apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-27 18:41:02 +08:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

[Feature] Deepseekv3 supports cudagraph (#3041)

Browse Source 
Co-authored-by: K11OntheBoat <“ruianmaidanglao@163.com”>
This commit is contained in:
K11OntheBoat
2025-07-28 17:12:54 +08:00
committed by GitHub
parent ec52d39e68
commit 83048bbe55
1 changed files with 12 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
fastdeploy/model_executor/models/deepseek_v3.py
View File

@@ -27,6 +27,9 @@ from paddleformers.utils.log import logger
from fastdeploy.config import FDConfig
from fastdeploy.distributed.communication import tensor_model_parallel_all_reduce
from fastdeploy.model_executor.forward_meta import ForwardMeta
from fastdeploy.model_executor.graph_optimization.decorator import (
support_graph_optimization,
)
from fastdeploy.model_executor.layers.activation import SiluAndMul
from fastdeploy.model_executor.layers.attention.attention import Attention
from fastdeploy.model_executor.layers.embeddings import VocabParallelEmbedding
@@ -501,6 +504,7 @@ class DeepSeekV3DecoderLayer(nn.Layer):
return hidden_states, residual
@support_graph_optimization
class DeepSeekV3Model(nn.Layer):
"""
DeepSeekV3Model
@@ -596,6 +600,10 @@ class DeepseekV3ForCausalLM(ModelForCasualLM):
num_embeddings=fd_config.model_config.vocab_size,
prefix="lm_head",
)
self.position_ids_buffer = paddle.empty([fd_config.parallel_config.max_num_batched_tokens], dtype=paddle.int32)
self.mask_encoder_batch_buffer = paddle.empty(
[fd_config.parallel_config.max_num_batched_tokens, 1], dtype=paddle.int32
)
@classmethod
def name(cls):
@@ -622,9 +630,10 @@ class DeepseekV3ForCausalLM(ModelForCasualLM):
seq_lens_encoder = forward_meta.seq_lens_encoder
seq_lens_decoder = forward_meta.seq_lens_decoder
seq_lens_this_time = forward_meta.seq_lens_this_time
position_ids_shape = paddle.sum(seq_lens_this_time)
position_ids = paddle.empty(shape=position_ids_shape, dtype=seq_lens_encoder.dtype)
mask_encoder_batch = paddle.empty(shape=position_ids_shape, dtype=seq_lens_encoder.dtype).unsqueeze(1)
current_total_tokens = paddle.sum(seq_lens_this_time)
position_ids = self.position_ids_buffer[:current_total_tokens]
mask_encoder_batch = self.mask_encoder_batch_buffer[:current_total_tokens]
get_position_ids_and_mask_encoder_batch(
seq_lens_encoder,
@@ -633,7 +642,6 @@ class DeepseekV3ForCausalLM(ModelForCasualLM):
position_ids,
mask_encoder_batch,
)
return position_ids, mask_encoder_batch
def forward(