddings.weight', data=None, dtype=torch.float32, local_shape=(50304, 768), global_shape=(50304, 768), global_offset=(0, 0), axis_fragmentations=(1, 1), replica_id=(0, 0, 0), prepend_axis_num=0, allow_shape_mismatch=True, flattened_range=None)), (6, ShardedTensor(key='model.embedding.word_embeddings.weight', data=None, dtype=torch.float32, local_shape=(50304, 768), global_shape=(50304, 768), global_offset=(0, 0), axis_fragmentations=(1, 1), replica_id=(0, 0, 0), prepend_axis_num=0, allow_shape_mismatch=True, flattened_range=None)), (7, ShardedTensor(key='model.embedding.word_embeddings.weight', data=None, dtype=torch.float32, local_shape=(50304, 768), global_shape=(50304, 768), global_offset=(0, 0), axis_fragmentations=(1, 1), replica_id=(0, 0, 0), prepend_axis_num=0, allow_shape_mismatch=True, flattened_range=None))]\r\n```\r\n\r\n```\r\n> torch.zeros(rank_sharding[0][1].axis_fragmentations)\r\ntensor([[0.]])\r\n```\r\n\r\nThe assertion is checking that the shard is access once on all ranks:\r\n```\r\n if not torch.all(shard_access_cnt == 1):\r\n logger.error(f'Invalid access pattern for {rank_sharding[0][1]}: {shard_access_cnt}')\r\n raise CheckpointingException(f'Invalid access pattern for {rank_sharding[0][1]}')\r\n```\r\n\r\nBut this is the behavior I observe:\r\n```\r\n> shard_access_cnt\r\ntensor([[8]], dtype=torch.int32)\r\n\r\n>for rank, sharding in rank_sharding:\r\n print(sharding.replica_id)\r\n(0, 0, 0)\r\n(0, 0, 0)\r\n(0, 0, 0)\r\n(0, 0, 0)\r\n(0, 0, 0)\r\n(0, 0, 0)\r\n(0, 0, 0)\r\n(0, 0, 0)\r\n\r\n```\r\n\r\n\r\n\r\nNotice that there is a TODO listed to check the shard_access_cnt of replicas as well: https://github.com/NVIDIA/Megatron-LM/blob/0fecd76e995c136021d478c6c52caa57c2f9aa25/megatron/core/dist_checkpointing/serialization.py#L444C1-L447C59\r\n\r\nBut there should only be one replica per rank, which is the one above. \r\n\r\nSince I'm setting tensor parallel to 1 here, and gbs to 8, I think that the embedding_weights should not be expected to be sharded since there is only a single TP group with worldsize=8. I believe that during training, this dist ckpt also only used tp=1. So it should be fine if the embedding weights are fully replicated across all of the ranks?\r\n\r\n\r\n\r\nOff topic: \r\nI also notice that when loading from the distributed checkpoint, it loads the saved optimizer states as well, which is probably expected behavior. Is there any reference or guidance on how to load without these saved optimizer states? Or does it not matter if it gets overwritten later on?", "Seem to have figured out the root cause.\r\n\r\n### Background\r\n\r\nWhen loading from distributed checkpoint, we call `NLPModel.load_from_checkpoint(..)` [ref](https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/nlp/models/nlp_model.py#L309)\r\nFor distributed checkpoint, loading the state_dict gets [deferred until the class is initialized](https://github.com/NVIDIA/NeMo/blob/main/nemo/collections/nlp/models/nlp_model.py#L384)\r\n\r\nThe current logic is as follows:\r\n```\r\n if 'cfg' in kwargs:\r\n model = ptl_load_state(cls, checkpoint, strict=strict, **kwargs)\r\n else:\r\n model = ptl_load_state(cls, checkpoint, strict=strict, cfg=cfg, **kwargs)\r\n # cfg = checkpoint[cls.CHECKPOINT_HYPER_PARAMS_KEY].cfg\r\n\r\n if checkpoint_dir is not None:\r\n sharded_state_dict = model.sharded_state_dict()\r\n checkpoint['state_dict'] = sharded_state_dict\r\n # dist checkpointing needs torch.distributed to load the checkpoint\r\n if parallel_state.is_unitialized():\r\n\r\n def dummy():\r\n return\r\n\r\n if model.trainer.strategy.launcher is not None:\r\n model.trainer.strategy.launcher.launch(dummy, trainer=model.trainer)\r\n model.trainer.strategy.setup_environment()\r\n```\r\n\r\nFor `MegatronGPTModel(MCoreGPTModel)`, this instantiates the `LanguageModelEmbedding()` if cfg.pre_process is true.\r\n\r\nOn this line `sharded_state_dict = model.sharded_state_dict()`, we call `MegatronGPTModel.sharded_state_dict(...)` which in turn calls `VocabParallelEmbedding.sharded_state_dict(...)`:\r\n\r\n```\r\n def sharded_state_dict(\r\n self, prefix: str = '', sharded_offsets: Tuple[Tuple[int, int, int]] = ()\r\n ) -> ShardedStateDict:\r\n \"\"\" Non-default implementation for embeddings due to `allow_shape_mismatch` param \"\"\"\r\n state_dict = self.state_dict(prefix='', keep_vars=True)\r\n\r\n weight_prefix = f'{prefix}weight'\r\n return {\r\n weight_prefix: make_tp_sharded_tensor_for_checkpoint(\r\n tensor=state_dict['weight'],\r\n key=weight_prefix,\r\n allow_shape_mismatch=True,\r\n prepend_offsets=sharded_offsets,\r\n )\r\n }\r\n```\r\n\r\n[make_tp_sharded_tensor_for_checkpoint](https://github.com/NVIDIA/Megatron-LM/blob/main/megatron/core/utils.py#L186):\r\n```\r\ndef make_tp_sharded_tensor_for_checkpoint(\r\n tensor, key, tp_axis=0, replica_id=None, prepend_offsets=(), **kwargs\r\n):\r\n prepend_axis_num = len(prepend_offsets)\r\n if replica_id is None:\r\n replica_id = (0, 0, parallel_state.get_data_parallel_rank(with_context_parallel=True))\r\n ...\r\n```\r\n\r\n### The Issue\r\n\r\nAt this point, parallel_state is not yet available since `torch.distributed.is_initialized() == False`, so any data_parallel ranks return as 0 and incorrectly sets all replica_ids as (0,0,0), which in turn causes the assertion mentioned above to fail when validating the sharded tensors.\r\n\r\n### Potential Resolution\r\n\r\nIt seems like an easy fix would be to move the initialization logic before the `model.sharded_state_dict()` call, but I'm not informed enough to know what the downstream impact would be.\r\n```\r\n if checkpoint_dir is not None:\r\n sharded_state_dict = model.sharded_state_dict()\r\n checkpoint['state_dict'] = sharded_state_dict\r\n # dist checkpointing needs torch.distributed to load the checkpoint\r\n if parallel_state.is_unitialized():\r\n\r\n def dummy():\r\n return\r\n\r\n if model.trainer.strategy.launcher is not None:\r\n model.trainer.strategy.launcher.launch(dummy, trainer=model.trainer)\r\n model.trainer.strategy.setup_environment()\r\n```\r\n\r\nCan you please let me know if this is a correct understanding? \r\n\r\n@dimapihtar @ericharper", "This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days.", "any updates on this issue?", "This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days.", "This issue was closed because it has been inactive for 7 days since being marked as stale." ]

[ { "id": 1485815557, "node_id": "MDU6TGFiZWwxNDg1ODE1NTU3", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/bug", "name": "bug", "color": "d73a4a", "default": true, "description": "Something isn't working" } ]

[]

[ "Probably related: https://github.com/pytorch/pytorch/issues/86566", "Yikes. Fyi @tango4j @nithinraok @artbataev ", "@dkurt What version of PyTorch do you use? What is the environment (OS, conda/pip, etc)?", "@artbataev, Extremely apologize. Segmentation fault reproduced with Python `3.11.0rc1` and PyTorch 2.1.0/2.2.0. After upgrade to Python `3.11.7` everything works fine." ]

[ { "id": 4811886210, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ogg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/NLP", "name": "NLP", "color": "CBF09A", "default": false, "description": "" } ]

[]

[ "jenkins", "jenkins" ]

[ { "id": 4585431311, "node_id": "LA_kwDOC_bI7s8AAAABEVAhDw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/stale", "name": "stale", "color": "ededed", "default": false, "description": null } ]

[ { "login": "okuchaiev", "id": 22035961, "node_id": "MDQ6VXNlcjIyMDM1OTYx", "avatar_url": "https://avatars.githubusercontent.com/u/22035961?v=4", "gravatar_id": "", "url": "https://api.github.com/users/okuchaiev", "html_url": "https://github.com/okuchaiev", "followers_url": "https://api.github.com/users/okuchaiev/followers", "following_url": "https://api.github.com/users/okuchaiev/following{/other_user}", "gists_url": "https://api.github.com/users/okuchaiev/gists{/gist_id}", "starred_url": "https://api.github.com/users/okuchaiev/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/okuchaiev/subscriptions", "organizations_url": "https://api.github.com/users/okuchaiev/orgs", "repos_url": "https://api.github.com/users/okuchaiev/repos", "events_url": "https://api.github.com/users/okuchaiev/events{/privacy}", "received_events_url": "https://api.github.com/users/okuchaiev/received_events", "type": "User", "user_view_type": "public", "site_admin": false } ]

[ "Hi, can you add me to the org/team ea-bignlp/ga-participants. I can't get any info about private repository, so i can't pull that container. Thanks you so much!", "This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days.", "This issue was closed because it has been inactive for 7 days since being marked as stale." ]

[ { "id": 4585431311, "node_id": "LA_kwDOC_bI7s8AAAABEVAhDw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/stale", "name": "stale", "color": "ededed", "default": false, "description": null } ]

[]

[ "This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days.", "This issue was closed because it has been inactive for 7 days since being marked as stale." ]

[ { "id": 4585431311, "node_id": "LA_kwDOC_bI7s8AAAABEVAhDw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/stale", "name": "stale", "color": "ededed", "default": false, "description": null } ]

[ { "login": "okuchaiev", "id": 22035961, "node_id": "MDQ6VXNlcjIyMDM1OTYx", "avatar_url": "https://avatars.githubusercontent.com/u/22035961?v=4", "gravatar_id": "", "url": "https://api.github.com/users/okuchaiev", "html_url": "https://github.com/okuchaiev", "followers_url": "https://api.github.com/users/okuchaiev/followers", "following_url": "https://api.github.com/users/okuchaiev/following{/other_user}", "gists_url": "https://api.github.com/users/okuchaiev/gists{/gist_id}", "starred_url": "https://api.github.com/users/okuchaiev/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/okuchaiev/subscriptions", "organizations_url": "https://api.github.com/users/okuchaiev/orgs", "repos_url": "https://api.github.com/users/okuchaiev/repos", "events_url": "https://api.github.com/users/okuchaiev/events{/privacy}", "received_events_url": "https://api.github.com/users/okuchaiev/received_events", "type": "User", "user_view_type": "public", "site_admin": false } ]

[ "This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days.", "This issue was closed because it has been inactive for 7 days since being marked as stale." ]

[ { "id": 1485815557, "node_id": "MDU6TGFiZWwxNDg1ODE1NTU3", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/bug", "name": "bug", "color": "d73a4a", "default": true, "description": "Something isn't working" }, { "id": 4585431311, "node_id": "LA_kwDOC_bI7s8AAAABEVAhDw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/stale", "name": "stale", "color": "ededed", "default": false, "description": null } ]

[]

[ "Also seeing this error.\r\n\r\nMy temporary workaround is to catch ValueErrors and just add a second or two of blank audio to the end of the file before re-processing which seems to work as a temporary stop-gap until this can be fixed.", "This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days.", "Adding comment to prevent this issue from closing.", "This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days.", "This issue was closed because it has been inactive for 7 days since being marked as stale.", "@bradmurray-dt Still facing this issue with 'parakeet-tdt-1.1b' and 'parakeet-tdt-ctc-1.1b':\r\n\r\n```\r\n File \"/lib/python3.10/site-packages/nemo/collections/asr/parts/submodules/rnnt_decoding.py\", line 510, in rnnt_decoder_predictions_tensor\r\n hypotheses[hyp_idx] = self.compute_rnnt_timestamps(hypotheses[hyp_idx], timestamp_type)\r\n File \"/lib/python3.10/site-packages/nemo/collections/asr/parts/submodules/rnnt_decoding.py\", line 753, in compute_rnnt_timestamps\r\n raise ValueError(\r\nValueError: `char_offsets`: [{'char': [tensor(386, dtype=torch.int32)], 'start_offset': 2, 'end_offset': 3},.....\r\nhave to be of the same length, but are: `len(offsets)`: 102 and `len(processed_tokens)`: 103\r\n```" ]

[]

[ { "login": "okuchaiev", "id": 22035961, "node_id": "MDQ6VXNlcjIyMDM1OTYx", "avatar_url": "https://avatars.githubusercontent.com/u/22035961?v=4", "gravatar_id": "", "url": "https://api.github.com/users/okuchaiev", "html_url": "https://github.com/okuchaiev", "followers_url": "https://api.github.com/users/okuchaiev/followers", "following_url": "https://api.github.com/users/okuchaiev/following{/other_user}", "gists_url": "https://api.github.com/users/okuchaiev/gists{/gist_id}", "starred_url": "https://api.github.com/users/okuchaiev/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/okuchaiev/subscriptions", "organizations_url": "https://api.github.com/users/okuchaiev/orgs", "repos_url": "https://api.github.com/users/okuchaiev/repos", "events_url": "https://api.github.com/users/okuchaiev/events{/privacy}", "received_events_url": "https://api.github.com/users/okuchaiev/received_events", "type": "User", "user_view_type": "public", "site_admin": false } ]

[ "For Parakeet, we have several tutorials in the Nemo tutorials directory - here is one to get started - https://github.com/NVIDIA/NeMo/blob/main/tutorials/asr/ASR_CTC_Language_Finetuning.ipynb\r\n\r\nMost of the concepts will apply to Parakeet CTC / RNNT/ TDT. \r\n\r\nFor Canary, it is a very new model, we are still adding documentation and tutorials for finetuning and scratch training ", "Also, if you want a finetune script, you can use this - https://github.com/NVIDIA/NeMo/blob/main/examples/asr/speech_to_text_finetune.py", "Thank @titu1994 for sharing these details", "> For Parakeet, we have several tutorials in the Nemo tutorials directory - here is one to get started - https://github.com/NVIDIA/NeMo/blob/main/tutorials/asr/ASR_CTC_Language_Finetuning.ipynb\n> \n> Most of the concepts will apply to Parakeet CTC / RNNT/ TDT.\n> \n> For Canary, it is a very new model, we are still adding documentation and tutorials for finetuning and scratch training\n\nHello, I would like to know about the progress on Canary finetuning. Thanks." ]

[ { "id": 4811884691, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ikw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/ASR", "name": "ASR", "color": "F3B33E", "default": false, "description": "" }, { "id": 4847373924, "node_id": "LA_kwDOC_bI7s8AAAABIO0OZA", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/common", "name": "common", "color": "c5def5", "default": false, "description": "" } ]

[]

[ "jenkins", "jenkins", "jenkins", "jenkins", "could you add unit tests that cover these edge cases? btw I didn't see that PR earlier and added the relative audio path support independently; we can either merge my PR or port the test I wrote to yours and close mine (up to you) https://github.com/NVIDIA/NeMo/pull/8464", "> could you add unit tests that cover these edge cases? btw I didn't see that PR earlier and added the relative audio path support independently; we can either merge my PR or port the test I wrote to yours and close mine (up to you) #8464\r\n\r\nSure I'll add the tests~I think it's better to use the `get_full_path` function to resolve the relative paths, since it's well tested.", "jenkins", "jenkins", "jenkins", "jenkins" ]

[ { "id": 4418253970, "node_id": "LA_kwDOC_bI7s8AAAABB1k0kg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/TTS", "name": "TTS", "color": "49AF28", "default": false, "description": "" }, { "id": 4585431311, "node_id": "LA_kwDOC_bI7s8AAAABEVAhDw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/stale", "name": "stale", "color": "ededed", "default": false, "description": null }, { "id": 4811884691, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ikw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/ASR", "name": "ASR", "color": "F3B33E", "default": false, "description": "" }, { "id": 4811886210, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ogg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/NLP", "name": "NLP", "color": "CBF09A", "default": false, "description": "" } ]

[]

[ "This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.", "This PR was closed because it has been inactive for 7 days since being marked as stale." ]

[ { "id": 4585431311, "node_id": "LA_kwDOC_bI7s8AAAABEVAhDw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/stale", "name": "stale", "color": "ededed", "default": false, "description": null }, { "id": 4811886210, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ogg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/NLP", "name": "NLP", "color": "CBF09A", "default": false, "description": "" } ]

[]

[ "This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.", "This PR was closed because it has been inactive for 7 days since being marked as stale." ]

[ { "id": 4585431311, "node_id": "LA_kwDOC_bI7s8AAAABEVAhDw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/stale", "name": "stale", "color": "ededed", "default": false, "description": null }, { "id": 4811886210, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ogg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/NLP", "name": "NLP", "color": "CBF09A", "default": false, "description": "" } ]

[ { "login": "JRD971000", "id": 71892896, "node_id": "MDQ6VXNlcjcxODkyODk2", "avatar_url": "https://avatars.githubusercontent.com/u/71892896?v=4", "gravatar_id": "", "url": "https://api.github.com/users/JRD971000", "html_url": "https://github.com/JRD971000", "followers_url": "https://api.github.com/users/JRD971000/followers", "following_url": "https://api.github.com/users/JRD971000/following{/other_user}", "gists_url": "https://api.github.com/users/JRD971000/gists{/gist_id}", "starred_url": "https://api.github.com/users/JRD971000/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/JRD971000/subscriptions", "organizations_url": "https://api.github.com/users/JRD971000/orgs", "repos_url": "https://api.github.com/users/JRD971000/repos", "events_url": "https://api.github.com/users/JRD971000/events{/privacy}", "received_events_url": "https://api.github.com/users/JRD971000/received_events", "type": "User", "user_view_type": "public", "site_admin": false } ]

[ "jenkins", "jenkins", "This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.", "This PR was closed because it has been inactive for 7 days since being marked as stale." ]

[ { "id": 4811886210, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ogg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/NLP", "name": "NLP", "color": "CBF09A", "default": false, "description": "" }, { "id": 4847373924, "node_id": "LA_kwDOC_bI7s8AAAABIO0OZA", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/common", "name": "common", "color": "c5def5", "default": false, "description": "" } ]

[ { "login": "JRD971000", "id": 71892896, "node_id": "MDQ6VXNlcjcxODkyODk2", "avatar_url": "https://avatars.githubusercontent.com/u/71892896?v=4", "gravatar_id": "", "url": "https://api.github.com/users/JRD971000", "html_url": "https://github.com/JRD971000", "followers_url": "https://api.github.com/users/JRD971000/followers", "following_url": "https://api.github.com/users/JRD971000/following{/other_user}", "gists_url": "https://api.github.com/users/JRD971000/gists{/gist_id}", "starred_url": "https://api.github.com/users/JRD971000/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/JRD971000/subscriptions", "organizations_url": "https://api.github.com/users/JRD971000/orgs", "repos_url": "https://api.github.com/users/JRD971000/repos", "events_url": "https://api.github.com/users/JRD971000/events{/privacy}", "received_events_url": "https://api.github.com/users/JRD971000/received_events", "type": "User", "user_view_type": "public", "site_admin": false } ]

[ "jenkins", "jenkins", "jenkins", "jenkins" ]

[ { "id": 4811886210, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ogg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/NLP", "name": "NLP", "color": "CBF09A", "default": false, "description": "" }, { "id": 4840216855, "node_id": "LA_kwDOC_bI7s8AAAABIH_ZFw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/CI", "name": "CI", "color": "1D76DB", "default": false, "description": "" } ]

[]

[ "jenkins", "jenkins", "jenkins" ]

[ { "id": 4556669155, "node_id": "LA_kwDOC_bI7s8AAAABD5lA4w", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/cherry-pick", "name": "cherry-pick", "color": "ededed", "default": false, "description": null }, { "id": 4811886210, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ogg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/NLP", "name": "NLP", "color": "CBF09A", "default": false, "description": "" } ]

[]

[ "jenkins", "jenkins", "jenkins" ]

[ { "id": 4556669155, "node_id": "LA_kwDOC_bI7s8AAAABD5lA4w", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/cherry-pick", "name": "cherry-pick", "color": "ededed", "default": false, "description": null } ]

[]

[ "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins" ]

[ { "id": 1485815557, "node_id": "MDU6TGFiZWwxNDg1ODE1NTU3", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/bug", "name": "bug", "color": "d73a4a", "default": true, "description": "Something isn't working" }, { "id": 4585431311, "node_id": "LA_kwDOC_bI7s8AAAABEVAhDw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/stale", "name": "stale", "color": "ededed", "default": false, "description": null } ]

[]

[ "Running into a similar issue training the conformer large model \r\nin a docker container with the latest `nvcr.io/nvidia/nemo:23.10` image, on p2.16xlarge (V100 instances). What is your training environment like?", "This is my docker container\r\n\r\n```\r\nnvcr.io/nvidia/clara/bionemo-framework:latest \"/workspace/bionemo/…\" 4 days ago Up 13 hours bionemo\r\n```", "This issue is stale because it has been open for 30 days with no activity. Remove stale label or comment or this will be closed in 7 days.", "This issue was closed because it has been inactive for 7 days since being marked as stale.", "I am facing the same issue on a multi node multi GPU and without docker.\r\nI am utilizing slurm to run the job." ]

[ { "id": 4418253970, "node_id": "LA_kwDOC_bI7s8AAAABB1k0kg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/TTS", "name": "TTS", "color": "49AF28", "default": false, "description": "" }, { "id": 4811884691, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ikw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/ASR", "name": "ASR", "color": "F3B33E", "default": false, "description": "" }, { "id": 4811886210, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ogg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/NLP", "name": "NLP", "color": "CBF09A", "default": false, "description": "" } ]

[]

[]

[ { "id": 4840216855, "node_id": "LA_kwDOC_bI7s8AAAABIH_ZFw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/CI", "name": "CI", "color": "1D76DB", "default": false, "description": "" } ]

[]

[ "jenkins" ]

[ { "id": 4585431311, "node_id": "LA_kwDOC_bI7s8AAAABEVAhDw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/stale", "name": "stale", "color": "ededed", "default": false, "description": null }, { "id": 4811884691, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ikw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/ASR", "name": "ASR", "color": "F3B33E", "default": false, "description": "" } ]

[]

[ "This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.", "@biscayan \r\n\r\nDo you remember how you come up with the model_stride_sec = 0.08? [Link to the line](https://github.com/biscayan/NeMo/blob/9f50dbdabed105d2649384e403c71147d57b4bae/nemo/collections/asr/parts/utils/decoder_timestamps_utils.py#L357C1-L358C1)\r\nWith 0.08 sec, fast conformer is not yielding the right timestamp. It is generating slightly longer sample length.\r\n\r\nThis needs more investigation on time-stamp generation on FastConformer Model. \r\n", "@tango4j \r\nThank you for reviewing.\r\nIn the fastconformer paper, it is explained that fastconformer model has downsampling schema with 80ms frame rate.\r\nIt means that the time of one frame is 80ms.\r\nI tested the extracted timestamps while listening the wav file. \r\nWhen I set the model_stride_sec=0.04, the timestamps do not match with the wavfile, and after changing the model_stride_sec from 0.04 to 0.08, timestamps accurately match with the wavfile\r\n", "jenkins", "@biscayan \r\nDid you check the cpWER ? or you just checked DER ?\r\nWhen I tested with CallHome data, the fast conformer was showing 50% ish cpWER while other CTC models are showing 18~20% errors.\r\nI think there is some issues with chunk-based ASR system. ", "@tango4j \r\nI didn't compute the cpWER for fastconformer model.\r\nI just compared the asr result and the transcript in clean speech data, and the result is almost accurate.\r\nHowever, I also faced the problem in real reverberated meeting data, the model has low asr performance. ", "@tango4j what are other CTC models you were comparing, are both of similar sizes? ", "@nithinraok \r\nModel size does not affect the timestamp issue with chunk based ASR. Only fast-conformer CTC models are returning the inaccurate timestamps with chunk based ASR.\r\nWe use ConformerCTC large `stt_en_conformer_ctc_large` for test purpose. \t", "@tango4j \r\nShould I do something more for the review?", "This PR is stale because it has been open for 14 days with no activity. Remove stale label or comment or update or this will be closed in 7 days.", "This PR was closed because it has been inactive for 7 days since being marked as stale.", "@KunalDhawan pls test this PR on a known set to compare against similar sized conformer model", "Also add a integration test for testing purposes! testing on English dataset is good enough. ", "Hi @KunalDhawan \r\nThank you for your report. \r\nI will develop the PR later. \r\nThank you." ]

[ { "id": 4556669155, "node_id": "LA_kwDOC_bI7s8AAAABD5lA4w", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/cherry-pick", "name": "cherry-pick", "color": "ededed", "default": false, "description": null }, { "id": 4811886210, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ogg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/NLP", "name": "NLP", "color": "CBF09A", "default": false, "description": "" } ]

[]

[]

[]

[ { "login": "okuchaiev", "id": 22035961, "node_id": "MDQ6VXNlcjIyMDM1OTYx", "avatar_url": "https://avatars.githubusercontent.com/u/22035961?v=4", "gravatar_id": "", "url": "https://api.github.com/users/okuchaiev", "html_url": "https://github.com/okuchaiev", "followers_url": "https://api.github.com/users/okuchaiev/followers", "following_url": "https://api.github.com/users/okuchaiev/following{/other_user}", "gists_url": "https://api.github.com/users/okuchaiev/gists{/gist_id}", "starred_url": "https://api.github.com/users/okuchaiev/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/okuchaiev/subscriptions", "organizations_url": "https://api.github.com/users/okuchaiev/orgs", "repos_url": "https://api.github.com/users/okuchaiev/repos", "events_url": "https://api.github.com/users/okuchaiev/events{/privacy}", "received_events_url": "https://api.github.com/users/okuchaiev/received_events", "type": "User", "user_view_type": "public", "site_admin": false } ]

[ "There's a reason those decoding classes aren't modules. When rnnt was being developed it turned out there was a memory leak due to pt graph tracking and Autoregressive calls to the decoder joint. That's why .freeze(), .as_frozen() was developed in NeMos core to deal with that. \r\n\r\nI dunno if that reason is still valid in pytorch after the inference_mode() decorator was added, maybe safe to make them modules again. \r\n\r\nAs you've mentioned, it is a lot of work to do this change, and I don't feel it is super high priority, but if someone wants to take a look that's fine. \r\n\r\nTo get device inside the decoding classes, the preferred mechanism is mostly `next(module.parameters()).device` and cache that value inside the inner forwards. \r\n\r\n> CUDA_VISIBLE_DEVICES environment variable appropriately before starting a process (and remove the cuda=1\r\n\r\nThis is not a good user experience and I want to avoid it. Users should not have to use any env flags for use Nemo ASR as a necessity. \r\n\r\n> Do we have any use for multiple cuda devices in a single process in NeMo?\r\n\r\nWell if you're training, yes, but it's handled by ptl. During inference, we stick to single GPU for now, but with larger models we may consider doing multi GPU in the future. ", "I frequently use in the notebooks `model.to(\"cuda:1\")`, and expect this to work (I have multiple GPUs). It is inconvenient to use `CUDA_VISIBLE_DEVICES` with notebooks.\r\n\r\nI think that there are two possible approaches:\r\n- if the decoder wants to store anything on the GPU (or MPS), it should be responsible for auto-transferring self-state to the device. This is due to `Infer` classes, which are not `nn.Module` instances (but have Joint/Prednet as submodules). In this case, everything will work, but the code's author should care about this.\r\n- redesign the RNN-T in the following way using `nn.Module`s\r\n - `DecodingWrapper` is an `nn.Module` instance and owns Joint and Prednet\r\n - decoding strategies inherit from `DecodingWrapper` (e.g., `BatchedGreedyDecodingWrapper`)\r\n - the core RNN-T model does not have joint/prednet members anymore, only through `DecodingWrapper`s subclasses\r\n - so, everything inherits `nn.Module`, no duplication of ownership of joint/prednet, buffers are automatically moved to the appropriate device\r\n - changing the decoding strategy implies changing a subclass of `DecodingWrapper` (transferring Joint/Prednet from the current wrapper)\r\n\r\nHowever, such a redesign will break the current checkpoints and will require a lot of work and testing.\r\n", "It's a few lines to manually manage device type from the encoder, decoder or joints param dtype, we shouldn't have such breaking changes for this ", "On top of this, yes the fact that parameter count would be double counted for decoder and joint with such a redesign is also bad. ", "I agree that, for now, such breaking changes are undesirable.\r\n\r\n> On top of this, yes the fact that parameter count would be double counted for decoder and joint with such a redesign is also bad\r\nNope, the concept aim is exactly to avoid parameter duplication.\r\n\r\n**Currently**, rnnt model owns:\r\n- encoder\r\n- joint\r\n- prediction network\r\n- `*Infer` (not nn.Module), owns\r\n - joint (duplicate)\r\n - prediction network (duplicate)\r\n\r\n**Proposed**, rnnt model owns:\r\n- encoder\r\n- `*DecodingWrapper`, owns\r\n - joint\r\n - prediction network", "That's just bad design, why merge the transcription and prediction network when in all literature it's denoted as separate modules. Also, you don't always call the prednet and decoder network with the same set of inputs (train time prepends blank, eval time it starts with blank as first token for autoregressive decoding). \r\n\r\nThis is not a viable proposal in my opinion ", "I don't really get what's the big issue with the decoding framework not being a neural module. It's responsibility is not to act as a parameter based operation on NN network as part of the forward, it's a agnostic layer that provides a stable interface (Hypothesis) to map the enc Dec Joint logprobs to text. It's a logical separation, not a module dependency. \r\n\r\nI understand that certain issues can arise due to current design, that's cause we're using more advanced design. However very simple solution to this exists which Daniel has already implemented and is trivial to do (and also recommended by pytorch btw - to base new tensors on the device of the current active ones that it will interact with). \r\n\r\nI don't see why such a thing requires a refactoring of the deciding framework. If there's a bug we fix it, we don't scrap the entire thing and do it all over cause of \"pytorch design pattern\" (which Nemo does not fully follow by design, we instead use PTL design pattern) ", "My overall conclusion is that I have encountered an edge case, and the right approach is just to recreate the appropriate state tensors anytime that the device of the input tensors changes from what was being used before.", "Closing this. The better way is lazy initialization given how nemo currently is." ]

[ { "id": 4840216855, "node_id": "LA_kwDOC_bI7s8AAAABIH_ZFw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/CI", "name": "CI", "color": "1D76DB", "default": false, "description": "" } ]

[]

[ "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins", "jenkins" ]

[ { "id": 1923495043, "node_id": "MDU6TGFiZWwxOTIzNDk1MDQz", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/core", "name": "core", "color": "f25e3a", "default": false, "description": "Changes to NeMo Core" }, { "id": 4811884691, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ikw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/ASR", "name": "ASR", "color": "F3B33E", "default": false, "description": "" }, { "id": 4811886210, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ogg", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/NLP", "name": "NLP", "color": "CBF09A", "default": false, "description": "" }, { "id": 4840216855, "node_id": "LA_kwDOC_bI7s8AAAABIH_ZFw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/CI", "name": "CI", "color": "1D76DB", "default": false, "description": "" }, { "id": 4847373924, "node_id": "LA_kwDOC_bI7s8AAAABIO0OZA", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/common", "name": "common", "color": "c5def5", "default": false, "description": "" } ]

[ { "login": "JRD971000", "id": 71892896, "node_id": "MDQ6VXNlcjcxODkyODk2", "avatar_url": "https://avatars.githubusercontent.com/u/71892896?v=4", "gravatar_id": "", "url": "https://api.github.com/users/JRD971000", "html_url": "https://github.com/JRD971000", "followers_url": "https://api.github.com/users/JRD971000/followers", "following_url": "https://api.github.com/users/JRD971000/following{/other_user}", "gists_url": "https://api.github.com/users/JRD971000/gists{/gist_id}", "starred_url": "https://api.github.com/users/JRD971000/starred{/owner}{/repo}", "subscriptions_url": "https://api.github.com/users/JRD971000/subscriptions", "organizations_url": "https://api.github.com/users/JRD971000/orgs", "repos_url": "https://api.github.com/users/JRD971000/repos", "events_url": "https://api.github.com/users/JRD971000/events{/privacy}", "received_events_url": "https://api.github.com/users/JRD971000/received_events", "type": "User", "user_view_type": "public", "site_admin": false } ]

[ "jenkins", "jenkins", "unrelated changes after rebasing with main" ]

[ { "id": 4556669155, "node_id": "LA_kwDOC_bI7s8AAAABD5lA4w", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/cherry-pick", "name": "cherry-pick", "color": "ededed", "default": false, "description": null }, { "id": 4811884691, "node_id": "LA_kwDOC_bI7s8AAAABHs-Ikw", "url": "https://api.github.com/repos/NVIDIA/NeMo/labels/ASR", "name": "ASR", "color": "F3B33E", "default": false, "description": "" } ]

[]

[ "jenkins" ]