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

Merge pull request #175 from VE-FORBRYDERNE/gptj-patch

Browse Source 
Fix GPT-J model loading in TPU Colab when `vocab_size` is not divisible by 8
This commit is contained in:
henk717
2022-11-03 00:13:50 +01:00
committed by GitHub
parent 7b5a766b4a b20d80ca2a
commit 09b5ffc09d
2 changed files with 3 additions and 3 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
maps/gptj.json
View File

@@ -9,11 +9,11 @@
}, },
"static_weights": { "static_weights": {
"transformer.wte.weight": {"mtj": {"module": "embedding_shard/~/linear", "param": "w", "transforms": ["no_transpose", "vocab_pad"]}}, "transformer.wte.weight": {"mtj": {"module": "embedding_shard/~/linear", "param": "w", "transforms": ["no_transpose", "vocab_pad"]}},
"transformer.wte.bias": {"mtj": {"module": "embedding_shard/~/linear", "param": "b"}}, "transformer.wte.bias": {"mtj": {"module": "embedding_shard/~/linear", "param": "b", "transforms": ["vocab_pad"]}},
"transformer.ln_f.weight": {"mtj": {"module": "projection_shard/~/replicated_layer_norm", "param": "scale"}}, "transformer.ln_f.weight": {"mtj": {"module": "projection_shard/~/replicated_layer_norm", "param": "scale"}},
"transformer.ln_f.bias": {"mtj": {"module": "projection_shard/~/replicated_layer_norm", "param": "offset"}}, "transformer.ln_f.bias": {"mtj": {"module": "projection_shard/~/replicated_layer_norm", "param": "offset"}},
"lm_head.weight": {"mtj": {"module": "projection_shard/~/linear", "param": "w", "transforms": ["vocab_pad"]}}, "lm_head.weight": {"mtj": {"module": "projection_shard/~/linear", "param": "w", "transforms": ["vocab_pad"]}},
"lm_head.bias": {"mtj": {"module": "projection_shard/~/linear", "param": "b"}} "lm_head.bias": {"mtj": {"module": "projection_shard/~/linear", "param": "b", "transforms": ["vocab_pad"]}}
}, },
"layer_weights": { "layer_weights": {
"transformer.h.{layer}.attn.bias": {}, "transformer.h.{layer}.attn.bias": {},

2
tpu_mtj_backend.py
View File

@@ -1304,7 +1304,7 @@ def load_model(path: str, driver_version="tpu_driver0.1_dev20210607", hf_checkpo
if "divide_by_shards" in transforms: if "divide_by_shards" in transforms:
tensor /= params["cores_per_replica"] tensor /= params["cores_per_replica"]
if "vocab_pad" in transforms: if "vocab_pad" in transforms:
tensor = torch.nn.functional.pad(tensor, (0, 0, 0, params["n_vocab_padding"])) tensor = torch.nn.functional.pad(tensor, (0,) * (tensor.ndim * 2 - 1) + (params["n_vocab_padding"],))
if "no_transpose" not in transforms and tensor.ndim == 2: if "no_transpose" not in transforms and tensor.ndim == 2:
tensor = tensor.T tensor = tensor.T
tensor.unsqueeze_(0) tensor.unsqueeze_(0)