mirrors/docling
1
0
Fork 0
mirror of https://github.com/DS4SD/docling.git synced 2025-07-26 20:14:47 +00:00
Code Issues Packages Projects Releases Wiki Activity

need to get Phi4 working again ...

Browse Source 
Signed-off-by: Peter Staar <taa@zurich.ibm.com>
This commit is contained in:
Peter Staar 2025-05-15 07:32:55 +02:00
parent 15a8f328c2
commit e2c95d09bc
2 changed files with 27 additions and 22 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

41
docling/models/hf_vlm_models/hf_vlm_model_AutoModelForCausalLM.py
View File

@ -73,21 +73,8 @@ class HuggingFaceVlmModel_AutoModelForCausalLM(BasePageModel):
artifacts_path,
trust_remote_code=self.trust_remote_code,
)
if not self.param_quantized:
self.vlm_model = AutoModelForCausalLM.from_pretrained(
artifacts_path,
device_map=self.device,
torch_dtype=torch.bfloat16,
_attn_implementation=(
"flash_attention_2"
if self.device.startswith("cuda")
and accelerator_options.cuda_use_flash_attention2
else "eager"
),
trust_remote_code=self.trust_remote_code,
).to(self.device)
else:
if self.param_quantized:
print("using quantized")
self.vlm_model = AutoModelForCausalLM.from_pretrained(
artifacts_path,
device_map=self.device,
@ -100,7 +87,21 @@ class HuggingFaceVlmModel_AutoModelForCausalLM(BasePageModel):
else "eager"
),
trust_remote_code=self.trust_remote_code,
).to(self.device)
) # .to(self.device)
else:
print("using original")
self.vlm_model = AutoModelForCausalLM.from_pretrained(
artifacts_path,
device_map=self.device,
torch_dtype="auto", # torch.bfloat16,
_attn_implementation=(
"flash_attention_2"
if self.device.startswith("cuda")
and accelerator_options.cuda_use_flash_attention2
else "eager"
),
trust_remote_code=self.trust_remote_code,
) # .to(self.device)
model_path = artifacts_path
@ -118,7 +119,8 @@ class HuggingFaceVlmModel_AutoModelForCausalLM(BasePageModel):
with TimeRecorder(conv_res, "vlm"):
assert page.size is not None
hi_res_image = page.get_image(scale=self.vlm_options.scale)
hi_res_image = page.get_image(scale=2) # self.vlm_options.scale)
# hi_res_image.show()
if hi_res_image is not None:
im_width, im_height = hi_res_image.size
@ -129,6 +131,7 @@ class HuggingFaceVlmModel_AutoModelForCausalLM(BasePageModel):
# Define prompt structure
prompt = self.formulate_prompt()
print(f"prompt: '{prompt}', size: {im_width}, {im_height}")
inputs = self.processor(
text=prompt, images=hi_res_image, return_tensors="pt"
@ -138,8 +141,8 @@ class HuggingFaceVlmModel_AutoModelForCausalLM(BasePageModel):
start_time = time.time()
generate_ids = self.vlm_model.generate(
**inputs,
max_new_tokens=self.max_new_tokens,
use_cache=self.use_cache, # Enables KV caching which can improve performance
max_new_tokens=4096, # self.max_new_tokens,
# use_cache=self.use_cache, # Enables KV caching which can improve performance
generation_config=self.generation_config,
num_logits_to_keep=1,
)

8
docs/examples/minimal_vlm_pipeline.py
View File

@ -6,13 +6,11 @@ from docling_core.types.doc import DocItemLabel, ImageRefMode
from docling_core.types.doc.document import DEFAULT_EXPORT_LABELS
from docling.datamodel.base_models import InputFormat
from docling.datamodel.pipeline_options import (
from docling.datamodel.pipeline_model_specializations import (
HuggingFaceVlmOptions,
InferenceFramework,
ResponseFormat,
VlmPipelineOptions,
granite_vision_vlm_conversion_options,
granite_vision_vlm_mlx_conversion_options,
granite_vision_vlm_ollama_conversion_options,
phi_vlm_conversion_options,
pixtral_12b_vlm_conversion_options,
@ -21,6 +19,9 @@ from docling.datamodel.pipeline_options import (
smoldocling_vlm_conversion_options,
smoldocling_vlm_mlx_conversion_options,
)
from docling.datamodel.pipeline_options import (
VlmPipelineOptions,
)
from docling.document_converter import DocumentConverter, PdfFormatOption
from docling.pipeline.vlm_pipeline import VlmPipeline
@ -49,6 +50,7 @@ pipeline_options.generate_page_images = True
# pipeline_options.vlm_options = granite_vision_vlm_conversion_options
pipeline_options.vlm_options = phi_vlm_conversion_options
# pipeline_options.vlm_options = qwen25_vl_3b_vlm_mlx_conversion_options
"""
pixtral_vlm_conversion_options = HuggingFaceVlmOptions(