Skip to content

【AutoParallel】support cast op from FP32 to low precision #60385

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

Merged
merged 2 commits into from
Jan 3, 2024
Merged

【AutoParallel】support cast op from FP32 to low precision #60385

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
418fcef
support cast op from FP32 to low precision
heavyrain-lzy Dec 27, 2023
37e7ea5
Merge remote-tracking branch 'upstream/develop' into support_cast_h2l
heavyrain-lzy Jan 2, 2024
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
5 changes: 5 additions & 0 deletions python/paddle/distributed/passes/auto_parallel_amp.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -232,6 +232,11 @@ def build_state(self):
return is_train

def _mark_black_white_ops(self, op, ops, block):
# deal auto_cast info
if not op.amp_options.enable:
self._op_fp16_dict[op.desc.original_id()] = False
return

# ernie inference trick
if op.type == "assign" and "array_" in op.input_arg_names[0]:
self._op_fp16_dict[op.desc.original_id()] = False
Expand Down
16 changes: 16 additions & 0 deletions python/paddle/distributed/passes/auto_parallel_fp16.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -209,6 +209,9 @@ def _build_state(self):
for block in self.program.blocks:
self.resolute_tensor_dtype(block)

for block in self.program.blocks:
self.resolute_cast_op(block)

# insert cast ops
for block in self.program.blocks:
self.cast_block(block)
Expand Down Expand Up @@ -296,6 +299,19 @@ def set_var_to_fp16(self, var_name, block):
if var.dtype == core.VarDesc.VarType.FP32:
var.desc.set_dtype(__target_dtype__)

def resolute_cast_op(self, block):
"""
Deal the "cast_op" from "FP32" to "FP16" or "BF16" in the model.
"""
for op in block.ops:
if op.type == "cast":
in_name = op.input('X')[0]
out_name = op.output('Out')[0]
in_var = block._find_var_recursive(in_name)
out_var = block._find_var_recursive(out_name)
op._set_attr("in_dtype", in_var.dtype)
op._set_attr("out_dtype", out_var.dtype)

def resolute_tensor_dtype(self, block):
for op in block.ops:
# 'amp_options' flag has highest priority
Expand Down