apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-06 17:17:14 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

cp dynamic Cfp8 (#4120)

Browse Source 
* supports dynamic Cfp8

* add unittest

* fix dynamic Cfp8 computing error

* fix Cfp8 for RL load

---------

Co-authored-by: carryyu <569782149@qq.com>
This commit is contained in:
Yuan Xiaolan
2025-09-17 11:55:47 +08:00
committed by GitHub
parent b6caf6e622
commit 25aa2d94aa
21 changed files with 1428 additions and 227 deletions
Download Patch File Download Diff File Expand all files Collapse all files

27
fastdeploy/model_executor/layers/attention/append_attn_backend.py
View File

@@ -233,6 +233,17 @@ class AppendAttentionBackend(AttentionBackend):
metadata.kv_signal_metadata,
layer.layer_id + self.start_layer_index,
)
cache_quant_type_str = getattr(layer, "cache_quant_type_str", "none")
if cache_quant_type_str == "block_wise_fp8":
cache_k = forward_meta.caches[4 * layer.layer_id]
cache_v = forward_meta.caches[4 * layer.layer_id + 1]
cache_k_scales = forward_meta.caches[4 * layer.layer_id + 2]
cache_v_scales = forward_meta.caches[4 * layer.layer_id + 3]
else:
cache_k = forward_meta.caches[2 * layer.layer_id]
cache_v = forward_meta.caches[2 * layer.layer_id + 1]
cache_k_scales = getattr(layer, "cache_k_scale", None)
cache_v_scales = getattr(layer, "cache_v_scale", None)
if self.use_output:
quant_max_bound = getattr(layer, "quant_max_bound", 0.0)
@@ -271,8 +282,8 @@ class AppendAttentionBackend(AttentionBackend):
append_attention_with_output(
qkv,
forward_meta.caches[2 * layer.layer_id],
forward_meta.caches[2 * layer.layer_id + 1],
cache_k,
cache_v,
forward_meta.seq_lens_encoder,
forward_meta.seq_lens_decoder,
forward_meta.seq_lens_this_time,
@@ -295,8 +306,8 @@ class AppendAttentionBackend(AttentionBackend):
metadata.attn_mask,
layer.qkv_bias,
layer.qkv_scale,
getattr(layer, "cache_k_scale", None),
getattr(layer, "cache_v_scale", None),
cache_k_scales,
cache_v_scales,
getattr(layer, "cache_k_out_scale", None),
getattr(layer, "cache_v_out_scale", None),
getattr(layer, "cache_k_zp", None),
@@ -327,8 +338,8 @@ class AppendAttentionBackend(AttentionBackend):
else:
res = append_attention(
qkv,
forward_meta.caches[2 * layer.layer_id],
forward_meta.caches[2 * layer.layer_id + 1],
cache_k,
cache_v,
forward_meta.seq_lens_encoder,
forward_meta.seq_lens_decoder,
forward_meta.seq_lens_this_time,
@@ -350,8 +361,8 @@ class AppendAttentionBackend(AttentionBackend):
metadata.attn_mask,
layer.qkv_bias,
layer.qkv_scale,
getattr(layer, "cache_k_scale", None),
getattr(layer, "cache_v_scale", None),
cache_k_scales,
cache_v_scales,
getattr(layer, "cache_k_out_scale", None),
getattr(layer, "cache_v_out_scale", None),
getattr(layer, "cache_k_zp", None),