apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-06 00:57:33 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[Feature] Support batched tokens for EP (#3415)

Browse Source 
* Support batched tokens for EP

* Support batched tokens for EP

* Support batched tokens for EP

* Support batched tokens for EP

* Support batched tokens for EP and fix bug

* Support batched tokens for EP and fix bug

* Support batched tokens for EP and fix bug

* Support batched tokens for EP and fix bug
This commit is contained in:
chenjian
2025-08-18 11:43:36 +08:00
committed by GitHub
parent 3f86ae0007
commit aba94169dc
9 changed files with 235 additions and 97 deletions
Download Patch File Download Diff File Expand all files Collapse all files

204
fastdeploy/model_executor/pre_and_post_process.py
View File

@@ -65,6 +65,7 @@ else:
update_inputs,
step_reschedule,
update_inputs_v1,
speculate_step_reschedule,
)
from fastdeploy.worker.output import ModelOutputData, ModelRunnerOutput, SamplerOutput
@@ -355,12 +356,11 @@ def step_cuda(
"""
if speculative_config.method is not None:
if enable_prefix_caching:
speculate_step_system_cache(
if DISABLE_RECOVER:
speculate_step_reschedule(
share_inputs["stop_flags"],
share_inputs["seq_lens_this_time"],
share_inputs["step_seq_lens_encoder"],
share_inputs["step_seq_lens_decoder"],
share_inputs["seq_lens_encoder"],
share_inputs["seq_lens_decoder"],
share_inputs["block_tables"],
@@ -386,64 +386,67 @@ def step_cuda(
speculative_config.num_speculative_tokens,
)
else:
speculate_step_paddle(
share_inputs["stop_flags"],
share_inputs["seq_lens_this_time"],
share_inputs["step_seq_lens_encoder"],
share_inputs["seq_lens_encoder"],
share_inputs["seq_lens_decoder"],
share_inputs["block_tables"],
share_inputs["encoder_block_lens"],
share_inputs["is_block_step"],
share_inputs["step_block_list"],
share_inputs["step_lens"],
share_inputs["recover_block_list"],
share_inputs["recover_lens"],
share_inputs["need_block_list"],
share_inputs["need_block_len"],
share_inputs["used_list_len"],
share_inputs["free_list"],
share_inputs["free_list_len"],
share_inputs["input_ids"],
share_inputs["pre_ids"],
share_inputs["step_idx"],
share_inputs["next_tokens"],
share_inputs["first_token_ids"],
share_inputs["accept_num"],
block_size,
enc_dec_block_num,
speculative_config.num_speculative_tokens,
)
if enable_prefix_caching:
speculate_step_system_cache(
share_inputs["stop_flags"],
share_inputs["seq_lens_this_time"],
share_inputs["step_seq_lens_encoder"],
share_inputs["step_seq_lens_decoder"],
share_inputs["seq_lens_encoder"],
share_inputs["seq_lens_decoder"],
share_inputs["block_tables"],
share_inputs["encoder_block_lens"],
share_inputs["is_block_step"],
share_inputs["step_block_list"],
share_inputs["step_lens"],
share_inputs["recover_block_list"],
share_inputs["recover_lens"],
share_inputs["need_block_list"],
share_inputs["need_block_len"],
share_inputs["used_list_len"],
share_inputs["free_list"],
share_inputs["free_list_len"],
share_inputs["input_ids"],
share_inputs["pre_ids"],
share_inputs["step_idx"],
share_inputs["next_tokens"],
share_inputs["first_token_ids"],
share_inputs["accept_num"],
block_size,
enc_dec_block_num,
speculative_config.num_speculative_tokens,
)
else:
speculate_step_paddle(
share_inputs["stop_flags"],
share_inputs["seq_lens_this_time"],
share_inputs["step_seq_lens_encoder"],
share_inputs["seq_lens_encoder"],
share_inputs["seq_lens_decoder"],
share_inputs["block_tables"],
share_inputs["encoder_block_lens"],
share_inputs["is_block_step"],
share_inputs["step_block_list"],
share_inputs["step_lens"],
share_inputs["recover_block_list"],
share_inputs["recover_lens"],
share_inputs["need_block_list"],
share_inputs["need_block_len"],
share_inputs["used_list_len"],
share_inputs["free_list"],
share_inputs["free_list_len"],
share_inputs["input_ids"],
share_inputs["pre_ids"],
share_inputs["step_idx"],
share_inputs["next_tokens"],
share_inputs["first_token_ids"],
share_inputs["accept_num"],
block_size,
enc_dec_block_num,
speculative_config.num_speculative_tokens,
)
else:
if enable_prefix_caching:
step_system_cache(
share_inputs["stop_flags"],
share_inputs["seq_lens_this_time"],
share_inputs["step_seq_lens_encoder"],
share_inputs["step_seq_lens_decoder"],
share_inputs["seq_lens_encoder"],
share_inputs["seq_lens_decoder"],
share_inputs["block_tables"],
share_inputs["encoder_block_lens"],
share_inputs["is_block_step"],
share_inputs["step_block_list"],
share_inputs["step_lens"],
share_inputs["recover_block_list"],
share_inputs["recover_lens"],
share_inputs["need_block_list"],
share_inputs["need_block_len"],
share_inputs["used_list_len"],
share_inputs["free_list"],
share_inputs["free_list_len"],
share_inputs["input_ids"],
share_inputs["pre_ids"],
share_inputs["step_idx"],
share_inputs["next_tokens"],
share_inputs["first_token_ids"],
block_size,
enc_dec_block_num,
)
elif DISABLE_RECOVER:
if DISABLE_RECOVER:
step_reschedule(
share_inputs["stop_flags"],
share_inputs["seq_lens_this_time"],
@@ -471,32 +474,61 @@ def step_cuda(
enc_dec_block_num,
)
else:
step_paddle(
share_inputs["stop_flags"],
share_inputs["seq_lens_this_time"],
share_inputs["step_seq_lens_encoder"],
share_inputs["seq_lens_encoder"],
share_inputs["seq_lens_decoder"],
share_inputs["block_tables"],
share_inputs["encoder_block_lens"],
share_inputs["is_block_step"],
share_inputs["step_block_list"],
share_inputs["step_lens"],
share_inputs["recover_block_list"],
share_inputs["recover_lens"],
share_inputs["need_block_list"],
share_inputs["need_block_len"],
share_inputs["used_list_len"],
share_inputs["free_list"],
share_inputs["free_list_len"],
share_inputs["input_ids"],
share_inputs["pre_ids"],
share_inputs["step_idx"],
share_inputs["next_tokens"],
share_inputs["first_token_ids"],
block_size,
enc_dec_block_num,
)
if enable_prefix_caching:
step_system_cache(
share_inputs["stop_flags"],
share_inputs["seq_lens_this_time"],
share_inputs["step_seq_lens_encoder"],
share_inputs["step_seq_lens_decoder"],
share_inputs["seq_lens_encoder"],
share_inputs["seq_lens_decoder"],
share_inputs["block_tables"],
share_inputs["encoder_block_lens"],
share_inputs["is_block_step"],
share_inputs["step_block_list"],
share_inputs["step_lens"],
share_inputs["recover_block_list"],
share_inputs["recover_lens"],
share_inputs["need_block_list"],
share_inputs["need_block_len"],
share_inputs["used_list_len"],
share_inputs["free_list"],
share_inputs["free_list_len"],
share_inputs["input_ids"],
share_inputs["pre_ids"],
share_inputs["step_idx"],
share_inputs["next_tokens"],
share_inputs["first_token_ids"],
block_size,
enc_dec_block_num,
)
else:
step_paddle(
share_inputs["stop_flags"],
share_inputs["seq_lens_this_time"],
share_inputs["step_seq_lens_encoder"],
share_inputs["seq_lens_encoder"],
share_inputs["seq_lens_decoder"],
share_inputs["block_tables"],
share_inputs["encoder_block_lens"],
share_inputs["is_block_step"],
share_inputs["step_block_list"],
share_inputs["step_lens"],
share_inputs["recover_block_list"],
share_inputs["recover_lens"],
share_inputs["need_block_list"],
share_inputs["need_block_len"],
share_inputs["used_list_len"],
share_inputs["free_list"],
share_inputs["free_list_len"],
share_inputs["input_ids"],
share_inputs["pre_ids"],
share_inputs["step_idx"],
share_inputs["next_tokens"],
share_inputs["first_token_ids"],
block_size,
enc_dec_block_num,
)
def rebuild_padding(