biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[ROCm][Quantization][1/N] Refactor quark_moe w_mxfp4 w/ oracle (#38774)

Browse Source 
Signed-off-by: Bowen Bao <bowenbao@amd.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
This commit is contained in:
Bowen Bao
2026-04-02 20:29:57 -07:00
committed by GitHub
parent 32e0c0bfa2
commit 103f0de565
6 changed files with 170 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
tests/evals/gpt_oss/configs/gpt-oss-20b-rocm-quark-mxfp4-bf16-aiter.yaml Normal file
View File

@@ -0,0 +1,8 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
model_name: amd/gpt-oss-20b-w-mxfp4-a-bf16
metric_threshold: 0.568
reasoning_effort: low
server_args: "--attention-backend ROCM_AITER_UNIFIED_ATTN --moe-backend aiter"
env:
VLLM_ROCM_USE_AITER: "1"

6
tests/evals/gpt_oss/configs/gpt-oss-20b-rocm-quark-mxfp4-bf16-triton.yaml Normal file
View File

@@ -0,0 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
model_name: amd/gpt-oss-20b-w-mxfp4-a-bf16
metric_threshold: 0.568
reasoning_effort: low
server_args: "--attention-backend ROCM_AITER_UNIFIED_ATTN --moe-backend triton"

0
tests/evals/gpt_oss/configs/gpt-oss-20b-rocm-mxfp4-fp8.yaml → tests/evals/gpt_oss/configs/gpt-oss-20b-rocm-quark-mxfp4-fp8-triton.yaml
View File

4
tests/evals/gpt_oss/configs/models-gfx950.txt
View File

@@ -1,4 +1,6 @@
# GFX950 model configurations for GPQA evaluation
# Tests different environment variable combinations
gpt-oss-20b-rocm-baseline.yaml
gpt-oss-20b-rocm-mxfp4-fp8.yaml
gpt-oss-20b-rocm-quark-mxfp4-bf16-aiter.yaml
gpt-oss-20b-rocm-quark-mxfp4-bf16-triton.yaml
gpt-oss-20b-rocm-quark-mxfp4-fp8-triton.yaml