A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 4, 18:45 UTC
[XPU] disable fusion pattern support on XPU platform (#39789)
Signed-off-by: Chaojun Zhang <chaojun.zhang@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
Chaojun Zhang committed
Apr 23, 2026 at 10:07 UTC
ccaf5ffaa3e1fb2a081b2c9e403ac0e4dfc142c8
1 file changed
+20
vllm/platforms/xpu.py
+20
@@ -213,6 +213,26 @@ class XPUPlatform(Platform):
213
"falling back to PIECEWISE graph mode on XPU platform."
214
)
215
216
+ # Disable fusion passes not yet supported on XPU.
217
+ pass_config = compilation_config.pass_config
218
+ fusion_passes_to_disable = {
219
+ "enable_sp": "Sequence parallelism",
220
+ "fuse_gemm_comms": "Async TP",
221
+ "fuse_allreduce_rms": "AllReduce + RMSNorm fusion",
222
+ "fuse_norm_quant": "RMSNorm + quant fusion",
223
+ "fuse_act_quant": "Activation + quant fusion",
224
+ "fuse_attn_quant": "Attention + quant fusion",
225
+ "fuse_act_padding": "Activation + padding fusion",
226
+ "fuse_rope_kvcache": "RoPE + KV cache fusion",
227
+ }
228
+ for flag, feature_name in fusion_passes_to_disable.items():
229
+ if getattr(pass_config, flag):
230
+ logger.warning(
231
+ "Feature %r is not yet supported on XPU and will be disabled.",
232
+ feature_name,
233
+ )
234
+ setattr(pass_config, flag, False)
235
+
236
# check and update parallel config
237
parallel_config = vllm_config.parallel_config
238
# Only override worker_cls if it's still the default "auto"