apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-31 11:56:44 +08:00
Code Issues Actions 6 Packages Projects Releases Wiki Activity

[FDConfig]Remove splitwise_role and engine_worker_queue_port in FDConfig (#4147)
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 
* remove splitwise_role and engine_worker_queue_port

* fix xpu

* fix xpu

* fix xpu

* fix unittest

* resolve conflct
This commit is contained in:
YuanRisheng
2025-09-19 17:01:52 +08:00
committed by GitHub
parent ee9d8a840a
commit 24180fba0a
23 changed files with 129 additions and 89 deletions
Download Patch File Download Diff File Expand all files Collapse all files

12
fastdeploy/model_executor/layers/moe/fused_moe_backend_base.py
View File

@@ -58,7 +58,7 @@ class MoEMethodBase(QuantMethodBase):
"top_k": layer.top_k,
"hidden_size": layer.hidden_size,
"num_experts": layer.num_experts,
"splitwise_role": layer.fd_config.parallel_config.splitwise_role,
"splitwise_role": layer.fd_config.scheduler_config.splitwise_role,
"num_max_dispatch_tokens_per_rank": layer.fd_config.model_config.num_max_dispatch_tokens_per_rank,
"ep_size": layer.ep_size,
"ep_rank": layer.ep_rank,
@@ -67,7 +67,7 @@ class MoEMethodBase(QuantMethodBase):
}
config = layer.fd_config
splitwise_role = config.parallel_config.splitwise_role
splitwise_role = config.scheduler_config.splitwise_role
load_strategy = config.load_config.load_strategy
# For "mixed" splitwise role: conditionally initialize both or none
@@ -81,7 +81,7 @@ class MoEMethodBase(QuantMethodBase):
return
# For non-mixed ep
phase = config.parallel_config.moe_phase.phase
phase = config.model_config.moe_phase.phase
if phase == "prefill":
self.ep_prefill_runner = EPPrefillRunner(**common_args)
else:
@@ -159,12 +159,12 @@ class MoEMethodBase(QuantMethodBase):
Paddle Cutlass compute Fused MoE.
"""
if layer.ep_size > 1:
if layer.fd_config.parallel_config.moe_phase.phase == "prefill":
if layer.fd_config.parallel_config.splitwise_role == "mixed":
if layer.fd_config.model_config.moe_phase.phase == "prefill":
if layer.fd_config.scheduler_config.splitwise_role == "mixed":
self.ep_prefill_runner.clean_low_latency_buffer()
return self.apply_ep_prefill(layer, x, gate)
else:
if layer.fd_config.parallel_config.splitwise_role == "mixed":
if layer.fd_config.scheduler_config.splitwise_role == "mixed":
self.ep_decoder_runner.clean_low_latency_buffer()
return self.apply_ep_decode(layer, x, gate)
else: