summer summer commited on
Upload 2 files
Browse files- modeling_loop_gqa_action.py +8 -8
- test_inference.py +11 -4
modeling_loop_gqa_action.py
CHANGED
|
@@ -457,9 +457,14 @@ class LoopTransformerLM(
|
|
| 457 |
self.token_embedding.weight
|
| 458 |
)
|
| 459 |
|
| 460 |
-
#
|
| 461 |
-
|
| 462 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 463 |
)
|
| 464 |
|
| 465 |
def _init_weights(
|
|
@@ -503,11 +508,6 @@ class LoopTransformerLM(
|
|
| 503 |
):
|
| 504 |
self.lm_head = new_embeddings
|
| 505 |
|
| 506 |
-
def tie_weights(self):
|
| 507 |
-
# Preserve the original exact shared parameter.
|
| 508 |
-
self.lm_head.weight = (
|
| 509 |
-
self.token_embedding.weight
|
| 510 |
-
)
|
| 511 |
|
| 512 |
def forward(
|
| 513 |
self,
|
|
|
|
| 457 |
self.token_embedding.weight
|
| 458 |
)
|
| 459 |
|
| 460 |
+
# Initialize Transformers bookkeeping for modern versions.
|
| 461 |
+
# This sets `all_tied_weights_keys` and performs standard
|
| 462 |
+
# PreTrainedModel post-initialization needed by from_pretrained().
|
| 463 |
+
self.post_init()
|
| 464 |
+
|
| 465 |
+
# Re-assert the exact shared parameter after post_init.
|
| 466 |
+
self.lm_head.weight = (
|
| 467 |
+
self.token_embedding.weight
|
| 468 |
)
|
| 469 |
|
| 470 |
def _init_weights(
|
|
|
|
| 508 |
):
|
| 509 |
self.lm_head = new_embeddings
|
| 510 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 511 |
|
| 512 |
def forward(
|
| 513 |
self,
|
test_inference.py
CHANGED
|
@@ -13,10 +13,17 @@ model = AutoModelForCausalLM.from_pretrained(
|
|
| 13 |
trust_remote_code=True,
|
| 14 |
device_map="auto",
|
| 15 |
)
|
|
|
|
| 16 |
model.eval()
|
| 17 |
|
| 18 |
-
print("
|
| 19 |
-
print("
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 20 |
|
| 21 |
prompt = "Task: Open Chrome and search for OpenAI.\nAction:"
|
| 22 |
|
|
@@ -26,7 +33,7 @@ inputs = tokenizer(
|
|
| 26 |
).to(model.device)
|
| 27 |
|
| 28 |
with torch.no_grad():
|
| 29 |
-
|
| 30 |
**inputs,
|
| 31 |
max_new_tokens=64,
|
| 32 |
do_sample=False,
|
|
@@ -34,4 +41,4 @@ with torch.no_grad():
|
|
| 34 |
pad_token_id=tokenizer.eos_token_id,
|
| 35 |
)
|
| 36 |
|
| 37 |
-
print(tokenizer.decode(
|
|
|
|
| 13 |
trust_remote_code=True,
|
| 14 |
device_map="auto",
|
| 15 |
)
|
| 16 |
+
|
| 17 |
model.eval()
|
| 18 |
|
| 19 |
+
print("class:", type(model).__name__)
|
| 20 |
+
print("has all_tied_weights_keys:", hasattr(model, "all_tied_weights_keys"))
|
| 21 |
+
print("all_tied_weights_keys:", getattr(model, "all_tied_weights_keys", None))
|
| 22 |
+
print(
|
| 23 |
+
"embedding tied:",
|
| 24 |
+
model.token_embedding.weight.data_ptr()
|
| 25 |
+
== model.lm_head.weight.data_ptr(),
|
| 26 |
+
)
|
| 27 |
|
| 28 |
prompt = "Task: Open Chrome and search for OpenAI.\nAction:"
|
| 29 |
|
|
|
|
| 33 |
).to(model.device)
|
| 34 |
|
| 35 |
with torch.no_grad():
|
| 36 |
+
output = model.generate(
|
| 37 |
**inputs,
|
| 38 |
max_new_tokens=64,
|
| 39 |
do_sample=False,
|
|
|
|
| 41 |
pad_token_id=tokenizer.eos_token_id,
|
| 42 |
)
|
| 43 |
|
| 44 |
+
print(tokenizer.decode(output[0], skip_special_tokens=True))
|