apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-25 17:40:35 +08:00
Code Issues Actions 4 Packages Projects Releases Wiki Activity

[Features] support hugging face qwen3 moe (#3649)
Some checks failed
CE Compile Job / ce_job_pre_check (push) Has been cancelled
Details
CE Compile Job / print_ce_job_pre_check_outputs (push) Has been cancelled
Details
CE Compile Job / FD-Clone-Linux (push) Has been cancelled
Details
CE Compile Job / Show Code Archive Output (push) Has been cancelled
Details
CE Compile Job / BUILD_SM8090 (push) Has been cancelled
Details
CE Compile Job / BUILD_SM8689 (push) Has been cancelled
Details
CE Compile Job / CE_UPLOAD (push) Has been cancelled
Details
Deploy GitHub Pages / deploy (push) Has been cancelled
Details

Browse Source 
* split ut

* qwen3-30B-A3B

* fix

* add test

* add test_torch_model.py

* fix test_torch_model.py

* delete print

* fix moe

* delete init.py

* fix

* fix

---------

Co-authored-by: bukejiyu <395822456@qq.com>
Co-authored-by: bukejiyu <52310069+bukejiyu@users.noreply.github.com>
This commit is contained in:
lizexu123
2025-08-30 15:26:05 +08:00
committed by GitHub
parent f206474cc7
commit 455205f991
9 changed files with 437 additions and 258 deletions
Download Patch File Download Diff File Expand all files Collapse all files

23
fastdeploy/model_executor/layers/moe/moe.py
View File

@@ -151,7 +151,9 @@ class FusedMoE(nn.Layer):
self.gate_correction_bias = gate_correction_bias
else:
self.gate_correction_bias = None
self.quant_method.create_weights(self, weight_loader=self.weight_loader)
self.quant_method.create_weights(
self, weight_loader=self.weight_loader, model_format=fd_config.model_config.model_format
)
logger.info(
f"{moe_tag}MoE config is {num_experts=}[{expert_id_offset}, {expert_id_offset + self.num_local_experts}), \
@@ -197,6 +199,9 @@ class FusedMoE(nn.Layer):
)
def _load_gate_up_weight(self, param, expert_id, loaded_weight, shard_id, shard_dim=None):
model_format = getattr(param, "model_format", "")
if model_format == "torch":
loaded_weight = loaded_weight.transpose([1, 0])
dim = -1 if shard_dim else 0
if self.tp_size > 1:
if isinstance(loaded_weight, (np.ndarray, paddle.Tensor)):
@@ -208,8 +213,6 @@ class FusedMoE(nn.Layer):
shard_size = (self.tp_rank + 1) * block_size
loaded_weight = slice_fn(loaded_weight, shard_dim, shard_offset, shard_size)
loaded_weight = get_tensor(loaded_weight)
expert_param = param[expert_id - self.expert_id_offset]
param_shard_size = expert_param.shape[dim] // 2
if shard_id == "gate":
@@ -229,14 +232,18 @@ class FusedMoE(nn.Layer):
)
# To ensure compatibility across backends, apply an extra transpose for GCU and XPU
if expert_param.shape != loaded_weight.shape:
loaded_weight = loaded_weight.transpose([1, 0])
if current_platform.is_xpu() or current_platform.is_gcu():
if expert_param.shape != loaded_weight.shape:
loaded_weight = loaded_weight.transpose([1, 0])
assert expert_param.shape == loaded_weight.shape, (
f"Attempted to load weight ({loaded_weight.shape}) " f"into parameter ({expert_param.shape})"
)
expert_param.copy_(loaded_weight, False)
def _load_down_weight(self, param, expert_id, loaded_weight, shard_id, shard_dim=None):
model_format = getattr(param, "model_format", "")
if model_format == "torch":
loaded_weight = loaded_weight.transpose([1, 0])
if self.tp_size > 1 and shard_dim is not None:
dim = -1 if shard_dim else 0
if isinstance(loaded_weight, (np.ndarray, paddle.Tensor)):
@@ -247,14 +254,14 @@ class FusedMoE(nn.Layer):
shard_offset = self.tp_rank * block_size
shard_size = (self.tp_rank + 1) * block_size
loaded_weight = slice_fn(loaded_weight, shard_dim, shard_offset, shard_size)
loaded_weight = get_tensor(loaded_weight)
expert_param = param[expert_id - self.expert_id_offset]
if hasattr(param, "tensor_track"):
# for dyn quant
param.tensor_track.mark(start=0, batch_id=expert_id - self.expert_id_offset)
# To ensure compatibility across backends, apply an extra transpose for GCU and XPU
if expert_param.shape != loaded_weight.shape:
loaded_weight = loaded_weight.transpose([1, 0])
if current_platform.is_xpu or current_platform.is_gcu():
if expert_param.shape != loaded_weight.shape:
loaded_weight = loaded_weight.transpose([1, 0])
assert expert_param.shape == loaded_weight.shape, (
f"Attempted to load weight ({loaded_weight.shape}) " f"into parameter ({expert_param.shape})"
)