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 19 Packages Projects Releases Wiki Activity

[Graph Optimization] Support CUDAGraph for P/PD mixed Batch using SOT subgraph spliting mode (#6196)

Browse Source 
* refine comment && refine variable name

* replace comment
This commit is contained in:
Ryan
2026-01-29 16:29:54 +08:00
committed by GitHub
parent e237313797
commit 5e78c1ac87
5 changed files with 118 additions and 21 deletions
Download Patch File Download Diff File Expand all files Collapse all files
fastdeploy/worker/gpu_worker.py
+17 -2
View File
@@ -213,13 +213,28 @@ class GpuWorker(WorkerBase):
def graph_optimize_and_warm_up_model(self) -> None:
"""
Perform the warm-up and the graph optimization
Perform the warm-up and the graph optimization.
Execution modes:
| Mode | Prefill + Mixed | Decode |
|-----------------------------------|--------------------------|--------------------------|
| Dynamic (graph_opt_level=0) | Dynamic | Dynamic + CUDAGraph |
| Static Full Graph (full=True) | Dynamic | Static + CUDAGraph |
| Static Split Graph (full=False) | Static + CUDAGraph | Dynamic + CUDAGraph |
"""
if self.fd_config.graph_opt_config.graph_opt_level >= 1 and not self.model_runner.use_cudagraph:
self.model_runner.sot_warmup()
if self.fd_config.graph_opt_config.graph_opt_level >= 1:
self.model_runner.vision_encoder_compile()
# Trigger cuda graph capture
# Static split graph mode: capture CUDAGraph for prefill/mixed phase
if (
self.fd_config.graph_opt_config.graph_opt_level >= 1
and not self.fd_config.graph_opt_config.full_cuda_graph
):
self.model_runner.capture_model_prefill_and_mixed()
# Capture CUDAGraph for decode phase (all modes)
self.model_runner.capture_model()
def check_health(self) -> bool: