apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 16:48:03 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[BugFix] Rename attention params of deepseekv3 (#2939)

Browse Source 
Co-authored-by: K11OntheBoat <“ruianmaidanglao@163.com”>
This commit is contained in:
K11OntheBoat
2025-07-22 14:01:30 +08:00
committed by GitHub
parent 56102e91e1
commit 8020927f50
7 changed files with 43 additions and 44 deletions
Download Patch File Download Diff File Expand all files Collapse all files

38
fastdeploy/model_executor/layers/attention/mla_attention_backend.py
View File

@@ -213,6 +213,10 @@ class MLAAttentionBackend(AttentionBackend):
self.attention_metadata: AttentionMetadata = metadata
forward_meta.decoder_batch_ids.copy_(metadata.decoder_batch_ids, False)
forward_meta.decoder_tile_ids_per_batch.copy_(
metadata.decoder_tile_ids_per_batch, False)
def get_attntion_meta(self) -> AttentionMetadata:
"""get_attntion_meta"""
return self.attention_metadata
@@ -259,8 +263,8 @@ class MLAAttentionBackend(AttentionBackend):
latent_cache,
forward_meta.seq_lens_encoder,
forward_meta.seq_lens_decoder,
forward_meta.padding_offset,
forward_meta.cum_offsets,
forward_meta.batch_id_per_token,
forward_meta.cu_seqlens_q,
metadata.block_tables,
"none",
getattr(forward_meta, "max_input_length", -1),
@@ -298,7 +302,7 @@ class MLAAttentionBackend(AttentionBackend):
"""
metadata = self.attention_metadata
if self.use_pd_disaggregation:
if self.pd_disaggregation_mode == "per_query":
metadata.kv_signal_data_list[layer.layer_id] = init_signal_layerwise(
metadata.kv_signal_metadata,
layer.layer_id + self.start_layer_index,
@@ -317,8 +321,8 @@ class MLAAttentionBackend(AttentionBackend):
latent_cache,
forward_meta.seq_lens_decoder,
forward_meta.seq_lens_encoder,
forward_meta.padding_offset,
forward_meta.cum_offsets,
forward_meta.batch_id_per_token,
forward_meta.cu_seqlens_q,
metadata.block_tables,
"none",
self.max_seq_len,
@@ -334,8 +338,7 @@ class MLAAttentionBackend(AttentionBackend):
forward_meta.seq_lens_decoder,
forward_meta.seq_lens_this_time,
forward_meta.cu_seqlens_q,
forward_meta.padding_offset,
forward_meta.cum_offsets,
forward_meta.batch_id_per_token,
metadata.block_tables,
metadata.encoder_batch_ids,
metadata.encoder_tile_ids_per_batch,
@@ -343,8 +346,8 @@ class MLAAttentionBackend(AttentionBackend):
metadata.kv_batch_ids,
metadata.kv_tile_ids_per_batch,
metadata.kv_num_blocks,
metadata.decoder_batch_ids,
metadata.decoder_tile_ids_per_batch,
forward_meta.decoder_batch_ids,
forward_meta.decoder_tile_ids_per_batch,
metadata.decoder_num_blocks,
metadata.decoder_num_blocks, # PaddleNLP 传入的是 decoder_num_blocks_cpu
metadata.max_enc_len_this_time,
@@ -394,7 +397,7 @@ class MLAAttentionBackend(AttentionBackend):
speculate_decoder = self.speculative_method is not None
speculate_max_tokens = self.speculate_max_draft_token_num
if self.use_pd_disaggregation:
if self.pd_disaggregation_mode == "per_query":
metadata.kv_signal_data_list[layer.layer_id] = init_signal_layerwise(
metadata.kv_signal_metadata,
layer.layer_id + self.start_layer_index,
@@ -409,8 +412,8 @@ class MLAAttentionBackend(AttentionBackend):
latent_cache,
forward_meta.seq_lens_encoder,
forward_meta.seq_lens_decoder,
forward_meta.padding_offset,
forward_meta.cum_offsets,
forward_meta.batch_id_per_token,
forward_meta.cu_seqlens_q,
metadata.block_tables,
"none",
self.max_seq_len,
@@ -440,8 +443,8 @@ class MLAAttentionBackend(AttentionBackend):
latent_cache,
forward_meta.seq_lens_decoder,
forward_meta.seq_lens_encoder,
forward_meta.padding_offset,
forward_meta.cum_offsets,
forward_meta.batch_id_per_token,
forward_meta.cu_seqlens_q,
metadata.block_tables,
"none",
self.max_seq_len,
@@ -457,8 +460,7 @@ class MLAAttentionBackend(AttentionBackend):
forward_meta.seq_lens_decoder,
forward_meta.seq_lens_this_time,
forward_meta.cu_seqlens_q,
forward_meta.padding_offset,
forward_meta.cum_offsets,
forward_meta.batch_id_per_token,
metadata.block_tables,
metadata.encoder_batch_ids,
metadata.encoder_tile_ids_per_batch,
@@ -466,8 +468,8 @@ class MLAAttentionBackend(AttentionBackend):
metadata.kv_batch_ids,
metadata.kv_tile_ids_per_batch,
metadata.kv_num_blocks,
metadata.decoder_batch_ids,
metadata.decoder_tile_ids_per_batch,
forward_meta.decoder_batch_ids,
forward_meta.decoder_tile_ids_per_batch,
metadata.decoder_num_blocks,
metadata.decoder_num_blocks, # PaddleNLP 传入的是 decoder_num_blocks_cpu
metadata.max_enc_len_this_time,