apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-28 10:51:39 +08:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

polish code with new pre-commit rule (#2923)

Browse Source
This commit is contained in:
Zero Rains
2025-07-19 23:19:27 +08:00
committed by GitHub
parent b8676d71a8
commit 25698d56d1
424 changed files with 14307 additions and 13518 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
fastdeploy/model_executor/layers/sample/ops/__init__.py
View File

@@ -15,7 +15,9 @@
"""
from .apply_penalty_multi_scores import (
apply_penalty_multi_scores, apply_speculative_penalty_multi_scores)
apply_penalty_multi_scores,
apply_speculative_penalty_multi_scores,
)
from .top_k_top_p_sampling import top_k_top_p_sampling
__all__ = [

27
fastdeploy/model_executor/layers/sample/ops/apply_penalty_multi_scores.py
View File

@@ -37,8 +37,8 @@ def apply_penalty_multi_scores(
apply_penalty_multi_scores
"""
if current_platform.is_cuda():
from fastdeploy.model_executor.ops.gpu import \
get_token_penalty_multi_scores
from fastdeploy.model_executor.ops.gpu import get_token_penalty_multi_scores
logits = get_token_penalty_multi_scores(
pre_token_ids,
prompt_ids,
@@ -54,8 +54,8 @@ def apply_penalty_multi_scores(
eos_token_ids,
)
elif current_platform.is_xpu():
from fastdeploy.model_executor.ops.xpu import \
get_token_penalty_multi_scores
from fastdeploy.model_executor.ops.xpu import get_token_penalty_multi_scores
logits = get_token_penalty_multi_scores(
pre_token_ids,
logits,
@@ -69,8 +69,10 @@ def apply_penalty_multi_scores(
eos_token_ids,
)
elif current_platform.is_iluvatar():
from fastdeploy.model_executor.ops.iluvatar import \
get_token_penalty_multi_scores
from fastdeploy.model_executor.ops.iluvatar import (
get_token_penalty_multi_scores,
)
logits = get_token_penalty_multi_scores(
pre_token_ids,
prompt_ids,
@@ -86,8 +88,8 @@ def apply_penalty_multi_scores(
eos_token_ids,
)
elif current_platform.is_gcu():
from fastdeploy.model_executor.ops.gcu import \
get_token_penalty_multi_scores
from fastdeploy.model_executor.ops.gcu import get_token_penalty_multi_scores
logits = get_token_penalty_multi_scores(
pre_token_ids,
logits,
@@ -101,7 +103,7 @@ def apply_penalty_multi_scores(
eos_token_ids,
)
else:
raise NotImplementedError()
raise NotImplementedError
return logits
@@ -126,8 +128,9 @@ def apply_speculative_penalty_multi_scores(
apply_speculative_penalty_multi_scores
"""
if current_platform.is_cuda():
from fastdeploy.model_executor.ops.gpu import \
speculate_get_token_penalty_multi_scores
from fastdeploy.model_executor.ops.gpu import (
speculate_get_token_penalty_multi_scores,
)
speculate_get_token_penalty_multi_scores(
pre_token_ids,
@@ -146,6 +149,6 @@ def apply_speculative_penalty_multi_scores(
max_len,
)
else:
raise NotImplementedError()
raise NotImplementedError
# inplace
return logits

60
fastdeploy/model_executor/layers/sample/ops/top_k_top_p_sampling.py
View File

@@ -22,8 +22,8 @@ from fastdeploy import envs
from fastdeploy.platforms import current_platform
if current_platform.is_gcu():
from fastdeploy.model_executor.ops.gcu import \
top_p_sampling as gcu_top_p_sampling
from fastdeploy.model_executor.ops.gcu import top_p_sampling as gcu_top_p_sampling
def top_k_top_p_sampling(
x: paddle.Tensor,
@@ -33,8 +33,8 @@ def top_k_top_p_sampling(
topp_seed: Optional[paddle.Tensor] = None,
seed: int = -1,
k: int = 0,
mode: Literal['truncated', 'non-truncated'] = "truncated",
order: Literal['top_k_first', 'joint'] = "top_k_first",
mode: Literal["truncated", "non-truncated"] = "truncated",
order: Literal["top_k_first", "joint"] = "top_k_first",
) -> tuple[paddle.Tensor, paddle.Tensor]:
"""
x(Tensor): An input 2-D Tensor with type float32, float16 and bfloat16.
@@ -61,35 +61,33 @@ def top_k_top_p_sampling(
"""
top_p_class = envs.FD_SAMPLING_CLASS.lower()
if top_p_class == "air":
_, ids = air_top_p_sampling(x,
top_p,
threshold,
topp_seed,
seed=seed,
k=k,
mode=mode)
_, ids = air_top_p_sampling(x, top_p, threshold, topp_seed, seed=seed, k=k, mode=mode)
elif top_p_class == "rejection":
ids = rejection_top_p_sampling(x, top_p, top_k, seed, order)
_ = None
elif top_p_class == "base_non_truncated":
_, ids = paddle.tensor.top_p_sampling(x,
top_p,
threshold=threshold,
topp_seed=topp_seed,
seed=seed,
k=k,
mode="non-truncated")
_, ids = paddle.tensor.top_p_sampling(
x,
top_p,
threshold=threshold,
topp_seed=topp_seed,
seed=seed,
k=k,
mode="non-truncated",
)
else:
if current_platform.is_gcu():
_, ids = gcu_top_p_sampling(x, top_p)
else:
_, ids = paddle.tensor.top_p_sampling(x,
top_p,
threshold=threshold,
topp_seed=topp_seed,
seed=seed,
k=k,
mode="truncated")
_, ids = paddle.tensor.top_p_sampling(
x,
top_p,
threshold=threshold,
topp_seed=topp_seed,
seed=seed,
k=k,
mode="truncated",
)
return _, ids
@@ -100,15 +98,15 @@ def air_top_p_sampling(
topp_seed: Optional[paddle.Tensor] = None,
seed: int = -1,
k: int = 0,
mode: Literal['truncated', 'non-truncated'] = "truncated",
mode: Literal["truncated", "non-truncated"] = "truncated",
) -> tuple[paddle.Tensor, paddle.Tensor]:
"""
air_top_p_sampling
"""
try:
from fastdeploy.model_executor.ops.gpu import air_top_p_sampling
out, ids = air_top_p_sampling(x, top_p, threshold, topp_seed, seed, k,
mode)
out, ids = air_top_p_sampling(x, top_p, threshold, topp_seed, seed, k, mode)
except ImportError:
raise RuntimeError("Cannot import air_top_p_sampling op.")
return out, ids
@@ -119,14 +117,16 @@ def rejection_top_p_sampling(
top_p: paddle.Tensor,
top_k: paddle.Tensor,
seed: int = -1,
order: Literal['top_k_first', 'joint'] = "top_k_first",
order: Literal["top_k_first", "joint"] = "top_k_first",
) -> paddle.Tensor:
"""
rejection_top_p_sampling
"""
try:
from fastdeploy.model_executor.ops.gpu import (
rejection_top_p_sampling, top_k_renorm_probs)
rejection_top_p_sampling,
top_k_renorm_probs,
)
if paddle.count_nonzero(top_k) == 0:
ids = rejection_top_p_sampling(