From 17de459fa3d9b2b82b192a9d2a11f263ffb76c41 Mon Sep 17 00:00:00 2001 From: khluu Date: Tue, 4 Aug 2026 03:40:41 -0700 Subject: [PATCH] [CI] Stabilize GLM-5.2 PCP evaluation Enable expandable CUDA allocator segments for the TP1/PCP4 GLM-5.2 eval so the 32K-token FlashInfer MoE workspace can reuse reserved memory. Co-authored-by: OpenAI Codex Signed-off-by: khluu --- tests/evals/gsm8k/configs/GLM-5.2-NVFP4-TP1-PCP4-EP.yaml | 1 + 1 file changed, 1 insertion(+) diff --git a/tests/evals/gsm8k/configs/GLM-5.2-NVFP4-TP1-PCP4-EP.yaml b/tests/evals/gsm8k/configs/GLM-5.2-NVFP4-TP1-PCP4-EP.yaml index 14b4f2c0d1da..d4ddfcc99fc6 100644 --- a/tests/evals/gsm8k/configs/GLM-5.2-NVFP4-TP1-PCP4-EP.yaml +++ b/tests/evals/gsm8k/configs/GLM-5.2-NVFP4-TP1-PCP4-EP.yaml @@ -13,5 +13,6 @@ server_args: >- --enable-expert-parallel --kv-cache-dtype fp8 env: + PYTORCH_CUDA_ALLOC_CONF: "expandable_segments:True" VLLM_LOGGING_LEVEL: "DEBUG" VLLM_USE_V2_MODEL_RUNNER: "1"