karylab_agents/vllm
1
0
Fork 0
forked from Karylab-cklius/vllm
Code Pull Requests 2 Actions 1 Packages Activity

[XPU] disable fusion pattern support on XPU platform (#39789)

Browse Source 
Signed-off-by: Chaojun Zhang <chaojun.zhang@intel.com>
Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
This commit is contained in:
Chaojun Zhang
2026-04-23 10:07:45 +08:00
committed by GitHub
co-authored by Kunshang Ji
parent 0283f303d8
commit ccaf5ffaa3
1 changed files with 20 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
vllm/platforms/xpu.py
+20
View File
@@ -213,6 +213,26 @@ class XPUPlatform(Platform):
"falling back to PIECEWISE graph mode on XPU platform."
)
# Disable fusion passes not yet supported on XPU.
pass_config = compilation_config.pass_config
fusion_passes_to_disable = {
"enable_sp": "Sequence parallelism",
"fuse_gemm_comms": "Async TP",
"fuse_allreduce_rms": "AllReduce + RMSNorm fusion",
"fuse_norm_quant": "RMSNorm + quant fusion",
"fuse_act_quant": "Activation + quant fusion",
"fuse_attn_quant": "Attention + quant fusion",
"fuse_act_padding": "Activation + padding fusion",
"fuse_rope_kvcache": "RoPE + KV cache fusion",
}
for flag, feature_name in fusion_passes_to_disable.items():
if getattr(pass_config, flag):
logger.warning(
"Feature %r is not yet supported on XPU and will be disabled.",
feature_name,
)
setattr(pass_config, flag, False)
# check and update parallel config
parallel_config = vllm_config.parallel_config
# Only override worker_cls if it's still the default "auto"