mirror of
https://github.com/docling-project/docling.git
synced 2026-05-17 13:10:38 +00:00
EliSchwartz
24f2d148d9
* feat(table-structure): swap VLM model to granite-vision-4.1-4b Updates GraniteVisionTableStructureModel to use the 4.1 model. The 4.1 weights are pre-merged, so merge_lora_adapters() is now hasattr-guarded. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> Signed-off-by: Eli Schwartz <eliyahu.schwartz@ibm.com> * feat(chart-extraction): swap V4 VLM model to granite-vision-4.1-4b Updates ChartExtractionModelGraniteVisionV4 to use the 4.1 model. hasattr-guards the merge_lora_adapters() call since 4.1 weights are pre-merged. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> Signed-off-by: Eli Schwartz <eliyahu.schwartz@ibm.com> * docs(example): mention granite-vision-4.1-4b in table-structure example Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> Signed-off-by: Eli Schwartz <eliyahu.schwartz@ibm.com> * docs(catalog): update Granite Vision entry to 4.1-4b Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> Signed-off-by: Eli Schwartz <eliyahu.schwartz@ibm.com> * feat(chart-extraction): honor cuda_use_flash_attention2 in V4 loader Mirrors the table-structure loader so ChartExtractionModelGraniteVisionV4 also passes _attn_implementation based on AcceleratorOptions. Without this the chart model falls back to the transformers SDPA default, which can hit cuDNN backend failures on some torch/cuDNN stacks while the table model (which already passed the flag) runs cleanly. Stores accelerator_options on the base class so subclasses can read it. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> Signed-off-by: Eli Schwartz <eliyahu.schwartz@ibm.com> * fix(model-downloader): update Granite Vision log message to 4.1 The log message in download_models still mentioned "Granite Vision 4.0" after the model swap. Correct it to match the current model version. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> Signed-off-by: Eli Schwartz <eliyahu.schwartz@ibm.com> * fix(chart-extraction): fall back to bare CSV when V4 model omits ```csv``` fence granite-vision-4.1-4b sometimes emits raw CSV without a ```csv``` code fence for the <chart2csv> prompt, which caused _extract_csv_to_dataframe to raise ValueError and drop the chart's tabular_chart metadata. Mirror the tolerant parsing already used by the v3 class: prefer a fenced block, otherwise strip any stray backtick prefix/suffix and parse the text as-is. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com> Signed-off-by: Eli Schwartz <eliyahu.schwartz@ibm.com> --------- Signed-off-by: Eli Schwartz <eliyahu.schwartz@ibm.com> Co-authored-by: Eli Schwartz <eliyahu.schwartz@ibm.com> Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>