Skip to content

【BugFix】add moe noaux_tc tatics in trition backend #3821

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
Jiang-Jia-Jun merged 3 commits into from
Sep 3, 2025
Merged

【BugFix】add moe noaux_tc tatics in trition backend #3821

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
686bea0
add moe noaux_tc tatics in trition backend
gzy19990617 Sep 2, 2025
4d1df7a
fix
gzy19990617 Sep 2, 2025
71b3288
add dp config
gzy19990617 Sep 2, 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
76 changes: 55 additions & 21 deletions fastdeploy/model_executor/layers/moe/fused_moe_triton_backend.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,6 +24,7 @@
from fastdeploy.utils import ceil_div

from ..quantization.quant_base import QuantMethodBase
from .ep import get_moe_scores

try:
from fastdeploy.model_executor.ops.gpu import tritonmoe_preprocess_func
Expand Down Expand Up @@ -167,13 +168,24 @@ def apply(
moe_intermediate_size = layer.moe_intermediate_size
hidden_size = layer.hidden_size

topk_ids, topk_weights = fastdeploy.model_executor.ops.gpu.moe_topk_select(
gate_out,
layer.gate_correction_bias,
top_k,
True, # apply_norm_weight,
False,
)
if layer.topk_method == "noaux_tc":
_, topk_weights, topk_ids = get_moe_scores(
gate_out,
layer.n_group,
layer.topk_group,
layer.top_k,
layer.routed_scaling_factor,
layer.gate_correction_bias,
)
else:
topk_ids, topk_weights = fastdeploy.model_executor.ops.gpu.moe_topk_select(
gate_out,
layer.gate_correction_bias,
layer.top_k,
True, # apply_norm_weight
False,
)

up_gate_proj_out = paddle.empty(
[token_num * top_k, moe_intermediate_size * 2],
dtype=x.dtype,
Expand Down Expand Up @@ -419,13 +431,25 @@ def apply(
moe_intermediate_size = layer.moe_intermediate_size
hidden_size = layer.hidden_size

topk_ids, topk_weights = fastdeploy.model_executor.ops.gpu.moe_topk_select(
gate_out,
layer.gate_correction_bias,
top_k,
True, # apply_norm_weight,
False,
)
if layer.topk_method == "noaux_tc":

_, topk_weights, topk_ids = get_moe_scores(
gate_out,
layer.n_group,
layer.topk_group,
layer.top_k,
layer.routed_scaling_factor,
layer.gate_correction_bias,
)
else:

topk_ids, topk_weights = fastdeploy.model_executor.ops.gpu.moe_topk_select(
gate_out,
layer.gate_correction_bias,
top_k,
True, # apply_norm_weight,
False,
)

up_gate_proj_out = paddle.empty(
[token_num * top_k, moe_intermediate_size * 2],
Expand Down Expand Up @@ -829,13 +853,23 @@ def apply(
E, N1, _ = getattr(layer, self.added_weight_attrs[0]).shape
N2 = getattr(layer, self.added_weight_attrs[1]).shape[1]

topk_ids, topk_weights = fastdeploy.model_executor.ops.gpu.moe_topk_select(
gate_out,
layer.gate_correction_bias,
layer.top_k,
True, # apply_norm_weight
False,
)
if layer.topk_method == "noaux_tc":
_, topk_weights, topk_ids = get_moe_scores(
gate_out,
layer.n_group,
layer.topk_group,
layer.top_k,
layer.routed_scaling_factor,
layer.gate_correction_bias,
)
else:
topk_ids, topk_weights = fastdeploy.model_executor.ops.gpu.moe_topk_select(
gate_out,
layer.gate_correction_bias,
layer.top_k,
True, # apply_norm_weight
False,
)

config = {
"BLOCK_SIZE_M": 64,
Expand Down
2 changes: 2 additions & 0 deletions fastdeploy/rl/rollout_config.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -60,6 +60,7 @@ def __init__(
early_stop_config: str = None,
local_rank: int = 0,
moba_attention_config: str = None,
data_parallel_size: int = 1,
):
# Required parameters
self.model = model_name_or_path
Expand Down Expand Up @@ -95,6 +96,7 @@ def __init__(
self.splitwise_role = splitwise_role
self.expert_parallel_size = expert_parallel_size
self.enable_expert_parallel = enable_expert_parallel
self.data_parallel_size = data_parallel_size
self.ori_vocab_size = ori_vocab_size
self.quantization = quantization
self.guided_decoding_backend = guided_decoding_backend
Expand Down
Loading