apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-12-24 13:28:13 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[Speculative Decoding][MTP]Support attn mask offset (#4641)
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 
* [MTP]Merge support attn (#4591)

* support mask_offset in speculate decoding

* fix dummpy run output

* add unit test

* fix unit test import

* support attn_mask_offset in mtp mode

* add update_attn_mask op

* fix unit test && fix code-style
This commit is contained in:
freeliuzc
2025-11-03 10:08:01 +08:00
committed by GitHub
parent f44f4bafd1
commit 11398790d3
13 changed files with 638 additions and 111 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
fastdeploy/worker/gpu_model_runner.py
View File

@@ -1208,6 +1208,8 @@ class GPUModelRunner(ModelRunnerBase):
self.share_inputs["logits_processors_args"] = [{} for _ in range(max_num_seqs)]
logger.info(f"Enabled logits processors: {self.share_inputs['logits_processors']}")
self.share_inputs["mask_rollback"] = paddle.full(shape=[max_num_seqs, 1], fill_value=0, dtype="int32")
def _prepare_inputs(self) -> None:
"""Prepare the model inputs"""
if envs.ENABLE_V1_KVCACHE_SCHEDULER:
@@ -1713,6 +1715,8 @@ class GPUModelRunner(ModelRunnerBase):
accept_num=(self.share_inputs["accept_num"] if self.speculative_decoding else None),
stop_token_ids=self.share_inputs["stop_seqs"],
stop_seqs_len=self.share_inputs["stop_seqs_len"],
prompt_lens=self.share_inputs["prompt_lens"],
mask_rollback=self.share_inputs["mask_rollback"],
)
post_process(
@@ -2223,6 +2227,7 @@ class GPUModelRunner(ModelRunnerBase):
stop_token_ids=self.share_inputs["stop_seqs"],
stop_seqs_len=self.share_inputs["stop_seqs_len"],
prompt_lens=self.share_inputs["prompt_lens"],
mask_rollback=self.share_inputs["mask_rollback"],
)
if self.speculative_config.method in ["mtp"] and self.scheduler_config.splitwise_role == "prefill":