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 6, 09:30 UTC
fix(xpu): Re-compute compile ranges after platform-specific config updates (#37523)
Signed-off-by: Yuxiang Liang <yuxiang.liang@intel.com> Signed-off-by: Yuxiang Liang <yuliang@habana.ai> Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Yuxiang Liang committed
Mar 20, 2026 at 11:52 UTC
638a872d77b51cc4c160e713a58a589671de3a0c
1 file changed
+4
-2
vllm/config/vllm.py
+4
-2
@@ -985,8 +985,6 @@ class VllmConfig:
985
"--kv-sharing-fast-prefill requires changes on model side for "
986
"correctness and to realize prefill savings."
987
)
988
- # TODO: Move after https://github.com/vllm-project/vllm/pull/26847 lands
989
- self._set_compile_ranges()
988
989
if (
990
self.model_config
@@ -1022,6 +1020,10 @@ class VllmConfig:
1020
)
1021
current_platform.check_and_update_config(self)
1022
1023
+ # Re-compute compile ranges after platform-specific config updates
1024
+ # (e.g., XPU may lower max_num_batched_tokens when MLA is enabled)
1025
+ self._set_compile_ranges()
1026
+
1027
# Do this after all the updates to compilation_config.mode
1028
effective_dp_size = (
1029
self.parallel_config.data_parallel_size