[AMD][CI][The Rock] Fix language models standard for The Rock on mi355 (#53837)

Signed-off-by: Randall Smith <Randall.Smith@amd.com>
Co-authored-by: Andreas Karatzas <akaratza@amd.com>
This commit is contained in:
rasmith
2026-09-17 22:13:13 +00:00
committed by GitHub
co-authored by Andreas Karatzas
parent e8a2a0a8bb
commit d3074acdda
2 changed files with 6 additions and 9 deletions
@@ -26,7 +26,7 @@ AITER_MODEL_LIST = [
"meta-llama/Llama-3.2-1B-Instruct",
"openbmb/MiniCPM3-4B",
"Qwen/Qwen2.5-0.5B-Instruct",
"TitanML/tiny-mixtral",
"axolotl-ai-co/tiny-mixtral-30m",
"Qwen/Qwen3-8B",
]
@@ -95,7 +95,7 @@ AITER_MODEL_LIST = [
pytest.param("stabilityai/stablelm-3b-4e1t"), # stablelm
pytest.param("bigcode/starcoder2-3b"), # starcoder2
pytest.param(
"TitanML/tiny-mixtral", # mixtral
"axolotl-ai-co/tiny-mixtral-30m", # mixtral
marks=[pytest.mark.core_model],
),
pytest.param("swiss-ai/Apertus-8B-Instruct-2509"), # apertus
@@ -126,12 +126,6 @@ def test_models(
model_info.check_available_online(on_fail="skip")
model_info.check_transformers_version(on_fail="skip")
if current_platform.is_rocm() and model == "TitanML/tiny-mixtral":
# Its single-token router selects LLMM1, whose low-precision
# accumulation can change the top-2 experts. Keep the optimized kernel
# enabled generally, but use the reference GEMM for this accuracy test.
monkeypatch.setenv("VLLM_ROCM_USE_SKINNY_GEMM", "0")
if use_rocm_aiter and (model in AITER_MODEL_LIST):
monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1")
elif use_rocm_aiter and model not in AITER_MODEL_LIST:
+4 -1
View File
@@ -447,7 +447,10 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
),
"MixtralForCausalLM": _HfExamplesInfo(
"mistralai/Mixtral-8x7B-Instruct-v0.1",
{"tiny": "TitanML/tiny-mixtral"},
{
"tiny": "axolotl-ai-co/tiny-mixtral-30m",
"tiny-random": "TitanML/tiny-mixtral",
},
),
"NemotronForCausalLM": _HfExamplesInfo("nvidia/Minitron-8B-Base"),
"NemotronHForCausalLM": _HfExamplesInfo(