minke.yu/sglang
1
0
Fork 0
Code Issues Pull Requests Actions 12 Packages Projects Releases Wiki Activity

Fix image (random multimodal) dataset token statistics (#22371)

Browse Source
This commit is contained in:
Bishwo Adhikari
2026-05-16 23:12:03 -07:00
committed by GitHub
parent 52875ab6f4
commit 46e0f5007d
1 changed files with 14 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
python/sglang/benchmark/datasets/image.py
+14
View File
@@ -293,6 +293,20 @@ def sample_image_requests(
else:
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(
f"\nCreated {len(dataset)} {image_content} {image_format} images with average {total_image_bytes // num_requests} bytes per request"
)