making converter additionally support deepseek-coder dense hf model #167
Conversation
add deepseek1b support
add two directions converter (hf->olmo, olmo->hf)
clean imports
replace hf to olmo conveter
| MODEL_CONFIG: TransformerConfig | ||
| if HF_MODEL == "meta-llama/Llama-3.2-1B": | ||
| if "Llama-3.2-1B" in HF_MODEL: |
There was a problem hiding this comment.
Do we also need this for bigger Llama models?
There was a problem hiding this comment.
I change it so that it works with local cache dir. We are encountering some additional weird bug related to conversion; and requires some help from you. We will submit a separate issue regarding that. Do you prefer we incorporate solution to that issue into this PR or should I separate that from this current PR?
There was a problem hiding this comment.
Many small PRs is always better, if possible.
# Conflicts: # src/examples/huggingface/convert_checkpoint_to_hf.py
|
@lingchensanwen , I merged current main into this. There were a fair number of conflicts, and I'm not sure I did it right. Can you check? |
@dirkgr, I think I've fixed this. Could you take a look at the new one? |
| SAVE_PATH = f"{os.environ['SHARE_RES_DIR']}/models/deepseek/olmo/deepseek-coder-1.3b-base" | ||
| SAVE_OVERWRITE = True |
There was a problem hiding this comment.
This is hardcoded to deepseek, and the SAVE_OVERWRITE was probably a leftover statement? It should default to the safe setting, so people don't overwrite their stuff by accident.
| ) | ||
|
|
||
| elif "deepseek-coder-1.3b-base" in HF_MODEL: |
There was a problem hiding this comment.
Does this have to happen for all DeepSeek models?
There was a problem hiding this comment.
yes, we defined class for this version deepseek following the olmo core setup. For other version, users would need to define their own class as well like this. Also, I can modify this to elif HF_MODEL.startswith("deepseek-coder-1.3b-base"): to follow the same format as repo did.
| ) | ||
| else: | ||
| # Extract RoPE scaling config | ||
| # from pdb import set_trace; set_trace() |
|
|
||
| with open(config_path, "r") as f: | ||
| olmo_config_dict = json.load(f)["model"] | ||
| # import pdb; pdb.set_trace() |
There was a problem hiding this comment.
resolved all above in new commit @dirkgr
Changes made:
@dirkgr