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 7, 11:30 UTC
[Rocm][CI] Fix LM Eval Large Models (H100) test group (#34750)
Signed-off-by: charlifu <charlifu@amd.com>
Charlie Fu committed
Mar 2, 2026 at 01:43 UTC
3fd1d4ec2c992689594f9c7ee0ac79597f74a2ef
2 files changed
+5
-4
.buildkite/lm-eval-harness/configs/models-large-rocm.txt
+1
@@ -1 +1,2 @@
1
Meta-Llama-4-Maverick-17B-128E-Instruct-FP8.yaml
2
+Qwen3-235B-A22B-Instruct-2507-FP8.yaml
.buildkite/test-amd.yaml
+4
-4
@@ -1544,8 +1544,8 @@ steps:
1544
- export VLLM_WORKER_MULTIPROC_METHOD=spawn
1545
- pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large.txt --tp-size=4
1546
1547
-##### H100 test #####
1548
-- label: LM Eval Large Models (H100) # optional
1547
+##### FP8 test #####
1548
+- label: LM Eval Large Models (H100) # optional, still use H100 for consistency
1549
gpu: h100
1550
optional: true
1551
mirror_hardwares: [amdexperimental, amdproduction]
@@ -1557,8 +1557,8 @@ steps:
1557
- csrc/
1558
- vllm/model_executor/layers/quantization
1559
commands:
1560
- - export VLLM_USE_DEEP_GEMM=0 # We found Triton is faster than DeepGEMM for H100
1561
- - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-hopper.txt --tp-size=4
1560
+ - export VLLM_USE_DEEP_GEMM=0
1561
+ - pytest -s -v test_lm_eval_correctness.py --config-list-file=configs/models-large-rocm.txt --tp-size=4
1562
1563
1564
##### H200 test #####