Update app.py
Browse files
app.py
CHANGED
|
@@ -7,9 +7,12 @@ from fastapi import FastAPI, UploadFile, File, HTTPException
|
|
| 7 |
# --- CPU OPTIMIZATION FLAGS ---
|
| 8 |
# Limit threads for HF Free Tier (2 vCPUs)
|
| 9 |
os.environ["OMP_NUM_THREADS"] = "2"
|
| 10 |
-
# MKLDNN works perfectly with PPStructureV3 for a massive Intel speedup
|
| 11 |
|
| 12 |
-
#
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
from paddleocr import PPStructureV3
|
| 14 |
|
| 15 |
app = FastAPI(
|
|
@@ -17,7 +20,7 @@ app = FastAPI(
|
|
| 17 |
description="Lightweight PP-StructureV3 extraction"
|
| 18 |
)
|
| 19 |
|
| 20 |
-
print("Initializing PP-StructureV3 (MKLDNN Enabled)...")
|
| 21 |
pipeline = PPStructureV3()
|
| 22 |
print("Pipeline ready!")
|
| 23 |
|
|
@@ -34,12 +37,11 @@ async def ingest_document(file: UploadFile = File(...)):
|
|
| 34 |
with open(file_path, "wb") as buffer:
|
| 35 |
shutil.copyfileobj(file.file, buffer)
|
| 36 |
|
| 37 |
-
# 2. Predict
|
| 38 |
output = pipeline.predict(file_path)
|
| 39 |
|
| 40 |
parsed_pages = []
|
| 41 |
for page_num, res in enumerate(output):
|
| 42 |
-
# We can save to the temp directory exactly like the VLM pipeline
|
| 43 |
md_path = os.path.join(temp_dir, f"page_{page_num + 1}.md")
|
| 44 |
json_path = os.path.join(temp_dir, f"page_{page_num + 1}.json")
|
| 45 |
|
|
|
|
| 7 |
# --- CPU OPTIMIZATION FLAGS ---
|
| 8 |
# Limit threads for HF Free Tier (2 vCPUs)
|
| 9 |
os.environ["OMP_NUM_THREADS"] = "2"
|
|
|
|
| 10 |
|
| 11 |
+
# THE MAGIC FIX: Disable the buggy PIR compiler but KEEP MKLDNN!
|
| 12 |
+
os.environ["FLAGS_enable_pir_api"] = "0"
|
| 13 |
+
os.environ["FLAGS_use_mkldnn"] = "1"
|
| 14 |
+
|
| 15 |
+
# IMPORT THE NEW V3 PIPELINE (Must be imported AFTER setting the flags)
|
| 16 |
from paddleocr import PPStructureV3
|
| 17 |
|
| 18 |
app = FastAPI(
|
|
|
|
| 20 |
description="Lightweight PP-StructureV3 extraction"
|
| 21 |
)
|
| 22 |
|
| 23 |
+
print("Initializing PP-StructureV3 (MKLDNN Enabled via AST Executor)...")
|
| 24 |
pipeline = PPStructureV3()
|
| 25 |
print("Pipeline ready!")
|
| 26 |
|
|
|
|
| 37 |
with open(file_path, "wb") as buffer:
|
| 38 |
shutil.copyfileobj(file.file, buffer)
|
| 39 |
|
| 40 |
+
# 2. Predict
|
| 41 |
output = pipeline.predict(file_path)
|
| 42 |
|
| 43 |
parsed_pages = []
|
| 44 |
for page_num, res in enumerate(output):
|
|
|
|
| 45 |
md_path = os.path.join(temp_dir, f"page_{page_num + 1}.md")
|
| 46 |
json_path = os.path.join(temp_dir, f"page_{page_num + 1}.json")
|
| 47 |
|