Updated template and image processing code to match native inference exactly

chat_template.jinja

@@ -1 +1 @@
-{% set DEMO_STYLES = ['point_count','pointing','cosyn_point','user_qa','long_caption','short_caption','video_long_caption','video_short_caption','video_point_track_per_frame','video_point_track_start_end','video_point_track_all_frames','video_single_point_track_start_end','video_transcript','video_clip_caption_start_end','video_clip_caption_start_end_in_seconds','video_clip_transcript_start_end','video_clip_transcript_start_end_in_seconds','video_frame_caption_timestamp','video_frame_caption_timestamp_in_seconds','correction_qa','text_sft','video_point','video_point_count','video_count','video_count_point','multi_image_pointing','multi_image_counting','multi_image_point_then_count','multi_image_count_then_point','demo','a_okvqa_mc','ai2_diagram_no_letter','ai2_diagram','science_qa','multi_image_mc','multi_image_mc_exp','mantis_instruct_mc','video_multiple_choice','video_multiple_choice_count_without_pointing','video_multiple_choice_multiple_correct','video_multiple_choice_w_subtitle'] %}{% set image_count = namespace(value=0) %}{% set video_count = namespace(value=0) %}{% set has_subtitle = messages and messages[0]['role'].lower() == 'subtitle' %}{% for message in messages %}{% if message['content'] is not string %}{% for content in message['content'] %}{% if content['type'] == 'image' or 'image' in content or 'image_url' in content %}{% set image_count.value = image_count.value + 1 %}{% elif content['type'] == 'video' or 'video' in content or 'video_url' in content %}{% set video_count.value = video_count.value + 1 %}{% endif %}{% endfor %}{% endif %}{% endfor %}{% if image_count.value == 1 %}{{ '<|image|>' }}{% elif image_count.value > 1 %}{% for i in range(image_count.value) %}{{ 'Image ' ~ (i + 1) ~ '<|image|>' }}{% endfor %}{% endif %}{% for _ in range(video_count.value) %}{{ '<|video|>' }}{% endfor %}{% if has_subtitle %}{{ messages[0]['content'] }}{% endif %}{% for message in messages %}{% set role = message['role'].lower() %}{% if role == 'subtitle' %}{% continue %}{% endif %}{% set conv_index = loop.index - (1 if has_subtitle else 0) %}{%- if (conv_index % 2 == 1 and role != 'user') or (conv_index % 2 == 0 and role != 'assistant') -%}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{%- endif -%}{% if message['content'] is string %}{% set text_content = message['content'] %}{% else %}{% set m = namespace(text='') %}{% for content in message['content'] %}{% if content['type'] == 'text' %}{% if content['style'] is defined and content['style'] not in DEMO_STYLES %}{% set seg = content['style'] ~ ': ' ~ content['text'] %}{% else %}{% set seg = content['text'] %}{% endif %}{% set m.text = m.text ~ ('' if not m.text else ' ') ~ seg %}{% endif %}{% endfor %}{% set text_content = m.text %}{% endif %}{% if role == 'user' %}{% if not (has_subtitle and loop.index == 2) and not (not has_subtitle and loop.first) %}{{ '<|im_end|>\n' }}{% endif %}{{ '<|im_start|>user\n' }}{{ text_content }}{{ '<|im_end|>\n' }}{% else %} {# assistant #}{{ '<|im_start|>assistant\n' }}{{ text_content }}{% endif %}{% endfor %}{% if add_generation_prompt %}{{ '<|im_start|>assistant\n' }}{% endif %}
+{% set DEMO_STYLES = ['point_count','pointing','cosyn_point','user_qa','long_caption','short_caption','video_long_caption','video_short_caption','video_point_track_per_frame','video_point_track_start_end','video_point_track_all_frames','video_single_point_track_start_end','video_transcript','video_clip_caption_start_end','video_clip_caption_start_end_in_seconds','video_clip_transcript_start_end','video_clip_transcript_start_end_in_seconds','video_frame_caption_timestamp','video_frame_caption_timestamp_in_seconds','correction_qa','text_sft','video_point','video_point_count','video_count','video_count_point','multi_image_pointing','multi_image_counting','multi_image_point_then_count','multi_image_count_then_point','demo','a_okvqa_mc','ai2_diagram_no_letter','ai2_diagram','science_qa','multi_image_mc','multi_image_mc_exp','mantis_instruct_mc','video_multiple_choice','video_multiple_choice_count_without_pointing','video_multiple_choice_multiple_correct','video_multiple_choice_w_subtitle'] %}{% set image_count = namespace(value=0) %}{% set video_count = namespace(value=0) %}{% set has_subtitle = messages and messages[0]['role'].lower() == 'subtitle' %}{% for message in messages %}{% if message['content'] is not string %}{% for content in message['content'] %}{% if content['type'] == 'image' or 'image' in content or 'image_url' in content %}{% set image_count.value = image_count.value + 1 %}{% elif content['type'] == 'video' or 'video' in content or 'video_url' in content %}{% set video_count.value = video_count.value + 1 %}{% endif %}{% endfor %}{% endif %}{% endfor %}{% if image_count.value == 1 %}{{ '<|image|>' }}{% elif image_count.value > 1 %}{% for i in range(image_count.value) %}{{ 'Image ' ~ (i + 1) ~ '<|image|>' }}{% endfor %}{% endif %}{% for _ in range(video_count.value) %}{{ '<|video|>' }}{% endfor %}{% if has_subtitle %}{{ messages[0]['content'] }}{% endif %}{% for message in messages %}{% set role = message['role'].lower() %}{% if role == 'subtitle' %}{% continue %}{% endif %}{% set conv_index = loop.index - (1 if has_subtitle else 0) %}{%- if (conv_index % 2 == 1 and role != 'user') or (conv_index % 2 == 0 and role != 'assistant') -%}{{ raise_exception('Conversation roles must alternate user/assistant/user/assistant/...') }}{%- endif -%}{% if message['content'] is string %}{% set text_content = message['content'] %}{% else %}{% set m = namespace(text='') %}{% for content in message['content'] %}{% if content['type'] == 'text' %}{% if content['style'] is defined and content['style'] not in DEMO_STYLES %}{% set seg = content['style'] ~ ': ' ~ content['text'] %}{% else %}{% set seg = content['text'] %}{% endif %}{% set m.text = m.text ~ ('' if not m.text else ' ') ~ seg %}{% endif %}{% endfor %}{% set text_content = m.text %}{% endif %}{% if role == 'user' %}User: {{ text_content }}{% else %} {{ text_content }}{% endif %}{% endfor %}{% if add_generation_prompt %} Assistant:{% endif %}

image_processing_molmo2.py

@@ -156,8 +156,8 @@ def build_overlapping_crops(
     # Decide how to tile the image, to account for the overlap margins we compute the tiling
     # as if we had an image without the margins and were using a crop size without the margins
     tiling = select_tiling(
-        original_image_h - total_margin_pixels,
-        original_image_w - total_margin_pixels,
+        max(original_image_h - total_margin_pixels, 1),
+        max(original_image_w - total_margin_pixels, 1),
         crop_window_size,
         max_crops,
     )