125m checkpoint outputting gibberish

[patrickvonplaten]

Converting the sharded checkpoints of 125m to a singleton checkpoint with #60:

    $ ls 125m
    dict.txt
    gpt2-merges.txt
    gpt2-vocab.json
    reshard-model_part-0.pt
    reshard-model_part-1.pt
    $ python -m metaseq.scripts.convert_to_singleton 125m

gives a new

restored.pt

file.

I then transformed the checkpoint into the same format as 350m to test some generation on it:

import torch
orig_state = torch.load("./reshard-model_part-0.pt")
model = torch.load("./restored.pt")

orig_state["model"] = model  # this format allows one to use the standard `checkpoint_utils.load_model_ensemble_and_task` function
orig_state["cfg"]["model"]._name = "transformer_lm"  # we change the architecture name to "transformer_lm" to be able to run it in a non-CUDA environment
torch.save(orig_state, "./reshard.pt")

I tried running an inference example on the model to see whether the generation works as expected.
Here the code:

import os

from transformers import GPT2Tokenizer
from metaseq import checkpoint_utils
import torch

path = "/home/patrick/add_opt"

"""
$ ls path
vocab.json
merges.txt
reshard.pt
"""


tokenizer = GPT2Tokenizer.from_pretrained("patrickvonplaten/opt_gpt2_tokenizer")
tokenizer.save_pretrained(path)

paths = [os.path.join(path, "reshard.pt")]

checkpoint = checkpoint_utils.load_model_ensemble_and_task(
    paths,
    arg_overrides={
        "vocab_filename": os.path.join(path, "vocab.json"),
        "merges_filename": os.path.join(path, "merges.txt"),
    }
)

model = checkpoint[0][0].eval()


# forward passes
def single_batch_forward_logits(prompts):
    input_ids = tokenizer(prompts, return_tensors="pt").input_ids
    input_ids = torch.cat([torch.tensor([[2]]), input_ids], dim=-1)
    logits = model(input_ids)[0]
    return logits


prompts = [
    "Today is a beautiful day and I want to",
    "In the city of",
    "Paris is the capital of France and",
    "Computers and mobile phones have taken",
]


print("Next word generation")
for prompt in prompts:
    print("-------------")
    print(f"Prompt: {prompt}...\n")
    logits = single_batch_forward_logits(prompt)
    pred_next_token = torch.argmax(logits[0, -1], -1)
    next_token = tokenizer.convert_ids_to_tokens([pred_next_token])
    next_token = next_token[0].replace("Ġ", "")
    print(f"Next word: {next_token}")
    print("-------------")

This sadly gives gibberish:

Next word generation
-------------
Prompt: Today is a beautiful day and I want to...

Next word: Robbins
-------------
-------------
Prompt: In the city of...

Next word: of
-------------
-------------
Prompt: Paris is the capital of France and...

Next word: Robbins
-------------
-------------
Prompt: Computers and mobile phones have taken...

Next word: Robbins
-------------

Note that this script works perfectly fine with the 350m checkpoint.

@stephenroller - any ideas?

[patrickvonplaten]

This holds also true when running the generation directly in the conversion script of #60

[patrickvonplaten]

The same holds true for the 1B3 checkpoint when using #72

Any ideas what could be the problem @stephenroller @suchenzang ?

Do you have a script that allows to use 125m or 1B3 in inference?

[Fengwills]

@patrickvonplaten hello, thanks for your share, can share the dict.txt? how to get it ?
can #60 be used on 350M model?
many thanks

[mrseeker]

@patrickvonplaten I have a conversion script that our team used to convert the fairseq models to XGLM. Can use it to test it out if that works? I am stuck on the script though, since it requires 2 GPU instances(?) to fully function.

[stephenroller]

Will try this in metaseq today.

[patrickvonplaten]

@mrseeker,

If it helps here is the "converted" fsq checkpoint . You should be able to run this on CPU

[patrickvonplaten]

For the conversion we use multi-GPU machines though

[suchenzang]

Seems to be addressed in huggingface/transformers#17088 - feel free to re-open this issue if not.