apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 08:37:06 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

supports dynamic Cfp8 (#3767)
Some checks failed
CE Compile Job / ce_job_pre_check (push) Has been cancelled
Details
CE Compile Job / print_ce_job_pre_check_outputs (push) Has been cancelled
Details
CE Compile Job / FD-Clone-Linux (push) Has been cancelled
Details
CE Compile Job / Show Code Archive Output (push) Has been cancelled
Details
CE Compile Job / BUILD_SM8090 (push) Has been cancelled
Details
CE Compile Job / BUILD_SM8689 (push) Has been cancelled
Details
CE Compile Job / CE_UPLOAD (push) Has been cancelled
Details
Deploy GitHub Pages / deploy (push) Has been cancelled
Details

Browse Source 
* supports dynamic Cfp8

* add unittest
This commit is contained in:
lzy
2025-09-08 11:41:29 +08:00
committed by GitHub
parent b5e20e3015
commit af49b81ffd
20 changed files with 1417 additions and 225 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

@@ -231,6 +231,17 @@ 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)
@@ -269,8 +280,8 @@ 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,
@@ -293,8 +304,8 @@ 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),
@@ -325,8 +336,8 @@ 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,
@@ -348,8 +359,8 @@ 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),

18
fastdeploy/model_executor/layers/quantization/kv_cache.py
View File

@@ -33,6 +33,7 @@ class KvCacheQuantzationTypes(str, Enum):
INT8 = "int8"
FP8 = "float8_e4m3fn"
BLOCK_WISE_FP8 = "block_wise_fp8"
INT8_ZP = "int8_zp"
INT4_ZP = "int4_zp"
FP8_ZP = "float8_e4m3fn_zp"
@@ -62,7 +63,11 @@ class KvCacheQuantConfig(QuantConfigBase):
if self.quant_type == KvCacheQuantzationTypes.INT8 or self.quant_type == KvCacheQuantzationTypes.INT8_ZP:
self.max_bound = 127.0
elif self.quant_type == KvCacheQuantzationTypes.FP8 or self.quant_type == KvCacheQuantzationTypes.FP8_ZP:
elif (
self.quant_type == KvCacheQuantzationTypes.FP8
or self.quant_type == KvCacheQuantzationTypes.FP8_ZP
or self.quant_type == KvCacheQuantzationTypes.BLOCK_WISE_FP8
):
self.max_bound = 448.0
elif self.quant_type == KvCacheQuantzationTypes.INT4_ZP:
self.max_bound = 7.0
@@ -178,12 +183,17 @@ class KVCacheMethodBase(QuantMethodBase):
layer.cache_quant_type_str = "cache_int4_zp"
layer.quant_max_bound = 7.0
layer.quant_min_bound = -7.0
elif self.cache_quant_config.quant_type == KvCacheQuantzationTypes.BLOCK_WISE_FP8:
layer.cache_quant_type_str = "block_wise_fp8"
layer.quant_max_bound = 448.0
layer.quant_min_bound = -448.0
else:
raise NotImplementedError(f"{self.cache_quant_config.quant_type} is not implemented")
self.load_scale(layer, state_dict)
if self.cache_quant_config.has_zero_point:
self.load_zp(layer, state_dict)
if "block_wise" not in layer.cache_quant_type_str:
self.load_scale(layer, state_dict)
if self.cache_quant_config.has_zero_point:
self.load_zp(layer, state_dict)
def apply(self, layer):
"""