HPI Supports paddle 3.2 (#4754)

* HPI allow untested Paddle versions for safe mode and update paddle version in HPS

* Update deps

* Limit CC>=8.0 for SDPA

Author: Bobholamovic

deploy/hps/server_env/Dockerfile

@@ -46,6 +46,7 @@ ENV PYTHONUNBUFFERED=1
 ENV PYTHONDONTWRITEBYTECODE=1
 ENV PIP_INDEX_URL=${PIP_INDEX_URL}
 
+RUN python -m pip install pip==25.2
 
 # Requirement collection
 FROM base AS rc

deploy/hps/server_env/cpu_version.txt

@@ -1 +1 @@
-0.3.9
+0.3.10

deploy/hps/server_env/gpu_version.txt

@@ -1 +1 @@
-0.3.10
+0.3.11

deploy/hps/server_env/requirements/app.in

@@ -4,4 +4,5 @@ numpy >= 1.24
 opencv-contrib-python == 4.10.0.84
 pycocotools >= 2
 pydantic >= 2
+safetensors @ https://paddle-whl.bj.bcebos.com/nightly/cu126/safetensors/safetensors-0.6.2.dev0-cp38-abi3-linux_x86_64.whl
 typing-extensions >= 4.11

deploy/hps/server_env/requirements/cpu.in

@@ -1 +1 @@
-paddlepaddle @ https://paddle-whl.bj.bcebos.com/stable/cpu/paddlepaddle/paddlepaddle-3.1.1-cp310-cp310-linux_x86_64.whl
+paddlepaddle @ https://paddle-whl.bj.bcebos.com/stable/cpu/paddlepaddle/paddlepaddle-3.2.1-cp310-cp310-linux_x86_64.whl

deploy/hps/server_env/requirements/cpu.txt

@@ -171,6 +171,7 @@ lxml==5.3.1
     # via
     #   paddlex (../../../setup.py)
     #   premailer
+    #   python-docx
 markupsafe==3.0.2
     # via jinja2
 marshmallow==3.26.1
@@ -238,7 +239,7 @@ packaging==24.2
     #   matplotlib
     #   paddlex (../../../setup.py)
     #   scikit-image
-paddlepaddle @ https://paddle-whl.bj.bcebos.com/stable/cpu/paddlepaddle/paddlepaddle-3.1.1-cp310-cp310-linux_x86_64.whl
+paddlepaddle @ https://paddle-whl.bj.bcebos.com/stable/cpu/paddlepaddle/paddlepaddle-3.2.1-cp310-cp310-linux_x86_64.whl
     # via -r requirements/cpu.in
 pandas==1.3.5
     # via paddlex (../../../setup.py)
@@ -295,6 +296,8 @@ python-dateutil==2.9.0.post0
     # via
     #   matplotlib
     #   pandas
+python-docx==1.2.0
+    # via paddlex (../../../setup.py)
 pytz==2025.1
     # via pandas
 pyyaml==6.0.2
@@ -326,8 +329,11 @@ ruamel-yaml==0.18.10
     # via paddlex (../../../setup.py)
 ruamel-yaml-clib==0.2.12
     # via ruamel-yaml
-safetensors==0.6.2
-    # via paddlex (../../../setup.py)
+safetensors @ https://paddle-whl.bj.bcebos.com/nightly/cu126/safetensors/safetensors-0.6.2.dev0-cp38-abi3-linux_x86_64.whl
+    # via
+    #   -r requirements/app.in
+    #   paddlepaddle
+    #   paddlex (../../../setup.py)
 scikit-image==0.24.0
     # via paddlex (../../../setup.py)
 scikit-learn==1.6.1
@@ -396,6 +402,7 @@ typing-extensions==4.12.2
     #   paddlex (../../../setup.py)
     #   pydantic
     #   pydantic-core
+    #   python-docx
     #   sqlalchemy
     #   typing-inspect
     #   uvicorn

deploy/hps/server_env/requirements/gpu.in

@@ -1 +1 @@
-paddlepaddle-gpu @ https://paddle-model-ecology.bj.bcebos.com/paddlex/PaddleX3.0/deploy/deps/paddlepaddle/paddlepaddle_gpu-3.1.1%2Bfc-cp310-cp310-linux_x86_64.whl
+paddlepaddle-gpu @ https://paddle-model-ecology.bj.bcebos.com/paddlex/PaddleX3.0/deploy/deps/paddlepaddle/paddlepaddle_gpu-3.2.1%2Bfc-cp310-cp310-linux_x86_64.whl

deploy/hps/server_env/requirements/gpu.txt

@@ -171,6 +171,7 @@ lxml==5.3.1
     # via
     #   paddlex (../../../setup.py)
     #   premailer
+    #   python-docx
 markupsafe==3.0.2
     # via jinja2
 marshmallow==3.26.1
@@ -238,7 +239,7 @@ packaging==24.2
     #   matplotlib
     #   paddlex (../../../setup.py)
     #   scikit-image
-paddlepaddle-gpu @ https://paddle-model-ecology.bj.bcebos.com/paddlex/PaddleX3.0/deploy/deps/paddlepaddle/paddlepaddle_gpu-3.1.1%2Bfc-cp310-cp310-linux_x86_64.whl
+paddlepaddle-gpu @ https://paddle-model-ecology.bj.bcebos.com/paddlex/PaddleX3.0/deploy/deps/paddlepaddle/paddlepaddle_gpu-3.2.1%2Bfc-cp310-cp310-linux_x86_64.whl
     # via -r requirements/gpu.in
 pandas==1.3.5
     # via paddlex (../../../setup.py)
@@ -295,6 +296,8 @@ python-dateutil==2.9.0.post0
     # via
     #   matplotlib
     #   pandas
+python-docx==1.2.0
+    # via paddlex (../../../setup.py)
 pytz==2025.1
     # via pandas
 pyyaml==6.0.2
@@ -326,8 +329,11 @@ ruamel-yaml==0.18.10
     # via paddlex (../../../setup.py)
 ruamel-yaml-clib==0.2.12
     # via ruamel-yaml
-safetensors==0.6.2
-    # via paddlex (../../../setup.py)
+safetensors @ https://paddle-whl.bj.bcebos.com/nightly/cu126/safetensors/safetensors-0.6.2.dev0-cp38-abi3-linux_x86_64.whl
+    # via
+    #   -r requirements/app.in
+    #   paddlepaddle-gpu
+    #   paddlex (../../../setup.py)
 scikit-image==0.24.0
     # via paddlex (../../../setup.py)
 scikit-learn==1.6.1
@@ -396,6 +402,7 @@ typing-extensions==4.12.2
     #   paddlex (../../../setup.py)
     #   pydantic
     #   pydantic-core
+    #   python-docx
     #   sqlalchemy
     #   starlette
     #   typing-inspect

paddlex/inference/models/doc_vlm/modeling/paddleocr_vl/_siglip.py

@@ -42,6 +42,7 @@
 import paddle.nn as nn
 import paddle.nn.functional as F
 
+from ......utils.env import get_gpu_compute_capability
 from ....common.vlm.activations import ACT2FN
 from ....common.vlm.transformers import PretrainedModel
 from ....common.vlm.transformers.model_outputs import (
@@ -137,6 +138,9 @@ def __init__(self, config):
         self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
         self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
 
+        cap = get_gpu_compute_capability()
+        self._supports_sdpa = cap >= (8, 0) if cap is not None else False
+
     def forward(
         self,
         hidden_states: paddle.Tensor,  # [B, L, D]
@@ -162,7 +166,7 @@ def forward(
             cos, sin = rope_emb
             q, k = apply_rotary_pos_emb_vision(q, k, cos, sin)
 
-        if q.dtype == paddle.float32:
+        if not self._supports_sdpa or q.dtype == paddle.float32:
             # → [B, H, L, Dh]
             q = q.transpose([0, 2, 1, 3])
             k = k.transpose([0, 2, 1, 3])

paddlex/inference/utils/hpi.py

@@ -24,6 +24,7 @@
 from pydantic import BaseModel, Field
 from typing_extensions import Annotated, TypeAlias
 
+from ...utils import logging
 from ...utils.deps import function_requires_deps, is_paddle2onnx_plugin_available
 from ...utils.env import get_paddle_cuda_version, get_paddle_version
 from ...utils.flags import USE_PIR_TRT
@@ -156,6 +157,14 @@ def suggest_inference_backend_and_config(
         return None, f"Inference backend {repr(hpi_config.backend)} is unavailable."
 
     paddle_version = get_paddle_version()
+
+    if paddle_version[:3] >= (3, 1, 0):
+        logging.debug(
+            "Paddle version %s is not supported yet. The prior knowledge of Paddle 3.1.1 will be used.",
+            paddle_version,
+        )
+        paddle_version = (3, 1, 1, None)
+
     if (3, 0) <= paddle_version[:2] <= (3, 1) and paddle_version[3] is None:
         if paddle_version[2] == 0:
             paddle_version = f"paddle{paddle_version[0]}{paddle_version[1]}"