Skip to content

[BugFix][Graph Optimization] Fix deterministic mode + CUDAGraph garbled output by making reset_share_inputs CUDAGraph-safe` #6746

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
gongweibao wants to merge 2 commits into
base: develop
Choose a base branch
from
Open

[BugFix][Graph Optimization] Fix deterministic mode + CUDAGraph garbled output by making reset_share_inputs CUDAGraph-safe` #6746

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
f5a77f5
fix crash
Mar 9, 2026
84186dd
fix comments
Mar 9, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
71 changes: 49 additions & 22 deletions fastdeploy/worker/input_batch.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,6 +17,7 @@
import paddle
from paddleformers.utils.log import logger

from fastdeploy import envs
from fastdeploy.config import CacheConfig, FDConfig, ModelConfig, SpeculativeConfig
from fastdeploy.model_executor.layers.rotary_embedding import get_rope
from fastdeploy.model_executor.logits_processor import build_logits_processors
Expand Down Expand Up @@ -577,7 +578,11 @@ def reset_share_inputs(self):
# Reset reasoning buffers
fill_paddle_tensor(self, "reasoning_status", 0)
# Reset reasoning allowed tokens (not using fill_paddle_tensor since it's a fixed tensor)
self.reasoning_allowed_tokens = paddle.to_tensor([100973, 100975], dtype="int64")
if envs.FD_DETERMINISTIC_MODE and hasattr(self, "reasoning_allowed_tokens"):
# In-place copy to preserve GPU address for CUDAGraph safety.
self.reasoning_allowed_tokens[:] = paddle.to_tensor([100973, 100975], dtype="int64")
else:
self.reasoning_allowed_tokens = paddle.to_tensor([100973, 100975], dtype="int64")

# Reset block tables
fill_paddle_tensor(self, "block_tables", -1)
Expand All @@ -590,8 +595,15 @@ def reset_share_inputs(self):
-1,
)
)
self.free_list = paddle.to_tensor(free_list, dtype="int32")
self.free_list_len = paddle.full([1], len(free_list), dtype="int32")
if envs.FD_DETERMINISTIC_MODE and hasattr(self, "free_list"):
# In-place copy to preserve GPU address for CUDAGraph safety.
self.free_list.fill_(0)
free_list_tensor = paddle.to_tensor(free_list, dtype="int32")
self.free_list[: len(free_list)] = free_list_tensor
self.free_list_len.fill_(len(free_list))
else:
self.free_list = paddle.to_tensor(free_list, dtype="int32")
self.free_list_len = paddle.full([1], len(free_list), dtype="int32")

# Reset stop sequences
fill_paddle_tensor(self, "stop_seqs_len", 0)
Expand Down Expand Up @@ -625,29 +637,44 @@ def reset_share_inputs(self):
else:
rope_head_dim = head_dim // 2

self.rope_emb = paddle.full(
shape=[
max_num_seqs,
2,
1,
self.model_config.max_model_len,
1,
rope_head_dim,
],
fill_value=0,
dtype="float32",
)
if envs.FD_DETERMINISTIC_MODE and hasattr(self, "rope_emb"):
# In-place fill to preserve GPU address for CUDAGraph safety.
self.rope_emb.fill_(0)
else:
self.rope_emb = paddle.full(
shape=[
max_num_seqs,
2,
1,
self.model_config.max_model_len,
1,
rope_head_dim,
],
fill_value=0,
dtype="float32",
)
self.image_features = None
self.image_features_list = None
else:
# Reset non-multimodal rope_emb
self.rope_emb = get_rope(
rotary_dim=self.model_config.head_dim,
position_ids=paddle.arange(self.model_config.max_model_len).reshape((1, -1)),
base=self.model_config.rope_theta,
model_config=self.model_config,
partial_rotary_factor=self.model_config.partial_rotary_factor,
)
if envs.FD_DETERMINISTIC_MODE and hasattr(self, "rope_emb"):
# In-place copy to preserve GPU address for CUDAGraph safety.
new_rope_emb = get_rope(
rotary_dim=self.model_config.head_dim,
position_ids=paddle.arange(self.model_config.max_model_len).reshape((1, -1)),
base=self.model_config.rope_theta,
model_config=self.model_config,
partial_rotary_factor=self.model_config.partial_rotary_factor,
)
self.rope_emb.copy_(new_rope_emb, False)
else:
self.rope_emb = get_rope(
rotary_dim=self.model_config.head_dim,
position_ids=paddle.arange(self.model_config.max_model_len).reshape((1, -1)),
base=self.model_config.rope_theta,
model_config=self.model_config,
partial_rotary_factor=self.model_config.partial_rotary_factor,
)

# Reset other miscellaneous tensors
fill_paddle_tensor(self, "mask_rollback", 0)
Expand Down
Loading