EngineX/xc-llm-ascend
3
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

[EPLB] Display the expert hotness comparison before and after eplb. (#6877)

Browse Source 
### What this PR does / why we need it?
To intuitively show the effect of the eplb algorithm, we print the
expert heat before and after eplb.

### Does this PR introduce _any_ user-facing change?

### How was this patch tested?

![Snipaste_2026-02-28_17-23-42](https://github.com/user-attachments/assets/db1dadd1-cf96-44da-af34-57d41ccf412f)


- vLLM version: v0.16.0
- vLLM main:
15d76f74e2

Signed-off-by: shenchuxiaofugui <1311027364@qq.com>
This commit is contained in:
LI SHENGYONG
2026-03-06 09:53:29 +08:00
committed by GitHub
parent 18b52afe2b
commit ccd00798f3
3 changed files with 42 additions and 38 deletions
Download Patch File Download Diff File Expand all files Collapse all files

37
vllm_ascend/eplb/eplb_updator.py
View File

@@ -34,7 +34,6 @@ class EplbUpdator:
self.eplb_loader = loader
self.eplb_process = eplb_process
self.shared_dict = self.eplb_process.shared_dict
self.moe_imbalance_dict: dict[int, float] = {}
self.comm_group = get_dynamic_eplb_group()
def set_adaptor(self, adaptor: VllmEplbAdaptor):
@@ -137,44 +136,8 @@ class EplbUpdator:
self.shared_dict["moe_load"] = moe_load.cpu()
logger.debug(f"[ModelRunner] Updated shared_dict['moe_load'] shape={moe_load.shape}")
if dist.get_rank() == 0:
self.compute_moe_imbalance(moe_load)
self.summarize_moe_imbalance()
return moe_load
def compute_moe_imbalance(self, moe_load: torch.Tensor):
self.moe_imbalance_dict.clear()
layer_card_load = moe_load.sum(dim=-1).cpu().float()
for layer_idx in range(layer_card_load.size(0)):
layer_load = layer_card_load[layer_idx]
mean_load = layer_load.mean().item()
max_load = layer_load.max().item()
moe_load_imbalance = max_load / (mean_load + 1e-6)
logger.debug(f"[ModelRunner][MOE_load_stats][Layer {layer_idx}] PAR={moe_load_imbalance:.4f}")
self.moe_imbalance_dict[layer_idx] = moe_load_imbalance
def summarize_moe_imbalance(self):
values = list(self.moe_imbalance_dict.values())
if not values:
logger.info("[MOE_load_stats] No data available.")
return
avg_imbalance = sum(values) / len(values)
max_imbalance = max(values)
min_imbalance = min(values)
logger.info(
f"[ModelRunner][MOE_load_stats] Peak-to-Average-Ratio: "
f"Mean={avg_imbalance:.4f}, Max={max_imbalance:.4f}, Min={min_imbalance:.4f}"
)
def warm_up_eplb(self):
self.shared_dict["expert_maps"] = self.adaptor.get_global_expert_map()
self.compute_and_set_moe_load()