apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-07 09:31:35 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

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

Browse Source 
* add moe noaux_tc tatics in trition backend

* fix

* add dp config
This commit is contained in:
gaoziyuan
2025-09-03 13:28:44 +08:00
committed by GitHub
parent 42402c80e9
commit 05b6591c23
2 changed files with 57 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files

76
fastdeploy/model_executor/layers/moe/fused_moe_triton_backend.py
View File

@@ -24,6 +24,7 @@ from fastdeploy.model_executor.utils import TensorTracker, set_weight_attrs
from fastdeploy.utils import ceil_div from fastdeploy.utils import ceil_div
from ..quantization.quant_base import QuantMethodBase from ..quantization.quant_base import QuantMethodBase
from .ep import get_moe_scores
try: try:
from fastdeploy.model_executor.ops.gpu import tritonmoe_preprocess_func from fastdeploy.model_executor.ops.gpu import tritonmoe_preprocess_func
@@ -167,13 +168,24 @@ class TritonWeightOnlyMoEMethod(QuantMethodBase):
moe_intermediate_size = layer.moe_intermediate_size moe_intermediate_size = layer.moe_intermediate_size
hidden_size = layer.hidden_size hidden_size = layer.hidden_size
topk_ids, topk_weights = fastdeploy.model_executor.ops.gpu.moe_topk_select( if layer.topk_method == "noaux_tc":
gate_out, _, topk_weights, topk_ids = get_moe_scores(
layer.gate_correction_bias, gate_out,
top_k, layer.n_group,
True, # apply_norm_weight, layer.topk_group,
False, 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( up_gate_proj_out = paddle.empty(
[token_num * top_k, moe_intermediate_size * 2], [token_num * top_k, moe_intermediate_size * 2],
dtype=x.dtype, dtype=x.dtype,
@@ -419,13 +431,25 @@ class TensorWiseFP8MoEMethod(QuantMethodBase):
moe_intermediate_size = layer.moe_intermediate_size moe_intermediate_size = layer.moe_intermediate_size
hidden_size = layer.hidden_size hidden_size = layer.hidden_size
topk_ids, topk_weights = fastdeploy.model_executor.ops.gpu.moe_topk_select( if layer.topk_method == "noaux_tc":
gate_out,
layer.gate_correction_bias, _, topk_weights, topk_ids = get_moe_scores(
top_k, gate_out,
True, # apply_norm_weight, layer.n_group,
False, 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( up_gate_proj_out = paddle.empty(
[token_num * top_k, moe_intermediate_size * 2], [token_num * top_k, moe_intermediate_size * 2],
@@ -829,13 +853,23 @@ class BlockWiseFP8MoEMethod(QuantMethodBase):
E, N1, _ = getattr(layer, self.added_weight_attrs[0]).shape E, N1, _ = getattr(layer, self.added_weight_attrs[0]).shape
N2 = getattr(layer, self.added_weight_attrs[1]).shape[1] N2 = getattr(layer, self.added_weight_attrs[1]).shape[1]
topk_ids, topk_weights = fastdeploy.model_executor.ops.gpu.moe_topk_select( if layer.topk_method == "noaux_tc":
gate_out, _, topk_weights, topk_ids = get_moe_scores(
layer.gate_correction_bias, gate_out,
layer.top_k, layer.n_group,
True, # apply_norm_weight layer.topk_group,
False, 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 = { config = {
"BLOCK_SIZE_M": 64, "BLOCK_SIZE_M": 64,

2
fastdeploy/rl/rollout_config.py
View File

@@ -60,6 +60,7 @@ class RolloutModelConfig:
early_stop_config: str = None, early_stop_config: str = None,
local_rank: int = 0, local_rank: int = 0,
moba_attention_config: str = None, moba_attention_config: str = None,
data_parallel_size: int = 1,
): ):
# Required parameters # Required parameters
self.model = model_name_or_path self.model = model_name_or_path
@@ -95,6 +96,7 @@ class RolloutModelConfig:
self.splitwise_role = splitwise_role self.splitwise_role = splitwise_role
self.expert_parallel_size = expert_parallel_size self.expert_parallel_size = expert_parallel_size
self.enable_expert_parallel = enable_expert_parallel self.enable_expert_parallel = enable_expert_parallel
self.data_parallel_size = data_parallel_size
self.ori_vocab_size = ori_vocab_size self.ori_vocab_size = ori_vocab_size
self.quantization = quantization self.quantization = quantization
self.guided_decoding_backend = guided_decoding_backend self.guided_decoding_backend = guided_decoding_backend