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 7, 16:30 UTC
[BugFix][XPU] Fix speculative decoding on Intel XPU due to bug with `IGC_ForceOCLSIMDWidth=16` (#35298)
Signed-off-by: Ofir Zafrir <ofir.zafrir@intel.com> Co-authored-by: Kunshang Ji <kunshang.ji@intel.com>
Ofir Zafrir committed
Feb 26, 2026 at 09:15 UTC
a07c4c59392aeffd619ac0a6a1c10fc364cf6840
1 file changed
-3
vllm/platforms/xpu.py
-3
@@ -201,9 +201,6 @@ class XPUPlatform(Platform):
201
202
if vllm_config.lora_config is not None:
203
compilation_config.mode = CompilationMode.NONE
204
- # decrease triton kernel compilation scratch space for speculative decoding
205
- if vllm_config.speculative_config is not None:
206
- os.environ["IGC_ForceOCLSIMDWidth"] = "16" # noqa: SIM112
204
# check and update parallel config
205
parallel_config = vllm_config.parallel_config
206
# Only override worker_cls if it's still the default "auto"