rixty/KoboldAI-Client
1
1
Fork 0
mirror of https://github.com/KoboldAI/KoboldAI-Client.git synced 2025-06-05 21:59:24 +02:00
Code Issues Projects Releases Wiki Activity

Add TPU support for OPT-350M

Browse Source 
The 350M model seems to have a different structure than the other ones ???
This commit is contained in:
Gnome Ann
2022-05-12 22:21:15 -04:00
parent dfa2aa7314
commit 4fa5f1cd6a
3 changed files with 9 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
aiserver.py
View File

@ -772,7 +772,7 @@ def spRequest(filename):
tensor = tensor.reshape(
tpu_mtj_backend.params["cores_per_replica"],
-1,
tpu_mtj_backend.params["d_model"],
tpu_mtj_backend.params.get("d_embed", tpu_mtj_backend.params["d_model"]),
)
vars.sp = tpu_mtj_backend.shard_xmap(np.float32(tensor))
else:
@ -1574,14 +1574,14 @@ else:
global np
if 'np' not in globals():
import numpy as np
tensor = np.zeros((1, tpu_mtj_backend.params["d_model"]), dtype=np.float32)
tensor = np.zeros((1, tpu_mtj_backend.params.get("d_embed", tpu_mtj_backend.params["d_model"])), dtype=np.float32)
rows = tensor.shape[0]
padding_amount = tpu_mtj_backend.params["seq"] - (tpu_mtj_backend.params["seq"] % -tpu_mtj_backend.params["cores_per_replica"]) - rows
tensor = np.pad(tensor, ((0, padding_amount), (0, 0)))
tensor = tensor.reshape(
tpu_mtj_backend.params["cores_per_replica"],
-1,
tpu_mtj_backend.params["d_model"],
tpu_mtj_backend.params.get("d_embed", tpu_mtj_backend.params["d_model"]),
)
vars.sp = tpu_mtj_backend.shard_xmap(tensor)
soft_tokens = np.arange(
@ -1682,7 +1682,7 @@ else:
loadmodelsettings()
loadsettings()
tpu_mtj_backend.load_model(vars.custmodpth, hf_checkpoint=vars.model not in ("TPUMeshTransformerGPTJ", "TPUMeshTransformerGPTNeoX") and vars.use_colab_tpu, **vars.modelconfig)
vars.modeldim = int(tpu_mtj_backend.params["d_model"])
vars.modeldim = int(tpu_mtj_backend.params.get("d_embed", tpu_mtj_backend.params["d_model"]))
tokenizer = tpu_mtj_backend.tokenizer
else:
loadsettings()