[ROCm] Enable V2 model runner for Kimi-K3 on ROCm (#51653)
Signed-off-by: vllmellm <vllm.ellm@embeddedllm.com>
vllmellm committed
Aug 13, 2026 at 22:34 UTC
f3c1638927eee6ea31ad5a66e86e5b7ed6ebaa02
1 file changed
-16
vllm/config/vllm.py
-16
@@ -83,26 +83,10 @@ DEFAULT_V2_MODEL_RUNNER_ARCHITECTURES = frozenset(
83
}
84
)
85
86
-# Architectures that default to V1 on ROCm: the V2 runner faults during the
87
-# profile run. VLLM_USE_V2_MODEL_RUNNER=1 still forces V2.
88
-# TODO: fix V2 enablement
89
-ROCM_EXCLUDED_V2_MODEL_RUNNER_ARCHITECTURES = frozenset(
90
- {
91
- "KimiK3ForConditionalGeneration",
92
- }
93
-)
94
-
86
87
@lru_cache
88
def default_v2_model_runner_architectures() -> frozenset[str]:
89
"""Architectures defaulting to the V2 model runner on this platform."""
99
- from vllm.platforms import current_platform
100
-
101
- if current_platform.is_rocm():
102
- return (
103
- DEFAULT_V2_MODEL_RUNNER_ARCHITECTURES
104
- - ROCM_EXCLUDED_V2_MODEL_RUNNER_ARCHITECTURES
105
- )
90
return DEFAULT_V2_MODEL_RUNNER_ARCHITECTURES
91
92