apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2026-04-23 00:17:25 +08:00
Code Issues Actions 19 Packages Projects Releases Wiki Activity

[Optimization] [OP] [Models] dsk del prefill mask (#7313)

Browse Source 
* dsk del prefill mask

* dsk support 1M+ seq_len rope

* update rope tests
This commit is contained in:
AIbin
2026-04-11 19:32:27 +08:00
committed by GitHub
parent 076ab07528
commit ba01d7a823
4 changed files with 83 additions and 49 deletions
Download Patch File Download Diff File Expand all files Collapse all files
fastdeploy/model_executor/models/deepseek_v3.py
+12 -2
View File
@@ -72,6 +72,7 @@ if current_platform.is_cuda():
from fastdeploy.model_executor.ops.gpu import (
cp_gather_indexer_k_quant_cache,
indexer_k_quant_and_cache,
merge_prefill_decode_output,
radix_topk_ragged_transform,
)
@@ -398,7 +399,6 @@ class DeepseekV3MLAAttention(nn.Layer):
fmha_out_prefill.reshape_([-1, self.num_attention_heads_tp, self.qk_head_dim])
fmha_out_prefill = fmha_out_prefill[:, :, : self.v_head_dim]
fmha_out_prefill.reshape_([-1, self.num_attention_heads_tp * self.v_head_dim])
fmha_out_prefill = fmha_out_prefill * forward_meta.mask_encoder_batch.cast(fmha_out_prefill.dtype)
fmha_out = fmha_out_prefill
if need_do_decode: # max_dec_len_this_time
@@ -433,7 +433,17 @@ class DeepseekV3MLAAttention(nn.Layer):
)
if need_do_prefill:
fmha_out += fmha_out_decode
merge_prefill_decode_output(
fmha_out,
fmha_out_decode,
forward_meta.seq_lens_encoder,
forward_meta.seq_lens_decoder,
forward_meta.seq_lens_this_time,
forward_meta.cu_seqlens_q,
self.num_attention_heads_tp,
self.v_head_dim,
1,
)
else:
fmha_out = fmha_out_decode