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 5, 16:00 UTC
[CI][ROCm] Add gpt-oss w4a8 in CI (#38292)
Signed-off-by: Bowen Bao <bowenbao@amd.com>
Bowen Bao committed
Apr 2, 2026 at 09:06 UTC
82a006beebf03c4f7bd600ab68f15b3325feb8e4
2 files changed
+10
-1
tests/evals/gpt_oss/configs/gpt-oss-20b-rocm-mxfp4-fp8.yaml
new
+8
@@ -0,0 +1,8 @@
1
+# SPDX-License-Identifier: Apache-2.0
2
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
3
+model_name: amd/gpt-oss-20b-MoE-Quant-W-MXFP4-A-FP8-KV-FP8
4
+metric_threshold: 0.568
5
+reasoning_effort: low
6
+server_args: "--attention-backend ROCM_AITER_UNIFIED_ATTN"
7
+env:
8
+ VLLM_ROCM_USE_AITER: "1"
\ No newline at end of file
tests/evals/gpt_oss/configs/models-gfx950.txt
+2
-1
@@ -1,3 +1,4 @@
1
# GFX950 model configurations for GPQA evaluation
2
# Tests different environment variable combinations
3
-gpt-oss-20b-rocm-baseline.yaml
\ No newline at end of file
3
+gpt-oss-20b-rocm-baseline.yaml
4
+gpt-oss-20b-rocm-mxfp4-fp8.yaml
\ No newline at end of file