apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2026-04-23 00:17:25 +08:00
Code Issues Actions 21 Packages Projects Releases Wiki Activity

[Optimization] Auto set num_max_dispatch_tokens_per_rank (#7237)

Browse Source 
* auto set num_max_dispatch_tokens_per_rank

* fix ci

* fix ci

* fix ci
This commit is contained in:
RichardWooSJTU
2026-04-15 19:13:38 +08:00
committed by GitHub
parent 3f84d8d893
commit dec0b060fc
1 changed files with 19 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
fastdeploy/config.py
+19
View File
@@ -2159,6 +2159,25 @@ class FDConfig:
self.speculative_config.num_speculative_tokens = 1 self.speculative_config.num_speculative_tokens = 1
self.speculative_config.num_model_steps = 1 self.speculative_config.num_model_steps = 1
# Auto-compute num_max_dispatch_tokens_per_rank from max_num_seqs and num_speculative_tokens
if self.speculative_config is not None and self.speculative_config.method is not None:
num_spec_tokens = self.speculative_config.num_speculative_tokens
auto_dispatch_tokens = self.scheduler_config.max_num_seqs * (num_spec_tokens + 1)
else:
auto_dispatch_tokens = self.scheduler_config.max_num_seqs
if (
getattr(self.model_config, "num_max_dispatch_tokens_per_rank", None)
and self.model_config.num_max_dispatch_tokens_per_rank != auto_dispatch_tokens
):
logger.info(
f"Auto-setting num_max_dispatch_tokens_per_rank from "
f"{self.model_config.num_max_dispatch_tokens_per_rank} to {auto_dispatch_tokens} "
f"(max_num_seqs={self.scheduler_config.max_num_seqs}"
f"{f', num_speculative_tokens={num_spec_tokens}' if self.speculative_config is not None and self.speculative_config.method is not None else ''})."
)
self.model_config.num_max_dispatch_tokens_per_rank = auto_dispatch_tokens
if self.scheduler_config.splitwise_role == "mixed": if self.scheduler_config.splitwise_role == "mixed":
self._disable_sequence_parallel_moe_if_needed("Mixed") self._disable_sequence_parallel_moe_if_needed("Mixed")
self.model_config.moe_phase = MoEPhase(phase="prefill") self.model_config.moe_phase = MoEPhase(phase="prefill")