Skip to content

Commit c044554

Browse files
BBC-EsqBBC-Esq
authored
v7.8.1
1 parent 51951c4 commit c044554

File tree

1 file changed

+51
-51
lines changed

1 file changed

+51
-51
lines changed

‎src/module_process_images.py‎

Lines changed: 51 additions & 51 deletions
Original file line numberDiff line numberDiff line change
@@ -311,18 +311,18 @@ def __init__(self, config):
311311
if self.device == "cuda":
312312
if native in ("float32", "bfloat16") and has_bfloat16_support():
313313
self.dtype = torch.bfloat16
314-
print(f"OVIS: Selected bfloat16 precision based on native={native}")
314+
# print(f"OVIS: Selected bfloat16 precision based on native={native}")
315315
elif native == "float32":
316316
self.dtype = torch.float16
317-
print(f"OVIS: Selected float16 precision based on native={native}")
317+
# print(f"OVIS: Selected float16 precision based on native={native}")
318318
else:
319319
self.dtype = torch.float16
320-
print(f"OVIS: Selected float16 precision based on native={native}")
320+
# print(f"OVIS: Selected float16 precision based on native={native}")
321321
else:
322322
self.dtype = torch.float32
323-
print(f"OVIS: Selected float32 precision for CPU based on native={native}")
323+
# print(f"OVIS: Selected float32 precision for CPU based on native={native}")
324324

325-
print(f"OVIS: Device={self.device}, Initial dtype selection={self.dtype}")
325+
# print(f"OVIS: Device={self.device}, Initial dtype selection={self.dtype}")
326326

327327
def initialize_model_and_tokenizer(self):
328328
chosen_model = self.config["vision"]["chosen_model"]
@@ -341,50 +341,50 @@ def initialize_model_and_tokenizer(self):
341341
cache_dir=cache_dir
342342
).to(self.device)
343343

344-
# Print model layers precision before eval
345-
print("OVIS: Model layer precisions after loading:")
346-
for name, module in model.named_modules():
347-
if isinstance(module, (torch.nn.Linear, torch.nn.Conv2d, torch.nn.LayerNorm)):
348-
if hasattr(module, "weight") and module.weight is not None:
349-
print(f" Layer {name}: {module.weight.dtype}")
344+
# # Print model layers precision before eval
345+
# print("OVIS: Model layer precisions after loading:")
346+
# for name, module in model.named_modules():
347+
# if isinstance(module, (torch.nn.Linear, torch.nn.Conv2d, torch.nn.LayerNorm)):
348+
# if hasattr(module, "weight") and module.weight is not None:
349+
# print(f" Layer {name}: {module.weight.dtype}")
350350

351351
model.eval()
352352

353-
# Print model layers precision after eval
354-
print("OVIS: Model layer precisions after eval():")
355-
for name, module in model.named_modules():
356-
if isinstance(module, (torch.nn.Linear, torch.nn.Conv2d, torch.nn.LayerNorm)):
357-
if hasattr(module, "weight") and module.weight is not None:
358-
print(f" Layer {name}: {module.weight.dtype}")
353+
# # Print model layers precision after eval
354+
# print("OVIS: Model layer precisions after eval():")
355+
# for name, module in model.named_modules():
356+
# if isinstance(module, (torch.nn.Linear, torch.nn.Conv2d, torch.nn.LayerNorm)):
357+
# if hasattr(module, "weight") and module.weight is not None:
358+
# print(f" Layer {name}: {module.weight.dtype}")
359359

360360
text_tokenizer = model.get_text_tokenizer()
361361
visual_tokenizer = model.get_visual_tokenizer()
362362

363-
# Print visual tokenizer layer info before conversion
364-
print("OVIS: Visual tokenizer layer precisions before conversion:")
365-
for name, module in visual_tokenizer.named_modules():
366-
if isinstance(module, torch.nn.Linear):
367-
if hasattr(module, "weight") and module.weight is not None:
368-
print(f" VT Layer {name}: {module.weight.dtype}")
363+
# # Print visual tokenizer layer info before conversion
364+
# print("OVIS: Visual tokenizer layer precisions before conversion:")
365+
# for name, module in visual_tokenizer.named_modules():
366+
# if isinstance(module, torch.nn.Linear):
367+
# if hasattr(module, "weight") and module.weight is not None:
368+
# print(f" VT Layer {name}: {module.weight.dtype}")
369369

370-
# Count modules before conversion
371-
linear_count = sum(1 for module in visual_tokenizer.modules()
372-
if isinstance(module, torch.nn.Linear))
373-
print(f"OVIS: Found {linear_count} Linear modules in visual_tokenizer")
374-
375-
for module in visual_tokenizer.modules():
376-
if isinstance(module, torch.nn.Linear):
377-
old_dtype = module.weight.dtype if hasattr(module, "weight") else "unknown"
378-
module.to(device=self.device, dtype=self.dtype)
379-
new_dtype = module.weight.dtype if hasattr(module, "weight") else "unknown"
380-
print(f"OVIS: Converting module from {old_dtype} to {self.dtype}, result={new_dtype}")
370+
# # Count modules before conversion
371+
# linear_count = sum(1 for module in visual_tokenizer.modules()
372+
# if isinstance(module, torch.nn.Linear))
373+
# print(f"OVIS: Found {linear_count} Linear modules in visual_tokenizer")
374+
375+
# for module in visual_tokenizer.modules():
376+
# if isinstance(module, torch.nn.Linear):
377+
# old_dtype = module.weight.dtype if hasattr(module, "weight") else "unknown"
378+
# module.to(device=self.device, dtype=self.dtype)
379+
# new_dtype = module.weight.dtype if hasattr(module, "weight") else "unknown"
380+
# print(f"OVIS: Converting module from {old_dtype} to {self.dtype}, result={new_dtype}")
381381

382-
# Print visual tokenizer layer info after conversion
383-
print("OVIS: Visual tokenizer layer precisions after conversion:")
384-
for name, module in visual_tokenizer.named_modules():
385-
if isinstance(module, torch.nn.Linear):
386-
if hasattr(module, "weight") and module.weight is not None:
387-
print(f" VT Layer {name}: {module.weight.dtype}")
382+
# # Print visual tokenizer layer info after conversion
383+
# print("OVIS: Visual tokenizer layer precisions after conversion:")
384+
# for name, module in visual_tokenizer.named_modules():
385+
# if isinstance(module, torch.nn.Linear):
386+
# if hasattr(module, "weight") and module.weight is not None:
387+
# print(f" VT Layer {name}: {module.weight.dtype}")
388388

389389
# Save model for process_single_image
390390
self.model = model
@@ -399,29 +399,29 @@ def process_single_image(self, raw_image):
399399
)
400400
query = f"<image>\n{prompt}"
401401

402-
print("OVIS: Starting image processing")
402+
# print("OVIS: Starting image processing")
403403
_, input_ids, pixel_values = self.model.preprocess_inputs(query, [raw_image])
404-
print(f"OVIS: After preprocess_inputs - pixel_values dtype={pixel_values.dtype}")
404+
# print(f"OVIS: After preprocess_inputs - pixel_values dtype={pixel_values.dtype}")
405405

406406
attention_mask = torch.ne(input_ids, self.tokenizer.pad_token_id)
407407

408408
# Batchify and move to the correct device & dtype
409409
input_ids = input_ids.unsqueeze(0).to(self.device)
410410
attention_mask = attention_mask.unsqueeze(0).to(self.device)
411411

412-
print(f"OVIS: Before pixel_values conversion - dtype={pixel_values.dtype}")
412+
# print(f"OVIS: Before pixel_values conversion - dtype={pixel_values.dtype}")
413413
pixel_values = pixel_values.to(device=self.device, dtype=self.dtype)
414-
print(f"OVIS: After pixel_values conversion - dtype={pixel_values.dtype}")
414+
# print(f"OVIS: After pixel_values conversion - dtype={pixel_values.dtype}")
415415

416416
pixel_values = [pixel_values] # wrap in list for generate()
417417

418-
# Check model precision during inference
419-
print("OVIS: Model layer precisions during inference:")
420-
for name, module in self.model.named_modules():
421-
if isinstance(module, (torch.nn.Linear, torch.nn.Conv2d)):
422-
if hasattr(module, "weight") and module.weight is not None:
423-
if name.startswith("transformer") or name.startswith("lm_head"):
424-
print(f" Inference layer {name}: {module.weight.dtype}")
418+
# # Check model precision during inference
419+
# print("OVIS: Model layer precisions during inference:")
420+
# for name, module in self.model.named_modules():
421+
# if isinstance(module, (torch.nn.Linear, torch.nn.Conv2d)):
422+
# if hasattr(module, "weight") and module.weight is not None:
423+
# if name.startswith("transformer") or name.startswith("lm_head"):
424+
# print(f" Inference layer {name}: {module.weight.dtype}")
425425

426426
gen_kwargs = {
427427
"max_new_tokens": 1024,

0 commit comments

Comments
 (0)