DulyRun public control vllm-top-p-61997-20261017-r1 Attribution: synthetic input and proposed counter update from Bowen Wang (abmfy), https://github.com/vllm-project/vllm/issues/61997 The retained run used one H100 PCIe (114 SMs), the official vLLM image: vllm/vllm-openai@sha256:a4a4c0437bf7240089da5f08aa370c4aee17ae5290f7a3b468825ee26c4c3a6b Actual versions: vLLM0.31.0 / torch2.13.0+cu130 / Triton3.7.1 / Python3.12.3. See the report for the retained pip-check failures. No clean dependency claim. In a directory containing reproduce.py, fetch the exact upstream files: curl -fL https://raw.githubusercontent.com/vllm-project/vllm/bb363db9a5ec2edc7b39e99b00af363a89d1fb81/vllm/v1/sample/ops/topk_topp_triton.py -o before-kernel.py curl -fL https://raw.githubusercontent.com/vllm-project/vllm/bf95f58d10891293aa6fde4e21f184d12ec0ba34/vllm/v1/sample/ops/topk_topp_triton.py -o introduced-kernel.py curl -fL https://raw.githubusercontent.com/vllm-project/vllm/fa75d3456d60a5ffe923818c4b495295675b38e1/vllm/v1/sample/ops/topk_topp_triton.py -o reported-kernel.py Verify their SHA-256 values against manifest.json, then run: python3 reproduce.py Exactly one visible compatible CUDA GPU is required. No weights are needed. Do not interpret setup failures as sampler defects. This procedure was executed only in the recorded environment; it is not a promise on other environments. The script runs B64 and B65 for four source arms twice; it saves each first masked-logit tensor plus metadata. Repeat tensors are saved only if unequal. The proposed update is an isolated control, not an upstream merge claim. outputs.tar.gz contains eight CPU tensors and per-row observations. Each tensor is float32, shaped (64 or 65, 131072); use torch.load(path, weights_only=True). The archive expands to about 258 MiB. Check the archive and tensor SHA-256 hashes in manifest.json. No weights, customer inputs, credentials or private MiMo rows. GPU resource was deleted after the run. Charge reconciliation was pending when this immutable report was published; reserved budget3USD is not an actual cost.