Fix image (random multimodal) dataset token statistics (#22371)
This commit is contained in:
@@ -293,6 +293,20 @@ def sample_image_requests(
|
|||||||
else:
|
else:
|
||||||
print(f"#Images per request: {image_count} (fixed)")
|
print(f"#Images per request: {image_count} (fixed)")
|
||||||
|
|
||||||
|
# Detailed token breakdown (derived from dataset + input_lens)
|
||||||
|
text_prompt_lens = np.array([r.text_prompt_len for r in dataset])
|
||||||
|
vision_prompt_lens = np.array([r.vision_prompt_len for r in dataset])
|
||||||
|
text_prompt_overheads = text_prompt_lens - input_lens
|
||||||
|
stat_fields = [
|
||||||
|
("Raw text prompt tokens (without overhead)", input_lens),
|
||||||
|
("Text prompt tokens (with chat template)", text_prompt_lens),
|
||||||
|
("Text prompt overhead", text_prompt_overheads),
|
||||||
|
("Vision tokens", vision_prompt_lens),
|
||||||
|
]
|
||||||
|
print("\n=== Token Breakdown (per request avg / total) ===")
|
||||||
|
for label, vals in stat_fields:
|
||||||
|
print(f" {label}: avg={np.mean(vals):.1f}, total={np.sum(vals)}")
|
||||||
|
|
||||||
print(
|
print(
|
||||||
f"\nCreated {len(dataset)} {image_content} {image_format} images with average {total_image_bytes // num_requests} bytes per request"
|
f"\nCreated {len(dataset)} {image_content} {image_format} images with average {total_image_bytes // num_requests} bytes per request"
|
||||||
)
|
)
|
||||||
|
|||||||
Reference in New Issue
Block a user