apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2026-04-23 08:21:53 +08:00
Code Issues Actions 19 Packages Projects Releases Wiki Activity

[DeepSeekV3.2][Graph Optimization]Remove synchronous operation to avoid capture fail and unnecessary contiguous in DSA Backend (#7253)

Browse Source 
* Delete contiguous ops.

* fix scale

* Delete unnecessary comments

* fix style
This commit is contained in:
ShaneGZhu
2026-04-09 11:00:13 +08:00
committed by GitHub
parent 48d2bbeb74
commit 7005404ce3
1 changed files with 5 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
fastdeploy/model_executor/layers/attention/dsa_attention_backend.py
+5 -4
View File
@@ -344,8 +344,7 @@ class DSAAttentionBackend(AttentionBackend):
from fastdeploy.model_executor.ops.gpu import dsk_attn_write_cache
k_range = paddle.tensor(200.0)
scale = paddle.abs(compressed_kv).max() / k_range
scale = paddle.abs(compressed_kv).max() / 200.0
slot_mapping = compute_slot_mapping(
forward_meta.block_tables,
@@ -378,10 +377,12 @@ class DSAAttentionBackend(AttentionBackend):
if forward_meta.max_len_tensor_cpu[2]: # max_enc_len_this_time
tile_scheduler_metadata, _ = flash_mla.get_mla_metadata()
new_cache_shape = latent_cache.shape
assert new_cache_shape[1] == 1
new_cache_shape[1], new_cache_shape[2] = new_cache_shape[2], new_cache_shape[1]
fmha_out_decode, _ = flash_mla.flash_mla_with_kvcache(
q.unsqueeze(1).contiguous(),
latent_cache.transpose([0, 2, 1, 3]).contiguous(),
latent_cache.view(new_cache_shape),
None, # forward_meta.block_tables,
None, # cache_seqlens
512, # self.qk_nope_head_dim,