[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>
rasmith committed
Sep 17, 2026 at 17:13 UTC
d3074acddade913b530e192143b3611f58938ce3
2 files changed
+6
-9
tests/models/language/generation/test_common.py
+2
-8
@@ -26,7 +26,7 @@ AITER_MODEL_LIST = [
26
"meta-llama/Llama-3.2-1B-Instruct",
27
"openbmb/MiniCPM3-4B",
28
"Qwen/Qwen2.5-0.5B-Instruct",
29
- "TitanML/tiny-mixtral",
29
+ "axolotl-ai-co/tiny-mixtral-30m",
30
"Qwen/Qwen3-8B",
31
]
32
@@ -95,7 +95,7 @@ AITER_MODEL_LIST = [
95
pytest.param("stabilityai/stablelm-3b-4e1t"), # stablelm
96
pytest.param("bigcode/starcoder2-3b"), # starcoder2
97
pytest.param(
98
- "TitanML/tiny-mixtral", # mixtral
98
+ "axolotl-ai-co/tiny-mixtral-30m", # mixtral
99
marks=[pytest.mark.core_model],
100
),
101
pytest.param("swiss-ai/Apertus-8B-Instruct-2509"), # apertus
@@ -126,12 +126,6 @@ def test_models(
126
model_info.check_available_online(on_fail="skip")
127
model_info.check_transformers_version(on_fail="skip")
128
129
- if current_platform.is_rocm() and model == "TitanML/tiny-mixtral":
130
- # Its single-token router selects LLMM1, whose low-precision
131
- # accumulation can change the top-2 experts. Keep the optimized kernel
132
- # enabled generally, but use the reference GEMM for this accuracy test.
133
- monkeypatch.setenv("VLLM_ROCM_USE_SKINNY_GEMM", "0")
134
-
129
if use_rocm_aiter and (model in AITER_MODEL_LIST):
130
monkeypatch.setenv("VLLM_ROCM_USE_AITER", "1")
131
elif use_rocm_aiter and model not in AITER_MODEL_LIST:
tests/models/registry.py
+4
-1
@@ -447,7 +447,10 @@ _TEXT_GENERATION_EXAMPLE_MODELS = {
447
),
448
"MixtralForCausalLM": _HfExamplesInfo(
449
"mistralai/Mixtral-8x7B-Instruct-v0.1",
450
- {"tiny": "TitanML/tiny-mixtral"},
450
+ {
451
+ "tiny": "axolotl-ai-co/tiny-mixtral-30m",
452
+ "tiny-random": "TitanML/tiny-mixtral",
453
+ },
454
),
455
"NemotronForCausalLM": _HfExamplesInfo("nvidia/Minitron-8B-Base"),
456
"NemotronHForCausalLM": _HfExamplesInfo(