[doc] add examples and minor updates #1071
Conversation
facebook-github-bot
added
the
CLA Signed
zyaoj
changed the title
| @@ -34,7 +34,7 @@ | |||
| " load_text_tokenizer,\n", | |||
| " setup_gangs,\n", | |||
| ")\n", | |||
| "from fairseq2.recipes.config import GangSection\n", | |||
| "from fairseq2.recipes.config import GangSection, ModelSection\n", | |||
There was a problem hiding this comment.
why defining it as
dataset_config.name = "gsm8k_sft"
dataset_config.path = Path("/path/to/gsm8k_data/sft")
and not directly like
dataset_config = InstructionFinetuneDatasetSection(name = "gsm8k_sft", path = Path("/path/to/gsm8k_data/sft"))
?
There was a problem hiding this comment.
same question for config = Config() # instantiate an object
There was a problem hiding this comment.
it would be interesting to say something about the expected data format in "/path/to/gsm8k_data/sft" (unless it's explained elsewhere) !
| "config = Config() # instantiate an object\n", | ||
| "config.gang = GangSection(tensor_parallel_size=1)\n", | ||
| "config.dataset = dataset_config\n", | ||
| "config.model = ModelSection(name=\"llama3_1_8b\")\n", |
There was a problem hiding this comment.
| "config = Config() # instantiate an object\n", | |
| "config.gang = GangSection(tensor_parallel_size=1)\n", | |
| "config.dataset = dataset_config\n", | |
| "config.model = ModelSection(name=\"llama3_1_8b\")\n", | |
| "config = Config(gang = GangSection(tensor_parallel_size=1), dataset = dataset_config, model = ModelSection(name=\"llama3_1_8b\")) |
would this work as well ?
| " dropout_p=0.1 # Dropout probability\n", | ||
| ")\n", | ||
| "\n", | ||
| "model = create_llama_model(custom_config)\n", |
There was a problem hiding this comment.
nit : maybe add a comment that this will init a model with some random weights.
I would show also device and dtype args that are important here
| "cell_type": "markdown", | ||
| "metadata": {}, | ||
| "source": [ | ||
| "You can also fetch some config presets from model hub." |
There was a problem hiding this comment.
| "You can also fetch some config presets from model hub." | |
| "You can also fetch some registered configs available in model hub." |
| "model_hub = get_llama_model_hub()\n", | ||
| "model_config = model_hub.load_config(\"llama3_1_8b_instruct\") # use llama3.1 8b preset as an example\n", | ||
| "\n", | ||
| "llama_model = create_llama_model(model_config)\n", |
There was a problem hiding this comment.
can you show also a path with llama_model = model_hub.load(model_card) ?
| "from fairseq2.context import get_runtime_context\n", | ||
| "context = get_runtime_context()\n", | ||
| "asset_store = context.asset_store" |
There was a problem hiding this comment.
maybe some comment about why context and how it's related the fs2 possible extensions.
| "outputs": [ | ||
| { | ||
| "data": { | ||
| "text/plain": [ |
There was a problem hiding this comment.
I think we DONT want to share our clusters info here ?it will not be relevant for general purpose usage
What does this PR do? Please describe:
Does your PR introduce any breaking changes? If yes, please list them:
N/A
Check list: