Skip to content

[XPU] Support V1 Loader in Bf16 #4746

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 8 commits into from
Nov 1, 2025
Merged

[XPU] Support V1 Loader in Bf16 #4746

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
077b2ec
add v1 support for bf16
iosmers Oct 31, 2025
19af292
update
iosmers Oct 31, 2025
817ac3f
update
iosmers Oct 31, 2025
30e0bf6
update
iosmers Oct 31, 2025
7143b33
update
iosmers Oct 31, 2025
336c419
update
iosmers Oct 31, 2025
ca133e5
update code
iosmers Oct 31, 2025
6cca8d2
Merge branch 'release/2.3' into cherry_pick_v1_loader
iosmers Oct 31, 2025
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
161 changes: 120 additions & 41 deletions fastdeploy/model_executor/layers/backends/xpu/moe/fused_moe.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,6 +29,8 @@
weight_quantize_xpu,
xpu_moe_layer,
)
from fastdeploy.model_executor.utils import default_weight_loader, set_weight_attrs
from fastdeploy.platforms import current_platform


class XPUMoEMethod(MoEMethodBase):
Expand Down Expand Up @@ -61,78 +63,155 @@ def create_weights(self, layer: nn.Layer, **extra_weight_attrs):
"""
create weight process.
"""
self.up_gate_proj_weight_shape = [
layer.num_local_experts,
layer.moe_intermediate_size * 2,
layer.hidden_size,
]
self.down_proj_weight_shape = [
layer.num_local_experts,
layer.hidden_size,
layer.moe_intermediate_size,
]
if self.moe_quant_type in ["weight_only_int4", "w4a8"]:
self.up_gate_proj_weight_shape[-1] //= 2
self.down_proj_weight_shape[-1] //= 2

setattr(
layer,
self.added_weight_attrs[0],
layer.create_parameter(
if layer.fd_config.load_config.load_choices == "default_v1" and self.moe_quant_type in ["w16a16"]:
if current_platform.is_cuda():
self.up_gate_proj_weight_shape = [
layer.num_local_experts,
layer.hidden_size,
layer.moe_intermediate_size * 2,
]
self.down_proj_weight_shape = [layer.num_local_experts, layer.moe_intermediate_size, layer.hidden_size]
extra_weight_attrs = {**extra_weight_attrs, "SHARD_ID_TO_SHARDED_DIM": {"gate": 1, "down": 0, "up": 1}}
else:
self.up_gate_proj_weight_shape = [
layer.num_local_experts,
layer.moe_intermediate_size * 2,
layer.hidden_size,
]
self.down_proj_weight_shape = [layer.num_local_experts, layer.hidden_size, layer.moe_intermediate_size]
extra_weight_attrs = {**extra_weight_attrs, "SHARD_ID_TO_SHARDED_DIM": {"gate": 0, "down": 1, "up": 0}}

layer.up_gate_proj_weight = layer.create_parameter(
shape=self.up_gate_proj_weight_shape,
dtype=self.weight_dtype,
dtype=layer.weight_dtype,
default_initializer=paddle.nn.initializer.Constant(0),
),
)
setattr(
layer,
self.added_weight_attrs[1],
layer.create_parameter(
)

layer.down_proj_weight = layer.create_parameter(
shape=self.down_proj_weight_shape,
dtype=self.weight_dtype,
dtype=layer.weight_dtype,
default_initializer=paddle.nn.initializer.Constant(0),
),
)
)

if self.moe_quant_type in ["weight_only_int8", "w8a8", "weight_only_int4", "w4a8"]:
self.up_gate_proj_scale_shape = [
set_weight_attrs(
layer.up_gate_proj_weight,
{
"weight_loader": extra_weight_attrs.get("weight_loader", default_weight_loader(layer.fd_config)),
"weight_need_transpose": extra_weight_attrs.get("model_format") == "torch",
},
)
set_weight_attrs(
layer.down_proj_weight,
{
"weight_loader": extra_weight_attrs.get("weight_loader", default_weight_loader(layer.fd_config)),
"weight_need_transpose": extra_weight_attrs.get("model_format") == "torch",
},
)

if layer.with_bias:
layer.up_gate_proj_bias = layer.create_parameter(
shape=[layer.num_experts, layer.moe_intermediate_size * 2],
dtype=layer.weight_dtype,
default_initializer=paddle.nn.initializer.Constant(0),
)

layer.down_proj_bias = layer.create_parameter(
shape=[layer.num_experts, layer.hidden_size],
dtype=layer.weight_dtype,
default_initializer=paddle.nn.initializer.Constant(0),
)
set_weight_attrs(
layer.up_gate_proj_bias,
{
"weight_loader": extra_weight_attrs.get(
"weight_loader", default_weight_loader(layer.fd_config)
),
"model_format": extra_weight_attrs.get("model_format", ""),
},
)
set_weight_attrs(
layer.down_proj_bias,
{
"weight_loader": extra_weight_attrs.get(
"weight_loader", default_weight_loader(layer.fd_config)
),
"model_format": extra_weight_attrs.get("model_format", ""),
},
)

else:
self.up_gate_proj_weight_shape = [
layer.num_local_experts,
layer.moe_intermediate_size * 2,
layer.hidden_size,
]
self.down_proj_scale_shape = [
self.down_proj_weight_shape = [
layer.num_local_experts,
layer.hidden_size,
layer.moe_intermediate_size,
]
if self.moe_quant_type in ["weight_only_int4", "w4a8"]:
self.up_gate_proj_weight_shape[-1] //= 2
self.down_proj_weight_shape[-1] //= 2

setattr(
layer,
self.added_scale_attrs[0],
self.added_weight_attrs[0],
layer.create_parameter(
shape=self.up_gate_proj_scale_shape,
dtype=self.scale_dtype,
shape=self.up_gate_proj_weight_shape,
dtype=self.weight_dtype,
default_initializer=paddle.nn.initializer.Constant(0),
),
)
setattr(
layer,
self.added_scale_attrs[1],
self.added_weight_attrs[1],
layer.create_parameter(
shape=self.down_proj_scale_shape,
dtype=self.scale_dtype,
shape=self.down_proj_weight_shape,
dtype=self.weight_dtype,
default_initializer=paddle.nn.initializer.Constant(0),
),
)

if self.moe_quant_type in ["w8a8", "w4a8"]:
for in_scale_name in self.added_in_scale_attrs:
if self.moe_quant_type in ["weight_only_int8", "w8a8", "weight_only_int4", "w4a8"]:
self.up_gate_proj_scale_shape = [
layer.num_local_experts,
layer.moe_intermediate_size * 2,
]
self.down_proj_scale_shape = [
layer.num_local_experts,
layer.hidden_size,
]
setattr(
layer,
in_scale_name,
self.added_scale_attrs[0],
layer.create_parameter(
shape=[layer.num_local_experts],
shape=self.up_gate_proj_scale_shape,
dtype=self.scale_dtype,
default_initializer=paddle.nn.initializer.Constant(0),
),
)
setattr(
layer,
self.added_scale_attrs[1],
layer.create_parameter(
shape=self.down_proj_scale_shape,
dtype=self.scale_dtype,
default_initializer=paddle.nn.initializer.Constant(0),
),
)

if self.moe_quant_type in ["w8a8", "w4a8"]:
for in_scale_name in self.added_in_scale_attrs:
setattr(
layer,
in_scale_name,
layer.create_parameter(
shape=[layer.num_local_experts],
dtype=self.scale_dtype,
default_initializer=paddle.nn.initializer.Constant(0),
),
)

def process_loaded_weights(self, layer: nn.Layer, state_dict):
up_gate_proj_weights, down_proj_weights, _, _ = layer.extract_moe_ffn_weights(state_dict)
Expand Down
4 changes: 2 additions & 2 deletions fastdeploy/model_executor/utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -261,8 +261,8 @@ def v1_loader_support(fd_config):
def _err_msg(msg: str) -> str:
logger.info(msg + "; fallback to the v0 loader for model loading.")

if not current_platform.is_cuda():
_err_msg("v1loader currently does not support backends other than CUDA")
if not (current_platform.is_cuda() or current_platform.is_xpu()):
_err_msg("v1loader currently only support backends gpu and xpu")
return False

if is_pre_sliced_weight(fd_config.model_config.model):
Expand Down
9 changes: 6 additions & 3 deletions scripts/run_ci_xpu.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -53,7 +53,8 @@ python -m fastdeploy.entrypoints.openai.api_server \
--num-gpu-blocks-override 16384 \
--max-model-len 32768 \
--max-num-seqs 128 \
--quantization wint4 > server.log 2>&1 &
--quantization wint4 \
--load-choices default > server.log 2>&1 &

sleep 60
# 探活
Expand Down Expand Up @@ -120,7 +121,8 @@ python -m fastdeploy.entrypoints.openai.api_server \
--num-gpu-blocks-override 16384 \
--max-model-len 32768 \
--max-num-seqs 64 \
--quantization "W4A8" > server.log 2>&1 &
--quantization "W4A8" \
--load-choices default > server.log 2>&1 &

sleep 60
# 探活
Expand Down Expand Up @@ -190,7 +192,8 @@ python -m fastdeploy.entrypoints.openai.api_server \
--enable-mm \
--mm-processor-kwargs '{"video_max_frames": 30}' \
--limit-mm-per-prompt '{"image": 10, "video": 3}' \
--reasoning-parser ernie-45-vl > server.log 2>&1 &
--reasoning-parser ernie-45-vl \
--load-choices default > server.log 2>&1 &

sleep 60
# 探活
Expand Down
1 change: 1 addition & 0 deletions tests/ci_use/XPU_45T/run_ep.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -44,6 +44,7 @@ def test_fd_ep():
quantization="wint4",
engine_worker_queue_port=engine_worker_queue_port,
max_num_seqs=8,
load_choices="default",
)

try:
Expand Down
Loading