Skip to content

[ROCm] Remove BF16 workarounds for PaddleOCR-VL on HIP #5096

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
austin1997 wants to merge 1 commit into
base: develop
Choose a base branch
from
+0 −15
Open

[ROCm] Remove BF16 workarounds for PaddleOCR-VL on HIP #5096

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 0 additions & 3 deletions paddlex/inference/models/doc_vlm/modeling/paddleocr_vl/_paddleocr_vl.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -65,9 +65,6 @@ class PaddleOCRVLForConditionalGeneration(Ernie4_5PretrainedModel):
_tied_weights_keys = ["lm_head.weight"]
config_class = PaddleOCRVLConfig
_no_split_modules = ["Ernie4_5DecoderLayer", "SiglipEncoderLayer"]
# Keep visual encoder in fp32 for ROCm stability (MIOpen bf16 conv has bugs)
# This also improves precision for vision processing
_keep_in_fp32_modules = ["visual", "mlp_AR"]
base_model_prefix = ""

def __init__(self, config):
Expand Down
12 changes: 0 additions & 12 deletions paddlex/inference/models/runners/paddle_static/runner.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -403,9 +403,6 @@ def _create(self):
config.set_optimization_level(3)
if self._model_name == "PP-DocLayoutV3":
config.delete_pass("matmul_add_act_fuse_pass")
if paddle.is_compiled_with_rocm():
config.delete_pass("conv2d_add_act_fuse_pass")
config.delete_pass("conv2d_add_fuse_pass")
elif self._config["device_type"] == "npu":
config.enable_custom_device("npu", self._config.get("device_id", 0))
if hasattr(config, "enable_new_ir"):
Expand Down Expand Up @@ -459,9 +456,6 @@ def _create(self):
config.disable_mkldnn()
if hasattr(config, "enable_new_executor"):
config.enable_new_executor()
if paddle.is_compiled_with_rocm():
config.delete_pass("conv2d_add_act_fuse_pass")
config.delete_pass("conv2d_add_fuse_pass")
elif self._config["device_type"] == "iluvatar_gpu":
config.enable_custom_device(
"iluvatar_gpu", int(self._config.get("device_id", 0))
Expand Down Expand Up @@ -493,18 +487,12 @@ def _create(self):
if hasattr(config, "enable_new_executor"):
config.enable_new_executor()
config.set_optimization_level(3)
if paddle.is_compiled_with_rocm():
config.delete_pass("conv2d_add_act_fuse_pass")
config.delete_pass("conv2d_add_fuse_pass")
config.enable_memory_optim()
for del_p in self._config.get("delete_pass", []):
config.delete_pass(del_p)

if not DEBUG:
config.disable_glog_info()
if paddle.is_compiled_with_rocm():
config.delete_pass("conv2d_add_act_fuse_pass")
config.delete_pass("conv2d_add_fuse_pass")

predictor = paddle_inference.create_predictor(config)

Expand Down